test_alt_diffusion_img2img.py

# coding=utf-8
# Copyright 2023 HuggingFace Inc.
#
# 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.

import gc
import random
import unittest

import numpy as np
import torch
from transformers import XLMRobertaTokenizer

from diffusers import (
    AltDiffusionImg2ImgPipeline,
    AutoencoderKL,
    PNDMScheduler,
    UNet2DConditionModel,
)
from diffusers.image_processor import VaeImageProcessor
from diffusers.pipelines.alt_diffusion.modeling_roberta_series import (
    RobertaSeriesConfig,
    RobertaSeriesModelWithTransformation,
)
from diffusers.utils import load_image
from diffusers.utils.testing_utils import (
    enable_full_determinism,
    floats_tensor,
    load_numpy,
    nightly,
    require_torch_gpu,
    torch_device,
)


enable_full_determinism()


class AltDiffusionImg2ImgPipelineFastTests(unittest.TestCase):
    def tearDown(self):
        # clean up the VRAM after each test
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    @property
    def dummy_image(self):
        batch_size = 1
        num_channels = 3
        sizes = (32, 32)

        image = floats_tensor((batch_size, num_channels) + sizes, rng=random.Random(0)).to(torch_device)
        return image

    @property
    def dummy_cond_unet(self):
        torch.manual_seed(0)
        model = UNet2DConditionModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
        return model

    @property
    def dummy_vae(self):
        torch.manual_seed(0)
        model = AutoencoderKL(
            block_out_channels=[32, 64],
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=4,
        )
        return model

    @property
    def dummy_text_encoder(self):
        torch.manual_seed(0)
        config = RobertaSeriesConfig(
            hidden_size=32,
            project_dim=32,
            intermediate_size=37,
            layer_norm_eps=1e-05,
            num_attention_heads=4,
            num_hidden_layers=5,
            pad_token_id=1,
            vocab_size=5006,
        )
        return RobertaSeriesModelWithTransformation(config)

    @property
    def dummy_extractor(self):
        def extract(*args, **kwargs):
            class Out:
                def __init__(self):
                    self.pixel_values = torch.ones([0])

                def to(self, device):
                    self.pixel_values.to(device)
                    return self

            return Out()

        return extract

    def test_stable_diffusion_img2img_default_case(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = XLMRobertaTokenizer.from_pretrained("hf-internal-testing/tiny-xlm-roberta")
        tokenizer.model_max_length = 77

        init_image = self.dummy_image.to(device)
        init_image = init_image / 2 + 0.5

        # make sure here that pndm scheduler skips prk
        alt_pipe = AltDiffusionImg2ImgPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
            image_encoder=None,
        )
        alt_pipe.image_processor = VaeImageProcessor(vae_scale_factor=alt_pipe.vae_scale_factor, do_normalize=True)
        alt_pipe = alt_pipe.to(device)
        alt_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.Generator(device=device).manual_seed(0)
        output = alt_pipe(
            [prompt],
            generator=generator,
            guidance_scale=6.0,
            num_inference_steps=2,
            output_type="np",
            image=init_image,
        )

        image = output.images

        generator = torch.Generator(device=device).manual_seed(0)
        image_from_tuple = alt_pipe(
            [prompt],
            generator=generator,
            guidance_scale=6.0,
            num_inference_steps=2,
            output_type="np",
            image=init_image,
            return_dict=False,
        )[0]

        image_slice = image[0, -3:, -3:, -1]
        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]

        assert image.shape == (1, 32, 32, 3)
        expected_slice = np.array([0.4427, 0.3731, 0.4249, 0.4941, 0.4546, 0.4148, 0.4193, 0.4666, 0.4499])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 5e-3
        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 5e-3

    @unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
    def test_stable_diffusion_img2img_fp16(self):
        """Test that stable diffusion img2img works with fp16"""
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = XLMRobertaTokenizer.from_pretrained("hf-internal-testing/tiny-xlm-roberta")
        tokenizer.model_max_length = 77

        init_image = self.dummy_image.to(torch_device)

        # put models in fp16
        unet = unet.half()
        vae = vae.half()
        bert = bert.half()

        # make sure here that pndm scheduler skips prk
        alt_pipe = AltDiffusionImg2ImgPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
            image_encoder=None,
        )
        alt_pipe.image_processor = VaeImageProcessor(vae_scale_factor=alt_pipe.vae_scale_factor, do_normalize=False)
        alt_pipe = alt_pipe.to(torch_device)
        alt_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.manual_seed(0)
        image = alt_pipe(
            [prompt],
            generator=generator,
            num_inference_steps=2,
            output_type="np",
            image=init_image,
        ).images

        assert image.shape == (1, 32, 32, 3)

    @unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
    def test_stable_diffusion_img2img_pipeline_multiple_of_8(self):
        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/img2img/sketch-mountains-input.jpg"
        )
        # resize to resolution that is divisible by 8 but not 16 or 32
        init_image = init_image.resize((760, 504))

        model_id = "BAAI/AltDiffusion"
        pipe = AltDiffusionImg2ImgPipeline.from_pretrained(
            model_id,
            safety_checker=None,
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "A fantasy landscape, trending on artstation"

        generator = torch.manual_seed(0)
        output = pipe(
            prompt=prompt,
            image=init_image,
            strength=0.75,
            guidance_scale=7.5,
            generator=generator,
            output_type="np",
        )
        image = output.images[0]

        image_slice = image[255:258, 383:386, -1]

        assert image.shape == (504, 760, 3)
        expected_slice = np.array([0.9358, 0.9397, 0.9599, 0.9901, 1.0000, 1.0000, 0.9882, 1.0000, 1.0000])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2


@nightly
@require_torch_gpu
class AltDiffusionImg2ImgPipelineIntegrationTests(unittest.TestCase):
    def tearDown(self):
        # clean up the VRAM after each test
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def test_stable_diffusion_img2img_pipeline_default(self):
        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/img2img/sketch-mountains-input.jpg"
        )
        init_image = init_image.resize((768, 512))
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/img2img/fantasy_landscape_alt.npy"
        )

        model_id = "BAAI/AltDiffusion"
        pipe = AltDiffusionImg2ImgPipeline.from_pretrained(
            model_id,
            safety_checker=None,
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "A fantasy landscape, trending on artstation"

        generator = torch.manual_seed(0)
        output = pipe(
            prompt=prompt,
            image=init_image,
            strength=0.75,
            guidance_scale=7.5,
            generator=generator,
            output_type="np",
        )
        image = output.images[0]

        assert image.shape == (512, 768, 3)
        # img2img is flaky across GPUs even in fp32, so using MAE here
        assert np.abs(expected_image - image).max() < 1e-2