import argparse import contextlib import gc import logging import math import os import random import shutil import ast import accelerate import numpy as np import torch import torch.nn.functional as F import transformers from PIL import Image from torchvision import transforms from tqdm.auto import tqdm import json import pickle from transformers import AutoTokenizer, PretrainedConfig import pandas as pd class RealV40CustomDataset(torch.utils.data.Dataset): def __init__(self, tokenizer, image_root_path, condition_root, txt_json_root, resolution=512, drop_txt_prob = 0.02, drop_all_prob = 0.02, ): self.local_tasks = 'segmentation' self.tokenizer = tokenizer self.resolution = resolution self.transform = transforms.Compose([ transforms.Resize(self.resolution, interpolation=transforms.InterpolationMode.BILINEAR), transforms.CenterCrop(self.resolution), transforms.ToTensor(), transforms.Normalize([0.5], [0.5]), ]) self.conditioning_image_transforms = transforms.Compose([ transforms.Resize(self.resolution, interpolation=transforms.InterpolationMode.BILINEAR), transforms.CenterCrop(self.resolution), transforms.ToTensor(), ]) self.img_path = image_root_path # self.img_name_list = os.listdir(self.img_path) self.txt_dict_global = json.load(open(txt_json_root['global'], 'r')) self.txt_dict_global = {os.path.basename(key):value for key, value in self.txt_dict_global.items()} self.txt_dict_shape = json.load(open(txt_json_root['shape'], 'r')) self.txt_dict_shape = {os.path.basename(key):value for key, value in self.txt_dict_shape.items()} self.condition_root = condition_root img_set = set(os.listdir(self.img_path)).intersection(set(os.listdir(self.condition_root))) img_set = img_set.intersection(set(self.txt_dict_global.keys())) img_set = img_set.intersection(set(self.txt_dict_shape.keys())) error_file = ['RealisticV4-0_90994_27412.png', 'RealisticV4-0_99115_23851.png', 'RealisticV4-0_82840_14492.png', 'RealisticV4-0_26257_6936.png', 'RealisticV4-0_915_24610.png', 'RealisticV4-0_49242_25603.png', 'RealisticV4-0_93324_19139.png', 'RealisticV4-0_79137_2021.png'] img_set = img_set - set(error_file) self.img_name_list = list(img_set) self.drop_txt_prob = drop_txt_prob self.drop_all_prob = drop_all_prob def __getitem__(self, index): if 1: img_name = self.img_name_list[index] # read image img = Image.open(os.path.join(self.img_path, img_name)).convert("RGB") img = self.transform(img) condition_image = Image.open(os.path.join(self.condition_root, img_name.replace('.jpg', '.png'))).convert("RGB") condition_image = self.conditioning_image_transforms(condition_image) text_global = self.txt_dict_global[img_name] drop_all_local_prob = 0.0 # randomly drop rand_num = random.random() text_shape = self.txt_dict_shape[img_name] if rand_num < self.drop_all_prob: text_shape = '' drop_all_local_prob = 1.0 elif rand_num < self.drop_all_prob + self.drop_txt_prob: text_shape = '' # drop conditions if drop_all_local_prob == 1.0: condition_image = torch.zeros_like(condition_image) # c_text = 'A face conditioned on segmentation' # text = text + " " + c_text.strip() # '/group_meiyan/meiyan_data/CelebA-HQ/CelebA-HQ-img/7738.jpg' return { "pixel_values": img, "img_name":img_name, "conditioning_pixel_values" : [condition_image], "input_ids_global": self.tokenize_captions([text_global]), "input_ids_shape": self.tokenize_captions([text_shape]), } # except: # return None def __len__(self): return len(self.img_name_list) def tokenize_captions(self, text, is_train=True): captions = [] for caption in text: if isinstance(caption, str): captions.append(caption) elif isinstance(caption, (list, np.ndarray)): # take a random caption if there are multiple captions.append(random.choice(caption) if is_train else caption[0]) else: raise ValueError( f"Caption column `{text}` should contain either strings or lists of strings." ) inputs = self.tokenizer( captions, max_length=self.tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt" ) return inputs.input_ids class RealV40CustomDataset_v4(torch.utils.data.Dataset): def __init__(self, tokenizer, TP_tokenizer, image_root_path, txt_root, resolution=512, drop_txt_prob = 0.02, drop_all_prob = 0.02, ): self.tokenizer = tokenizer self.TP_tokenizer = TP_tokenizer self.resolution = resolution self.transform = transforms.Compose([ transforms.Resize(self.resolution, interpolation=transforms.InterpolationMode.BILINEAR), transforms.CenterCrop(self.resolution), transforms.ToTensor(), transforms.Normalize([0.5], [0.5]), ]) self.img_path = image_root_path # self.img_name_list = os.listdir(self.img_path) self.txt_dict_global = json.load(open(txt_root['global'], 'r')) self.txt_dict_global = {os.path.basename(key):value for key, value in self.txt_dict_global.items()} self.txt_dict_shape = pd.read_csv(txt_root['fine_grained']).set_index('img_path')['caption'].to_dict() self.txt_dict_shape = {os.path.basename(key):ast.literal_eval(value) for key, value in self.txt_dict_shape.items()} [value.pop('lianxing') for key, value in self.txt_dict_shape.items()] ## lianxing暂时去掉 img_set = set(os.listdir(image_root_path)) img_set = img_set.intersection(set(self.txt_dict_global.keys())) img_set = img_set.intersection(set(self.txt_dict_shape.keys())) error_file = ['RealisticV4-0_90994_27412.png', 'RealisticV4-0_99115_23851.png', 'RealisticV4-0_82840_14492.png', 'RealisticV4-0_26257_6936.png', 'RealisticV4-0_915_24610.png', 'RealisticV4-0_49242_25603.png', 'RealisticV4-0_93324_19139.png', 'RealisticV4-0_79137_2021.png'] img_set = img_set - set(error_file) self.img_name_list = list(img_set) self.drop_txt_prob = drop_txt_prob self.drop_all_prob = drop_all_prob def __getitem__(self, index): if 1: img_name = self.img_name_list[index] # read image img = Image.open(os.path.join(self.img_path, img_name)).convert("RGB") img = self.transform(img) text_global = self.txt_dict_global[img_name] drop_all_local_prob = 0.0 text_list_finegrained = [] caption_dict = self.txt_dict_shape[img_name] for key in caption_dict: caption = caption_dict[key].strip(" .") if caption != "": text_list_finegrained.append(caption) # if len(text_list_finegrained) > 1 and random.random() < 0.2: # text_list_finegrained.pop(random.randint(0,len(text_list_finegrained) - 1)) text_shape = random.choice(text_list_finegrained) # text_shape = '. '.join(text_list_finegrained) # if text_shape != '': # text_shape +='.' # randomly drop rand_num = random.random() if rand_num < self.drop_txt_prob: text_shape = '' elif rand_num < self.drop_txt_prob + self.drop_all_prob: text_shape, text_global= '', '' return { "pixel_values": img, "img_name":img_name, "input_ids_global": self.tokenize_captions([text_global]), "input_ids_shape": self.tokenize_finegrained_captions([text_shape])[0], } # except: # return None def __len__(self): return len(self.img_name_list) def tokenize_captions(self, text, is_train=True): captions = [] for caption in text: if isinstance(caption, str): captions.append(caption) elif isinstance(caption, (list, np.ndarray)): # take a random caption if there are multiple captions.append(random.choice(caption) if is_train else caption[0]) else: raise ValueError( f"Caption column `{text}` should contain either strings or lists of strings." ) inputs = self.tokenizer( captions, max_length=self.tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt" ) return inputs.input_ids def tokenize_finegrained_captions(self, text): inputs = self.TP_tokenizer(text) return inputs prompt_list = { "fuzhi": { "freckles": [ "freckled skin", "a person with freckled skin", "freckles on the face", "obvious freckles on the face", ], "fair": [ "smooth skin", "a person with smooth skin" ], "wrinkles": [ "wrinkled skin", "serve wrinkles on the face", "a person with wrinkled face", "wrinkles on the face", ], }, "hair color": [ "The color of the hair is {}", "{} hair.", "the hair is in {}.", "A person with {} color hair.", ], "hair curliness": [ "hair that is {}.", "hair which is {}.", "a {} hair is obvious.", "A person with {} hair." ], "eyeglasses": [ "a person with {}", "wearing {}", "the person is accessorized with {}." ], "beards": [ "a person with {}", "wearing {}", "the person is accessorized with {}." ], "hat": [ "a person with {}", "wearing {}", "the person is accessorized with {}." ], "necklace": [ "a person with {}", "wearing {}", "the person is accessorized with {}." ], "makeup": { "makeup":[ "a person with makeup", "wearing makeup", "the human has a makeup look" ], "no makeup": [ "The bare face is the obvious.", "the person has a bare face", "a person with bare face." ], }, } prompt_list2 = { "fuzhi": { "freckles": [ "freckled skin", "freckled face", ], "fair": [ "smooth skin", "smooth face" ], "wrinkles": [ "wrinkled skin", "wrinkles face", ], }, "hair color": [ "{} hair.", "hair in {}.", ], "hair curliness": [ "{} hair", "hair is {}.", ], "eyeglasses": [ "a person with {}", "wearing {}", ], "beards": [ "a person with {}", "wearing {}", ], "hat": [ "a person with {}", "wearing {}", ], "necklace": [ "a person with {}", "wearing {}", ], "makeup": { "makeup":[ "wearing makeup", ], "no makeup": [ "natural face.", ], "red lips": [ "red lips", "the lips is red", ] }, "style": [ "{}", "{} image", "{} human image", ], "expression": [ "a person is {}", "with {} facial expression", "{}", ], "scene object": [ "{}", "a person {}", ], "cloth": [ "wearing {}", "{}", "a person with {}", ] } class RealV40CustomDataset_v7(torch.utils.data.Dataset): def __init__(self, tokenizer, TP_tokenizer, data_csv_root, img_key, caption_key, resolution=512, drop_txt_prob = 0.02, drop_all_prob = 0.02, ): self.tokenizer = tokenizer self.TP_tokenizer = TP_tokenizer self.resolution = resolution self.transform = transforms.Compose([ transforms.Resize(self.resolution, interpolation=transforms.InterpolationMode.BILINEAR), transforms.CenterCrop(self.resolution), transforms.ToTensor(), transforms.Normalize([0.5], [0.5]), ]) # self.img_path = image_root_path # self.img_name_list = os.listdir(self.img_path) # self.txt_dict_global = json.load(open(txt_root['global'], 'r')) # self.txt_dict_global = {os.path.basename(key):value for key, value in self.txt_dict_global.items()} # self.txt_dict_shape = pd.read_csv(txt_root['fine_grained']).set_index('img_path')['caption'].to_dict() # self.txt_dict_shape = {os.path.basename(key):ast.literal_eval(value) for key, value in self.txt_dict_shape.items()} # [value.pop('lianxing') for key, value in self.txt_dict_shape.items()] ## lianxing暂时去掉 # img_set = set(os.listdir(image_root_path)) # img_set = img_set.intersection(set(self.txt_dict_global.keys())) # img_set = img_set.intersection(set(self.txt_dict_shape.keys())) df = pd.read_csv(data_csv_root) df[caption_key] = df[caption_key].apply(ast.literal_eval) for i in range(len(df)): data = df.iloc[i] new_caption = {k: v for k, v in data[caption_key].items() if v != '' and ('not ' not in v or k == 'global_prompt')} df.at[i, caption_key] = new_caption df = df[df[caption_key].apply(lambda d: len(d) > 1)] # df = df[df[caption_key].apply(lambda x: 'cloth' in x and x.get('cloth', '') != '')] self.images = df[img_key].tolist() self.captions = df[caption_key].tolist() self.prompt_list = prompt_list2 self.drop_txt_prob = drop_txt_prob self.drop_all_prob = drop_all_prob self.fill_dimension = ["hair color", "hair curliness", 'eyeglasses', 'beards', 'hat', 'necklace', 'style', 'expression', 'scene object', 'cloth'] self.unfill_dimension = ["makeup", "fuzhi"] # self.check_data() def __getitem__(self, index): if 1: img_name = os.path.basename(self.images[index]) img = Image.open(str(self.images[index].strip().replace("crop_img", "img"))).convert("RGB") # read image img = self.transform(img) text_global = self.captions[index]['global_prompt'] drop_all_loclal_prob = 0.0 caption_dict = self.captions[index].copy() new_caption = caption_dict.pop("global_prompt") text_list_finegrained = [] if len(caption_dict)> 1: attr_num = random.choices([1,2], [0.8, 0.2])[0] attr_list = random.sample(caption_dict.keys(), attr_num) else: attr_list = [random.choice(list(caption_dict.keys()))] for attr in attr_list: label = caption_dict[attr] if attr == 'cloth': text_global = text_global.replace(label, "") text_global = text_global.replace(', ,', ",") if attr in self.fill_dimension: caption = random.choice(self.prompt_list[attr]) caption = caption.format(label) text_list_finegrained.append(caption.strip(" .")) elif attr in self.unfill_dimension: caption = random.choice(self.prompt_list[attr][label]) text_list_finegrained.append(caption.strip(" .")) random.shuffle(text_list_finegrained) text_shape = ', '.join(text_list_finegrained).strip(' ,.') + "." text_global = text_global.strip(' ,.') + "." # randomly drop rand_num = random.random() if rand_num < self.drop_txt_prob: text_shape = '' elif rand_num < self.drop_txt_prob + self.drop_all_prob: text_shape, text_global= '', '' return { "pixel_values": img, "img_name":img_name, "input_ids_global": self.tokenize_captions([text_global]), "input_ids_shape": self.TP_tokenize_captions([text_shape]), } # except: # return None def check_data(self): for index in tqdm(range(len(self.images))): # img_name = os.path.basename(self.images[index]) # img = Image.open(str(self.images[index].strip().replace("crop_img", "img"))).convert("RGB") text_global = self.captions[index]['global_prompt'] def __len__(self): return len(self.images) def tokenize_captions(self, text, is_train=True): captions = [] for caption in text: if isinstance(caption, str): captions.append(caption) elif isinstance(caption, (list, np.ndarray)): # take a random caption if there are multiple captions.append(random.choice(caption) if is_train else caption[0]) else: raise ValueError( f"Caption column `{text}` should contain either strings or lists of strings." ) inputs = self.tokenizer( captions, max_length=self.tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt" ) return inputs.input_ids def TP_tokenize_captions(self, text, is_train=True): captions = [] for caption in text: if isinstance(caption, str): captions.append(caption) elif isinstance(caption, (list, np.ndarray)): # take a random caption if there are multiple captions.append(random.choice(caption) if is_train else caption[0]) else: raise ValueError( f"Caption column `{text}` should contain either strings or lists of strings." ) inputs = self.TP_tokenizer( captions, max_length=self.tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt" ) return inputs.input_ids class RealV40CustomDataset_v7_stage2(torch.utils.data.Dataset): def __init__(self, tokenizer, TP_tokenizer, clip_processor, data_csv_root, img_key, caption_key, resolution=512, drop_txt_prob = 0.02, drop_all_prob = 0.02, ): self.tokenizer = tokenizer self.TP_tokenizer = TP_tokenizer self.clip_processor = clip_processor self.resolution = resolution self.transform = transforms.Compose([ transforms.Resize(self.resolution, interpolation=transforms.InterpolationMode.BILINEAR), transforms.CenterCrop(self.resolution), transforms.ToTensor(), transforms.Normalize([0.5], [0.5]), ]) df = pd.read_csv(data_csv_root) # df = df.tail(len(df) - 107626) df[caption_key] = df[caption_key].apply(ast.literal_eval) df[caption_key] = df[caption_key].apply(lambda x: {k: v for k, v in x.items() if v != '' or k == 'global_prompt'}) # for i in range(len(df)): # data = df.iloc[i] # new_caption = {k: v for k, v in data[caption_key].items() if v != ''} # df.at[i, caption_key] = new_caption df = df[df[caption_key].apply(lambda d: len(d) > 1)] self.images = df[img_key].tolist() self.captions = df[caption_key].tolist() self.prompt_list = prompt_list2 self.drop_txt_prob = drop_txt_prob self.drop_all_prob = drop_all_prob self.fill_dimension = ["hair color", "hair curliness", 'eyeglasses', 'beards', 'hat', 'necklace', 'style', 'expression', 'scene object', 'cloth'] self.unfill_dimension = ["makeup", "fuzhi"] def __getitem__(self, index): if 1: img_name = os.path.basename(self.images[index]) img = Image.open(str(self.images[index].strip().replace("crop_img", "img"))).convert("RGB") # read image clip_img = self.clip_processor(images=img, return_tensors="pt").pixel_values img = self.transform(img) text_global = self.captions[index]['global_prompt'] drop_all_loclal_prob = 0.0 caption_dict = self.captions[index].copy() new_caption = caption_dict.pop("global_prompt") text_list_finegrained = [] if len(caption_dict)> 1: attr_num = random.choices([1,2], [0.8, 0.2])[0] attr_list = random.sample(caption_dict.keys(), attr_num) else: attr_list = [random.choice(list(caption_dict.keys()))] for attr in attr_list: label = caption_dict[attr] if attr in self.fill_dimension: caption = random.choice(self.prompt_list[attr]) caption = caption.format(label) text_list_finegrained.append(caption.strip(" .")) elif attr in self.unfill_dimension: caption = random.choice(self.prompt_list[attr][label]) text_list_finegrained.append(caption.strip(" .")) random.shuffle(text_list_finegrained) text_shape = '. '.join(text_list_finegrained) + "." # randomly drop rand_num = random.random() if rand_num < self.drop_txt_prob: text_shape = '' elif rand_num < self.drop_txt_prob + self.drop_all_prob: text_shape, text_global= '', '' return { "pixel_values": img, "img_name":img_name, "input_ids_global": self.tokenize_captions([text_global]), "input_ids_shape": self.TP_tokenize_captions([text_shape]), "clip_img":clip_img, } # except: # return None def __len__(self): return len(self.images) def tokenize_captions(self, text, is_train=True): captions = [] for caption in text: if isinstance(caption, str): captions.append(caption) elif isinstance(caption, (list, np.ndarray)): # take a random caption if there are multiple captions.append(random.choice(caption) if is_train else caption[0]) else: raise ValueError( f"Caption column `{text}` should contain either strings or lists of strings." ) inputs = self.tokenizer( captions, max_length=self.tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt" ) return inputs.input_ids def TP_tokenize_captions(self, text, is_train=True): captions = [] for caption in text: if isinstance(caption, str): captions.append(caption) elif isinstance(caption, (list, np.ndarray)): # take a random caption if there are multiple captions.append(random.choice(caption) if is_train else caption[0]) else: raise ValueError( f"Caption column `{text}` should contain either strings or lists of strings." ) inputs = self.TP_tokenizer( captions, max_length=self.tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt" ) return inputs.input_ids class RealV40CustomDataset_v7_stage2_drop_head(torch.utils.data.Dataset): def __init__(self, tokenizer, TP_tokenizer, clip_processor, data_csv_root, img_key, caption_key, resolution=512, drop_txt_prob = 0.02, drop_all_prob = 0.02, ): self.tokenizer = tokenizer self.TP_tokenizer = TP_tokenizer self.clip_processor = clip_processor self.resolution = resolution self.transform = transforms.Compose([ transforms.Resize(self.resolution, interpolation=transforms.InterpolationMode.BILINEAR), transforms.CenterCrop(self.resolution), transforms.ToTensor(), transforms.Normalize([0.5], [0.5]), ]) df = pd.read_csv(data_csv_root) df = df.tail(len(df) - 32000) df[caption_key] = df[caption_key].apply(ast.literal_eval) # df = df[df[caption_key].apply(lambda x: "white" in x.get('hair color', '') )] df[caption_key] = df[caption_key].apply(lambda x: {k: v for k, v in x.items() if v != '' and ('not ' not in v or k == 'global_prompt') }) # df = df[df[caption_key].apply(lambda x: x.get('scene object', '') != '' )] df = df[df[caption_key].apply(lambda d: len(d) > 1)] self.images = df[img_key].tolist() self.captions = df[caption_key].tolist() self.prompt_list = prompt_list2 self.drop_txt_prob = drop_txt_prob self.drop_all_prob = drop_all_prob self.fill_dimension = ["hair color", "hair curliness", 'eyeglasses', 'beards', 'hat', 'necklace', 'style', 'expression', 'scene object', 'cloth'] self.unfill_dimension = ["makeup", "fuzhi"] def __getitem__(self, index): if 1: img_name = os.path.basename(self.images[index]) img = Image.open(str(self.images[index].strip().replace("crop_img", "img"))).convert("RGB") # read image clip_img = self.clip_processor(images=img, return_tensors="pt").pixel_values img = self.transform(img) text_global = self.captions[index]['global_prompt'] drop_all_loclal_prob = 0.0 caption_dict = self.captions[index].copy() new_caption = caption_dict.pop("global_prompt") text_list_finegrained = [] if len(caption_dict)> 1: attr_num = random.choices([1,2], [0.8, 0.2])[0] attr_list = random.sample(caption_dict.keys(), attr_num) else: attr_list = [random.choice(list(caption_dict.keys()))] for attr in attr_list: label = caption_dict[attr] if attr == 'cloth': text_global = text_global.replace(label + ", ", "") if attr in self.fill_dimension: caption = random.choice(self.prompt_list[attr]) caption = caption.format(label) text_list_finegrained.append(caption.strip(" .")) elif attr in self.unfill_dimension: caption = random.choice(self.prompt_list[attr][label]) text_list_finegrained.append(caption.strip(" .")) random.shuffle(text_list_finegrained) text_shape = '. '.join(text_list_finegrained).strip(' ,.') + "." # randomly drop rand_num = random.random() if rand_num < self.drop_txt_prob: text_shape = '' elif rand_num < self.drop_txt_prob + self.drop_all_prob: text_shape, text_global= '', '' return { "pixel_values": img, "img_name":img_name, "input_ids_global": self.tokenize_captions([text_global]), "input_ids_global_indice": self.tokenize_captions(text_global.split(", ")), "input_ids_shape": self.TP_tokenize_captions([text_shape]), "clip_img":clip_img, } # except: # return None def __len__(self): return len(self.images) def tokenize_captions(self, text, is_train=True): captions = [] for caption in text: if isinstance(caption, str): captions.append(caption) elif isinstance(caption, (list, np.ndarray)): # take a random caption if there are multiple captions.append(random.choice(caption) if is_train else caption[0]) else: raise ValueError( f"Caption column `{text}` should contain either strings or lists of strings." ) inputs = self.tokenizer( captions, max_length=self.tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt" ) return inputs.input_ids def TP_tokenize_captions(self, text, is_train=True): captions = [] for caption in text: if isinstance(caption, str): captions.append(caption) elif isinstance(caption, (list, np.ndarray)): # take a random caption if there are multiple captions.append(random.choice(caption) if is_train else caption[0]) else: raise ValueError( f"Caption column `{text}` should contain either strings or lists of strings." ) inputs = self.TP_tokenizer( captions, max_length=self.tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt" ) return inputs.input_ids