always rederive the predicted noise from the clipped x0 for ddim + predict noise objective

fix the previous commit which assumes open_clip is installed

README.md

@@ -1298,4 +1298,14 @@ For detailed information on training the diffusion prior, please refer to the [d
 }
 ```
 
+```bibtex
+@article{Salimans2022ProgressiveDF,
+    title   = {Progressive Distillation for Fast Sampling of Diffusion Models},
+    author  = {Tim Salimans and Jonathan Ho},
+    journal = {ArXiv},
+    year    = {2022},
+    volume  = {abs/2202.00512}
+}
+```
+
 *Creating noise from data is easy; creating data from noise is generative modeling.* - <a href="https://arxiv.org/abs/2011.13456">Yang Song's paper</a>

dalle2_pytorch/__init__.py

@@ -1,6 +1,6 @@
 from dalle2_pytorch.version import __version__
 from dalle2_pytorch.dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder
-from dalle2_pytorch.dalle2_pytorch import OpenAIClipAdapter
+from dalle2_pytorch.dalle2_pytorch import OpenAIClipAdapter, OpenClipAdapter
 from dalle2_pytorch.trainer import DecoderTrainer, DiffusionPriorTrainer
 
 from dalle2_pytorch.vqgan_vae import VQGanVAE

dalle2_pytorch/dalle2_pytorch.py

@@ -100,6 +100,9 @@ def eval_decorator(fn):
         return out
     return inner
 
+def is_float_dtype(dtype):
+    return any([dtype == float_dtype for float_dtype in (torch.float64, torch.float32, torch.float16, torch.bfloat16)])
+
 def is_list_str(x):
     if not isinstance(x, (list, tuple)):
         return False
@@ -357,6 +360,7 @@ class OpenAIClipAdapter(BaseClipAdapter):
         is_eos_id = (text == self.eos_id)
         text_mask_excluding_eos = is_eos_id.cumsum(dim = -1) == 0
         text_mask = F.pad(text_mask_excluding_eos, (1, -1), value = True)
+        text_mask = text_mask & (text != 0)
         assert not self.cleared
 
         text_embed = self.clip.encode_text(text)
@@ -386,6 +390,8 @@ class OpenClipAdapter(BaseClipAdapter):
         self.eos_id = 49407
 
         text_attention_final = self.find_layer('ln_final')
+        self._dim_latent = text_attention_final.weight.shape[0]
+
         self.handle = text_attention_final.register_forward_hook(self._hook)
         self.clip_normalize = preprocess.transforms[-1]
         self.cleared = False
@@ -405,7 +411,7 @@ class OpenClipAdapter(BaseClipAdapter):
 
     @property
     def dim_latent(self):
-        return 512
+        return self._dim_latent
 
     @property
     def image_size(self):
@@ -429,6 +435,7 @@ class OpenClipAdapter(BaseClipAdapter):
         is_eos_id = (text == self.eos_id)
         text_mask_excluding_eos = is_eos_id.cumsum(dim = -1) == 0
         text_mask = F.pad(text_mask_excluding_eos, (1, -1), value = True)
+        text_mask = text_mask & (text != 0)
         assert not self.cleared
 
         text_embed = self.clip.encode_text(text)
@@ -614,7 +621,7 @@ class NoiseScheduler(nn.Module):
         posterior_log_variance_clipped = extract(self.posterior_log_variance_clipped, t, x_t.shape)
         return posterior_mean, posterior_variance, posterior_log_variance_clipped
 
-    def q_sample(self, x_start, t, noise=None):
+    def q_sample(self, x_start, t, noise = None):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
         return (
@@ -622,6 +629,12 @@ class NoiseScheduler(nn.Module):
             extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
         )
 
+    def calculate_v(self, x_start, t, noise = None):
+        return (
+            extract(self.sqrt_alphas_cumprod, t, x_start.shape) * noise -
+            extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * x_start
+        )
+
     def q_sample_from_to(self, x_from, from_t, to_t, noise = None):
         shape = x_from.shape
         noise = default(noise, lambda: torch.randn_like(x_from))
@@ -633,6 +646,12 @@ class NoiseScheduler(nn.Module):
 
         return x_from * (alpha_next / alpha) + noise * (sigma_next * alpha - sigma * alpha_next) / alpha
 
+    def predict_start_from_v(self, x_t, t, v):
+        return (
+            extract(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t -
+            extract(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v
+        )
+
     def predict_start_from_noise(self, x_t, t, noise):
         return (
             extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
@@ -968,6 +987,8 @@ class DiffusionPriorNetwork(nn.Module):
             Rearrange('b (n d) -> b n d', n = num_text_embeds)
         )
 
+        self.continuous_embedded_time = not exists(num_timesteps)
+
         self.to_time_embeds = nn.Sequential(
             nn.Embedding(num_timesteps, dim * num_time_embeds) if exists(num_timesteps) else nn.Sequential(SinusoidalPosEmb(dim), MLP(dim, dim * num_time_embeds)), # also offer a continuous version of timestep embeddings, with a 2 layer MLP
             Rearrange('b (n d) -> b n d', n = num_time_embeds)
@@ -1095,12 +1116,15 @@ class DiffusionPriorNetwork(nn.Module):
         # whether text embedding is used for conditioning depends on whether text encodings are available for attention (for classifier free guidance, even though it seems from the paper it was not used in the prior ddpm, as the objective is different)
         # but let's just do it right
 
+        if self.continuous_embedded_time:
+            diffusion_timesteps = diffusion_timesteps.type(dtype)
+
         time_embed = self.to_time_embeds(diffusion_timesteps)
 
         learned_queries = repeat(self.learned_query, 'd -> b 1 d', b = batch)
 
         if self.self_cond:
-            learned_queries = torch.cat((image_embed, self_cond), dim = -2)
+            learned_queries = torch.cat((self_cond, learned_queries), dim = -2)
 
         tokens = torch.cat((
             text_encodings,
@@ -1136,6 +1160,7 @@ class DiffusionPrior(nn.Module):
         image_cond_drop_prob = None,
         loss_type = "l2",
         predict_x_start = True,
+        predict_v = False,
         beta_schedule = "cosine",
         condition_on_text_encodings = True,  # the paper suggests this is needed, but you can turn it off for your CLIP preprocessed text embed -> image embed training
         sampling_clamp_l2norm = False,       # whether to l2norm clamp the image embed at each denoising iteration (analogous to -1 to 1 clipping for usual DDPMs)
@@ -1187,6 +1212,7 @@ class DiffusionPrior(nn.Module):
         # in paper, they do not predict the noise, but predict x0 directly for image embedding, claiming empirically better results. I'll just offer both.
 
         self.predict_x_start = predict_x_start
+        self.predict_v = predict_v # takes precedence over predict_x_start
 
         # @crowsonkb 's suggestion - https://github.com/lucidrains/DALLE2-pytorch/issues/60#issue-1226116132
 
@@ -1216,7 +1242,9 @@ class DiffusionPrior(nn.Module):
 
         pred = self.net.forward_with_cond_scale(x, t, cond_scale = cond_scale, self_cond = self_cond, **text_cond)
 
-        if self.predict_x_start:
+        if self.predict_v:
+            x_start = self.noise_scheduler.predict_start_from_v(x, t = t, v = pred)
+        elif self.predict_x_start:
             x_start = pred
         else:
             x_start = self.noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
@@ -1289,10 +1317,12 @@ class DiffusionPrior(nn.Module):
 
             # derive x0
 
-            if self.predict_x_start:
+            if self.predict_v:
+                x_start = self.noise_scheduler.predict_start_from_v(image_embed, t = time_cond, v = pred)
+            elif self.predict_x_start:
                 x_start = pred
             else:
-                x_start = self.noise_scheduler.predict_start_from_noise(image_embed, t = time_cond, noise = pred_noise)
+                x_start = self.noise_scheduler.predict_start_from_noise(image_embed, t = time_cond, noise = pred)
 
             # clip x0 before maybe predicting noise
 
@@ -1304,10 +1334,7 @@ class DiffusionPrior(nn.Module):
 
             # predict noise
 
-            if self.predict_x_start:
-                pred_noise = self.noise_scheduler.predict_noise_from_start(image_embed, t = time_cond, x0 = x_start)
-            else:
-                pred_noise = pred
+            pred_noise = self.noise_scheduler.predict_noise_from_start(image_embed, t = time_cond, x0 = x_start)
 
             if time_next < 0:
                 image_embed = x_start
@@ -1362,7 +1389,12 @@ class DiffusionPrior(nn.Module):
         if self.predict_x_start and self.training_clamp_l2norm:
             pred = self.l2norm_clamp_embed(pred)
 
-        target = noise if not self.predict_x_start else image_embed
+        if self.predict_v:
+            target = self.noise_scheduler.calculate_v(image_embed, times, noise)
+        elif self.predict_x_start:
+            target = image_embed
+        else:
+            target = noise
 
         loss = self.noise_scheduler.loss_fn(pred, target)
         return loss
@@ -1432,7 +1464,7 @@ class DiffusionPrior(nn.Module):
         **kwargs
     ):
         assert exists(text) ^ exists(text_embed), 'either text or text embedding must be supplied'
-        assert exists(image) ^ exists(image_embed), 'either text or text embedding must be supplied'
+        assert exists(image) ^ exists(image_embed), 'either image or image embedding must be supplied'
         assert not (self.condition_on_text_encodings and (not exists(text_encodings) and not exists(text))), 'text encodings must be present if you specified you wish to condition on it on initialization'
 
         if exists(image):
@@ -1538,6 +1570,8 @@ class SinusoidalPosEmb(nn.Module):
 
     def forward(self, x):
         dtype, device = x.dtype, x.device
+        assert is_float_dtype(dtype), 'input to sinusoidal pos emb must be a float type'
+
         half_dim = self.dim // 2
         emb = math.log(10000) / (half_dim - 1)
         emb = torch.exp(torch.arange(half_dim, device = device, dtype = dtype) * -emb)
@@ -2436,6 +2470,7 @@ class Decoder(nn.Module):
         loss_type = 'l2',
         beta_schedule = None,
         predict_x_start = False,
+        predict_v = False,
         predict_x_start_for_latent_diffusion = False,
         image_sizes = None,                         # for cascading ddpm, image size at each stage
         random_crop_sizes = None,                   # whether to random crop the image at that stage in the cascade (super resoluting convolutions at the end may be able to generalize on smaller crops)
@@ -2458,7 +2493,7 @@ class Decoder(nn.Module):
         dynamic_thres_percentile = 0.95,
         p2_loss_weight_gamma = 0.,                  # p2 loss weight, from https://arxiv.org/abs/2204.00227 - 0 is equivalent to weight of 1 across time - 1. is recommended
         p2_loss_weight_k = 1,
-        ddim_sampling_eta = 1.                      # can be set to 0. for deterministic sampling afaict
+        ddim_sampling_eta = 0.                      # can be set to 0. for deterministic sampling afaict
     ):
         super().__init__()
 
@@ -2608,6 +2643,10 @@ class Decoder(nn.Module):
 
         self.predict_x_start = cast_tuple(predict_x_start, len(unets)) if not predict_x_start_for_latent_diffusion else tuple(map(lambda t: isinstance(t, VQGanVAE), self.vaes))
 
+        # predict v
+
+        self.predict_v = cast_tuple(predict_v, len(unets))
+
         # input image range
 
         self.input_image_range = (-1. if not auto_normalize_img else 0., 1.)
@@ -2719,14 +2758,16 @@ class Decoder(nn.Module):
         x = x.clamp(-s, s) / s
         return x
 
-    def p_mean_variance(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, lowres_cond_img = None, self_cond = None, clip_denoised = True, predict_x_start = False, learned_variance = False, cond_scale = 1., model_output = None, lowres_noise_level = None):
+    def p_mean_variance(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, lowres_cond_img = None, self_cond = None, clip_denoised = True, predict_x_start = False, predict_v = False, learned_variance = False, cond_scale = 1., model_output = None, lowres_noise_level = None):
         assert not (cond_scale != 1. and not self.can_classifier_guidance), 'the decoder was not trained with conditional dropout, and thus one cannot use classifier free guidance (cond_scale anything other than 1)'
 
         model_output = default(model_output, lambda: unet.forward_with_cond_scale(x, t, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, self_cond = self_cond, lowres_noise_level = lowres_noise_level))
 
         pred, var_interp_frac_unnormalized = self.parse_unet_output(learned_variance, model_output)
 
-        if predict_x_start:
+        if predict_v:
+            x_start = noise_scheduler.predict_start_from_v(x, t = t, v = pred)
+        elif predict_x_start:
             x_start = pred
         else:
             x_start = noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
@@ -2753,9 +2794,9 @@ class Decoder(nn.Module):
         return model_mean, posterior_variance, posterior_log_variance, x_start
 
     @torch.no_grad()
-    def p_sample(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, cond_scale = 1., lowres_cond_img = None, self_cond = None, predict_x_start = False, learned_variance = False, clip_denoised = True, lowres_noise_level = None):
+    def p_sample(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, cond_scale = 1., lowres_cond_img = None, self_cond = None, predict_x_start = False, predict_v = False, learned_variance = False, clip_denoised = True, lowres_noise_level = None):
         b, *_, device = *x.shape, x.device
-        model_mean, _, model_log_variance, x_start = self.p_mean_variance(unet, x = x, t = t, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, self_cond = self_cond, clip_denoised = clip_denoised, predict_x_start = predict_x_start, noise_scheduler = noise_scheduler, learned_variance = learned_variance, lowres_noise_level = lowres_noise_level)
+        model_mean, _, model_log_variance, x_start = self.p_mean_variance(unet, x = x, t = t, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, self_cond = self_cond, clip_denoised = clip_denoised, predict_x_start = predict_x_start, predict_v = predict_v, noise_scheduler = noise_scheduler, learned_variance = learned_variance, lowres_noise_level = lowres_noise_level)
         noise = torch.randn_like(x)
         # no noise when t == 0
         nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
@@ -2770,6 +2811,7 @@ class Decoder(nn.Module):
         image_embed,
         noise_scheduler,
         predict_x_start = False,
+        predict_v = False,
         learned_variance = False,
         clip_denoised = True,
         lowres_cond_img = None,
@@ -2828,6 +2870,7 @@ class Decoder(nn.Module):
                     lowres_cond_img = lowres_cond_img,
                     lowres_noise_level = lowres_noise_level,
                     predict_x_start = predict_x_start,
+                    predict_v = predict_v,
                     noise_scheduler = noise_scheduler,
                     learned_variance = learned_variance,
                     clip_denoised = clip_denoised
@@ -2853,6 +2896,7 @@ class Decoder(nn.Module):
         timesteps,
         eta = 1.,
         predict_x_start = False,
+        predict_v = False,
         learned_variance = False,
         clip_denoised = True,
         lowres_cond_img = None,
@@ -2914,7 +2958,9 @@ class Decoder(nn.Module):
 
                 # predict x0
 
-                if predict_x_start:
+                if predict_v:
+                    x_start = noise_scheduler.predict_start_from_v(img, t = time_cond, v = pred)
+                elif predict_x_start:
                     x_start = pred
                 else:
                     x_start = noise_scheduler.predict_start_from_noise(img, t = time_cond, noise = pred)
@@ -2926,10 +2972,7 @@ class Decoder(nn.Module):
 
                 # predict noise
 
-                if predict_x_start:
-                    pred_noise = noise_scheduler.predict_noise_from_start(img, t = time_cond, x0 = pred)
-                else:
-                    pred_noise = pred
+                pred_noise = noise_scheduler.predict_noise_from_start(img, t = time_cond, x0 = x_start)
 
                 c1 = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
                 c2 = ((1 - alpha_next) - torch.square(c1)).sqrt()
@@ -2963,7 +3006,7 @@ class Decoder(nn.Module):
 
         return self.p_sample_loop_ddim(*args, noise_scheduler = noise_scheduler, timesteps = timesteps, **kwargs)
 
-    def p_losses(self, unet, x_start, times, *, image_embed, noise_scheduler, lowres_cond_img = None, text_encodings = None, predict_x_start = False, noise = None, learned_variance = False, clip_denoised = False, is_latent_diffusion = False, lowres_noise_level = None):
+    def p_losses(self, unet, x_start, times, *, image_embed, noise_scheduler, lowres_cond_img = None, text_encodings = None, predict_x_start = False, predict_v = False, noise = None, learned_variance = False, clip_denoised = False, is_latent_diffusion = False, lowres_noise_level = None):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
         # normalize to [-1, 1]
@@ -3008,7 +3051,12 @@ class Decoder(nn.Module):
 
         pred, _ = self.parse_unet_output(learned_variance, unet_output)
 
-        target = noise if not predict_x_start else x_start
+        if predict_v:
+            target = noise_scheduler.calculate_v(x_start, times, noise)
+        elif predict_x_start:
+            target = x_start
+        else:
+            target = noise
 
         loss = noise_scheduler.loss_fn(pred, target, reduction = 'none')
         loss = reduce(loss, 'b ... -> b (...)', 'mean')
@@ -3094,7 +3142,7 @@ class Decoder(nn.Module):
         num_unets = self.num_unets
         cond_scale = cast_tuple(cond_scale, num_unets)
 
-        for unet_number, unet, vae, channel, image_size, predict_x_start, learned_variance, noise_scheduler, lowres_cond, sample_timesteps, unet_cond_scale in tqdm(zip(range(1, num_unets + 1), self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start, self.learned_variance, self.noise_schedulers, self.lowres_conds, self.sample_timesteps, cond_scale)):
+        for unet_number, unet, vae, channel, image_size, predict_x_start, predict_v, learned_variance, noise_scheduler, lowres_cond, sample_timesteps, unet_cond_scale in tqdm(zip(range(1, num_unets + 1), self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start, self.predict_v, self.learned_variance, self.noise_schedulers, self.lowres_conds, self.sample_timesteps, cond_scale)):
             if unet_number < start_at_unet_number:
                 continue  # It's the easiest way to do it
 
@@ -3130,6 +3178,7 @@ class Decoder(nn.Module):
                     text_encodings = text_encodings,
                     cond_scale = unet_cond_scale,
                     predict_x_start = predict_x_start,
+                    predict_v = predict_v,
                     learned_variance = learned_variance,
                     clip_denoised = not is_latent_diffusion,
                     lowres_cond_img = lowres_cond_img,
@@ -3169,6 +3218,7 @@ class Decoder(nn.Module):
         lowres_conditioner  = self.lowres_conds[unet_index]
         target_image_size   = self.image_sizes[unet_index]
         predict_x_start     = self.predict_x_start[unet_index]
+        predict_v           = self.predict_v[unet_index]
         random_crop_size    = self.random_crop_sizes[unet_index]
         learned_variance    = self.learned_variance[unet_index]
         b, c, h, w, device, = *image.shape, image.device
@@ -3207,7 +3257,7 @@ class Decoder(nn.Module):
             image = vae.encode(image)
             lowres_cond_img = maybe(vae.encode)(lowres_cond_img)
 
-        losses = self.p_losses(unet, image, times, image_embed = image_embed, text_encodings = text_encodings, lowres_cond_img = lowres_cond_img, predict_x_start = predict_x_start, learned_variance = learned_variance, is_latent_diffusion = is_latent_diffusion, noise_scheduler = noise_scheduler, lowres_noise_level = lowres_noise_level)
+        losses = self.p_losses(unet, image, times, image_embed = image_embed, text_encodings = text_encodings, lowres_cond_img = lowres_cond_img, predict_x_start = predict_x_start, predict_v = predict_v, learned_variance = learned_variance, is_latent_diffusion = is_latent_diffusion, noise_scheduler = noise_scheduler, lowres_noise_level = lowres_noise_level)
 
         if not return_lowres_cond_image:
             return losses

dalle2_pytorch/train_configs.py

@@ -4,11 +4,13 @@ from pydantic import BaseModel, validator, root_validator
 from typing import List, Optional, Union, Tuple, Dict, Any, TypeVar
 
 from x_clip import CLIP as XCLIP
+from open_clip import list_pretrained
 from coca_pytorch import CoCa
 
 from dalle2_pytorch.dalle2_pytorch import (
     CoCaAdapter,
     OpenAIClipAdapter,
+    OpenClipAdapter,
     Unet,
     Decoder,
     DiffusionPrior,
@@ -117,6 +119,10 @@ class AdapterConfig(BaseModel):
     def create(self):
         if self.make == "openai":
             return OpenAIClipAdapter(self.model)
+        elif self.make == "open_clip":
+            pretrained = dict(list_pretrained())
+            checkpoint = pretrained[self.model]
+            return OpenClipAdapter(name=self.model, pretrained=checkpoint)
         elif self.make == "x-clip":
             return XClipAdapter(XCLIP(**self.base_model_kwargs))
         elif self.make == "coca":
@@ -307,6 +313,7 @@ class DecoderTrainConfig(BaseModel):
     wd: SingularOrIterable[float] = 0.01
     warmup_steps: Optional[SingularOrIterable[int]] = None
     find_unused_parameters: bool = True
+    static_graph: bool = True
     max_grad_norm: SingularOrIterable[float] = 0.5
     save_every_n_samples: int = 100000
     n_sample_images: int = 6                       # The number of example images to produce when sampling the train and test dataset

dalle2_pytorch/trainer.py

@@ -236,7 +236,7 @@ class DiffusionPriorTrainer(nn.Module):
         )
 
         if exists(cosine_decay_max_steps):
-            self.scheduler = CosineAnnealingLR(optimizer, T_max = cosine_decay_max_steps)
+            self.scheduler = CosineAnnealingLR(self.optimizer, T_max = cosine_decay_max_steps)
         else:
             self.scheduler = LambdaLR(self.optimizer, lr_lambda = lambda _: 1.0)
         

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.10.7'
+__version__ = '1.12.3'

setup.py

@@ -26,7 +26,8 @@ setup(
   install_requires=[
     'accelerate',
     'click',
-    'clip-anytorch>=2.4.0',
+    'open-clip-torch>=2.0.0,<3.0.0',
+    'clip-anytorch>=2.5.2',
     'coca-pytorch>=0.0.5',
     'ema-pytorch>=0.0.7',
     'einops>=0.4',

train_decoder.py

@@ -556,7 +556,7 @@ def initialize_training(config: TrainDecoderConfig, config_path):
     torch.manual_seed(config.seed)
 
     # Set up accelerator for configurable distributed training
-    ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=config.train.find_unused_parameters)
+    ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=config.train.find_unused_parameters, static_graph=config.train.static_graph)
     init_kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=60*60))
     accelerator = Accelerator(kwargs_handlers=[ddp_kwargs, init_kwargs])