merge

models/vision/ddim/README.md

+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Denoising Diffusion Implicit Models (DDIM)
+
+## Overview
+
+DDPM was proposed in [Denoising Diffusion Implicit Models](https://arxiv.org/abs/2010.02502) by *Jiaming Song, Chenlin Meng, Stefano Ermon*
+
+The abstract from the paper is the following:
+
+*Denoising diffusion probabilistic models (DDPMs) have achieved high quality image generation without adversarial training, yet they require simulating a Markov chain for many steps to produce a sample. To accelerate sampling, we present denoising diffusion implicit models (DDIMs), a more efficient class of iterative implicit probabilistic models with the same training procedure as DDPMs. In DDPMs, the generative process is defined as the reverse of a Markovian diffusion process. We construct a class of non-Markovian diffusion processes that lead to the same training objective, but whose reverse process can be much faster to sample from. We empirically demonstrate that DDIMs can produce high quality samples 10× to 50× faster in terms of wall-clock time compared to DDPMs, allow us to trade off computation for sample quality, and can perform semantically meaningful image interpolation directly in the latent space.*
+
+Tips:
+
+- ...
+- ...
+
+This model was contributed by [???](https://huggingface.co/???). The original code can be found [here](https://github.com/hojonathanho/diffusion).

models/vision/ddim/example.py

+#!/usr/bin/env python3
+import os
+import pathlib
+from modeling_ddim import DDIM
+import PIL.Image
+import numpy as np
+
+model_ids = ["ddim-celeba-hq", "ddim-lsun-church", "ddim-lsun-bedroom"]
+
+for model_id in model_ids:
+    path = os.path.join("/home/patrick/images/hf", model_id)
+    pathlib.Path(path).mkdir(parents=True, exist_ok=True)
+
+    ddpm = DDIM.from_pretrained("fusing/" + model_id)
+    image = ddpm(batch_size=4)
+
+    image_processed = image.cpu().permute(0, 2, 3, 1)
+    image_processed = (image_processed + 1.0) * 127.5
+    image_processed = image_processed.numpy().astype(np.uint8)
+
+    for i in range(image_processed.shape[0]):
+        image_pil = PIL.Image.fromarray(image_processed[i])
+        image_pil.save(os.path.join(path, f"image_{i}.png"))

models/vision/ddim/modeling_ddim.py

+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+
+# limitations under the License.
+
+
+from diffusers import DiffusionPipeline
+import tqdm
+import torch
+
+
+class DDIM(DiffusionPipeline):
+
+    def __init__(self, unet, noise_scheduler):
+        super().__init__()
+        self.register_modules(unet=unet, noise_scheduler=noise_scheduler)
+
+    def __call__(self, batch_size=1, generator=None, torch_device=None, eta=0.0, num_inference_steps=50):
+        # eta corresponds to η in paper and should be between [0, 1]
+        if torch_device is None:
+            torch_device = "cuda" if torch.cuda.is_available() else "cpu"
+
+        num_trained_timesteps = self.noise_scheduler.num_timesteps
+        inference_step_times = range(0, num_trained_timesteps, num_trained_timesteps // num_inference_steps)
+
+        self.unet.to(torch_device)
+        image = self.noise_scheduler.sample_noise((batch_size, self.unet.in_channels, self.unet.resolution, self.unet.resolution), device=torch_device, generator=generator)
+
+        for t in tqdm.tqdm(reversed(range(num_inference_steps)), total=num_inference_steps):
+            # get actual t and t-1
+            train_step = inference_step_times[t]
+            prev_train_step = inference_step_times[t - 1] if t > 0 else - 1
+
+            # compute alphas
+            alpha_prod_t = self.noise_scheduler.get_alpha_prod(train_step)
+            alpha_prod_t_prev = self.noise_scheduler.get_alpha_prod(prev_train_step)
+            alpha_prod_t_rsqrt = 1 / alpha_prod_t.sqrt()
+            alpha_prod_t_prev_rsqrt = 1 / alpha_prod_t_prev.sqrt()
+            beta_prod_t_sqrt = (1 - alpha_prod_t).sqrt()
+            beta_prod_t_prev_sqrt = (1 - alpha_prod_t_prev).sqrt()
+
+            # compute relevant coefficients
+            coeff_1 = (alpha_prod_t_prev - alpha_prod_t).sqrt() * alpha_prod_t_prev_rsqrt * beta_prod_t_prev_sqrt / beta_prod_t_sqrt * eta
+            coeff_2 = ((1 - alpha_prod_t_prev) - coeff_1 ** 2).sqrt()
+
+            # model forward
+            with torch.no_grad():
+                noise_residual = self.unet(image, train_step)
+
+            # predict mean of prev image
+            pred_mean = alpha_prod_t_rsqrt * (image - beta_prod_t_sqrt * noise_residual)
+            pred_mean = torch.clamp(pred_mean, -1, 1)
+            pred_mean = (1 / alpha_prod_t_prev_rsqrt) * pred_mean + coeff_2 * noise_residual
+
+            # if eta > 0.0 add noise. Note eta = 1.0 essentially corresponds to DDPM
+            if eta > 0.0:
+                noise = self.noise_scheduler.sample_noise(image.shape, device=image.device, generator=generator)
+                image = pred_mean + coeff_1 * noise
+            else:
+                image = pred_mean
+
+        return image

models/vision/ddim/run_ddpm.py

+#!/usr/bin/env python3
+import torch
+
+from diffusers import GaussianDDPMScheduler, UNetModel
+
+
+model = UNetModel(dim=64, dim_mults=(1, 2, 4, 8))
+
+diffusion = GaussianDDPMScheduler(model, image_size=128, timesteps=1000, loss_type="l1")  # number of steps  # L1 or L2
+
+training_images = torch.randn(8, 3, 128, 128)  # your images need to be normalized from a range of -1 to +1
+loss = diffusion(training_images)
+loss.backward()
+# after a lot of training
+
+sampled_images = diffusion.sample(batch_size=4)
+sampled_images.shape  # (4, 3, 128, 128)

models/vision/ddim/run_inference.py

+#!/usr/bin/env python3
+# !pip install diffusers
+from modeling_ddim import DDIM
+import PIL.Image
+import numpy as np
+
+model_id = "fusing/ddpm-cifar10"
+model_id = "fusing/ddpm-lsun-bedroom"
+
+# load model and scheduler
+ddpm = DDIM.from_pretrained(model_id)
+
+# run pipeline in inference (sample random noise and denoise)
+image = ddpm()
+
+# process image to PIL
+image_processed = image.cpu().permute(0, 2, 3, 1)
+image_processed = (image_processed + 1.0) * 127.5
+image_processed = image_processed.numpy().astype(np.uint8)
+image_pil = PIL.Image.fromarray(image_processed[0])
+
+# save image
+image_pil.save("/home/patrick/images/show.png")

models/vision/ddpm/modeling_ddpm.py

@@ -21,8 +21,6 @@ import torch
 
 class DDPM(DiffusionPipeline):
 
-    modeling_file = "modeling_ddpm.py"
-
     def __init__(self, unet, noise_scheduler):
         super().__init__()
         self.register_modules(unet=unet, noise_scheduler=noise_scheduler)

models/vision/glide/README.md

+# References
+
+[GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models](https://arxiv.org/pdf/2112.10741.pdf)
+[Diffusion Models Beat GANs on Image Synthesis](https://arxiv.org/pdf/2105.05233.pdf)
\ No newline at end of file

models/vision/glide/convert_weights.py

-import argparse
-
 import torch
 from torch import nn
 
-from transformers import CLIPTextConfig, CLIPTextModel, GPT2Tokenizer
+from diffusers import ClassifierFreeGuidanceScheduler, GlideDDIMScheduler, CLIPTextModel, GLIDETextToImageUNetModel, GLIDESuperResUNetModel
+from modeling_glide import GLIDE
+from transformers import CLIPTextConfig, GPT2Tokenizer
+
 
 # wget https://openaipublic.blob.core.windows.net/diffusion/dec-2021/base.pt
 state_dict = torch.load("base.pt", map_location="cpu")
 state_dict = {k: nn.Parameter(v) for k, v in state_dict.items()}
+
+### Convert the text encoder
+
 config = CLIPTextConfig(
+    vocab_size=50257,
+    max_position_embeddings=128,
     hidden_size=512,
     intermediate_size=2048,
     num_hidden_layers=16,
     num_attention_heads=8,
-    max_position_embeddings=128
+    use_padding_embeddings=True,
 )
 model = CLIPTextModel(config).eval()
-tokenizer = GPT2Tokenizer("./glide-base/vocab.json", "./glide-base/merges.txt", pad_token="<|endoftext|>")
-tokenizer.save_pretrained("./glide-base")
+tokenizer = GPT2Tokenizer("./glide-base/tokenizer/vocab.json", "./glide-base/tokenizer/merges.txt", pad_token="<|endoftext|>")
 
 hf_encoder = model.text_model
 
@@ -30,15 +35,8 @@ hf_encoder.final_layer_norm.bias = state_dict["final_ln.bias"]
 
 for layer_idx in range(config.num_hidden_layers):
     hf_layer = hf_encoder.encoder.layers[layer_idx]
-    q_proj, k_proj, v_proj = state_dict[f"transformer.resblocks.{layer_idx}.attn.c_qkv.weight"].chunk(3, dim=0)
-    q_proj_bias, k_proj_bias, v_proj_bias = state_dict[f"transformer.resblocks.{layer_idx}.attn.c_qkv.bias"].chunk(3, dim=0)
-
-    hf_layer.self_attn.q_proj.weight.data = q_proj
-    hf_layer.self_attn.q_proj.bias.data = q_proj_bias
-    hf_layer.self_attn.k_proj.weight.data = k_proj
-    hf_layer.self_attn.k_proj.bias.data = k_proj_bias
-    hf_layer.self_attn.v_proj.weight.data = v_proj
-    hf_layer.self_attn.v_proj.bias.data = v_proj_bias
+    hf_layer.self_attn.qkv_proj.weight = state_dict[f"transformer.resblocks.{layer_idx}.attn.c_qkv.weight"]
+    hf_layer.self_attn.qkv_proj.bias = state_dict[f"transformer.resblocks.{layer_idx}.attn.c_qkv.bias"]
 
     hf_layer.self_attn.out_proj.weight = state_dict[f"transformer.resblocks.{layer_idx}.attn.c_proj.weight"]
     hf_layer.self_attn.out_proj.bias = state_dict[f"transformer.resblocks.{layer_idx}.attn.c_proj.bias"]
@@ -53,8 +51,56 @@ for layer_idx in range(config.num_hidden_layers):
     hf_layer.mlp.fc2.weight = state_dict[f"transformer.resblocks.{layer_idx}.mlp.c_proj.weight"]
     hf_layer.mlp.fc2.bias = state_dict[f"transformer.resblocks.{layer_idx}.mlp.c_proj.bias"]
 
-inputs = tokenizer(["an oil painting of a corgi", ""], padding="max_length", max_length=128, return_tensors="pt")
-with torch.no_grad():
-    outputs = model(**inputs)
+### Convert the Text-to-Image UNet
+
+text2im_model = GLIDETextToImageUNetModel(
+    in_channels=3,
+    model_channels=192,
+    out_channels=6,
+    num_res_blocks=3,
+    attention_resolutions=(2, 4, 8),
+    dropout=0.1,
+    channel_mult=(1, 2, 3, 4),
+    num_heads=1,
+    num_head_channels=64,
+    num_heads_upsample=1,
+    use_scale_shift_norm=True,
+    resblock_updown=True,
+    transformer_dim=512,
+)
+
+text2im_model.load_state_dict(state_dict, strict=False)
+
+text_scheduler = ClassifierFreeGuidanceScheduler(timesteps=1000, beta_schedule="squaredcos_cap_v2")
+
+### Convert the Super-Resolution UNet
+
+# wget https://openaipublic.blob.core.windows.net/diffusion/dec-2021/upsample.pt
+ups_state_dict = torch.load("upsample.pt", map_location="cpu")
+
+superres_model = GLIDESuperResUNetModel(
+    in_channels=6,
+    model_channels=192,
+    out_channels=6,
+    num_res_blocks=2,
+    attention_resolutions=(8, 16, 32),
+    dropout=0.1,
+    channel_mult=(1, 1, 2, 2, 4, 4),
+    num_heads=1,
+    num_head_channels=64,
+    num_heads_upsample=1,
+    use_scale_shift_norm=True,
+    resblock_updown=True,
+)
+
+superres_model.load_state_dict(ups_state_dict, strict=False)
+
+upscale_scheduler = GlideDDIMScheduler(timesteps=1000, beta_schedule="linear")
+
+glide = GLIDE(text_unet=text2im_model, text_noise_scheduler=text_scheduler, text_encoder=model, tokenizer=tokenizer,
+              upscale_unet=superres_model, upscale_noise_scheduler=upscale_scheduler)
+
+glide.save_pretrained("./glide-base")
+
+
 
-model.save_pretrained("./glide-base")
\ No newline at end of file

models/vision/glide/modeling_glide.py

@@ -14,46 +14,215 @@
 # limitations under the License.
 
 
-from diffusers import DiffusionPipeline
-from diffusers import UNetGLIDEModel
+import numpy as np
+import torch
 
 import tqdm
-import torch
+from diffusers import ClassifierFreeGuidanceScheduler, GlideDDIMScheduler, CLIPTextModel, DiffusionPipeline, GLIDETextToImageUNetModel, GLIDESuperResUNetModel
+from transformers import GPT2Tokenizer
+
+
+def _extract_into_tensor(arr, timesteps, broadcast_shape):
+    """
+    Extract values from a 1-D numpy array for a batch of indices.
+
+    :param arr: the 1-D numpy array.
+    :param timesteps: a tensor of indices into the array to extract.
+    :param broadcast_shape: a larger shape of K dimensions with the batch
+                            dimension equal to the length of timesteps.
+    :return: a tensor of shape [batch_size, 1, ...] where the shape has K dims.
+    """
+    res = torch.from_numpy(arr).to(device=timesteps.device)[timesteps].float()
+    while len(res.shape) < len(broadcast_shape):
+        res = res[..., None]
+    return res + torch.zeros(broadcast_shape, device=timesteps.device)
 
 
 class GLIDE(DiffusionPipeline):
-    def __init__(self, unet: UNetGLIDEModel, noise_scheduler):
+    def __init__(
+        self,
+        text_unet: GLIDETextToImageUNetModel,
+        text_noise_scheduler: ClassifierFreeGuidanceScheduler,
+        text_encoder: CLIPTextModel,
+        tokenizer: GPT2Tokenizer,
+        upscale_unet: GLIDESuperResUNetModel,
+        upscale_noise_scheduler: GlideDDIMScheduler
+    ):
         super().__init__()
-        self.register_modules(unet=unet, noise_scheduler=noise_scheduler)
+        self.register_modules(
+            text_unet=text_unet, text_noise_scheduler=text_noise_scheduler, text_encoder=text_encoder, tokenizer=tokenizer,
+            upscale_unet=upscale_unet, upscale_noise_scheduler=upscale_noise_scheduler
+        )
+
+    def q_posterior_mean_variance(self, scheduler, x_start, x_t, t):
+        """
+        Compute the mean and variance of the diffusion posterior:
+
+            q(x_{t-1} | x_t, x_0)
+
+        """
+        assert x_start.shape == x_t.shape
+        posterior_mean = (
+            _extract_into_tensor(scheduler.posterior_mean_coef1, t, x_t.shape) * x_start
+            + _extract_into_tensor(scheduler.posterior_mean_coef2, t, x_t.shape) * x_t
+        )
+        posterior_variance = _extract_into_tensor(scheduler.posterior_variance, t, x_t.shape)
+        posterior_log_variance_clipped = _extract_into_tensor(
+            scheduler.posterior_log_variance_clipped, t, x_t.shape
+        )
+        assert (
+            posterior_mean.shape[0]
+            == posterior_variance.shape[0]
+            == posterior_log_variance_clipped.shape[0]
+            == x_start.shape[0]
+        )
+        return posterior_mean, posterior_variance, posterior_log_variance_clipped
+
+    def p_mean_variance(self, model, scheduler, x, t, transformer_out=None, low_res=None, clip_denoised=True):
+        """
+        Apply the model to get p(x_{t-1} | x_t), as well as a prediction of
+        the initial x, x_0.
+
+        :param model: the model, which takes a signal and a batch of timesteps
+                      as input.
+        :param x: the [N x C x ...] tensor at time t.
+        :param t: a 1-D Tensor of timesteps.
+        :param clip_denoised: if True, clip the denoised signal into [-1, 1].
+        :param model_kwargs: if not None, a dict of extra keyword arguments to
+            pass to the model. This can be used for conditioning.
+        :return: a dict with the following keys:
+                 - 'mean': the model mean output.
+                 - 'variance': the model variance output.
+                 - 'log_variance': the log of 'variance'.
+                 - 'pred_xstart': the prediction for x_0.
+        """
+
+        B, C = x.shape[:2]
+        assert t.shape == (B,)
+        if transformer_out is None:
+            # super-res model
+            model_output = model(x, t, low_res)
+        else:
+            # text2image model
+            model_output = model(x, t, transformer_out)
+
+        assert model_output.shape == (B, C * 2, *x.shape[2:])
+        model_output, model_var_values = torch.split(model_output, C, dim=1)
+        min_log = _extract_into_tensor(scheduler.posterior_log_variance_clipped, t, x.shape)
+        max_log = _extract_into_tensor(np.log(scheduler.betas), t, x.shape)
+        # The model_var_values is [-1, 1] for [min_var, max_var].
+        frac = (model_var_values + 1) / 2
+        model_log_variance = frac * max_log + (1 - frac) * min_log
+        model_variance = torch.exp(model_log_variance)
+
+        pred_xstart = self._predict_xstart_from_eps(scheduler, x_t=x, t=t, eps=model_output)
+        if clip_denoised:
+            pred_xstart = pred_xstart.clamp(-1, 1)
+        model_mean, _, _ = self.q_posterior_mean_variance(scheduler, x_start=pred_xstart, x_t=x, t=t)
 
-    def __call__(self, generator=None, torch_device=None):
+        assert model_mean.shape == model_log_variance.shape == pred_xstart.shape == x.shape
+        return model_mean, model_variance, model_log_variance, pred_xstart
+
+    def _predict_xstart_from_eps(self, scheduler, x_t, t, eps):
+        assert x_t.shape == eps.shape
+        return (
+            _extract_into_tensor(scheduler.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t
+            - _extract_into_tensor(scheduler.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * eps
+        )
+
+    def _predict_eps_from_xstart(self, scheduler, x_t, t, pred_xstart):
+        return (
+            _extract_into_tensor(scheduler.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart
+        ) / _extract_into_tensor(scheduler.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
+
+    @torch.no_grad()
+    def __call__(self, prompt, generator=None, torch_device=None):
         torch_device = "cuda" if torch.cuda.is_available() else "cpu"
 
-        self.unet.to(torch_device)
+        self.text_unet.to(torch_device)
+        self.text_encoder.to(torch_device)
+        self.upscale_unet.to(torch_device)
+
+        # Create a classifier-free guidance sampling function
+        guidance_scale = 3.0
+
+        def text_model_fn(x_t, ts, transformer_out, **kwargs):
+            half = x_t[: len(x_t) // 2]
+            combined = torch.cat([half, half], dim=0)
+            model_out = self.text_unet(combined, ts, transformer_out, **kwargs)
+            eps, rest = model_out[:, :3], model_out[:, 3:]
+            cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
+            half_eps = uncond_eps + guidance_scale * (cond_eps - uncond_eps)
+            eps = torch.cat([half_eps, half_eps], dim=0)
+            return torch.cat([eps, rest], dim=1)
+
         # 1. Sample gaussian noise
-        image = self.noise_scheduler.sample_noise((1, self.unet.in_channels, self.unet.resolution, self.unet.resolution), device=torch_device, generator=generator)
-        for t in tqdm.tqdm(reversed(range(len(self.noise_scheduler))), total=len(self.noise_scheduler)):
+        batch_size = 2  # second image is empty for classifier-free guidance
+        image = self.text_noise_scheduler.sample_noise(
+            (batch_size, self.text_unet.in_channels, 64, 64), device=torch_device, generator=generator
+        )
+
+        # 2. Encode tokens
+        # an empty input is needed to guide the model away from (
+        inputs = self.tokenizer([prompt, ""], padding="max_length", max_length=128, return_tensors="pt")
+        input_ids = inputs["input_ids"].to(torch_device)
+        attention_mask = inputs["attention_mask"].to(torch_device)
+        transformer_out = self.text_encoder(input_ids, attention_mask).last_hidden_state
+
+        # 3. Run the text2image generation step
+        num_timesteps = len(self.text_noise_scheduler)
+        for i in tqdm.tqdm(reversed(range(num_timesteps)), total=num_timesteps):
+            t = torch.tensor([i] * image.shape[0], device=torch_device)
+            mean, variance, log_variance, pred_xstart = self.p_mean_variance(
+                text_model_fn, self.text_noise_scheduler, image, t, transformer_out=transformer_out
+            )
+            noise = self.text_noise_scheduler.sample_noise(image.shape, device=torch_device, generator=generator)
+            nonzero_mask = (t != 0).float().view(-1, *([1] * (len(image.shape) - 1)))  # no noise when t == 0
+            image = mean + nonzero_mask * torch.exp(0.5 * log_variance) * noise
+
+        # 4. Run the upscaling step
+        batch_size = 1
+        image = image[:1]
+        low_res = ((image + 1) * 127.5).round() / 127.5 - 1
+        eta = 0.0
+
+        # Tune this parameter to control the sharpness of 256x256 images.
+        # A value of 1.0 is sharper, but sometimes results in grainy artifacts.
+        upsample_temp = 0.997
+
+        image = self.upscale_noise_scheduler.sample_noise(
+            (batch_size, 3, 256, 256), device=torch_device, generator=generator
+        ) * upsample_temp
+
+        num_timesteps = len(self.upscale_noise_scheduler)
+        for t in tqdm.tqdm(reversed(range(len(self.upscale_noise_scheduler))), total=len(self.upscale_noise_scheduler)):
             # i) define coefficients for time step t
-            clip_image_coeff = 1 / torch.sqrt(self.noise_scheduler.get_alpha_prod(t))
-            clip_noise_coeff = torch.sqrt(1 / self.noise_scheduler.get_alpha_prod(t) - 1)
-            image_coeff = (1 - self.noise_scheduler.get_alpha_prod(t - 1)) * torch.sqrt(self.noise_scheduler.get_alpha(t)) / (1 - self.noise_scheduler.get_alpha_prod(t))
-            clip_coeff = torch.sqrt(self.noise_scheduler.get_alpha_prod(t - 1)) * self.noise_scheduler.get_beta(t) / (1 - self.noise_scheduler.get_alpha_prod(t))
+            clipped_image_coeff = 1 / torch.sqrt(self.upscale_noise_scheduler.get_alpha_prod(t))
+            clipped_noise_coeff = torch.sqrt(1 / self.upscale_noise_scheduler.get_alpha_prod(t) - 1)
+            image_coeff = (1 - self.upscale_noise_scheduler.get_alpha_prod(t - 1)) * torch.sqrt(
+                self.upscale_noise_scheduler.get_alpha(t)) / (1 - self.upscale_noise_scheduler.get_alpha_prod(t))
+            clipped_coeff = torch.sqrt(self.upscale_noise_scheduler.get_alpha_prod(t - 1)) * self.upscale_noise_scheduler.get_beta(
+                t) / (1 - self.upscale_noise_scheduler.get_alpha_prod(t))
 
             # ii) predict noise residual
-            with torch.no_grad():
-                noise_residual = self.unet(image, t)
+            time_input = torch.tensor([t] * image.shape[0], device=torch_device)
+            model_output = self.upscale_unet(image, time_input, low_res)
+            noise_residual, pred_variance = torch.split(model_output, 3, dim=1)
 
             # iii) compute predicted image from residual
             # See 2nd formula at https://github.com/hojonathanho/diffusion/issues/5#issue-896554416 for comparison
-            pred_mean = clip_image_coeff * image - clip_noise_coeff * noise_residual
+            pred_mean = clipped_image_coeff * image - clipped_noise_coeff * noise_residual
             pred_mean = torch.clamp(pred_mean, -1, 1)
-            prev_image = clip_coeff * pred_mean + image_coeff * image
+            prev_image = clipped_coeff * pred_mean + image_coeff * image
 
             # iv) sample variance
-            prev_variance = self.noise_scheduler.sample_variance(t, prev_image.shape, device=torch_device, generator=generator)
+            prev_variance = self.upscale_noise_scheduler.sample_variance(t, prev_image.shape, device=torch_device,
+                                                                 generator=generator)
 
             # v) sample  x_{t-1} ~ N(prev_image, prev_variance)
             sampled_prev_image = prev_image + prev_variance
             image = sampled_prev_image
 
+        image = image[0].permute(1, 2, 0)
+
         return image

models/vision/glide/run_glide.py

 import torch
-from .modeling_glide import GLIDE
-from diffusers import UNetGLIDEModel, GaussianDDPMScheduler
+from diffusers import DiffusionPipeline
+import PIL.Image
 
 generator = torch.Generator()
 generator = generator.manual_seed(0)
 
-# 1. Load models
+model_id = "fusing/glide-base"
 
-scheduler = GaussianDDPMScheduler.from_config("fusing/glide-base")
-model = UNetGLIDEModel.from_pretrained("fusing/glide-base")
+# load model and scheduler
+pipeline = DiffusionPipeline.from_pretrained(model_id)
 
-pipeline = GLIDE(model, scheduler)
+# run inference (text-conditioned denoising + upscaling)
+img = pipeline("a clip art of a hugging face", generator)
 
-img = pipeline(generator)
+# process image to PIL
+img = ((img + 1)*127.5).round().clamp(0, 255).to(torch.uint8).cpu().numpy()
+image_pil = PIL.Image.fromarray(img)
 
-print(img)
+# save image
+image_pil.save("test.png")
\ No newline at end of file

src/diffusers/__init__.py

@@ -5,7 +5,12 @@
 __version__ = "0.0.1"
 
 from .modeling_utils import ModelMixin
+from .models.clip_text_transformer import CLIPTextModel
 from .models.unet import UNetModel
-from .models.unet_glide import UNetGLIDEModel
+from .models.unet_glide import GLIDETextToImageUNetModel, GLIDESuperResUNetModel
+from .models.unet_ldm import UNetLDMModel
+from .models.vqvae import VQModel
 from .pipeline_utils import DiffusionPipeline
+from .schedulers.classifier_free_guidance import ClassifierFreeGuidanceScheduler
 from .schedulers.gaussian_ddpm import GaussianDDPMScheduler
+from .schedulers.glide_ddim import GlideDDIMScheduler

src/diffusers/configuration_utils.py

@@ -215,6 +215,7 @@ class ConfigMixin:
                 init_dict[key] = config_dict.pop(key)
 
         unused_kwargs = config_dict.update(kwargs)
+
         passed_keys = set(init_dict.keys())
         if len(expected_keys - passed_keys) > 0:
             logger.warn(

src/diffusers/models/__init__.py

@@ -16,5 +16,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .clip_text_transformer import CLIPTextModel
 from .unet import UNetModel
-from .unet_glide import UNetGLIDEModel
+from .unet_glide import GLIDETextToImageUNetModel, GLIDESuperResUNetModel
+from .unet_ldm import UNetLDMModel
+from .vqvae import VQModel
\ No newline at end of file

src/diffusers/models/clip_text_transformer.py

+# coding=utf-8
+# Copyright 2021 The OpenAI Team Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch CLIP model."""
+
+import math
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from transformers import CLIPConfig, CLIPModel, CLIPTextConfig, CLIPVisionConfig
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "openai/clip-vit-base-patch32"
+
+CLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "openai/clip-vit-base-patch32",
+    # See all CLIP models at https://huggingface.co/models?filter=clip
+]
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+# contrastive loss function, adapted from
+# https://sachinruk.github.io/blog/pytorch/pytorch%20lightning/loss%20function/gpu/2021/03/07/CLIP.html
+def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
+    return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))
+
+
+def clip_loss(similarity: torch.Tensor) -> torch.Tensor:
+    caption_loss = contrastive_loss(similarity)
+    image_loss = contrastive_loss(similarity.T)
+    return (caption_loss + image_loss) / 2.0
+
+
+@dataclass
+class CLIPOutput(ModelOutput):
+    """
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
+            Contrastive loss for image-text similarity.
+        logits_per_image:(`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
+            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
+            similarity scores.
+        logits_per_text:(`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
+            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
+            similarity scores.
+        text_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of [`CLIPTextModel`].
+        image_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The image embeddings obtained by applying the projection layer to the pooled output of [`CLIPVisionModel`].
+        text_model_output(`BaseModelOutputWithPooling`):
+            The output of the [`CLIPTextModel`].
+        vision_model_output(`BaseModelOutputWithPooling`):
+            The output of the [`CLIPVisionModel`].
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits_per_image: torch.FloatTensor = None
+    logits_per_text: torch.FloatTensor = None
+    text_embeds: torch.FloatTensor = None
+    image_embeds: torch.FloatTensor = None
+    text_model_output: BaseModelOutputWithPooling = None
+    vision_model_output: BaseModelOutputWithPooling = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+class CLIPVisionEmbeddings(nn.Module):
+    def __init__(self, config: CLIPVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(torch.randn(self.embed_dim))
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size, bias=False
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+        self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)))
+
+    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+        batch_size = pixel_values.shape[0]
+        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, width, grid, grid]
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+        embeddings = embeddings + self.position_embedding(self.position_ids)
+        return embeddings
+
+
+class CLIPTextEmbeddings(nn.Module):
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__()
+        embed_dim = config.hidden_size
+
+        self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
+        self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
+        self.use_padding_embeddings = config.use_padding_embeddings
+        if self.use_padding_embeddings:
+            self.padding_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if inputs_embeds is None:
+            inputs_embeds = self.token_embedding(input_ids)
+
+        position_embeddings = self.position_embedding(position_ids)
+        embeddings = inputs_embeds + position_embeddings
+
+        if self.use_padding_embeddings and attention_mask is not None:
+            padding_embeddings = self.padding_embedding(position_ids)
+            embeddings = torch.where(attention_mask.bool().unsqueeze(-1), embeddings, padding_embeddings)
+
+        return embeddings
+
+
+class CLIPAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = 1 / math.sqrt(math.sqrt(self.head_dim))
+
+        self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        qkv_states = self.qkv_proj(hidden_states)
+        qkv_states = qkv_states.view(bsz, tgt_len, self.num_heads, -1)
+        query_states, key_states, value_states = torch.split(qkv_states, self.head_dim, dim=-1)
+
+        attn_weights = torch.einsum("bthc,bshc->bhts", query_states * self.scale, key_states * self.scale)
+
+        wdtype = attn_weights.dtype
+        attn_weights = nn.functional.softmax(attn_weights.float(), dim=-1).type(wdtype)
+
+        attn_output = torch.einsum("bhts,bshc->bthc", attn_weights, value_states)
+        attn_output = attn_output.reshape(bsz, tgt_len, -1)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class CLIPMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class CLIPEncoderLayer(nn.Module):
+    def __init__(self, config: CLIPConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = CLIPAttention(config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim)
+        self.mlp = CLIPMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        causal_attention_mask: torch.Tensor,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class CLIPPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = CLIPConfig
+    base_model_prefix = "clip"
+    supports_gradient_checkpointing = True
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_factor
+        if isinstance(module, CLIPTextEmbeddings):
+            module.token_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
+            module.position_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
+            if hasattr(module, "padding_embedding"):
+                module.padding_embedding.weight.data.normal_(mean=0.0, std=factor * 0.02)
+        elif isinstance(module, CLIPVisionEmbeddings):
+            factor = self.config.initializer_factor
+            nn.init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor)
+            nn.init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor)
+            nn.init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor)
+        elif isinstance(module, CLIPAttention):
+            factor = self.config.initializer_factor
+            in_proj_std = (module.embed_dim**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
+            out_proj_std = (module.embed_dim**-0.5) * factor
+            nn.init.normal_(module.qkv_proj.weight, std=in_proj_std)
+            nn.init.normal_(module.out_proj.weight, std=out_proj_std)
+        elif isinstance(module, CLIPMLP):
+            factor = self.config.initializer_factor
+            in_proj_std = (
+                (module.config.hidden_size**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor
+            )
+            fc_std = (2 * module.config.hidden_size) ** -0.5 * factor
+            nn.init.normal_(module.fc1.weight, std=fc_std)
+            nn.init.normal_(module.fc2.weight, std=in_proj_std)
+        elif isinstance(module, CLIPModel):
+            nn.init.normal_(
+                module.text_projection.weight,
+                std=module.text_embed_dim**-0.5 * self.config.initializer_factor,
+            )
+            nn.init.normal_(
+                module.visual_projection.weight,
+                std=module.vision_embed_dim**-0.5 * self.config.initializer_factor,
+            )
+
+        if isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, CLIPEncoder):
+            module.gradient_checkpointing = value
+
+
+CLIP_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it
+    as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`CLIPConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+CLIP_TEXT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`CLIPTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+CLIP_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`CLIPFeatureExtractor`]. See [`CLIPFeatureExtractor.__call__`] for details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+CLIP_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`CLIPTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`CLIPFeatureExtractor`]. See [`CLIPFeatureExtractor.__call__`] for details.
+        return_loss (`bool`, *optional*):
+            Whether or not to return the contrastive loss.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class CLIPEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`CLIPEncoderLayer`].
+
+    Args:
+        config: CLIPConfig
+    """
+
+    def __init__(self, config: CLIPConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([CLIPEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        inputs_embeds,
+        attention_mask: Optional[torch.Tensor] = None,
+        causal_attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Causal mask for the text model. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(encoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    causal_attention_mask,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    causal_attention_mask,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class CLIPTextTransformer(nn.Module):
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+        self.embeddings = CLIPTextEmbeddings(config)
+        self.encoder = CLIPEncoder(config)
+        self.final_layer_norm = nn.LayerNorm(embed_dim)
+
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is None:
+            raise ValueError("You have to specify either input_ids")
+
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+
+        hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids, attention_mask=attention_mask)
+
+        bsz, seq_len = input_shape
+        # CLIP's text model uses causal mask, prepare it here.
+        # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
+        causal_attention_mask = self._build_causal_attention_mask(bsz, seq_len).to(hidden_states.device)
+
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _expand_mask(attention_mask, hidden_states.dtype)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            attention_mask=None,
+            causal_attention_mask=None,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.final_layer_norm(last_hidden_state)
+
+        # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        pooled_output = last_hidden_state[torch.arange(last_hidden_state.shape[0]), input_ids.argmax(dim=-1)]
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def _build_causal_attention_mask(self, bsz, seq_len):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(bsz, seq_len, seq_len)
+        mask.fill_(torch.tensor(float("-inf")))
+        mask.triu_(1)  # zero out the lower diagonal
+        mask = mask.unsqueeze(1)  # expand mask
+        return mask
+
+
+class CLIPTextModel(CLIPPreTrainedModel):
+    config_class = CLIPTextConfig
+
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__(config)
+        self.text_model = CLIPTextTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.text_model.embeddings.token_embedding
+
+    def set_input_embeddings(self, value):
+        self.text_model.embeddings.token_embedding = value
+
+    @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import CLIPTokenizer, CLIPTextModel
+
+        >>> model = CLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
+        >>> tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.last_hidden_state
+        >>> pooled_output = outputs.pooler_output  # pooled (EOS token) states
+        ```"""
+        return self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )

src/diffusers/models/unet_glide.py

@@ -388,7 +388,7 @@ class QKVAttention(nn.Module):
         return a.reshape(bs, -1, length)
 
 
-class UNetGLIDEModel(ModelMixin, ConfigMixin):
+class GLIDEUNetModel(ModelMixin, ConfigMixin):
     """
     The full UNet model with attention and timestep embedding.
 
@@ -419,11 +419,11 @@ class UNetGLIDEModel(ModelMixin, ConfigMixin):
 
     def __init__(
         self,
-        in_channels,
-        model_channels,
-        out_channels,
-        num_res_blocks,
-        attention_resolutions,
+        in_channels=3,
+        model_channels=192,
+        out_channels=6,
+        num_res_blocks=3,
+        attention_resolutions=(2, 4, 8),
         dropout=0,
         channel_mult=(1, 2, 4, 8),
         conv_resample=True,
@@ -435,28 +435,9 @@ class UNetGLIDEModel(ModelMixin, ConfigMixin):
         num_heads_upsample=-1,
         use_scale_shift_norm=False,
         resblock_updown=False,
-        encoder_channels=None,
+        transformer_dim=None,
     ):
         super().__init__()
-        self.register(
-            in_channels=in_channels,
-            model_channels=model_channels,
-            out_channels=out_channels,
-            num_res_blocks=num_res_blocks,
-            attention_resolutions=attention_resolutions,
-            dropout=dropout,
-            channel_mult=channel_mult,
-            conv_resample=conv_resample,
-            dims=dims,
-            use_checkpoint=use_checkpoint,
-            use_fp16=use_fp16,
-            num_heads=num_heads,
-            num_head_channels=num_head_channels,
-            num_heads_upsample=num_heads_upsample,
-            use_scale_shift_norm=use_scale_shift_norm,
-            resblock_updown=resblock_updown,
-            encoder_channels=encoder_channels,
-        )
 
         if num_heads_upsample == -1:
             num_heads_upsample = num_heads
@@ -470,7 +451,7 @@ class UNetGLIDEModel(ModelMixin, ConfigMixin):
         self.channel_mult = channel_mult
         self.conv_resample = conv_resample
         self.use_checkpoint = use_checkpoint
-        self.dtype = torch.float16 if use_fp16 else torch.float32
+        # self.dtype = torch.float16 if use_fp16 else torch.float32
         self.num_heads = num_heads
         self.num_head_channels = num_head_channels
         self.num_heads_upsample = num_heads_upsample
@@ -508,7 +489,7 @@ class UNetGLIDEModel(ModelMixin, ConfigMixin):
                             use_checkpoint=use_checkpoint,
                             num_heads=num_heads,
                             num_head_channels=num_head_channels,
-                            encoder_channels=encoder_channels,
+                            encoder_channels=transformer_dim,
                         )
                     )
                 self.input_blocks.append(TimestepEmbedSequential(*layers))
@@ -551,7 +532,7 @@ class UNetGLIDEModel(ModelMixin, ConfigMixin):
                 use_checkpoint=use_checkpoint,
                 num_heads=num_heads,
                 num_head_channels=num_head_channels,
-                encoder_channels=encoder_channels,
+                encoder_channels=transformer_dim,
             ),
             ResBlock(
                 ch,
@@ -587,7 +568,7 @@ class UNetGLIDEModel(ModelMixin, ConfigMixin):
                             use_checkpoint=use_checkpoint,
                             num_heads=num_heads_upsample,
                             num_head_channels=num_head_channels,
-                            encoder_channels=encoder_channels,
+                            encoder_channels=transformer_dim,
                         )
                     )
                 if level and i == num_res_blocks:
@@ -633,7 +614,7 @@ class UNetGLIDEModel(ModelMixin, ConfigMixin):
         self.middle_block.apply(convert_module_to_f32)
         self.output_blocks.apply(convert_module_to_f32)
 
-    def forward(self, x, timesteps, transformer_out):
+    def forward(self, x, timesteps):
         """
         Apply the model to an input batch.
 
@@ -642,18 +623,184 @@ class UNetGLIDEModel(ModelMixin, ConfigMixin):
         :param y: an [N] Tensor of labels, if class-conditional.
         :return: an [N x C x ...] Tensor of outputs.
         """
-        assert (y is not None) == (
-            self.num_classes is not None
-        ), "must specify y if and only if the model is class-conditional"
 
         hs = []
         emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
 
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb)
+            hs.append(h)
+        h = self.middle_block(h, emb)
+        for module in self.output_blocks:
+            h = torch.cat([h, hs.pop()], dim=1)
+            h = module(h, emb)
+        h = h.type(x.dtype)
+        return self.out(h)
+
+
+class GLIDETextToImageUNetModel(GLIDEUNetModel):
+    """
+    A UNetModel that performs super-resolution.
+
+    Expects an extra kwarg `low_res` to condition on a low-resolution image.
+    """
+
+    def __init__(
+             self,
+             in_channels=3,
+             model_channels=192,
+             out_channels=6,
+             num_res_blocks=3,
+             attention_resolutions=(2, 4, 8),
+             dropout=0,
+             channel_mult=(1, 2, 4, 8),
+             conv_resample=True,
+             dims=2,
+             use_checkpoint=False,
+             use_fp16=False,
+             num_heads=1,
+             num_head_channels=-1,
+             num_heads_upsample=-1,
+             use_scale_shift_norm=False,
+             resblock_updown=False,
+             transformer_dim=512
+    ):
+        super().__init__(
+            in_channels=in_channels,
+            model_channels=model_channels,
+            out_channels=out_channels,
+            num_res_blocks=num_res_blocks,
+            attention_resolutions=attention_resolutions,
+            dropout=dropout,
+            channel_mult=channel_mult,
+            conv_resample=conv_resample,
+            dims=dims,
+            use_checkpoint=use_checkpoint,
+            use_fp16=use_fp16,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            num_heads_upsample=num_heads_upsample,
+            use_scale_shift_norm=use_scale_shift_norm,
+            resblock_updown=resblock_updown,
+            transformer_dim=transformer_dim
+        )
+        self.register(
+            in_channels=in_channels,
+            model_channels=model_channels,
+            out_channels=out_channels,
+            num_res_blocks=num_res_blocks,
+            attention_resolutions=attention_resolutions,
+            dropout=dropout,
+            channel_mult=channel_mult,
+            conv_resample=conv_resample,
+            dims=dims,
+            use_checkpoint=use_checkpoint,
+            use_fp16=use_fp16,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            num_heads_upsample=num_heads_upsample,
+            use_scale_shift_norm=use_scale_shift_norm,
+            resblock_updown=resblock_updown,
+            transformer_dim=transformer_dim
+        )
+
+        self.transformer_proj = nn.Linear(transformer_dim, self.model_channels * 4)
+
+    def forward(self, x, timesteps, transformer_out=None):
+        hs = []
+        emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+
         # project the last token
         transformer_proj = self.transformer_proj(transformer_out[:, -1])
         transformer_out = transformer_out.permute(0, 2, 1)  # NLC -> NCL
 
-        h = x.type(self.dtype)
+        emb = emb + transformer_proj.to(emb)
+
+        h = x
+        for module in self.input_blocks:
+            h = module(h, emb, transformer_out)
+            hs.append(h)
+        h = self.middle_block(h, emb, transformer_out)
+        for module in self.output_blocks:
+            other = hs.pop()
+            h = torch.cat([h, other], dim=1)
+            h = module(h, emb, transformer_out)
+        return self.out(h)
+
+
+class GLIDESuperResUNetModel(GLIDEUNetModel):
+    """
+    A UNetModel that performs super-resolution.
+
+    Expects an extra kwarg `low_res` to condition on a low-resolution image.
+    """
+
+    def __init__(
+            self,
+            in_channels=3,
+            model_channels=192,
+            out_channels=6,
+            num_res_blocks=3,
+            attention_resolutions=(2, 4, 8),
+            dropout=0,
+            channel_mult=(1, 2, 4, 8),
+            conv_resample=True,
+            dims=2,
+            use_checkpoint=False,
+            use_fp16=False,
+            num_heads=1,
+            num_head_channels=-1,
+            num_heads_upsample=-1,
+            use_scale_shift_norm=False,
+            resblock_updown=False,
+    ):
+        super().__init__(
+            in_channels=in_channels,
+            model_channels=model_channels,
+            out_channels=out_channels,
+            num_res_blocks=num_res_blocks,
+            attention_resolutions=attention_resolutions,
+            dropout=dropout,
+            channel_mult=channel_mult,
+            conv_resample=conv_resample,
+            dims=dims,
+            use_checkpoint=use_checkpoint,
+            use_fp16=use_fp16,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            num_heads_upsample=num_heads_upsample,
+            use_scale_shift_norm=use_scale_shift_norm,
+            resblock_updown=resblock_updown,
+        )
+        self.register(
+            in_channels=in_channels,
+            model_channels=model_channels,
+            out_channels=out_channels,
+            num_res_blocks=num_res_blocks,
+            attention_resolutions=attention_resolutions,
+            dropout=dropout,
+            channel_mult=channel_mult,
+            conv_resample=conv_resample,
+            dims=dims,
+            use_checkpoint=use_checkpoint,
+            use_fp16=use_fp16,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            num_heads_upsample=num_heads_upsample,
+            use_scale_shift_norm=use_scale_shift_norm,
+            resblock_updown=resblock_updown,
+        )
+
+    def forward(self, x, timesteps, low_res=None):
+        _, _, new_height, new_width = x.shape
+        upsampled = F.interpolate(low_res, (new_height, new_width), mode="bilinear")
+        x = torch.cat([x, upsampled], dim=1)
+
+        hs = []
+        emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+
+        h = x
         for module in self.input_blocks:
             h = module(h, emb)
             hs.append(h)
@@ -661,5 +808,5 @@ class UNetGLIDEModel(ModelMixin, ConfigMixin):
         for module in self.output_blocks:
             h = torch.cat([h, hs.pop()], dim=1)
             h = module(h, emb)
-        h = h.type(x.dtype)
-        return self.out(h)
+
+        return self.out(h)
\ No newline at end of file

src/diffusers/models/unet_ldm.py

+from inspect import isfunction
+from abc import abstractmethod
+import math
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+try:
+    from einops import repeat, rearrange
+except:
+    print("Einops is not installed")
+    pass
+
+from ..configuration_utils import ConfigMixin
+from ..modeling_utils import ModelMixin
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return{el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
+        k = k.softmax(dim=-1)  
+        context = torch.einsum('bhdn,bhen->bhde', k, v)
+        out = torch.einsum('bhde,bhdn->bhen', context, q)
+        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        return self.to_out(out)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = rearrange(q, 'b c h w -> b (h w) c')
+        k = rearrange(k, 'b c h w -> b c (h w)')
+        w_ = torch.einsum('bij,bjk->bik', q, k)
+
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, 'b c h w -> b c (h w)')
+        w_ = rearrange(w_, 'b i j -> b j i')
+        h_ = torch.einsum('bij,bjk->bik', v, w_)
+        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, context=None, mask=None):
+        h = self.heads
+
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+
+        sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
+
+        if exists(mask):
+            mask = rearrange(mask, 'b ... -> b (...)')
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+
+        # attention, what we cannot get enough of
+        attn = sim.softmax(dim=-1)
+
+        out = torch.einsum('b i j, b j d -> b i d', attn, v)
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True):
+        super().__init__()
+        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout)  # is a self-attention
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
+                                    heads=n_heads, dim_head=d_head, dropout=dropout)  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None):
+        x = self.attn1(self.norm1(x)) + x
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    """
+    def __init__(self, in_channels, n_heads, d_head,
+                 depth=1, dropout=0., context_dim=None):
+        super().__init__()
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = Normalize(in_channels)
+
+        self.proj_in = nn.Conv2d(in_channels,
+                                 inner_dim,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim)
+                for d in range(depth)]
+        )
+
+        self.proj_out = zero_module(nn.Conv2d(inner_dim,
+                                              in_channels,
+                                              kernel_size=1,
+                                              stride=1,
+                                              padding=0))
+
+    def forward(self, x, context=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c')
+        for block in self.transformer_blocks:
+            x = block(x, context=context)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
+        x = self.proj_out(x)
+        return x + x_in
+
+def convert_module_to_f16(l):
+    """
+    Convert primitive modules to float16.
+    """
+    if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Conv3d)):
+        l.weight.data = l.weight.data.half()
+        if l.bias is not None:
+            l.bias.data = l.bias.data.half()
+
+
+def convert_module_to_f32(l):
+    """
+    Convert primitive modules to float32, undoing convert_module_to_f16().
+    """
+    if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Conv3d)):
+        l.weight.data = l.weight.data.float()
+        if l.bias is not None:
+            l.bias.data = l.bias.data.float()
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+class GroupNorm32(nn.GroupNorm):
+    def __init__(self, num_groups, num_channels, swish, eps=1e-5):
+        super().__init__(num_groups=num_groups, num_channels=num_channels, eps=eps)
+        self.swish = swish
+
+    def forward(self, x):
+        y = super().forward(x.float()).to(x.dtype)
+        if self.swish == 1.0:
+            y = F.silu(y)
+        elif self.swish:
+            y = y * F.sigmoid(y * float(self.swish))
+        return y
+
+
+def normalization(channels, swish=0.0):
+    """
+    Make a standard normalization layer, with an optional swish activation.
+
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    return GroupNorm32(num_channels=channels, num_groups=32, swish=swish)
+
+
+def timestep_embedding(timesteps, dim, max_period=10000):
+    """
+    Create sinusoidal timestep embeddings.
+
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
+        device=timesteps.device
+    )
+    args = timesteps[:, None].float() * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+## go
+class AttentionPool2d(nn.Module):
+    """
+    Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
+    """
+
+    def __init__(
+        self,
+        spacial_dim: int,
+        embed_dim: int,
+        num_heads_channels: int,
+        output_dim: int = None,
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(torch.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5)
+        self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
+        self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
+        self.num_heads = embed_dim // num_heads_channels
+        self.attention = QKVAttention(self.num_heads)
+
+    def forward(self, x):
+        b, c, *_spatial = x.shape
+        x = x.reshape(b, c, -1)  # NC(HW)
+        x = torch.cat([x.mean(dim=-1, keepdim=True), x], dim=-1)  # NC(HW+1)
+        x = x + self.positional_embedding[None, :, :].to(x.dtype)  # NC(HW+1)
+        x = self.qkv_proj(x)
+        x = self.attention(x)
+        x = self.c_proj(x)
+        return x[:, :, 0]
+
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb)
+            elif isinstance(layer, SpatialTransformer):
+                x = layer(x, context)
+            else:
+                x = layer(x)
+        return x
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(
+                x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
+            )
+        else:
+            x = F.interpolate(x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        up=False,
+        down=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
+            ),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, 3, padding=1
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = torch.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    An attention block that allows spatial positions to attend to each other.
+    Originally ported from here, but adapted to the N-d case.
+    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
+    """
+
+    def __init__(
+        self,
+        channels,
+        num_heads=1,
+        num_head_channels=-1,
+        use_checkpoint=False,
+        use_new_attention_order=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        if num_head_channels == -1:
+            self.num_heads = num_heads
+        else:
+            assert (
+                channels % num_head_channels == 0
+            ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
+            self.num_heads = channels // num_head_channels
+        self.use_checkpoint = use_checkpoint
+        self.norm = normalization(channels)
+        self.qkv = conv_nd(1, channels, channels * 3, 1)
+        # split heads before split qkv
+        self.attention = QKVAttentionLegacy(self.num_heads)
+
+        self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
+
+    def forward(self, x):
+        b, c, *spatial = x.shape
+        x = x.reshape(b, c, -1)
+        qkv = self.qkv(self.norm(x))
+        h = self.attention(qkv)
+        h = self.proj_out(h)
+        return (x + h).reshape(b, c, *spatial)
+
+
+class QKVAttention(nn.Module):
+    """
+    A module which performs QKV attention and splits in a different order.
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.chunk(3, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = torch.einsum(
+            "bct,bcs->bts",
+            (q * scale).view(bs * self.n_heads, ch, length),
+            (k * scale).view(bs * self.n_heads, ch, length),
+        )  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = torch.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+def count_flops_attn(model, _x, y):
+    """
+    A counter for the `thop` package to count the operations in an
+    attention operation.
+    Meant to be used like:
+        macs, params = thop.profile(
+            model,
+            inputs=(inputs, timestamps),
+            custom_ops={QKVAttention: QKVAttention.count_flops},
+        )
+    """
+    b, c, *spatial = y[0].shape
+    num_spatial = int(np.prod(spatial))
+    # We perform two matmuls with the same number of ops.
+    # The first computes the weight matrix, the second computes
+    # the combination of the value vectors.
+    matmul_ops = 2 * b * (num_spatial ** 2) * c
+    model.total_ops += torch.DoubleTensor([matmul_ops])
+
+
+class QKVAttentionLegacy(nn.Module):
+    """
+    A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = torch.einsum(
+            "bct,bcs->bts", q * scale, k * scale
+        )  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = torch.einsum("bts,bcs->bct", weight, v)
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class UNetLDMModel(ModelMixin, ConfigMixin):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+    ):
+        super().__init__()
+        
+        # register all __init__ params with self.register
+        self.register(
+            image_size=image_size,
+            in_channels=in_channels,
+            model_channels=model_channels,
+            out_channels=out_channels,
+            num_res_blocks=num_res_blocks,
+            attention_resolutions=attention_resolutions,
+            dropout=dropout,
+            channel_mult=channel_mult,
+            conv_resample=conv_resample,
+            dims=dims,
+            num_classes=num_classes,
+            use_checkpoint=use_checkpoint,
+            use_fp16=use_fp16,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            num_heads_upsample=num_heads_upsample,
+            use_scale_shift_norm=use_scale_shift_norm,
+            resblock_updown=resblock_updown,
+            use_new_attention_order=use_new_attention_order,
+            use_spatial_transformer=use_spatial_transformer,
+            transformer_depth=transformer_depth,
+            context_dim=context_dim,
+            n_embed=n_embed,
+            legacy=legacy,
+        )
+
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype_ = torch.float16 if use_fp16 else torch.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        if self.num_classes is not None:
+            self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=dim_head,
+                            use_new_attention_order=use_new_attention_order,
+                        ) if not use_spatial_transformer else SpatialTransformer(
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        )
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=dim_head,
+                use_new_attention_order=use_new_attention_order,
+            ) if not use_spatial_transformer else SpatialTransformer(
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(num_res_blocks + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads_upsample,
+                            num_head_channels=dim_head,
+                            use_new_attention_order=use_new_attention_order,
+                        ) if not use_spatial_transformer else SpatialTransformer(
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        )
+                    )
+                if level and i == num_res_blocks:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+            normalization(ch),
+            conv_nd(dims, model_channels, n_embed, 1),
+            #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+        )
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+        self.output_blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+        self.output_blocks.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps=None, context=None, y=None,**kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+
+        if self.num_classes is not None:
+            assert y.shape == (x.shape[0],)
+            emb = emb + self.label_emb(y)
+
+        h = x.type(self.dtype_)
+        for module in self.input_blocks:
+            h = module(h, emb, context)
+            hs.append(h)
+        h = self.middle_block(h, emb, context)
+        for module in self.output_blocks:
+            h = torch.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)
+
+
+class EncoderUNetModel(nn.Module):
+    """
+    The half UNet model with attention and timestep embedding.
+    For usage, see UNet.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        pool="adaptive",
+        *args,
+        **kwargs
+    ):
+        super().__init__()
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.use_checkpoint = use_checkpoint
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=num_head_channels,
+                            use_new_attention_order=use_new_attention_order,
+                        )
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=num_head_channels,
+                use_new_attention_order=use_new_attention_order,
+            ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+        self.pool = pool
+        if pool == "adaptive":
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                nn.AdaptiveAvgPool2d((1, 1)),
+                zero_module(conv_nd(dims, ch, out_channels, 1)),
+                nn.Flatten(),
+            )
+        elif pool == "attention":
+            assert num_head_channels != -1
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                AttentionPool2d(
+                    (image_size // ds), ch, num_head_channels, out_channels
+                ),
+            )
+        elif pool == "spatial":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                nn.ReLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        elif pool == "spatial_v2":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                normalization(2048),
+                nn.SiLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        else:
+            raise NotImplementedError(f"Unexpected {pool} pooling")
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :return: an [N x K] Tensor of outputs.
+        """
+        emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+
+        results = []
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb)
+            if self.pool.startswith("spatial"):
+                results.append(h.type(x.dtype).mean(dim=(2, 3)))
+        h = self.middle_block(h, emb)
+        if self.pool.startswith("spatial"):
+            results.append(h.type(x.dtype).mean(dim=(2, 3)))
+            h = torch.cat(results, axis=-1)
+            return self.out(h)
+        else:
+            h = h.type(x.dtype)
+            return self.out(h)
+

src/diffusers/models/vqvae.py

+# pytorch_diffusion + derived encoder decoder
+import math
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from ..configuration_utils import ConfigMixin
+from ..modeling_utils import ModelMixin
+
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
+    return emb
+
+
+def nonlinearity(x):
+    # swish
+    return x * torch.sigmoid(x)
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0, 1, 0, 1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.k = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.v = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = q.reshape(b, c, h * w)
+        q = q.permute(0, 2, 1)  # b,hw,c
+        k = k.reshape(b, c, h * w)  # b,c,hw
+        w_ = torch.bmm(q, k)  # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b, c, h * w)
+        w_ = w_.permute(0, 2, 1)  # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v, w_)  # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b, c, h, w)
+
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+class Model(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        use_timestep=True,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = self.ch * 4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList(
+                [
+                    torch.nn.Linear(self.ch, self.temb_ch),
+                    torch.nn.Linear(self.temb_ch, self.temb_ch),
+                ]
+            )
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(AttnBlock(block_in))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
+        )
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            skip_in = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch * in_ch_mult[i_level]
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in + skip_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(AttnBlock(block_in))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x, t=None):
+        # assert x.shape[2] == x.shape[3] == self.resolution
+
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](torch.cat([h, hs.pop()], dim=1), temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        double_z=True,
+        **ignore_kwargs,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(AttnBlock(block_in))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
+        )
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
+        )
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in, 2 * z_channels if double_z else z_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, x):
+        # assert x.shape[2] == x.shape[3] == self.resolution, "{}, {}, {}".format(x.shape[2], x.shape[3], self.resolution)
+
+        # timestep embedding
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        give_pre_end=False,
+        **ignorekwargs,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,) + tuple(ch_mult)
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+        print("Working with z of shape {} = {} dimensions.".format(self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
+        )
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(AttnBlock(block_in))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, z):
+        # assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class VectorQuantizer(nn.Module):
+    """
+    Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly
+    avoids costly matrix multiplications and allows for post-hoc remapping of indices.
+    """
+
+    # NOTE: due to a bug the beta term was applied to the wrong term. for
+    # backwards compatibility we use the buggy version by default, but you can
+    # specify legacy=False to fix it.
+    def __init__(self, n_e, e_dim, beta, remap=None, unknown_index="random", sane_index_shape=False, legacy=True):
+        super().__init__()
+        self.n_e = n_e
+        self.e_dim = e_dim
+        self.beta = beta
+        self.legacy = legacy
+
+        self.embedding = nn.Embedding(self.n_e, self.e_dim)
+        self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)
+
+        self.remap = remap
+        if self.remap is not None:
+            self.register_buffer("used", torch.tensor(np.load(self.remap)))
+            self.re_embed = self.used.shape[0]
+            self.unknown_index = unknown_index  # "random" or "extra" or integer
+            if self.unknown_index == "extra":
+                self.unknown_index = self.re_embed
+                self.re_embed = self.re_embed + 1
+            print(
+                f"Remapping {self.n_e} indices to {self.re_embed} indices. "
+                f"Using {self.unknown_index} for unknown indices."
+            )
+        else:
+            self.re_embed = n_e
+
+        self.sane_index_shape = sane_index_shape
+
+    def remap_to_used(self, inds):
+        ishape = inds.shape
+        assert len(ishape) > 1
+        inds = inds.reshape(ishape[0], -1)
+        used = self.used.to(inds)
+        match = (inds[:, :, None] == used[None, None, ...]).long()
+        new = match.argmax(-1)
+        unknown = match.sum(2) < 1
+        if self.unknown_index == "random":
+            new[unknown] = torch.randint(0, self.re_embed, size=new[unknown].shape).to(device=new.device)
+        else:
+            new[unknown] = self.unknown_index
+        return new.reshape(ishape)
+
+    def unmap_to_all(self, inds):
+        ishape = inds.shape
+        assert len(ishape) > 1
+        inds = inds.reshape(ishape[0], -1)
+        used = self.used.to(inds)
+        if self.re_embed > self.used.shape[0]:  # extra token
+            inds[inds >= self.used.shape[0]] = 0  # simply set to zero
+        back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds)
+        return back.reshape(ishape)
+
+    def forward(self, z, temp=None, rescale_logits=False, return_logits=False):
+        assert temp is None or temp == 1.0, "Only for interface compatible with Gumbel"
+        assert rescale_logits == False, "Only for interface compatible with Gumbel"
+        assert return_logits == False, "Only for interface compatible with Gumbel"
+        # reshape z -> (batch, height, width, channel) and flatten
+        z = rearrange(z, "b c h w -> b h w c").contiguous()
+        z_flattened = z.view(-1, self.e_dim)
+        # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
+
+        d = (
+            torch.sum(z_flattened**2, dim=1, keepdim=True)
+            + torch.sum(self.embedding.weight**2, dim=1)
+            - 2 * torch.einsum("bd,dn->bn", z_flattened, rearrange(self.embedding.weight, "n d -> d n"))
+        )
+
+        min_encoding_indices = torch.argmin(d, dim=1)
+        z_q = self.embedding(min_encoding_indices).view(z.shape)
+        perplexity = None
+        min_encodings = None
+
+        # compute loss for embedding
+        if not self.legacy:
+            loss = self.beta * torch.mean((z_q.detach() - z) ** 2) + torch.mean((z_q - z.detach()) ** 2)
+        else:
+            loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * torch.mean((z_q - z.detach()) ** 2)
+
+        # preserve gradients
+        z_q = z + (z_q - z).detach()
+
+        # reshape back to match original input shape
+        z_q = rearrange(z_q, "b h w c -> b c h w").contiguous()
+
+        if self.remap is not None:
+            min_encoding_indices = min_encoding_indices.reshape(z.shape[0], -1)  # add batch axis
+            min_encoding_indices = self.remap_to_used(min_encoding_indices)
+            min_encoding_indices = min_encoding_indices.reshape(-1, 1)  # flatten
+
+        if self.sane_index_shape:
+            min_encoding_indices = min_encoding_indices.reshape(z_q.shape[0], z_q.shape[2], z_q.shape[3])
+
+        return z_q, loss, (perplexity, min_encodings, min_encoding_indices)
+
+    def get_codebook_entry(self, indices, shape):
+        # shape specifying (batch, height, width, channel)
+        if self.remap is not None:
+            indices = indices.reshape(shape[0], -1)  # add batch axis
+            indices = self.unmap_to_all(indices)
+            indices = indices.reshape(-1)  # flatten again
+
+        # get quantized latent vectors
+        z_q = self.embedding(indices)
+
+        if shape is not None:
+            z_q = z_q.view(shape)
+            # reshape back to match original input shape
+            z_q = z_q.permute(0, 3, 1, 2).contiguous()
+
+        return z_q
+
+
+class VQModel(ModelMixin, ConfigMixin):
+    def __init__(
+        self,
+        ch,
+        out_ch,
+        num_res_blocks,
+        attn_resolutions,
+        in_channels,
+        resolution,
+        z_channels,
+        n_embed,
+        embed_dim,
+        remap=None,
+        sane_index_shape=False,  # tell vector quantizer to return indices as bhw
+        ch_mult=(1, 2, 4, 8),
+        dropout=0.0,
+        double_z=True,
+        resamp_with_conv=True,
+        give_pre_end=False,
+    ):
+        super().__init__()
+
+        # register all __init__ params with self.register
+        self.register(
+            ch=ch,
+            out_ch=out_ch,
+            num_res_blocks=num_res_blocks,
+            attn_resolutions=attn_resolutions,
+            in_channels=in_channels,
+            resolution=resolution,
+            z_channels=z_channels,
+            n_embed=n_embed,
+            embed_dim=embed_dim,
+            remap=remap,
+            sane_index_shape=sane_index_shape,
+            ch_mult=ch_mult,
+            dropout=dropout,
+            double_z=double_z,
+            resamp_with_conv=resamp_with_conv,
+            give_pre_end=give_pre_end,
+        )
+
+        # pass init params to Encoder
+        self.encoder = Encoder(
+            ch=ch,
+            out_ch=out_ch,
+            num_res_blocks=num_res_blocks,
+            attn_resolutions=attn_resolutions,
+            in_channels=in_channels,
+            resolution=resolution,
+            z_channels=z_channels,
+            ch_mult=ch_mult,
+            dropout=dropout,
+            resamp_with_conv=resamp_with_conv,
+            double_z=double_z,
+            give_pre_end=give_pre_end,
+        )
+
+        self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25, remap=remap, sane_index_shape=sane_index_shape)
+
+        # pass init params to Decoder
+        self.decoder = Decoder(
+            ch=ch,
+            out_ch=out_ch,
+            num_res_blocks=num_res_blocks,
+            attn_resolutions=attn_resolutions,
+            in_channels=in_channels,
+            resolution=resolution,
+            z_channels=z_channels,
+            ch_mult=ch_mult,
+            dropout=dropout,
+            resamp_with_conv=resamp_with_conv,
+            give_pre_end=give_pre_end,
+        )
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, h, force_not_quantize=False):
+        # also go through quantization layer
+        if not force_not_quantize:
+            quant, emb_loss, info = self.quantize(h)
+        else:
+            quant = h
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec

src/diffusers/pipeline_utils.py

@@ -17,6 +17,7 @@
 import importlib
 import os
 from typing import Optional, Union
+
 from huggingface_hub import snapshot_download
 
 from .utils import logging, DIFFUSERS_CACHE
@@ -34,10 +35,13 @@ logger = logging.get_logger(__name__)
 LOADABLE_CLASSES = {
     "diffusers": {
         "ModelMixin": ["save_pretrained", "from_pretrained"],
+        "CLIPTextModel": ["save_pretrained", "from_pretrained"],  # TODO (Anton): move to transformers
         "GaussianDDPMScheduler": ["save_config", "from_config"],
+        "ClassifierFreeGuidanceScheduler": ["save_config", "from_config"],
+        "GlideDDIMScheduler": ["save_config", "from_config"],
     },
     "transformers": {
-        "ModelMixin": ["save_pretrained", "from_pretrained"],
+        "PreTrainedTokenizer": ["save_pretrained", "from_pretrained"],
     },
 }
 
@@ -50,6 +54,10 @@ class DiffusionPipeline(ConfigMixin):
         for name, module in kwargs.items():
             # retrive library
             library = module.__module__.split(".")[0]
+            # if library is not in LOADABLE_CLASSES, then it is a custom module
+            if library not in LOADABLE_CLASSES:
+                library = module.__module__.split(".")[-1]
+
             # retrive class_name
             class_name = module.__class__.__name__
 
@@ -61,7 +69,7 @@ class DiffusionPipeline(ConfigMixin):
             # set models
             setattr(self, name, module)
 
-        register_dict = {"_module" : self.__module__.split(".")[-1] + ".py"}
+        register_dict = {"_module": self.__module__.split(".")[-1] + ".py"}
         self.register(**register_dict)
 
     def save_pretrained(self, save_directory: Union[str, os.PathLike]):
@@ -123,26 +131,41 @@ class DiffusionPipeline(ConfigMixin):
 
         module = config_dict["_module"]
         class_name_ = config_dict["_class_name"]
+        module_candidate = config_dict["_module"]
+        module_candidate_name = module_candidate.replace(".py", "")
 
-        if class_name_ == cls.__name__:
+        # if we load from explicit class, let's use it
+        if cls != DiffusionPipeline:
             pipeline_class = cls
         else:
+            # else we need to load the correct module from the Hub
+            class_name_ = config_dict["_class_name"]
+            module = module_candidate
             pipeline_class = get_class_from_dynamic_module(cached_folder, module, class_name_, cached_folder)
 
         init_dict, _ = pipeline_class.extract_init_dict(config_dict, **kwargs)
 
         init_kwargs = {}
 
-        for name, (library_name, class_name) in init_dict.items():
-            importable_classes = LOADABLE_CLASSES[library_name]
-
-            if library_name == module:
-                # TODO(Suraj)
-                pass
+        # get all importable classes to get the load method name for custom models/components
+        # here we enforce that custom models/components should always subclass from base classes in tansformers and diffusers
+        all_importable_classes = {}
+        for library in LOADABLE_CLASSES:
+            all_importable_classes.update(LOADABLE_CLASSES[library])
 
-            library = importlib.import_module(library_name)
-            class_obj = getattr(library, class_name)
-            class_candidates = {c: getattr(library, c) for c in importable_classes.keys()}
+        for name, (library_name, class_name) in init_dict.items():
+            # if the model is not in diffusers or transformers, we need to load it from the hub
+            # assumes that it's a subclass of ModelMixin
+            if library_name == module_candidate_name:
+                class_obj = get_class_from_dynamic_module(cached_folder, module, class_name, cached_folder)
+                # since it's not from a library, we need to check class candidates for all importable classes
+                importable_classes = all_importable_classes
+                class_candidates = {c: class_obj for c in all_importable_classes}
+            else:
+                library = importlib.import_module(library_name)
+                class_obj = getattr(library, class_name)
+                importable_classes = LOADABLE_CLASSES[library_name]
+                class_candidates = {c: getattr(library, c) for c in importable_classes.keys()}
 
             load_method_name = None
             for class_name, class_candidate in class_candidates.items():

src/diffusers/schedulers/__init__.py

@@ -16,4 +16,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .classifier_free_guidance import ClassifierFreeGuidanceScheduler
 from .gaussian_ddpm import GaussianDDPMScheduler
+from .glide_ddim import GlideDDIMScheduler