from __future__ import annotations import math import random import sys from argparse import ArgumentParser import einops import k_diffusion as K import numpy as np import torch import torch.nn as nn from einops import rearrange from omegaconf import OmegaConf from PIL import Image, ImageOps from torch import autocast from PIL import Image, ImageDraw, ImageFont import textwrap import json import os from dataset_loading import get_dataset from edit_dataset import EditITMDataset from torch.utils.data import DataLoader from tqdm import tqdm import clip from collections import defaultdict sys.path.append("./stable_diffusion") from stable_diffusion.ldm.util import instantiate_from_config # Assuming CFGDenoiser and other dependencies are correctly set up, # no changes needed there for image aspect ratio handling. def calculate_clip_similarity(generated_images, original_image, clip_model, preprocess, device): original_image_processed = preprocess(original_image).unsqueeze(0).to(device) with torch.no_grad(): original_features = clip_model.encode_image(original_image_processed) similarities = [] for img in generated_images: img_processed = preprocess(img).unsqueeze(0).to(device) with torch.no_grad(): generated_features = clip_model.encode_image(img_processed) similarity = torch.nn.functional.cosine_similarity(generated_features, original_features, dim=-1) similarities.append(similarity.item()) #concat both img and original_image for visualization # original_image_np = np.array(original_image) # img_np = np.array(img) # both = np.concatenate((original_image_np, img_np), axis=1) # both = Image.fromarray(both) # if not os.path.exists('eval_output/edit_itm/flickr_edit_clip_sim/'): # os.makedirs('eval_output/edit_itm/flickr_edit_clip_sim/') # random_id = random.randint(0, 100000) # both.save(f'eval_output/edit_itm/flickr_edit_clip_sim/{similarity.item()}_{random_id}.png') # average_similarity = sum(similarities) / len(similarities) # dist = 1 - average_similarity dists = [1 - sim for sim in similarities] return dists class CFGDenoiser(nn.Module): def __init__(self, model): super().__init__() self.inner_model = model def forward(self, z, sigma, cond, uncond, text_cfg_scale, image_cfg_scale, conditional_only=False): # cfg_z = einops.repeat(z, "1 ... -> n ...", n=3) cfg_z = z.repeat(3, 1, 1, 1) # cfg_sigma = einops.repeat(sigma, "1 ... -> n ...", n=3) cfg_sigma = sigma.repeat(3) cfg_cond = { "c_crossattn": [torch.cat([cond["c_crossattn"][0], uncond["c_crossattn"][0], uncond["c_crossattn"][0]])], "c_concat": [torch.cat([cond["c_concat"][0], cond["c_concat"][0], uncond["c_concat"][0]])], } out_cond, out_img_cond, out_uncond = self.inner_model(cfg_z, cfg_sigma, cond=cfg_cond).chunk(3) if conditional_only: return out_cond else: return out_uncond + text_cfg_scale * (out_cond - out_img_cond) + image_cfg_scale * (out_img_cond - out_uncond) def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False): print(f"Loading model from {ckpt}") pl_sd = torch.load(ckpt, map_location="cpu") if "global_step" in pl_sd: print(f"Global Step: {pl_sd['global_step']}") sd = pl_sd["state_dict"] if vae_ckpt is not None: print(f"Loading VAE from {vae_ckpt}") vae_sd = torch.load(vae_ckpt, map_location="cpu")["state_dict"] sd = { k: vae_sd[k[len("first_stage_model.") :]] if k.startswith("first_stage_model.") else v for k, v in sd.items() } model = instantiate_from_config(config.model) m, u = model.load_state_dict(sd, strict=False) if len(m) > 0 and verbose: print("missing keys:") print(m) if len(u) > 0 and verbose: print("unexpected keys:") print(u) return model def calculate_accuracy(losses): correct_count = 0 for loss in losses: if loss[0] < min(loss[1:]): correct_count += 1 return correct_count, len(losses) # Return counts for aggregation def main(): parser = ArgumentParser() parser.add_argument("--config", default="configs/generate.yaml", type=str) parser.add_argument("--ckpt", default="aurora-mixratio-15-15-1-1-42k-steps.ckpt", type=str) parser.add_argument("--vae-ckpt", default=None, type=str) parser.add_argument("--task", default='flickr_edit', type=str) parser.add_argument("--batchsize", default=1, type=int) parser.add_argument("--samples", default=4, type=int) parser.add_argument("--size", default=512, type=int) parser.add_argument("--steps", default=20, type=int) parser.add_argument("--cfg-text", default=7.5, type=float) parser.add_argument("--cfg-image", default=1.5, type=float) parser.add_argument('--targets', type=str, nargs='*', help="which target groups for mmbias",default='') parser.add_argument("--device", default=0, type=int, help="GPU device index") parser.add_argument("--log_imgs", action="store_true") parser.add_argument("--conditional_only", action="store_true") parser.add_argument("--metric", default="latent", type=str) parser.add_argument("--split", default='test', type=str) parser.add_argument("--skip", default=1, type=int) args = parser.parse_args() device = torch.device(f"cuda:{args.device}" if torch.cuda.is_available() else "cpu") config = OmegaConf.load(args.config) model = load_model_from_config(config, args.ckpt, args.vae_ckpt) model.eval() model.to(dtype=torch.float) model = model.to(device) model_wrap = K.external.CompVisDenoiser(model) model_wrap_cfg = CFGDenoiser(model_wrap) null_token = model.get_learned_conditioning([""]) clip_model, preprocess = clip.load("ViT-B/32", device=device) dataset = EditITMDataset(split=args.split, task=args.task, min_resize_res=args.size, max_resize_res=args.size, crop_res=args.size) dataloader= DataLoader(dataset,batch_size=args.batchsize,num_workers=1,worker_init_fn=None,shuffle=False, persistent_workers=True) if os.path.exists(f'itm_evaluation/{args.split}/{args.task}/{args.ckpt.replace("/", "_")}_results.json'): with open(f'itm_evaluation/{args.split}/{args.task}/{args.ckpt.replace("/", "_")}_results.json', 'r') as f: results = json.load(f) results = defaultdict(dict, results) else: results = defaultdict(dict) for i, batch in tqdm(enumerate(dataloader), total=len(dataloader)): if len(batch['input'][0].shape) < 3: continue for j, prompt in enumerate(batch['texts']): # check if we already have results for this image img_id = batch['path'][0] + f'_{i}' # if img_id in results and ('pos' in results[img_id] and 'neg' in results[img_id]): # continue with torch.no_grad(), autocast("cuda"), model.ema_scope(): prompt = prompt[0] cond = {} cond["c_crossattn"] = [model.get_learned_conditioning([prompt])] input_image = batch['input'][0].to(device) cond["c_concat"] = [model.encode_first_stage(input_image.unsqueeze(0)).mode()] scaled_input = model.scale_factor * input_image uncond = {} uncond["c_crossattn"] = [null_token] uncond["c_concat"] = [torch.zeros_like(cond["c_concat"][0])] sigmas = model_wrap.get_sigmas(args.steps) # move everything to the device cond = {k: [v.to(device) for v in vs] for k, vs in cond.items()} uncond = {k: [v.to(device) for v in vs] for k, vs in uncond.items()} cond["c_concat"][0] = cond["c_concat"][0].repeat(args.samples, 1, 1, 1) cond["c_crossattn"][0] = cond["c_crossattn"][0].repeat(args.samples, 1, 1) uncond["c_concat"][0] = uncond["c_concat"][0].repeat(args.samples, 1, 1, 1) uncond["c_crossattn"][0] = uncond["c_crossattn"][0].repeat(args.samples, 1, 1) extra_args = { "cond": cond, "uncond": uncond, "text_cfg_scale": args.cfg_text, "image_cfg_scale": args.cfg_image, "conditional_only": args.conditional_only, } # torch.manual_seed(i) torch.manual_seed(42) z = torch.randn_like(cond["c_concat"][0]) * sigmas[0] z = K.sampling.sample_euler_ancestral(model_wrap_cfg, z, sigmas, extra_args=extra_args, disable=True) x = model.decode_first_stage(z) x = torch.clamp((x + 1.0) / 2.0, min=0.0, max=1.0) ######## LOG IMAGES ######## input_image_pil = ((input_image + 1) * 0.5).clamp(0, 1) input_image_pil = input_image_pil.permute(1, 2, 0) # Change from CxHxW to HxWxC for PIL input_image_pil = (input_image_pil * 255).type(torch.uint8).cpu().numpy() for k in range(2): x_ = 255.0 * rearrange(x[k], "c h w -> h w c") edited_image = x_.type(torch.uint8).cpu().numpy() both = np.concatenate((input_image_pil, edited_image), axis=1) both = Image.fromarray(both) out_base = f'itm_evaluation/{args.split}/{args.task}/{args.ckpt.replace("/", "_")}' if not os.path.exists(out_base): os.makedirs(out_base) prompt_str = prompt.replace(' ', '_')[0:100] both.save(f'{out_base}/{i}_{"correct" if j == 0 else "incorrect"}_sample{k}_{prompt}.png') ######## CLIP ######## edited_images = [] for k in range(args.samples): x_ = 255.0 * rearrange(x[k], "c h w -> h w c") edited_image = Image.fromarray(x_.type(torch.uint8).cpu().numpy()) edited_images.append(edited_image) input_image_pil = ((input_image + 1) * 0.5).clamp(0, 1) input_image_pil = input_image_pil.permute(1, 2, 0) # Change from CxHxW to HxWxC for PIL input_image_pil = (input_image_pil * 255).type(torch.uint8).cpu().numpy() input_image_pil = Image.fromarray(input_image_pil) dists_clip = calculate_clip_similarity(edited_images, input_image_pil, clip_model, preprocess, device) ######## LATENT ######## z = z.flatten(1) original_latent = cond["c_concat"][0].flatten(1) dists_latent = torch.norm(z - original_latent, dim=1, p=2).cpu().numpy().tolist() cos_sim = torch.nn.functional.cosine_similarity(z, original_latent, dim=1).cpu().numpy().tolist() cos_dists_latent = [1 - sim for sim in cos_sim] ######## SAVE RESULTS ######## img_id = batch['path'][0] + f'_{i}' results[img_id]['pos' if j == 0 else 'neg'] = { "prompt" : prompt, "clip": dists_clip, "latent_l2": dists_latent, "latent_cosine": cos_dists_latent } with open(f'itm_evaluation/{args.split}/{args.task}/{args.ckpt.replace("/", "_")}_results.json', 'w') as f: json.dump(results, f, indent=2) if __name__ == "__main__": main()