Update preprocess_utils.py

preprocess_utils.py

@@ -1,5 +1,6 @@
 from transformers import CLIPTextModel, CLIPTokenizer, logging
 from diffusers import AutoencoderKL, UNet2DConditionModel, DDIMScheduler
+from diffusers.utils.torch_utils import randn_tensor
 # suppress partial model loading warning
 logging.set_verbosity_error()
 
@@ -12,6 +13,8 @@ from torchvision.io import write_video
 from pathlib import Path
 from utils import *
 import torchvision.transforms as T
+import cv2
+import numpy as np
 
 
 def get_timesteps(scheduler, num_inference_steps, strength, device):
@@ -64,7 +67,9 @@ class Preprocess(nn.Module):
         self.text_encoder = text_encoder
         self.unet = unet
         self.scheduler=scheduler
+        
         self.total_inverted_latents = {}
+        self.noise_total = None # will contain all zs if inversion == 'ddpm', var name chosen to match the save path of zs used in pr      https://github.com/omerbt/TokenFlow/pull/24/files#
         
         self.paths, self.frames, self.latents = self.get_data(self.config["data_path"], self.config["n_frames"])
         print("self.frames", self.frames.shape)
@@ -163,14 +168,34 @@ class Preprocess(nn.Module):
         )[0]
         return noise_pred
     
+    @torch.no_grad()
+    def encode_text(self, prompts, device=None):
+        if device is None:
+            device = self.device
+        text_inputs = self.tokenizer(
+            prompts,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+
+        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length:])
+            print(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+        text_embeddings = self.text_encoder(text_input_ids.to(device))[0]
+
+        return text_embeddings
+
     @torch.no_grad()
     def get_text_embeds(self, prompt, negative_prompt, device="cuda"):
-        text_input = self.tokenizer(prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
-                                    truncation=True, return_tensors='pt')
-        text_embeddings = self.text_encoder(text_input.input_ids.to(device))[0]
-        uncond_input = self.tokenizer(negative_prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
-                                      return_tensors='pt')
-        uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(device))[0]
+        text_embeddings = self.encode_text(prompt, device=device)
+        uncond_embeddings = self.encode_text(negative_prompt, device=device)
+
         text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
         return text_embeddings
 
@@ -192,7 +217,7 @@ class Preprocess(nn.Module):
         for i in range(0, len(imgs), batch_size):
             posterior = self.vae.encode(imgs[i:i + batch_size]).latent_dist
             latent = posterior.mean if deterministic else posterior.sample()
-            latents.append(latent * 0.18215)
+            latents.append(latent * self.vae.config.scaling_factor)
         latents = torch.cat(latents)
         return latents
 
@@ -264,6 +289,137 @@ class Preprocess(nn.Module):
             self.total_inverted_latents[f'noisy_latents_{t}'] = latent_frames.clone()
 
         return latent_frames
+    
+    @torch.no_grad()
+    def ddpm_inversion(self, cond, 
+                       latent_frames, 
+                       batch_size, 
+                       num_inversion_steps,
+                       save_path=None,
+                       save_latents=True, 
+                       eta: float = 1.0, 
+                       skip_steps=20):
+        timesteps = self.scheduler.timesteps
+        return_inverted_latents = self.config["frames"] is not None
+
+        variance_noise_shape = (
+            num_inversion_steps,
+            *latent_frames.shape)
+        x0 = latent_frames
+
+        t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
+        xts = torch.zeros(size=variance_noise_shape, device=self.device, dtype=cond.dtype)
+
+        for t in reversed(timesteps):
+            idx = t_to_idx[int(t)]
+            for b in range(0, x0.shape[0], batch_size):
+                x_batch = x0[b:b + batch_size]
+
+                noise = randn_tensor(shape=x_batch.shape, device=self.device, dtype=x0.dtype)
+                xts[idx, b:b + batch_size] = self.scheduler.add_noise(x_batch, noise, t)
+
+        xts = torch.cat([xts, x0.unsqueeze(0)], dim=0)
+
+        zs = torch.zeros(size=variance_noise_shape, device=self.device, dtype=cond.dtype)
+
+        for t in tqdm(timesteps):
+            idx = t_to_idx[int(t)]
+            # 1. predict noise residual
+            for b in range(0, x0.shape[0], batch_size):
+                xt = xts[idx, b:b + batch_size]
+
+                cond_batch = cond.repeat(xt.shape[0], 1, 1)
+                noise_pred = self.unet(xt, timestep=t, encoder_hidden_states=cond_batch).sample
+
+                xtm1 = xts[idx + 1, b:b + batch_size]
+                z, xtm1_corrected = compute_noise(self.scheduler, xtm1, xt, t, noise_pred, eta)
+                zs[idx, b:b + batch_size] = z
+
+                # correction to avoid error accumulation
+                xts[idx + 1, b:b + batch_size] = xtm1_corrected
+
+            if save_latents:
+                torch.save(xts[idx], os.path.join(save_path, 'latents', f'noisy_latents_{t}.pt'))
+            
+            if return_inverted_latents:
+                self.total_inverted_latents[f'noisy_latents_{t}'] = xts[idx].clone()
+
+        if save_path:
+            torch.save(xts[idx], os.path.join(save_path, 'latents', f'noisy_latents_{t}.pt'))
+            torch.save(zs, os.path.join(save_path, 'latents', f'noise_total.pt'))
+        
+        if return_inverted_latents:
+                self.total_inverted_latents[f'noisy_latents_{t}'] = xts[idx].clone()
+                self.noise_total = zs.clone()
+
+        return xts[skip_steps].expand(latent_frames.shape[0], -1, -1, -1), zs
+    
+    def prepare_extra_step_kwargs(self, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        return extra_step_kwargs
+
+    @torch.no_grad()
+    def ddpm_sample(self, init_latents, cond, batch_size, num_inversion_steps, skip_steps, eta, zs_all,
+                    guidance_scale=0):
+        use_ddpm = True
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        total_latents = init_latents
+        self.scheduler.set_timesteps(num_inversion_steps, device=device)
+        timesteps = self.scheduler.timesteps
+        zs_total = zs_all[skip_steps:]
+
+        if use_ddpm:
+            t_to_idx = {int(v): k for k, v in enumerate(timesteps[-zs_total.shape[0]:])}
+            timesteps = timesteps[-zs_total.shape[0]:]
+
+        num_warmup_steps = len(timesteps) - num_inversion_steps * self.scheduler.order
+        extra_step_kwargs = self.prepare_extra_step_kwargs(eta)
+
+        for i, t in enumerate(tqdm(timesteps)):
+            for b in range(0, total_latents.shape[0], batch_size):
+                latents = total_latents[b:b + batch_size]
+                if do_classifier_free_guidance:
+                    latent_model_input = torch.cat([latents] * 2)
+                else:
+                    latent_model_input = latents
+                cond_batch = cond.repeat(latents.shape[0], 1, 1)
+
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=cond_batch,
+                    return_dict=False,
+                )[0]
+
+                if do_classifier_free_guidance:
+                    noise_pred_out = noise_pred.chunk(2)  # [b,4, 64, 64]
+                    noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1]
+
+                    # default text guidance
+                    noise_guidance = guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                    noise_pred = noise_pred_uncond + noise_guidance
+
+                idx = t_to_idx[int(t)]
+                zs = zs_total[idx, b:b + batch_size]
+                latents = self.scheduler.step(noise_pred, t, latents, variance_noise=zs,
+                                              **extra_step_kwargs).prev_sample
+                total_latents[b:b + batch_size] = latents
+        return total_latents
 
     @torch.no_grad()
     def ddim_sample(self, x, cond, batch_size):
@@ -295,6 +451,8 @@ class Preprocess(nn.Module):
                 pred_x0 = (x_batch - sigma * eps) / mu
                 x[b:b + batch_size] = mu_prev * pred_x0 + sigma_prev * eps
         return x
+    
+
 
     @torch.no_grad()
     def extract_latents(self, 
@@ -303,31 +461,89 @@ class Preprocess(nn.Module):
                         batch_size,
                         timesteps_to_save,
                         inversion_prompt='',
-                       reconstruct=False):
+                        skip_steps=20,
+                        inversion_type='ddim', 
+                        eta=1.0, 
+                        reconstruction=False):
+        
         self.scheduler.set_timesteps(num_steps)
         cond = self.get_text_embeds(inversion_prompt, "")[1].unsqueeze(0)
         latent_frames = self.latents
-        print("latent_frames", latent_frames.shape)
-  
-        inverted_x= self.ddim_inversion(cond,
-                                         latent_frames,
-                                         save_path,
-                                         batch_size=batch_size,
-                                         save_latents=True if save_path else False,
-                                         timesteps_to_save=timesteps_to_save)
+        
+        if inversion_type == 'ddim':
+            inverted_x= self.ddim_inversion(cond,
+                                             latent_frames,
+                                             save_path,
+                                             batch_size=batch_size,
+                                             save_latents=True if save_path else False,
+                                             timesteps_to_save=timesteps_to_save)
+
+            if reconstruction:
+                latent_reconstruction = self.ddim_sample(inverted_x, cond, batch_size=batch_size)
+
+                rgb_reconstruction = self.decode_latents(latent_reconstruction)
+                return (self.frames, self.latents, self.total_inverted_latents), rgb_reconstruction
+
+            else:
+                return (self.frames, self.latents, self.total_inverted_latents), None
+        
+        elif inversion_type == 'ddpm':
+            inverted_x, zs = self.ddpm_inversion(cond,
+                                                 latent_frames,
+                                                 save_path= save_path,
+                                                 batch_size=batch_size,
+                                                 save_latents=True if save_path else False,
+                                                 num_inversion_steps=num_steps,
+                                                 eta=eta,
+                                                 skip_steps=skip_steps)
+        
+            cond = self.encode_text(inversion_prompt)
+            if reconstruction:
+                latent_reconstruction = self.ddpm_sample(init_latents=inverted_x,
+                                                         cond=cond, batch_size=batch_size,
+                                                         num_inversion_steps=num_steps, skip_steps=skip_steps,
+                                                         eta=eta, zs_all=zs)
+                rgb_reconstruction = self.decode_latents(latent_reconstruction)
+                return (self.frames, self.latents, self.total_inverted_latents, self.noise_total), rgb_reconstruction
+            else:
+                return (self.frames, self.latents, self.total_inverted_latents, self.noise_total), None
 
+        else:
+            raise NotImplementedError()
 
-       
-        # print("total_inverted_latents", len(total_inverted_latents.keys()))
-        
-        if reconstruct:
-            latent_reconstruction = self.ddim_sample(inverted_x, cond, batch_size=batch_size)
+def compute_noise(scheduler, prev_latents, latents, timestep, noise_pred, eta):
+    # 1. get previous step value (=t-1)
+    prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps
+
+    # 2. compute alphas, betas
+    alpha_prod_t = scheduler.alphas_cumprod[timestep]
+    alpha_prod_t_prev = (
+        scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod
+    )
+
+    beta_prod_t = 1 - alpha_prod_t
 
-            rgb_reconstruction = self.decode_latents(latent_reconstruction)
-            return self.frames, self.latents, self.total_inverted_latents, rgb_reconstruction
+    # 3. compute predicted original sample from predicted noise also called
+    # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+    pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5)
 
-        return self.frames, self.latents, self.total_inverted_latents, None
+    # 4. Clip "predicted x_0"
+    if scheduler.config.clip_sample:
+        pred_original_sample = torch.clamp(pred_original_sample, -1, 1)
 
+    # 5. compute variance: "sigma_t(η)" -> see formula (16)
+    # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
+    variance = scheduler._get_variance(timestep, prev_timestep)
+    std_dev_t = eta * variance ** (0.5)
+
+    # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+    pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t ** 2) ** (0.5) * noise_pred
+
+    # modifed so that updated xtm1 is returned as well (to avoid error accumulation)
+    mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+    noise = (prev_latents - mu_xt) / (variance ** (0.5) * eta)
+
+    return noise, mu_xt + (eta * variance ** 0.5) * noise
 
 def prep(opt):
     # timesteps to save
@@ -348,11 +564,14 @@ def prep(opt):
     seed_everything(opt["seed"])
     if not opt["frames"]: # original non demo setting
         save_path = os.path.join(opt["save_dir"],
+                                 f'inversion_{opt[inversion]}',
                                  f'sd_{opt["sd_version"]}',
                                  Path(opt["data_path"]).stem,
                                  f'steps_{opt["steps"]}',
                                  f'nframes_{opt["n_frames"]}') 
         os.makedirs(os.path.join(save_path, f'latents'), exist_ok=True)
+        if opt[inversion] == 'ddpm':
+            os.makedirs(os.path.join(save_path, f'latents'), exist_ok=True)
         add_dict_to_yaml_file(os.path.join(opt["save_dir"], 'inversion_prompts.yaml'), Path(opt["data_path"]).stem, opt["inversion_prompt"])    
         # save inversion prompt in a txt file
         with open(os.path.join(save_path, 'inversion_prompt.txt'), 'w') as f:
@@ -360,16 +579,31 @@ def prep(opt):
     else:
         save_path = None
     
-    model = Preprocess(device, opt)
+    model = Preprocess(device, 
+                       config,
+                      vae=vae,
+                      text_encoder=text_encoder,
+                      scheduler=scheduler,
+                      tokenizer=tokenizer,
+                      unet=unet)
   
-    frames, latents, total_inverted_latents, rgb_reconstruction = model.extract_latents(
+    frames_and_latents, rgb_reconstruction = model.extract_latents(
                                          num_steps=model.config["steps"],
                                          save_path=save_path,
                                          batch_size=model.config["batch_size"],
                                          timesteps_to_save=timesteps_to_save,
                                          inversion_prompt=model.config["inversion_prompt"],
+                                         inversion_type=model.config[inversion],
+                                         skip_steps=model.config[skip_steps],
+                                         reconstruction=model.config[reconstruct]
     )
 
-    
-    return frames, latents, total_inverted_latents, rgb_reconstruction
+    if model.config[inversion] == 'ddpm':
+        frames, latents, total_inverted_latents, zs = frames_and_latents
+        return frames, latents, total_inverted_latents, zs, rgb_reconstruction
+    else:
+        frames, latents, total_inverted_latents = frames_and_latents
+        return frames, latents, total_inverted_latents, rgb_reconstruction
+
+