import torch import torch.nn as nn import os import json import torch.nn.functional as F import random from torch.utils.data import Dataset from transformers import AutoTokenizer from glob import glob import math from PIL import Image device = torch.device('cuda') import numpy as np from typing import Any, Dict, List, Optional, Tuple, Union import torch import torch.nn as nn from diffusers.utils import logging from diffusers.models.embeddings import PatchEmbed from diffusers.models.modeling_outputs import Transformer2DModelOutput from diffusers.models.attention import BasicTransformerBlock from diffusers.models.normalization import AdaLayerNormContinuous from torchvision import transforms def add_hook_to_module(model, module_name): outputs = [] def hook(module, input, output): outputs.append(output) module = dict(model.named_modules()).get(module_name) if module is None: raise ValueError(f"can't find module {module_name}") hook_handle = module.register_forward_hook(hook) return hook_handle, outputs class PromptSD35Net(nn.Module): def __init__(self, sample_size: int = 128, patch_size: int = 2, in_channels: int = 16, num_layers: int = 8, attention_head_dim: int = 64, num_attention_heads: int = 24, out_channels: int = 16, pos_embed_max_size: int = 192 ): super().__init__() self.sample_size = sample_size self.patch_size = patch_size self.in_channels = in_channels self.num_layers = num_layers self.attention_head_dim = attention_head_dim self.num_attention_heads = num_attention_heads self.out_channels = out_channels self.pos_embed_max_size = pos_embed_max_size self.inner_dim = self.num_attention_heads * self.attention_head_dim self.pos_embed = PatchEmbed( height=self.sample_size, width=self.sample_size, patch_size=self.patch_size, in_channels=self.in_channels, embed_dim=self.inner_dim, pos_embed_max_size=pos_embed_max_size ) self.transformer_blocks = nn.ModuleList( [ BasicTransformerBlock( dim=self.inner_dim, num_attention_heads=self.num_attention_heads, attention_head_dim=self.attention_head_dim, ff_inner_dim=2*self.inner_dim # mult should be 4 by default ) for i in range(self.num_layers) ] ) self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6) self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True) self.noise_shape = (1, 16, 128, 128) # (667, 4096) self.pre8_linear = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.pre16_linear = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.pre24_linear = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.pre8_linear2 = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.pre16_linear2 = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.pre24_linear2 = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.last_linear = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) # self.last_linear2 = nn.Sequential(nn.Linear(667, 32)) self.skip_connection2 = nn.Linear(4096, 1, bias=False) self.skip_connection = nn.Linear(667, 32, bias=False) self.trans_linear = nn.Linear(666+1+4096, 1536, bias=False) nn.init.constant_(self.skip_connection.weight.data, 0) nn.init.constant_(self.trans_linear.weight.data, 0) nn.init.constant_(self.trans_linear.weight.data, 0) nn.init.constant_(self.pre8_linear[-1].weight.data, 0) nn.init.constant_(self.pre16_linear[-1].weight.data, 0) nn.init.constant_(self.pre24_linear[-1].weight.data, 0) nn.init.constant_(self.pre8_linear2[-1].weight.data, 0) nn.init.constant_(self.pre16_linear2[-1].weight.data, 0) nn.init.constant_(self.pre24_linear2[-1].weight.data, 0) def forward(self, noise: torch.Tensor, _s, _v, _d, _pool_embedding) -> torch.Tensor: assert noise is not None _ori_v = _v.clone() _v = torch.stack([torch.diag(_v[jj]) for jj in range(_v.shape[0])], dim=0) positive_embedding = _s.permute(0, 2, 1) @ _v @ _d # [2, 64, 666] [2, 64] [2, 64, 4096] pool_embedding = _pool_embedding[:, None, :] embedding = torch.cat([positive_embedding, pool_embedding], dim=1) bs = noise.shape[0] height, width = noise.shape[-2:] embed_8 = embedding embed_16 = embedding embed_24 = embedding scale_8 = self.pre8_linear2(embed_8).mean(1) scale_16 = self.pre16_linear2(embed_16).mean(1) scale_24 = self.pre24_linear2(embed_24).mean(1) embed_8 = self.pre8_linear(embed_8).mean(1) embed_16 = self.pre16_linear(embed_16).mean(1) embed_24 = self.pre24_linear(embed_24).mean(1) embed_last = self.last_linear(embedding).mean(1) embed_trans = self.trans_linear(torch.cat([_s, _ori_v[...,None], _d], dim=2)).mean(1) skip_embedding = self.skip_connection(self.skip_connection2(embedding).permute(0,2,1)).permute(0,2,1) scale_skip, embed_skip = skip_embedding.chunk(2,dim=1) ori_noise = noise * (scale_skip[...,None]) + embed_skip[...,None] noise = self.pos_embed(noise) noise = noise * (1 + scale_8[:, None, :] + embed_trans[:, None, :]) + embed_8[:, None, :] scale_list = [scale_16, scale_24] embed_list = [embed_16, embed_24] for _ii, block in enumerate(self.transformer_blocks): noise = block(noise) if len(scale_list)!=0 and len(embed_list)!=0: noise = noise * (1 + scale_list[int(_ii//4)][:, None, :] + embed_trans[:, None, :]) + embed_list[int(_ii//4)][:, None, :] hidden_states = noise hidden_states = self.norm_out(hidden_states, embed_last) hidden_states = self.proj_out(hidden_states) # unpatchify patch_size = self.patch_size height = height // patch_size width = width // patch_size hidden_states = hidden_states.reshape( shape=(hidden_states.shape[0], height, width, patch_size, patch_size, self.out_channels) ) hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) output = hidden_states.reshape( shape=(hidden_states.shape[0], self.out_channels, height * patch_size, width * patch_size) ) return output + ori_noise def weak_load_state_dict(self, state_dict: os.Mapping[str, torch.any], strict: bool = True, assign: bool = False): return load_filtered_state_dict(self, state_dict) class PromptSDXLNet(nn.Module): def __init__(self, sample_size: int = 128, patch_size: int = 2, in_channels: int = 4, num_layers: int = 4, attention_head_dim: int = 64, num_attention_heads: int = 24, out_channels: int = 4, pos_embed_max_size: int = 192 ): super().__init__() self.sample_size = sample_size self.patch_size = patch_size self.in_channels = in_channels self.num_layers = num_layers self.attention_head_dim = attention_head_dim self.num_attention_heads = num_attention_heads self.out_channels = out_channels self.pos_embed_max_size = pos_embed_max_size self.inner_dim = self.num_attention_heads * self.attention_head_dim self.pos_embed = PatchEmbed( height=self.sample_size, width=self.sample_size, patch_size=self.patch_size, in_channels=self.in_channels, embed_dim=self.inner_dim, pos_embed_max_size=pos_embed_max_size ) self.transformer_blocks = nn.ModuleList( [ BasicTransformerBlock( dim=self.inner_dim, num_attention_heads=self.num_attention_heads, attention_head_dim=self.attention_head_dim, ff_inner_dim=2*self.inner_dim # mult should be 4 by default ) for i in range(self.num_layers) ] ) self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6) self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True) self.noise_shape = (1, 4, 128, 128) self.pre8_linear = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.pre16_linear = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.pre24_linear = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.pre8_linear2 = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.pre16_linear2 = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.pre24_linear2 = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) self.last_linear = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536)) # self.last_linear2 = nn.Sequential(nn.Linear(667, 32)) self.skip_connection2 = nn.Linear(2048, 1, bias=False) self.skip_connection = nn.Linear(154+1, 8, bias=False) self.trans_linear = nn.Linear(154+1+2048, 1536, bias=False) self.pool_prompt_linear = nn.Linear(2560, 2048, bias=False) nn.init.constant_(self.skip_connection.weight.data, 0) nn.init.constant_(self.trans_linear.weight.data, 0) nn.init.constant_(self.trans_linear.weight.data, 0) nn.init.constant_(self.pre8_linear[-1].weight.data, 0) nn.init.constant_(self.pre16_linear[-1].weight.data, 0) nn.init.constant_(self.pre24_linear[-1].weight.data, 0) nn.init.constant_(self.pre8_linear2[-1].weight.data, 0) nn.init.constant_(self.pre16_linear2[-1].weight.data, 0) nn.init.constant_(self.pre24_linear2[-1].weight.data, 0) def forward(self, noise: torch.Tensor, _s, _v, _d, _pool_embedding) -> torch.Tensor: assert noise is not None _ori_v = _v.clone() _v = torch.stack([torch.diag(_v[jj]) for jj in range(_v.shape[0])], dim=0) positive_embedding = _s.permute(0, 2, 1) @ _v @ _d # [2, 64, 154] [2, 64] [2, 64, 2048] pool_embedding = self.pool_prompt_linear(_pool_embedding[:, None, :]) embedding = torch.cat([positive_embedding, pool_embedding], dim=1) bs = noise.shape[0] height, width = noise.shape[-2:] embed_8 = embedding embed_16 = embedding embed_24 = embedding scale_8 = self.pre8_linear2(embed_8).mean(1) scale_16 = self.pre16_linear2(embed_16).mean(1) scale_24 = self.pre24_linear2(embed_24).mean(1) embed_8 = self.pre8_linear(embed_8).mean(1) embed_16 = self.pre16_linear(embed_16).mean(1) embed_24 = self.pre24_linear(embed_24).mean(1) embed_last = self.last_linear(embedding).mean(1) embed_trans = self.trans_linear(torch.cat([_s, _ori_v[...,None], _d], dim=2)).mean(1) skip_embedding = self.skip_connection(self.skip_connection2(embedding).permute(0,2,1)).permute(0,2,1) scale_skip, embed_skip = skip_embedding.chunk(2,dim=1) ori_noise = noise * (scale_skip[...,None]) + embed_skip[...,None] noise = self.pos_embed(noise) noise = noise * (1 + scale_8[:, None, :] + embed_trans[:, None, :]) + embed_8[:, None, :] scale_list = [scale_16, scale_24] embed_list = [embed_16, embed_24] for _ii, block in enumerate(self.transformer_blocks): noise = block(noise) if len(scale_list)!=0 and len(embed_list)!=0: noise = noise * (1 + scale_list[int(_ii//4)][:, None, :] + embed_trans[:, None, :]) + embed_list[int(_ii//4)][:, None, :] hidden_states = noise hidden_states = self.norm_out(hidden_states, embed_last) hidden_states = self.proj_out(hidden_states) # unpatchify patch_size = self.patch_size height = height // patch_size width = width // patch_size hidden_states = hidden_states.reshape( shape=(hidden_states.shape[0], height, width, patch_size, patch_size, self.out_channels) ) hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) output = hidden_states.reshape( shape=(hidden_states.shape[0], self.out_channels, height * patch_size, width * patch_size) ) return output + ori_noise def weak_load_state_dict(self, state_dict: os.Mapping[str, torch.any], strict: bool = True, assign: bool = False): return load_filtered_state_dict(self, state_dict) class NoisePromptDataset(Dataset): def __init__(self, if_weight=False): self.if_weight = if_weight json_list = glob('/home/xiedian/total_datacollect/json/*.json') self.original_score = [] self.optim_score = [] self.prompt = [] self.noise_paths = [] self.mask_conditions = [] self.embeddings = [] counter = 0 for i in range(len(json_list)): with open('//home/xiedian/total_datacollect/json/new{:06d}.json'.format(i), 'r') as f: data = json.load(f) self.original_score.append(data['original_score_list']) self.optim_score.append(data['optimized_score_list']) if data['optimized_score_list']>data['original_score_list']: counter += 1 self.prompt.append(data['caption']) self.noise_paths.append('/home/xiedian/total_datacollect/latents/{:06d}.pt'.format(i)) self.embeddings.append('/data/xiedian/hg_reward/newdatacollect/total_datacollect_step_1_1/embedding/embeds_{:06d}.pt'.format(i)) z = [0, 1] * ((512+77+77) // 2) self.mask_conditions.append(data['mid_token_ids'] if 'mid_token_ids' in data else z) # while counter * 2 > len(self.prompt): # p = random.randint(0,len(self.prompt)-1) # if self.original_score[p] > self.optim_score[p]: # self.optim_score.append(self.optim_score[p]) # self.original_score.append(self.original_score[p]) # self.mask_conditions.append(self.mask_conditions[p]) # self.noise_paths.append(self.noise_paths[p]) # self.prompt.append(self.prompt[p]) # while counter * 2 < len(self.prompt): # p = random.randint(0,len(self.prompt)-1) # if self.original_score[p] > self.optim_score[p]: # self.optim_score.append(self.optim_score[p]) # self.original_score.append(self.original_score[p]) # self.mask_conditions.append(self.mask_conditions[p]) # self.noise_paths.append(self.noise_paths[p]) # self.prompt.append(self.prompt[p]) self.original_score = torch.Tensor(self.original_score) self.optim_score = torch.Tensor(self.optim_score) def __len__(self): return len(self.prompt) def __getitem__(self, index): try: noise = torch.load(self.noise_paths[index], map_location='cpu').squeeze(0).float() noise_pred_uncond, mid_noise_pred, noise_pred_text = noise.chunk(3,dim=0) prompt = self.prompt[index] original_score = self.original_score[index] optim_score = self.optim_score[index] embedding = torch.load(self.embeddings[index], map_location='cpu') _s, _v, _d, _pool_embedding = embedding['_s'], embedding['_v'], embedding['_d'], embedding['pooled_prompt_embeds'] _s = _s.detach().float() _v = _v.detach().float() _d = _d.detach().float() _pool_embedding = _pool_embedding.detach().float() if original_score > optim_score: noise_pred = noise_pred_uncond + 4.5 * (noise_pred_text - noise_pred_uncond) else: guidance_scale = 4.5 * 1.6 diff_text = torch.norm(noise_pred_text - noise_pred_uncond) mid_guidance_scale = (diff_text / torch.norm(noise_pred_text - mid_noise_pred)).item() guidance_scale_mid = guidance_scale / (2.4 + 1) guidance_scale_all = guidance_scale * 2.4 / (2.4 + 1) all_mid = (noise_pred_text - mid_noise_pred) * mid_guidance_scale all_null = noise_pred_text - noise_pred_uncond noise_pred = all_mid * guidance_scale_mid + all_null * guidance_scale_all + (mid_noise_pred + noise_pred_uncond) / 2 except: print("error", index) return self.__getitem__((index+1)%len(self.prompt)) if self.if_weight: return noise_pred_text, prompt, noise_pred, 2 / (1+ math.exp((-abs(original_score) + 26.5)/1.7)), _s, _v, _d, _pool_embedding else: return noise_pred_text, prompt, noise_pred, _s, _v, _d, _pool_embedding class NoisePromptDataset_2_0(Dataset): def __init__(self, if_weight=False): self.if_weight = if_weight json_list = glob('/data/xiedian/hg_reward/newdatacollect/total_datacollect_step_1_1/json/*.json') self.original_score = [] self.quick_score = [] self.slow_score = [] self.prompt = [] self.noise_paths = [] self.mask_conditions = [] self.img_list = [] self.embeddings = [] counter = 0 for i in range(len(json_list)): with open('/data/xiedian/hg_reward/newdatacollect/total_datacollect_step_1_1/json/new{:06d}.json'.format(i), 'r') as f: data = json.load(f) if (not os.path.exists('/data/xiedian/hg_reward/newdatacollect/total_datacollect_step_1_1/latents/{:06d}.pt'.format(i))) or \ max(data['original_score_list'], data['quick_score_list'], data['slow_score_list']) != data['original_score_list']: continue self.original_score.append(data['original_score_list']) self.quick_score.append(data['quick_score_list']) self.slow_score.append(data['slow_score_list']) self.prompt.append(data['caption']) self.noise_paths.append('/data/xiedian/hg_reward/newdatacollect/total_datacollect_step_1_1/latents/{:06d}.pt'.format(i)) z = [0, 1] * ((512+77+77) // 2) self.mask_conditions.append(data['mid_token_ids'] if 'mid_token_ids' in data else z) if data['original_score_list'] >= max(data['quick_score_list'], data['slow_score_list']): self.img_list.append('/data/xiedian/hg_reward/newdatacollect/total_datacollect_step_1_1/optim/original{:06d}.png'.format(i)) elif data['quick_score_list'] >= max(data['original_score_list'], data['slow_score_list']): self.img_list.append('/data/xiedian/hg_reward/newdatacollect/total_datacollect_step_1_1/optim/quick{:06d}.png'.format(i)) else: self.img_list.append('/data/xiedian/hg_reward/newdatacollect/total_datacollect_step_1_1/optim/slow{:06d}.png'.format(i)) self.embeddings.append('/home/xiedian/total_datacollect/embedding/embeds_{:06d}.pt'.format(i)) self.original_score = torch.Tensor(self.original_score) self.quick_score = torch.Tensor(self.quick_score) self.slow_score = torch.Tensor(self.slow_score) def __len__(self): return len(self.prompt) def __getitem__(self, index): try: original_score = self.original_score[index] quick_score = self.quick_score[index] slow_score = self.slow_score[index] original_score = max(max(quick_score, slow_score), original_score) embedding = torch.load(self.embeddings[index], map_location='cpu') _s, _v, _d, _pool_embedding = embedding['_s'], embedding['_v'], embedding['_d'], embedding['pooled_prompt_embeds'] _s = _s.detach().float() _v = _v.detach().float() _d = _d.detach().float() _pool_embedding = _pool_embedding.detach().float() noise = torch.load(self.noise_paths[index], map_location='cpu').squeeze(0).float() noise_pred_text, noise_pred = noise.chunk(2,dim=0) prompt = self.prompt[index] except: print("error", index) return self.__getitem__((index+1)%len(self.prompt)) if self.if_weight: return noise_pred_text, prompt, noise_pred, 2 / (1+ math.exp((-abs(original_score) + 26.5)/1.7)), _s, _v, _d, _pool_embedding else: return noise_pred_text, prompt, noise_pred, _s, _v, _d, _pool_embedding class NoisePromptDataset_3_0(Dataset): def __init__(self, if_weight=False): self.if_weight = if_weight json_list = glob('/data/xiedian/hg_reward/CFG_TOTAL/total_datacollect/json/*.json') self.score = [] self.prompt = [] self.noise_paths = [] self.mask_conditions = [] self.img_list = [] self.embeddings = [] print(len(json_list)) for i in range(len(json_list)): if (not os.path.exists('/data/xiedian/hg_reward/CFG_TOTAL/total_datacollect/embedding/{:06d}.pt'.format(i))): continue with open('/data/xiedian/hg_reward/CFG_TOTAL/total_datacollect/json/new{:06d}.json'.format(i), 'r') as f: data = json.load(f) if data['original_score_list'] > data['optimized_score_list']: tag = 0 if (not os.path.exists('/data/xiedian/hg_reward/CFG_TOTAL/total_datacollect/latents/original{:06d}.pt'.format(i))): continue else: tag = 1 if (not os.path.exists('/data/xiedian/hg_reward/CFG_TOTAL/total_datacollect/latents/new{:06d}.pt'.format(i))): continue if tag == 1: self.score.append(data['optimized_score_list']) self.noise_paths.append('/data/xiedian/hg_reward/CFG_TOTAL/total_datacollect/latents/new{:06d}.pt'.format(i)) else: self.score.append(data['original_score_list']) self.noise_paths.append('/data/xiedian/hg_reward/CFG_TOTAL/total_datacollect/latents/original{:06d}.pt'.format(i)) self.prompt.append(data['caption']) self.embeddings.append('/data/xiedian/hg_reward/CFG_TOTAL/total_datacollect/embedding/{:06d}.pt'.format(i)) self.score = torch.Tensor(self.score) def __len__(self): return len(self.prompt) def __getitem__(self, index): try: embedding = torch.load(self.embeddings[index], map_location='cpu') _s, _v, _d, _pool_embedding = embedding['_s'], embedding['_v'], embedding['_d'], embedding['_pooled_prompt_embeds'] _s = _s.detach().float() _v = _v.detach().float() _d = _d.detach().float() _pool_embedding = _pool_embedding.detach().float() noise = torch.load(self.noise_paths[index], map_location='cpu').float() # [2XT, 16, 128, 128] prompt = self.prompt[index] # [ori, target, ori] score = self.score[index] except: print("error", index) return self.__getitem__((index+1)%len(self.prompt)) if self.if_weight: return noise, prompt, 2 / (1+ math.exp((-abs(score) + 26.5)/1.7)), _s, _v, _d, _pool_embedding else: return noise, prompt, _s, _v, _d, _pool_embedding def load_filtered_state_dict(model, state_dict): model_state_dict = model.state_dict() filtered_state_dict = {} for k, v in state_dict.items(): if k in model_state_dict: if model_state_dict[k].size() == v.size(): filtered_state_dict[k] = v else: print(f"Skipping {k}: shape mismatch ({model_state_dict[k].size()} vs {v.size()})") else: print(f"Skipping {k}: not found in model's state_dict.") model.load_state_dict(filtered_state_dict, strict=False) return model def custom_collate_fn_2_0(batch): noise_pred_texts, prompts, noise_preds, max_scores = zip(*batch) noise_pred_texts = torch.stack(noise_pred_texts) noise_preds = torch.stack(noise_preds) max_scores = torch.stack(max_scores) return noise_pred_texts, prompts, noise_preds, max_scores if __name__ == "__main__": dataset = NoisePromptDataset(if_weight=True) weights = [] for i, (noise, prompt, gt, weight) in enumerate(dataset): weights.append(weight) weights = torch.from_numpy(np.array(weights)).cuda() print(weights.mean(), weights.std(dim=0))