dynamicrafter

scripts/evaluation/funcs.py

+import os, sys, glob
+import numpy as np
+from collections import OrderedDict
+from decord import VideoReader, cpu
+import cv2
+import torch
+import torchvision
+sys.path.insert(1, os.path.join(sys.path[0], '..', '..'))
+from lvdm.models.samplers.ddim import DDIMSampler
+from einops import rearrange
+
+
+def batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\
+                        cfg_scale=1.0, temporal_cfg_scale=None, **kwargs):
+    ddim_sampler = DDIMSampler(model)
+    uncond_type = model.uncond_type
+    batch_size = noise_shape[0]
+    fs = cond["fs"]
+    del cond["fs"]
+    if noise_shape[-1] == 32:
+        timestep_spacing = "uniform"
+        guidance_rescale = 0.0
+    else:
+        timestep_spacing = "uniform_trailing"
+        guidance_rescale = 0.7
+    ## construct unconditional guidance
+    if cfg_scale != 1.0:
+        if uncond_type == "empty_seq":
+            prompts = batch_size * [""]
+            #prompts = N * T * [""]  ## if is_imgbatch=True
+            uc_emb = model.get_learned_conditioning(prompts)
+        elif uncond_type == "zero_embed":
+            c_emb = cond["c_crossattn"][0] if isinstance(cond, dict) else cond
+            uc_emb = torch.zeros_like(c_emb)
+                
+        ## process image embedding token
+        if hasattr(model, 'embedder'):
+            uc_img = torch.zeros(noise_shape[0],3,224,224).to(model.device)
+            ## img: b c h w >> b l c
+            uc_img = model.embedder(uc_img)
+            uc_img = model.image_proj_model(uc_img)
+            uc_emb = torch.cat([uc_emb, uc_img], dim=1)
+        
+        if isinstance(cond, dict):
+            uc = {key:cond[key] for key in cond.keys()}
+            uc.update({'c_crossattn': [uc_emb]})
+        else:
+            uc = uc_emb
+    else:
+        uc = None
+    
+    x_T = None
+    batch_variants = []
+
+    for _ in range(n_samples):
+        if ddim_sampler is not None:
+            kwargs.update({"clean_cond": True})
+            samples, _ = ddim_sampler.sample(S=ddim_steps,
+                                            conditioning=cond,
+                                            batch_size=noise_shape[0],
+                                            shape=noise_shape[1:],
+                                            verbose=False,
+                                            unconditional_guidance_scale=cfg_scale,
+                                            unconditional_conditioning=uc,
+                                            eta=ddim_eta,
+                                            temporal_length=noise_shape[2],
+                                            conditional_guidance_scale_temporal=temporal_cfg_scale,
+                                            x_T=x_T,
+                                            fs=fs,
+                                            timestep_spacing=timestep_spacing,
+                                            guidance_rescale=guidance_rescale,
+                                            precision=16,
+                                            **kwargs
+                                            )
+        ## reconstruct from latent to pixel space
+        batch_images = model.decode_first_stage(samples)
+        batch_variants.append(batch_images)
+    ## batch, <samples>, c, t, h, w
+    batch_variants = torch.stack(batch_variants, dim=1)
+    return batch_variants
+
+
+def get_filelist(data_dir, ext='*'):
+    file_list = glob.glob(os.path.join(data_dir, '*.%s'%ext))
+    file_list.sort()
+    return file_list
+
+def get_dirlist(path):
+    list = []
+    if (os.path.exists(path)):
+        files = os.listdir(path)
+        for file in files:
+            m = os.path.join(path,file)
+            if (os.path.isdir(m)):
+                list.append(m)
+    list.sort()
+    return list
+
+
+def load_model_checkpoint(model, ckpt):
+    def load_checkpoint(model, ckpt, full_strict):
+        state_dict = torch.load(ckpt, map_location="cpu")
+        if "state_dict" in list(state_dict.keys()):
+            state_dict = state_dict["state_dict"]
+            try:
+                model.load_state_dict(state_dict, strict=full_strict)
+            except:
+                ## rename the keys for 256x256 model
+                new_pl_sd = OrderedDict()
+                for k,v in state_dict.items():
+                    new_pl_sd[k] = v
+
+                for k in list(new_pl_sd.keys()):
+                    if "framestride_embed" in k:
+                        new_key = k.replace("framestride_embed", "fps_embedding")
+                        new_pl_sd[new_key] = new_pl_sd[k]
+                        del new_pl_sd[k]
+                model.load_state_dict(new_pl_sd, strict=full_strict)
+        else:
+            ## deepspeed
+            new_pl_sd = OrderedDict()
+            for key in state_dict['module'].keys():
+                new_pl_sd[key[16:]]=state_dict['module'][key]
+            model.load_state_dict(new_pl_sd, strict=full_strict)
+
+        return model
+    load_checkpoint(model, ckpt, full_strict=True)
+    print('>>> model checkpoint loaded.')
+    return model
+
+
+def load_prompts(prompt_file):
+    f = open(prompt_file, 'r')
+    prompt_list = []
+    for idx, line in enumerate(f.readlines()):
+        l = line.strip()
+        if len(l) != 0:
+            prompt_list.append(l)
+        f.close()
+    return prompt_list
+
+
+def load_video_batch(filepath_list, frame_stride, video_size=(256,256), video_frames=16):
+    '''
+    Notice about some special cases:
+    1. video_frames=-1 means to take all the frames (with fs=1)
+    2. when the total video frames is less than required, padding strategy will be used (repreated last frame)
+    '''
+    fps_list = []
+    batch_tensor = []
+    assert frame_stride > 0, "valid frame stride should be a positive interge!"
+    for filepath in filepath_list:
+        padding_num = 0
+        vidreader = VideoReader(filepath, ctx=cpu(0), width=video_size[1], height=video_size[0])
+        fps = vidreader.get_avg_fps()
+        total_frames = len(vidreader)
+        max_valid_frames = (total_frames-1) // frame_stride + 1
+        if video_frames < 0:
+            ## all frames are collected: fs=1 is a must
+            required_frames = total_frames
+            frame_stride = 1
+        else:
+            required_frames = video_frames
+        query_frames = min(required_frames, max_valid_frames)
+        frame_indices = [frame_stride*i for i in range(query_frames)]
+
+        ## [t,h,w,c] -> [c,t,h,w]
+        frames = vidreader.get_batch(frame_indices)
+        frame_tensor = torch.tensor(frames.asnumpy()).permute(3, 0, 1, 2).float()
+        frame_tensor = (frame_tensor / 255. - 0.5) * 2
+        if max_valid_frames < required_frames:
+            padding_num = required_frames - max_valid_frames
+            frame_tensor = torch.cat([frame_tensor, *([frame_tensor[:,-1:,:,:]]*padding_num)], dim=1)
+            print(f'{os.path.split(filepath)[1]} is not long enough: {padding_num} frames padded.')
+        batch_tensor.append(frame_tensor)
+        sample_fps = int(fps/frame_stride)
+        fps_list.append(sample_fps)
+    
+    return torch.stack(batch_tensor, dim=0)
+
+from PIL import Image
+def load_image_batch(filepath_list, image_size=(256,256)):
+    batch_tensor = []
+    for filepath in filepath_list:
+        _, filename = os.path.split(filepath)
+        _, ext = os.path.splitext(filename)
+        if ext == '.mp4':
+            vidreader = VideoReader(filepath, ctx=cpu(0), width=image_size[1], height=image_size[0])
+            frame = vidreader.get_batch([0])
+            img_tensor = torch.tensor(frame.asnumpy()).squeeze(0).permute(2, 0, 1).float()
+        elif ext == '.png' or ext == '.jpg':
+            img = Image.open(filepath).convert("RGB")
+            rgb_img = np.array(img, np.float32)
+            #bgr_img = cv2.imread(filepath, cv2.IMREAD_COLOR)
+            #bgr_img = cv2.cvtColor(bgr_img, cv2.COLOR_BGR2RGB)
+            rgb_img = cv2.resize(rgb_img, (image_size[1],image_size[0]), interpolation=cv2.INTER_LINEAR)
+            img_tensor = torch.from_numpy(rgb_img).permute(2, 0, 1).float()
+        else:
+            print(f'ERROR: <{ext}> image loading only support format: [mp4], [png], [jpg]')
+            raise NotImplementedError
+        img_tensor = (img_tensor / 255. - 0.5) * 2
+        batch_tensor.append(img_tensor)
+    return torch.stack(batch_tensor, dim=0)
+
+
+def save_videos(batch_tensors, savedir, filenames, fps=10):
+    # b,samples,c,t,h,w
+    n_samples = batch_tensors.shape[1]
+    for idx, vid_tensor in enumerate(batch_tensors):
+        video = vid_tensor.detach().cpu()
+        video = torch.clamp(video.float(), -1., 1.)
+        video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w
+        frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n_samples)) for framesheet in video] #[3, 1*h, n*w]
+        grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w]
+        grid = (grid + 1.0) / 2.0
+        grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
+        savepath = os.path.join(savedir, f"{filenames[idx]}.mp4")
+        torchvision.io.write_video(savepath, grid, fps=fps, video_codec='h264', options={'crf': '10'})
+
+
+def get_latent_z(model, videos):
+    b, c, t, h, w = videos.shape
+    x = rearrange(videos, 'b c t h w -> (b t) c h w')
+    z = model.encode_first_stage(x)
+    z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t)
+    return z
\ No newline at end of file

scripts/evaluation/inference.py

+import argparse, os, sys, glob
+import datetime, time
+from omegaconf import OmegaConf
+from tqdm import tqdm
+from einops import rearrange, repeat
+from collections import OrderedDict
+
+import torch
+import torchvision
+import torchvision.transforms as transforms
+from pytorch_lightning import seed_everything
+from PIL import Image
+sys.path.insert(1, os.path.join(sys.path[0], '..', '..'))
+from lvdm.models.samplers.ddim import DDIMSampler
+from lvdm.models.samplers.ddim_multiplecond import DDIMSampler as DDIMSampler_multicond
+from utils.utils import instantiate_from_config
+
+
+def get_filelist(data_dir, postfixes):
+    patterns = [os.path.join(data_dir, f"*.{postfix}") for postfix in postfixes]
+    file_list = []
+    for pattern in patterns:
+        file_list.extend(glob.glob(pattern))
+    file_list.sort()
+    return file_list
+
+def load_model_checkpoint(model, ckpt):
+    state_dict = torch.load(ckpt, map_location="cpu")
+    if "state_dict" in list(state_dict.keys()):
+        state_dict = state_dict["state_dict"]
+        try:
+            model.load_state_dict(state_dict, strict=True)
+        except:
+            ## rename the keys for 256x256 model
+            new_pl_sd = OrderedDict()
+            for k,v in state_dict.items():
+                new_pl_sd[k] = v
+
+            for k in list(new_pl_sd.keys()):
+                if "framestride_embed" in k:
+                    new_key = k.replace("framestride_embed", "fps_embedding")
+                    new_pl_sd[new_key] = new_pl_sd[k]
+                    del new_pl_sd[k]
+            model.load_state_dict(new_pl_sd, strict=True)
+    else:
+        # deepspeed
+        new_pl_sd = OrderedDict()
+        for key in state_dict['module'].keys():
+            new_pl_sd[key[16:]]=state_dict['module'][key]
+        model.load_state_dict(new_pl_sd)
+    print('>>> model checkpoint loaded.')
+    return model
+
+def load_prompts(prompt_file):
+    f = open(prompt_file, 'r')
+    prompt_list = []
+    for idx, line in enumerate(f.readlines()):
+        l = line.strip()
+        if len(l) != 0:
+            prompt_list.append(l)
+        f.close()
+    return prompt_list
+
+def load_data_prompts(data_dir, video_size=(256,256), video_frames=16, gfi=False):
+    transform = transforms.Compose([
+        transforms.Resize(min(video_size)),
+        transforms.CenterCrop(video_size),
+        transforms.ToTensor(),
+        transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))])
+    ## load prompts
+    prompt_file = get_filelist(data_dir, ['txt'])
+    assert len(prompt_file) > 0, "Error: found NO prompt file!"
+    ###### default prompt
+    default_idx = 0
+    default_idx = min(default_idx, len(prompt_file)-1)
+    if len(prompt_file) > 1:
+        print(f"Warning: multiple prompt files exist. The one {os.path.split(prompt_file[default_idx])[1]} is used.")
+    ## only use the first one (sorted by name) if multiple exist
+    
+    ## load video
+    file_list = get_filelist(data_dir, ['jpg', 'png', 'jpeg', 'JPEG', 'PNG'])
+    # assert len(file_list) == n_samples, "Error: data and prompts are NOT paired!"
+    data_list = []
+    filename_list = []
+    prompt_list = load_prompts(prompt_file[default_idx])
+    n_samples = len(prompt_list)
+    for idx in range(n_samples):
+        image = Image.open(file_list[idx]).convert('RGB')
+        image_tensor = transform(image).unsqueeze(1) # [c,1,h,w]
+        frame_tensor = repeat(image_tensor, 'c t h w -> c (repeat t) h w', repeat=video_frames)
+
+        data_list.append(frame_tensor)
+        _, filename = os.path.split(file_list[idx])
+        filename_list.append(filename)
+        
+    return filename_list, data_list, prompt_list
+
+
+def save_results(prompt, samples, filename, fakedir, fps=8, loop=False):
+    filename = filename.split('.')[0]+'.mp4'
+    prompt = prompt[0] if isinstance(prompt, list) else prompt
+
+    ## save video
+    videos = [samples]
+    savedirs = [fakedir]
+    for idx, video in enumerate(videos):
+        if video is None:
+            continue
+        # b,c,t,h,w
+        video = video.detach().cpu()
+        video = torch.clamp(video.float(), -1., 1.)
+        n = video.shape[0]
+        video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w
+        if loop:
+            video = video[:-1,...]
+        
+        frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n), padding=0) for framesheet in video] #[3, 1*h, n*w]
+        grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, h, n*w]
+        grid = (grid + 1.0) / 2.0
+        grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
+        path = os.path.join(savedirs[idx], filename)
+        torchvision.io.write_video(path, grid, fps=fps, video_codec='h264', options={'crf': '10'}) ## crf indicates the quality
+
+
+def save_results_seperate(prompt, samples, filename, fakedir, fps=10, loop=False):
+    prompt = prompt[0] if isinstance(prompt, list) else prompt
+
+    ## save video
+    videos = [samples]
+    savedirs = [fakedir]
+    for idx, video in enumerate(videos):
+        if video is None:
+            continue
+        # b,c,t,h,w
+        video = video.detach().cpu()
+        if loop: # remove the last frame
+            video = video[:,:,:-1,...]
+        video = torch.clamp(video.float(), -1., 1.)
+        n = video.shape[0]
+        for i in range(n):
+            grid = video[i,...]
+            grid = (grid + 1.0) / 2.0
+            grid = (grid * 255).to(torch.uint8).permute(1, 2, 3, 0) #thwc
+            path = os.path.join(savedirs[idx].replace('samples', 'samples_separate'), f'{filename.split(".")[0]}_sample{i}.mp4')
+            torchvision.io.write_video(path, grid, fps=fps, video_codec='h264', options={'crf': '10'})
+
+def get_latent_z(model, videos):
+    b, c, t, h, w = videos.shape
+    x = rearrange(videos, 'b c t h w -> (b t) c h w')
+    z = model.encode_first_stage(x)
+    z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t)
+    return z
+
+
+def image_guided_synthesis(model, prompts, videos, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1., \
+                        unconditional_guidance_scale=1.0, cfg_img=None, fs=None, text_input=False, multiple_cond_cfg=False, loop=False, gfi=False, timestep_spacing='uniform', guidance_rescale=0.0, **kwargs):
+    ddim_sampler = DDIMSampler(model) if not multiple_cond_cfg else DDIMSampler_multicond(model)
+    batch_size = noise_shape[0]
+    fs = torch.tensor([fs] * batch_size, dtype=torch.long, device=model.device)
+
+    if not text_input:
+        prompts = [""]*batch_size
+
+    img = videos[:,:,0] #bchw
+    img_emb = model.embedder(img) ## blc
+    img_emb = model.image_proj_model(img_emb)
+
+    cond_emb = model.get_learned_conditioning(prompts)
+    cond = {"c_crossattn": [torch.cat([cond_emb,img_emb], dim=1)]}
+    if model.model.conditioning_key == 'hybrid':
+        z = get_latent_z(model, videos) # b c t h w
+        if loop or gfi:
+            img_cat_cond = torch.zeros_like(z)
+            img_cat_cond[:,:,0,:,:] = z[:,:,0,:,:]
+            img_cat_cond[:,:,-1,:,:] = z[:,:,-1,:,:]
+        else:
+            img_cat_cond = z[:,:,:1,:,:]
+            img_cat_cond = repeat(img_cat_cond, 'b c t h w -> b c (repeat t) h w', repeat=z.shape[2])
+        cond["c_concat"] = [img_cat_cond] # b c 1 h w
+    
+    if unconditional_guidance_scale != 1.0:
+        if model.uncond_type == "empty_seq":
+            prompts = batch_size * [""]
+            uc_emb = model.get_learned_conditioning(prompts)
+        elif model.uncond_type == "zero_embed":
+            uc_emb = torch.zeros_like(cond_emb)
+        uc_img_emb = model.embedder(torch.zeros_like(img)) ## b l c
+        uc_img_emb = model.image_proj_model(uc_img_emb)
+        uc = {"c_crossattn": [torch.cat([uc_emb,uc_img_emb],dim=1)]}
+        if model.model.conditioning_key == 'hybrid':
+            uc["c_concat"] = [img_cat_cond]
+    else:
+        uc = None
+
+    ## we need one more unconditioning image=yes, text=""
+    if multiple_cond_cfg and cfg_img != 1.0:
+        uc_2 = {"c_crossattn": [torch.cat([uc_emb,img_emb],dim=1)]}
+        if model.model.conditioning_key == 'hybrid':
+            uc_2["c_concat"] = [img_cat_cond]
+        kwargs.update({"unconditional_conditioning_img_nonetext": uc_2})
+    else:
+        kwargs.update({"unconditional_conditioning_img_nonetext": None})
+
+    z0 = None
+    cond_mask = None
+
+    batch_variants = []
+    for _ in range(n_samples):
+
+        if z0 is not None:
+            cond_z0 = z0.clone()
+            kwargs.update({"clean_cond": True})
+        else:
+            cond_z0 = None
+        if ddim_sampler is not None:
+
+            samples, _ = ddim_sampler.sample(S=ddim_steps,
+                                            conditioning=cond,
+                                            batch_size=batch_size,
+                                            shape=noise_shape[1:],
+                                            verbose=False,
+                                            unconditional_guidance_scale=unconditional_guidance_scale,
+                                            unconditional_conditioning=uc,
+                                            eta=ddim_eta,
+                                            cfg_img=cfg_img, 
+                                            mask=cond_mask,
+                                            x0=cond_z0,
+                                            fs=fs,
+                                            timestep_spacing=timestep_spacing,
+                                            guidance_rescale=guidance_rescale,
+                                            **kwargs
+                                            )
+
+        ## reconstruct from latent to pixel space
+        batch_images = model.decode_first_stage(samples)
+        batch_variants.append(batch_images)
+    ## variants, batch, c, t, h, w
+    batch_variants = torch.stack(batch_variants)
+    return batch_variants.permute(1, 0, 2, 3, 4, 5)
+
+
+def run_inference(args, gpu_num, gpu_no):
+    ## model config
+    config = OmegaConf.load(args.config)
+    model_config = config.pop("model", OmegaConf.create())
+    
+    ## set use_checkpoint as False as when using deepspeed, it encounters an error "deepspeed backend not set"
+    model_config['params']['unet_config']['params']['use_checkpoint'] = False
+    model = instantiate_from_config(model_config)
+    model = model.cuda(gpu_no)
+    model.perframe_ae = args.perframe_ae
+    assert os.path.exists(args.ckpt_path), "Error: checkpoint Not Found!"
+    model = load_model_checkpoint(model, args.ckpt_path)
+    model.eval()
+
+    ## run over data
+    assert (args.height % 16 == 0) and (args.width % 16 == 0), "Error: image size [h,w] should be multiples of 16!"
+    assert args.bs == 1, "Current implementation only support [batch size = 1]!"
+    ## latent noise shape
+    h, w = args.height // 8, args.width // 8
+    channels = model.model.diffusion_model.out_channels
+    n_frames = args.video_length
+    print(f'Inference with {n_frames} frames')
+    noise_shape = [args.bs, channels, n_frames, h, w]
+
+    fakedir = os.path.join(args.savedir, "samples")
+    fakedir_separate = os.path.join(args.savedir, "samples_separate")
+
+    # os.makedirs(fakedir, exist_ok=True)
+    os.makedirs(fakedir_separate, exist_ok=True)
+
+    ## prompt file setting
+    assert os.path.exists(args.prompt_dir), "Error: prompt file Not Found!"
+    filename_list, data_list, prompt_list = load_data_prompts(args.prompt_dir, video_size=(args.height, args.width), video_frames=n_frames, gfi=args.gfi)
+    num_samples = len(prompt_list)
+    samples_split = num_samples // gpu_num
+    print('Prompts testing [rank:%d] %d/%d samples loaded.'%(gpu_no, samples_split, num_samples))
+    #indices = random.choices(list(range(0, num_samples)), k=samples_per_device)
+    indices = list(range(samples_split*gpu_no, samples_split*(gpu_no+1)))
+    prompt_list_rank = [prompt_list[i] for i in indices]
+    data_list_rank = [data_list[i] for i in indices]
+    filename_list_rank = [filename_list[i] for i in indices]
+
+    start = time.time()
+    with torch.no_grad(), torch.cuda.amp.autocast():
+        for idx, indice in tqdm(enumerate(range(0, len(prompt_list_rank), args.bs)), desc='Sample Batch'):
+            prompts = prompt_list_rank[indice:indice+args.bs]
+            videos = data_list_rank[indice:indice+args.bs]
+            filenames = filename_list_rank[indice:indice+args.bs]
+            if isinstance(videos, list):
+                videos = torch.stack(videos, dim=0).to("cuda")
+            else:
+                videos = videos.unsqueeze(0).to("cuda")
+
+            batch_samples = image_guided_synthesis(model, prompts, videos, noise_shape, args.n_samples, args.ddim_steps, args.ddim_eta, \
+                                args.unconditional_guidance_scale, args.cfg_img, args.frame_stride, args.text_input, args.multiple_cond_cfg, args.loop, args.gfi, args.timestep_spacing, args.guidance_rescale)
+
+            ## save each example individually
+            for nn, samples in enumerate(batch_samples):
+                ## samples : [n_samples,c,t,h,w]
+                prompt = prompts[nn]
+                filename = filenames[nn]
+                # save_results(prompt, samples, filename, fakedir, fps=8, loop=args.loop)
+                save_results_seperate(prompt, samples, filename, fakedir, fps=8, loop=args.loop)
+
+    print(f"Saved in {args.savedir}. Time used: {(time.time() - start):.2f} seconds")
+
+
+def get_parser():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--savedir", type=str, default=None, help="results saving path")
+    parser.add_argument("--ckpt_path", type=str, default=None, help="checkpoint path")
+    parser.add_argument("--config", type=str, help="config (yaml) path")
+    parser.add_argument("--prompt_dir", type=str, default=None, help="a data dir containing videos and prompts")
+    parser.add_argument("--n_samples", type=int, default=1, help="num of samples per prompt",)
+    parser.add_argument("--ddim_steps", type=int, default=50, help="steps of ddim if positive, otherwise use DDPM",)
+    parser.add_argument("--ddim_eta", type=float, default=1.0, help="eta for ddim sampling (0.0 yields deterministic sampling)",)
+    parser.add_argument("--bs", type=int, default=1, help="batch size for inference, should be one")
+    parser.add_argument("--height", type=int, default=512, help="image height, in pixel space")
+    parser.add_argument("--width", type=int, default=512, help="image width, in pixel space")
+    parser.add_argument("--frame_stride", type=int, default=3, help="frame stride control for 256 model (larger->larger motion), FPS control for 512 or 1024 model (smaller->larger motion)")
+    parser.add_argument("--unconditional_guidance_scale", type=float, default=1.0, help="prompt classifier-free guidance")
+    parser.add_argument("--seed", type=int, default=123, help="seed for seed_everything")
+    parser.add_argument("--video_length", type=int, default=16, help="inference video length")
+    parser.add_argument("--negative_prompt", action='store_true', default=False, help="negative prompt")
+    parser.add_argument("--text_input", action='store_true', default=False, help="input text to I2V model or not")
+    parser.add_argument("--multiple_cond_cfg", action='store_true', default=False, help="use multi-condition cfg or not")
+    parser.add_argument("--cfg_img", type=float, default=None, help="guidance scale for image conditioning")
+    parser.add_argument("--timestep_spacing", type=str, default="uniform", help="The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.")
+    parser.add_argument("--guidance_rescale", type=float, default=0.0, help="guidance rescale in [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://huggingface.co/papers/2305.08891)")
+    parser.add_argument("--perframe_ae", action='store_true', default=False, help="if we use per-frame AE decoding, set it to True to save GPU memory, especially for the model of 576x1024")
+
+    ## currently not support looping video and generative frame interpolation
+    parser.add_argument("--loop", action='store_true', default=False, help="generate looping videos or not")
+    parser.add_argument("--gfi", action='store_true', default=False, help="generate generative frame interpolation (gfi) or not")
+    return parser
+
+
+if __name__ == '__main__':
+    now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
+    print("@DynamiCrafter cond-Inference: %s"%now)
+    parser = get_parser()
+    args = parser.parse_args()
+    
+    seed_everything(args.seed)
+    rank, gpu_num = 0, 1
+    run_inference(args, gpu_num, rank)
\ No newline at end of file

scripts/gradio/i2v_test.py

+import os
+import time
+from omegaconf import OmegaConf
+import torch
+from scripts.evaluation.funcs import load_model_checkpoint, save_videos, batch_ddim_sampling, get_latent_z
+from utils.utils import instantiate_from_config
+from huggingface_hub import hf_hub_download
+from einops import repeat
+import torchvision.transforms as transforms
+from pytorch_lightning import seed_everything
+
+
+class Image2Video():
+    def __init__(self,result_dir='./tmp/',gpu_num=1,resolution='256_256', **kwargs) -> None:
+        self.resolution = (int(resolution.split('_')[0]), int(resolution.split('_')[1])) #hw
+        self.download_model()
+        
+        self.result_dir = result_dir
+        if not os.path.exists(self.result_dir):
+            os.mkdir(self.result_dir)
+        ckpt_path='checkpoints/dynamicrafter_'+resolution.split('_')[1]+'_v1/model.ckpt'
+        config_file='configs/inference_'+resolution.split('_')[1]+'_v1.0.yaml'
+        config = OmegaConf.load(config_file)
+        model_config = config.pop("model", OmegaConf.create())
+        model_config['params']['unet_config']['params']['use_checkpoint']=False   
+        model_list = []
+        for gpu_id in range(gpu_num):
+            model = instantiate_from_config(model_config)
+            # model = model.cuda(gpu_id)
+            assert os.path.exists(ckpt_path), "Error: checkpoint Not Found!"
+            model = load_model_checkpoint(model, ckpt_path)
+            model.eval()
+            model_list.append(model)
+        self.model_list = model_list
+        self.save_fps = 8
+        
+        self.kwargs = kwargs
+
+    def get_image(self, image, prompt, steps=50, cfg_scale=7.5, eta=1.0, fs=3, seed=123):
+        seed_everything(seed)
+        transform = transforms.Compose([
+            transforms.Resize(min(self.resolution)),
+            transforms.CenterCrop(self.resolution),
+            ])
+        torch.cuda.empty_cache()
+        print('start:', prompt, time.strftime('%Y-%m-%d %H:%M:%S',time.localtime(time.time())))
+        start = time.time()
+        gpu_id=0
+        if steps > 60:
+            steps = 60 
+        model = self.model_list[gpu_id]
+        model = model.cuda()
+        batch_size=1
+        channels = model.model.diffusion_model.out_channels
+        frames = model.temporal_length
+        h, w = self.resolution[0] // 8, self.resolution[1] // 8
+        noise_shape = [batch_size, channels, frames, h, w]
+
+        # text cond
+        text_emb = model.get_learned_conditioning([prompt])
+
+        # img cond
+        img_tensor = torch.from_numpy(image).permute(2, 0, 1).float().to(model.device)
+        img_tensor = (img_tensor / 255. - 0.5) * 2
+
+        image_tensor_resized = transform(img_tensor) #3,h,w
+        videos = image_tensor_resized.unsqueeze(0) # bchw
+        
+        z = get_latent_z(model, videos.unsqueeze(2)) #bc,1,hw
+        
+        img_tensor_repeat = repeat(z, 'b c t h w -> b c (repeat t) h w', repeat=frames)
+
+        cond_images = model.embedder(img_tensor.unsqueeze(0)) ## blc
+        img_emb = model.image_proj_model(cond_images)
+
+        imtext_cond = torch.cat([text_emb, img_emb], dim=1)
+
+        fs = torch.tensor([fs], dtype=torch.long, device=model.device)
+        cond = {"c_crossattn": [imtext_cond], "fs": fs, "c_concat": [img_tensor_repeat]}
+        
+        ## inference
+        with torch.no_grad(), torch.cuda.amp.autocast():
+            batch_samples = batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=steps, ddim_eta=eta, cfg_scale=cfg_scale)
+        ## b,samples,c,t,h,w
+        prompt_str = prompt.replace("/", "_slash_") if "/" in prompt else prompt
+        prompt_str = prompt_str.replace(" ", "_") if " " in prompt else prompt_str
+        prompt_str=prompt_str[:40]
+        if len(prompt_str) == 0:
+            prompt_str = 'empty_prompt'
+
+        save_videos(batch_samples, self.result_dir, filenames=[prompt_str], fps=self.save_fps)
+        print(f"Saved in {prompt_str}. Time used: {(time.time() - start):.2f} seconds")
+        model = model.cpu()
+        return os.path.join(self.result_dir, f"{prompt_str}.mp4")
+    
+    def download_model(self):
+        REPO_ID = 'Doubiiu/DynamiCrafter_'+str(self.resolution[1]) if self.resolution[1]!=256 else 'Doubiiu/DynamiCrafter'
+        filename_list = ['model.ckpt']
+        if not os.path.exists('./checkpoints/dynamicrafter_'+str(self.resolution[1])+'_v1/'):
+            os.makedirs('./checkpoints/dynamicrafter_'+str(self.resolution[1])+'_v1/')
+        for filename in filename_list:
+            local_file = os.path.join('./checkpoints/dynamicrafter_'+str(self.resolution[1])+'_v1/', filename)
+            if not os.path.exists(local_file):
+                hf_hub_download(repo_id=REPO_ID, filename=filename, local_dir='./checkpoints/dynamicrafter_'+str(self.resolution[1])+'_v1/', local_dir_use_symlinks=False)
+    
+if __name__ == '__main__':
+    i2v = Image2Video()
+    video_path = i2v.get_image('prompts/art.png','man fishing in a boat at sunset')
+    print('done', video_path)
\ No newline at end of file

scripts/run.sh

+version=$1 ##1024, 512, 256
+seed=123
+name=dynamicrafter_$1_seed${seed}
+
+ckpt=checkpoints/dynamicrafter_$1_v1/model.ckpt
+config=configs/inference_$1_v1.0.yaml
+
+prompt_dir=prompts/$1/
+res_dir="results"
+
+if [ "$1" == "256" ]; then
+    H=256
+    FS=3  ## This model adopts frame stride=3, range recommended: 1-6 (larger value -> larger motion)
+elif [ "$1" == "512" ]; then
+    H=320
+    FS=24 ## This model adopts FPS=24, range recommended: 15-30 (smaller value -> larger motion)
+elif [ "$1" == "1024" ]; then
+    H=576
+    FS=10 ## This model adopts FPS=10, range recommended: 15-5 (smaller value -> larger motion)
+else
+    echo "Invalid input. Please enter 256, 512, or 1024."
+    exit 1
+fi
+
+if [ "$1" == "256" ]; then
+CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/inference.py \
+--seed ${seed} \
+--ckpt_path $ckpt \
+--config $config \
+--savedir $res_dir/$name \
+--n_samples 1 \
+--bs 1 --height ${H} --width $1 \
+--unconditional_guidance_scale 7.5 \
+--ddim_steps 50 \
+--ddim_eta 1.0 \
+--prompt_dir $prompt_dir \
+--text_input \
+--video_length 16 \
+--frame_stride ${FS}
+else
+CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/inference.py \
+--seed ${seed} \
+--ckpt_path $ckpt \
+--config $config \
+--savedir $res_dir/$name \
+--n_samples 1 \
+--bs 1 --height ${H} --width $1 \
+--unconditional_guidance_scale 7.5 \
+--ddim_steps 50 \
+--ddim_eta 1.0 \
+--prompt_dir $prompt_dir \
+--text_input \
+--video_length 16 \
+--frame_stride ${FS} \
+--timestep_spacing 'uniform_trailing' --guidance_rescale 0.7 --perframe_ae
+fi
+
+
+## multi-cond CFG: the <unconditional_guidance_scale> is s_txt, <cfg_img> is s_img
+#--multiple_cond_cfg --cfg_img 7.5
+#--loop
\ No newline at end of file

scripts/run_mp.sh

+version=$1 ##1024, 512, 256
+seed=123
+
+name=dynamicrafter_$1_mp_seed${seed}
+
+ckpt=checkpoints/dynamicrafter_$1_v1/model.ckpt
+config=configs/inference_$1_v1.0.yaml
+
+prompt_dir=prompts/$1/
+res_dir="results"
+
+if [ "$1" == "256" ]; then
+    H=256
+    FS=3  ## This model adopts frame stride=3
+elif [ "$1" == "512" ]; then
+    H=320
+    FS=24 ## This model adopts FPS=24
+elif [ "$1" == "1024" ]; then
+    H=576
+    FS=10 ## This model adopts FPS=10
+else
+    echo "Invalid input. Please enter 256, 512, or 1024."
+    exit 1
+fi
+
+# if [ "$1" == "256" ]; then
+# CUDA_VISIBLE_DEVICES=2 python3 scripts/evaluation/inference.py \
+# --seed 123 \
+# --ckpt_path $ckpt \
+# --config $config \
+# --savedir $res_dir/$name \
+# --n_samples 1 \
+# --bs 1 --height ${H} --width $1 \
+# --unconditional_guidance_scale 7.5 \
+# --ddim_steps 50 \
+# --ddim_eta 1.0 \
+# --prompt_dir $prompt_dir \
+# --text_input \
+# --video_length 16 \
+# --frame_stride ${FS}
+# else
+# CUDA_VISIBLE_DEVICES=2 python3 scripts/evaluation/inference.py \
+# --seed 123 \
+# --ckpt_path $ckpt \
+# --config $config \
+# --savedir $res_dir/$name \
+# --n_samples 1 \
+# --bs 1 --height ${H} --width $1 \
+# --unconditional_guidance_scale 7.5 \
+# --ddim_steps 50 \
+# --ddim_eta 1.0 \
+# --prompt_dir $prompt_dir \
+# --text_input \
+# --video_length 16 \
+# --frame_stride ${FS} \
+# --timestep_spacing 'uniform_trailing' --guidance_rescale 0.7
+# fi
+
+
+## multi-cond CFG: the <unconditional_guidance_scale> is s_txt, <cfg_img> is s_img
+#--multiple_cond_cfg --cfg_img 7.5
+#--loop
+
+## inference using single node with multi-GPUs:
+if [ "$1" == "256" ]; then
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m torch.distributed.launch \
+--nproc_per_node=8 --nnodes=1 --master_addr=127.0.0.1 --master_port=23456 --node_rank=0 \
+scripts/evaluation/ddp_wrapper.py \
+--module 'inference' \
+--seed ${seed} \
+--ckpt_path $ckpt \
+--config $config \
+--savedir $res_dir/$name \
+--n_samples 1 \
+--bs 1 --height ${H} --width $1 \
+--unconditional_guidance_scale 7.5 \
+--ddim_steps 50 \
+--ddim_eta 1.0 \
+--prompt_dir $prompt_dir \
+--text_input \
+--video_length 16 \
+--frame_stride ${FS}
+else
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m torch.distributed.launch \
+--nproc_per_node=8 --nnodes=1 --master_addr=127.0.0.1 --master_port=23456 --node_rank=0 \
+scripts/evaluation/ddp_wrapper.py \
+--module 'inference' \
+--seed ${seed} \
+--ckpt_path $ckpt \
+--config $config \
+--savedir $res_dir/$name \
+--n_samples 1 \
+--bs 1 --height ${H} --width $1 \
+--unconditional_guidance_scale 7.5 \
+--ddim_steps 50 \
+--ddim_eta 1.0 \
+--prompt_dir $prompt_dir \
+--text_input \
+--video_length 16 \
+--frame_stride ${FS} \
+--timestep_spacing 'uniform_trailing' --guidance_rescale 0.7 --perframe_ae
+fi
\ No newline at end of file

utils/utils.py

+import importlib
+import numpy as np
+import cv2
+import torch
+import torch.distributed as dist
+
+
+def count_params(model, verbose=False):
+    total_params = sum(p.numel() for p in model.parameters())
+    if verbose:
+        print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.")
+    return total_params
+
+
+def check_istarget(name, para_list):
+    """ 
+    name: full name of source para
+    para_list: partial name of target para 
+    """
+    istarget=False
+    for para in para_list:
+        if para in name:
+            return True
+    return istarget
+
+
+def instantiate_from_config(config):
+    if not "target" in config:
+        if config == '__is_first_stage__':
+            return None
+        elif config == "__is_unconditional__":
+            return None
+        raise KeyError("Expected key `target` to instantiate.")
+    return get_obj_from_str(config["target"])(**config.get("params", dict()))
+
+
+def get_obj_from_str(string, reload=False):
+    module, cls = string.rsplit(".", 1)
+    if reload:
+        module_imp = importlib.import_module(module)
+        importlib.reload(module_imp)
+    return getattr(importlib.import_module(module, package=None), cls)
+
+
+def load_npz_from_dir(data_dir):
+    data = [np.load(os.path.join(data_dir, data_name))['arr_0'] for data_name in os.listdir(data_dir)]
+    data = np.concatenate(data, axis=0)
+    return data
+
+
+def load_npz_from_paths(data_paths):
+    data = [np.load(data_path)['arr_0'] for data_path in data_paths]
+    data = np.concatenate(data, axis=0)
+    return data   
+
+
+def resize_numpy_image(image, max_resolution=512 * 512, resize_short_edge=None):
+    h, w = image.shape[:2]
+    if resize_short_edge is not None:
+        k = resize_short_edge / min(h, w)
+    else:
+        k = max_resolution / (h * w)
+        k = k**0.5
+    h = int(np.round(h * k / 64)) * 64
+    w = int(np.round(w * k / 64)) * 64
+    image = cv2.resize(image, (w, h), interpolation=cv2.INTER_LANCZOS4)
+    return image
+
+
+def setup_dist(args):
+    if dist.is_initialized():
+        return
+    torch.cuda.set_device(args.local_rank)
+    torch.distributed.init_process_group(
+        'nccl',
+        init_method='env://'
+    )
\ No newline at end of file