init

.gitignore

+*.png
\ No newline at end of file

diffusers-sd2-new/DDIM_diffusers.py

+# import tomesd
+import torch
+import time
+import os
+import pandas as pd
+from diffusers import StableDiffusionPipeline, DDIMScheduler
+#from xformers.ops import MemoryEfficientAttentionFlashAttentionOp, MemoryEfficientAttentionTritonFwdFlashBwOp
+
+import torch._dynamo
+torch._dynamo.config.suppress_errors = True
+
+ddim_sample_path = "./DDIM-sample"
+os.makedirs(ddim_sample_path, exist_ok=True)
+
+generator = torch.manual_seed(2024)
+#model_id = "/diffusers-sd2/stable-diffusion-2-1-base"
+model_id = "/data1/models/stablediffusion/stable-diffusion-2-1-base"
+text_file = "PartiPrompts.tsv"
+
+df = pd.read_csv(text_file, sep='\t')
+prompts = df['Prompt']
+num_inference_steps = 50
+pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
+pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+pipe = pipe.to("cuda")
+
+from DeepCache import DeepCacheSDHelper
+helper = DeepCacheSDHelper(pipe=pipe)
+helper.set_params(
+     cache_interval=2,
+     cache_branch_id=0,
+ )
+helper.enable()
+###################################
+#pipe.unet = torch.compile(pipe.unet,mode="max-autotune-no-cudagraphs")
+# pipe.vae = torch.compile(pipe.vae,mode="max-autotune-no-cudagraphs")
+###################################
+
+base_count = 0
+print("======================================start DDIM ==================================")
+for prompt in prompts:
+    start = time.time()
+    image = pipe(prompt, 512, 512, num_inference_steps=num_inference_steps, num_images_per_prompt=1, generator=generator).images[0]
+    print(f"the {base_count} text-to-image use time {time.time()-start}")
+    image.save(os.path.join(ddim_sample_path, f"{base_count:05}.png"))
+    base_count += 1
+    if base_count == 50:
+        break
+print(f"Your samples are ready and waiting for you here\n{ddim_sample_path} \n"
+          f" \nEnjoy.")

diffusers-sd2-new/DPM_diffusers.py

+#import tomesd
+import torch
+import time
+import os
+import pandas as pd
+from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler, StableDiffusionXLControlNetPipeline, DiffusionPipeline
+#from xformers.ops import MemoryEfficientAttentionFlashAttentionOp, MemoryEfficientAttentionTritonFwdFlashBwOp
+
+# import torch._dynamo
+# torch._dynamo.config.suppress_errors = True
+
+# torch.backends.cuda.matmul.allow_tf32 = True
+# torch.backends.cudnn.allow_tf32 = True
+dpm_sample_path = "./DPM-sample"
+os.makedirs(dpm_sample_path, exist_ok=True)
+
+generator = torch.manual_seed(2024)
+# model_id = "/data1/models/stablediffusion/stable-diffusion-xl-base-1.0"
+model_id = "/data1/models/stablediffusion/stable-diffusion-2-1-base"
+text_file = "PartiPrompts.tsv"
+
+df = pd.read_csv(text_file, sep='\t')
+prompts = df['Prompt']
+num_inference_steps = 20
+pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
+# pipe = StableDiffusionXLControlNetPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
+# pipe = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
+pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+pipe = pipe.to("cuda")
+
+from DeepCache import DeepCacheSDHelper
+helper = DeepCacheSDHelper(pipe=pipe)
+helper.set_params(
+     cache_interval=2,
+     cache_branch_id=0,
+ )
+helper.enable()
+###################################
+#pipe.unet = torch.compile(pipe.unet,mode="max-autotune-no-cudagraphs")
+# pipe.vae = torch.compile(pipe.vae,mode="max-autotune-no-cudagraphs")
+###################################
+
+base_count = 0
+print("======================================start DPM ==================================")
+for prompt in prompts:
+    start = time.time()
+    image = pipe(prompt, 512, 512, num_inference_steps=num_inference_steps, num_images_per_prompt=1, generator=generator).images[0]
+    # image = pipe(prompt).images[0]
+    print(f"the {base_count} text-to-image use time {time.time()-start}")
+    image.save(os.path.join(dpm_sample_path, f"{base_count:05}.png"))
+    base_count += 1
+    if base_count == 50:
+        break
+print(f"Your samples are ready and waiting for you here\n{dpm_sample_path} \n"
+          f" \nEnjoy.")

diffusers-sd2-new/EULER_diffusers.py

+# import tomesd
+import torch
+import time
+import os
+import pandas as pd
+from diffusers import StableDiffusionPipeline, EulerDiscreteScheduler
+#from xformers.ops import MemoryEfficientAttentionFlashAttentionOp, MemoryEfficientAttentionTritonFwdFlashBwOp
+
+import torch._dynamo
+torch._dynamo.config.suppress_errors = True
+
+euler_sample_path = "./Euler-sample"
+os.makedirs(euler_sample_path, exist_ok=True)
+
+generator = torch.manual_seed(2024)
+#model_id = "/diffusers-sd2/stable-diffusion-2-1-base"
+model_id = "/data1/models/stablediffusion/stable-diffusion-2-1-base"
+text_file = "PartiPrompts.tsv"
+
+df = pd.read_csv(text_file, sep='\t')
+prompts = df['Prompt']
+num_inference_steps = 50
+pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+pipe = pipe.to("cuda")
+from DeepCache import DeepCacheSDHelper
+helper = DeepCacheSDHelper(pipe=pipe)
+helper.set_params(
+     cache_interval=2,
+     cache_branch_id=0,
+ )
+helper.enable()
+###################################
+#pipe.unet = torch.compile(pipe.unet,mode="max-autotune-no-cudagraphs")
+#pipe.vae = torch.compile(pipe.vae,mode="max-autotune-no-cudagraphs")
+###################################
+
+base_count = 0
+print("======================================start EULER ==================================")
+for prompt in prompts:
+    start = time.time()
+    image = pipe(prompt, 512, 512, num_inference_steps=num_inference_steps, num_images_per_prompt=1, generator=generator).images[0]
+    print(f"the {base_count} text-to-image use time {time.time()-start}")
+    image.save(os.path.join(euler_sample_path, f"{base_count:05}.png"))
+    base_count += 1
+    if base_count == 50:
+        break
+
+print(f"Your samples are ready and waiting for you here\n{euler_sample_path} \n"
+          f" \nEnjoy.")
+

diffusers-sd2-new/README.md

+- pip uninstall torch -y
+- pip uninstall torchvision -y
+- pip install ./whl/torch-2.1.0+git93ce03f.abi0.dtk2404-cp310-cp310-manylinux2014_x86_64.whl -i https://pypi.doubanio.com/simple
+- pip install ./whl/torchvision-0.16.0+das1.0+gitc9e7141.abi0.dtk2404.torch2.1-cp310-cp310-manylinux2014_x86_64.whl -i https://pypi.doubanio.com/simple
+- pip install -r requirements.txt -i https://pypi.doubanio.com/simple
+
+## 训练启动
+
+- 复制open_clip_pytorch_model.bin 到当前文件下
+- bash sd-run.sh
\ No newline at end of file

diffusers-sd2-new/clip-score.py

+import numpy as np
+import pandas as pd
+import torch
+import sys
+from tqdm import tqdm
+import open_clip
+import argparse, os
+import torch.nn.functional as F
+from PIL import Image
+from torch.utils.data import Dataset
+from transformers import CLIPTokenizer
+
+class TextImagePairDataset_all(Dataset):
+    def __init__(self, text_file, image_dir, tokenizer, transform=None):
+        self.image_dir = image_dir
+        self.text_file = text_file
+        self.tokenizer = tokenizer
+        self.transform = transform
+
+        df = pd.read_csv(text_file, sep='\t')
+        self.image_paths = [os.path.join(image_dir, f"{f:05}.png") for f in range(len(df))]
+        # df = pd.read_csv(text_file, sep='\t')
+        self.prompts = df['Prompt']
+
+        assert len(self.image_paths) == len(self.prompts), "The number of images and texts must be the same."
+
+    def __len__(self):
+        return len(self.image_paths)
+
+    def __getitem__(self, idx):
+        if torch.is_tensor(idx):
+            idx = idx.tolist()
+        image_path = self.image_paths[idx]
+        text = self.prompts[idx]
+        image = Image.open(image_path).convert('RGB')
+        image = self.transform(image) #.unsqueeze(0)
+        tokens = self.tokenizer(text)
+
+        return tokens, image
+
+class TextImagePairDataset(Dataset):  
+    def __init__(self, text_file, image_dir, tokenizer, transform=None):  
+        self.image_dir = image_dir  
+        self.text_file = text_file  
+        self.tokenizer = tokenizer
+        self.transform = transform
+
+        self.image_paths = [os.path.join(image_dir, f) for f in os.listdir(image_dir) if f.endswith(('.png','.jpg','.jpeg','.tiff','.bmp','.gif'))]  
+        self.texts = []  
+        with open(text_file, 'r') as f:  
+            for line in f:  
+                self.texts.append(line.strip())  
+          
+        assert len(self.image_paths) == len(self.texts), "The number of images and texts must be the same."  
+          
+    def __len__(self):  
+        return len(self.image_paths)  
+  
+    def __getitem__(self, idx):
+        if torch.is_tensor(idx):
+            idx = idx.tolist()
+
+        image_path = self.image_paths[idx]
+        text = self.texts[idx]  
+        image = Image.open(image_path) # .convert('RGB')  
+        image = self.transform(image) #.unsqueeze(0)
+        tokens = self.tokenizer(text)
+
+        return tokens, image
+
+def calculate_clip_score(texts_file, images_dir, batch_size, device, num_workers, output):
+    model_clip, _, preprocess_clip = open_clip.create_model_and_transforms('ViT-H-14', device=device, pretrained='laion2b_s32b_b79k')
+    tokenizer = open_clip.get_tokenizer('ViT-H-14')
+    dataset = TextImagePairDataset_all(texts_file, images_dir, tokenizer=tokenizer, transform=preprocess_clip)
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=False, drop_last=False, num_workers=num_workers)
+    all_scores = []
+    all_scores_cpu = []
+    all_similarity = []
+    print(len(dataloader))
+    for texts, imgs in tqdm(dataloader):
+        texts = texts.reshape(texts.shape[1],texts.shape[2]).to(device)
+        texts = texts.to(device)
+        imgs = imgs.to(device)
+        with torch.no_grad():
+           img_fts = model_clip.encode_image(imgs)
+           text_fts = model_clip.encode_text(texts)
+           scores = F.cosine_similarity(img_fts, text_fts).squeeze()
+           all_scores.append(scores)
+    
+    results_name = f"{output}.txt"
+    if os.path.exists(results_name):
+        os.remove(results_name)
+        print("delete old results")
+    for i in range(len(all_scores)):
+        with open(results_name, 'a') as f:
+            f.write(str(all_scores[i].cpu().numpy()) + '\n')
+        all_scores_cpu.append(all_scores[i].cpu().numpy())
+    average_score = np.mean(all_scores_cpu)
+    return average_score
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--texts",
+        type=str,
+        nargs="?",
+        default="./PartiPrompts.tsv", #数据集路径
+        # default="texts/text.txt",
+    )
+    parser.add_argument(
+        "--images",
+        type=str,
+        nargs="?",
+        default="./DPM-sample"  #保存图片的文件夹路径
+    )
+    parser.add_argument(
+        "--output",
+        type=str,
+        nargs="?",
+        default="./DMP_all_scores"  #保存图片的文件夹路径
+    )
+    parser.add_argument(
+        "--batch_size",
+        type=int,
+        default=1
+    )
+    parser.add_argument(
+        "--num_workers",
+        type=int,
+        default=1
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+    )
+    args = parser.parse_args()
+    if args.device is None:
+        device = torch.device('cuda' if (torch.cuda.is_available()) else 'cpu')
+    else:
+        device = torch.device(args.device)
+    clip_score = calculate_clip_score(args.texts, args.images, args.batch_size, device, args.num_workers, args.output)
+    print('CLIP-score: ', clip_score)
+
+if __name__ == '__main__':
+    main()

diffusers-sd2-new/requirements.txt

+accelerate==0.28.0
+certifi==2024.7.4
+charset-normalizer==3.3.2
+DeepCache==0.1.1
+diffusers==0.29.2
+filelock==3.15.4
+fsspec==2024.6.1
+ftfy==6.2.0
+huggingface-hub==0.24.0
+idna==3.7
+importlib_metadata==8.0.0
+Jinja2==3.1.4
+MarkupSafe==2.1.5
+mpmath==1.3.0
+networkx==3.3
+numpy==1.24.0
+open-clip-torch==2.24.0
+packaging==24.1
+pandas==2.2.2
+peft==0.9.0
+pillow==10.4.0
+protobuf==5.27.2
+psutil==6.0.0
+python-dateutil==2.9.0.post0
+pytz==2024.1
+PyYAML==6.0.1
+regex==2024.5.15
+requests==2.32.3
+safetensors==0.4.3
+sentencepiece==0.2.0
+six==1.16.0
+sympy==1.13.0
+timm==1.0.7
+tokenizers==0.15.2
+#torch 
+#torchvision
+tqdm==4.66.4
+transformers==4.38.1
+typing_extensions==4.12.2
+tzdata==2024.1
+urllib3==2.2.2
+wcwidth==0.2.13
+zipp==3.19.2

diffusers-sd2-new/sd-run.sh

+export HIP_VISIBLE_DEVICES=0
+unset ROCBLAS_TENSILE_LIBPATH
+export LD_LIBRARY_PATH=rocblas-install-0513/lib:$LD_LIBRARY_PATH
+# export MIOPEN_PRECISION_FP32_FP32_FP32_TF32_FP32=1
+#  推理
+# hipprof --hip-trace 
+python ./DPM_diffusers.py
+python ./DDIM_diffusers.py
+python ./EULER_diffusers.py
+
+# # clip-score
+
+# python ./clip-score.py --texts ./PartiPrompts.tsv --images DPM-sample --output DMP_all_scores
+# python ./clip-score.py --texts ./PartiPrompts.tsv --images DDIM-sample --output DDIM_all_scores
+# python ./clip-score.py --texts ./PartiPrompts.tsv --images Euler-sample --output EULER_all_scores