makes more sense for blur augmentation to happen before the upsampling

added diffusion-gan thoughts

README.md

@@ -1,8 +1,10 @@
 <img src="./dalle2.png" width="450px"></img>
 
-## DALL-E 2 - Pytorch (wip)
+## DALL-E 2 - Pytorch
 
-Implementation of <a href="https://openai.com/dall-e-2/">DALL-E 2</a>, OpenAI's updated text-to-image synthesis neural network, in Pytorch. <a href="https://youtu.be/RJwPN4qNi_Y?t=555">Yannic Kilcher summary</a>
+Implementation of <a href="https://openai.com/dall-e-2/">DALL-E 2</a>, OpenAI's updated text-to-image synthesis neural network, in Pytorch.
+
+<a href="https://youtu.be/RJwPN4qNi_Y?t=555">Yannic Kilcher summary</a> | <a href="https://www.youtube.com/watch?v=F1X4fHzF4mQ">AssemblyAI explainer</a>
 
 The main novelty seems to be an extra layer of indirection with the prior network (whether it is an autoregressive transformer or a diffusion network), which predicts an image embedding based on the text embedding from CLIP. Specifically, this repository will only build out the diffusion prior network, as it is the best performing variant (but which incidentally involves a causal transformer as the denoising network 😂)
 
@@ -12,9 +14,7 @@ It may also explore an extension of using <a href="https://huggingface.co/spaces
 
 Please join <a href="https://discord.gg/xBPBXfcFHd"><img alt="Join us on Discord" src="https://img.shields.io/discord/823813159592001537?color=5865F2&logo=discord&logoColor=white"></a> if you are interested in helping out with the replication
 
-Do let me know if anyone is interested in a Jax version https://github.com/lucidrains/DALLE2-pytorch/discussions/8
-
-For all of you emailing me (there is a lot), the best way to contribute is through pull requests. Everything is open sourced after all. All my thoughts are public. This is your moment to participate.
+There was enough interest for a Jax version. It will be completed after the Pytorch version shows signs of life on my toy tasks. <a href="https://github.com/lucidrains/dalle2-jax">Placeholder repository</a>. I will also eventually extend this to <a href="https://github.com/lucidrains/dalle2-video">text to video</a>, once the repository is in a good place.
 
 ## Install
 
@@ -109,7 +109,7 @@ unet = Unet(
 # decoder, which contains the unet and clip
 
 decoder = Decoder(
-    net = unet,
+    unet = unet,
     clip = clip,
     timesteps = 100,
     cond_drop_prob = 0.2
@@ -182,7 +182,82 @@ loss.backward()
 # now the diffusion prior can generate image embeddings from the text embeddings
 ```
 
-Finally, to generate the DALL-E2 images from text. Insert the trained `DiffusionPrior` as well as the `Decoder` (which both contains `CLIP`, a unet, and a causal transformer)
+In the paper, they actually used a <a href="https://cascaded-diffusion.github.io/">recently discovered technique</a>, from <a href="http://www.jonathanho.me/">Jonathan Ho</a> himself (original author of DDPMs, the core technique used in DALL-E v2) for high resolution image synthesis.
+
+This can easily be used within this framework as so
+
+```python
+import torch
+from dalle2_pytorch import Unet, Decoder, CLIP
+
+# trained clip from step 1
+
+clip = CLIP(
+    dim_text = 512,
+    dim_image = 512,
+    dim_latent = 512,
+    num_text_tokens = 49408,
+    text_enc_depth = 6,
+    text_seq_len = 256,
+    text_heads = 8,
+    visual_enc_depth = 6,
+    visual_image_size = 256,
+    visual_patch_size = 32,
+    visual_heads = 8
+).cuda()
+
+# 2 unets for the decoder (a la cascading DDPM)
+
+unet1 = Unet(
+    dim = 32,
+    image_embed_dim = 512,
+    cond_dim = 128,
+    channels = 3,
+    dim_mults = (1, 2, 4, 8)
+).cuda()
+
+unet2 = Unet(
+    dim = 32,
+    image_embed_dim = 512,
+    cond_dim = 128,
+    channels = 3,
+    dim_mults = (1, 2, 4, 8, 16)
+).cuda()
+
+# decoder, which contains the unet(s) and clip
+
+decoder = Decoder(
+    clip = clip,
+    unet = (unet1, unet2),            # insert both unets in order of low resolution to highest resolution (you can have as many stages as you want here)
+    image_sizes = (256, 512),         # resolutions, 256 for first unet, 512 for second. these must be unique and in ascending order (matches with the unets passed in)
+    timesteps = 1000,
+    cond_drop_prob = 0.2
+).cuda()
+
+# mock images (get a lot of this)
+
+images = torch.randn(4, 3, 512, 512).cuda()
+
+# feed images into decoder, specifying which unet you want to train
+# each unet can be trained separately, which is one of the benefits of the cascading DDPM scheme
+
+loss = decoder(images, unet_number = 1)
+loss.backward()
+
+loss = decoder(images, unet_number = 2)
+loss.backward()
+
+# do the above for many steps for both unets
+
+# then it will learn to generate images based on the CLIP image embeddings
+
+# chaining the unets from lowest resolution to highest resolution (thus cascading)
+
+mock_image_embed = torch.randn(1, 512).cuda()
+images = decoder.sample(mock_image_embed) # (1, 3, 512, 512)
+```
+
+Finally, to generate the DALL-E2 images from text. Insert the trained `DiffusionPrior` as well as the `Decoder` (which wraps `CLIP`, the causal transformer, and unet(s))
 
 ```python
 from dalle2_pytorch import DALLE2
@@ -261,7 +336,7 @@ loss.backward()
 
 # decoder (with unet)
 
-unet = Unet(
+unet1 = Unet(
     dim = 128,
     image_embed_dim = 512,
     cond_dim = 128,
@@ -269,15 +344,25 @@ unet = Unet(
     dim_mults=(1, 2, 4, 8)
 ).cuda()
 
+unet2 = Unet(
+    dim = 16,
+    image_embed_dim = 512,
+    cond_dim = 128,
+    channels = 3,
+    dim_mults = (1, 2, 4, 8, 16)
+).cuda()
+
 decoder = Decoder(
-    net = unet,
+    unet = (unet1, unet2),
+    image_sizes = (128, 256),
     clip = clip,
     timesteps = 100,
     cond_drop_prob = 0.2
 ).cuda()
 
-loss = decoder(images) # this can optionally be decoder(images, text) if you wish to condition on the text encodings as well, though it was hinted in the paper it didn't do much
-loss.backward()
+for unet_number in (1, 2):
+    loss = decoder(images, unet_number = unet_number) # this can optionally be decoder(images, text) if you wish to condition on the text encodings as well, though it was hinted in the paper it didn't do much
+    loss.backward()
 
 # do above for many steps
 
@@ -291,11 +376,13 @@ images = dalle2(
     cond_scale = 2. # classifier free guidance strength (> 1 would strengthen the condition)
 )
 
-# save your image
+# save your image (in this example, of size 256x256)
 ```
 
 Everything in this readme should run without error
 
+You can also train the decoder on images of greater than the size (say 512x512) at which CLIP was trained (256x256). The images will be resized to CLIP image resolution for the image embeddings
+
 For the layperson, no worries, training will all be automated into a CLI tool, at least for small scale training.
 
 ## CLI Usage (work in progress)
@@ -320,12 +407,16 @@ Offer training wrappers
 - [x] add what was proposed in the paper, where DDPM objective for image latent embedding predicts x0 directly (reread vq-diffusion paper and get caught up on that line of work)
 - [x] make sure it works end to end to produce an output tensor, taking a single gradient step
 - [x] augment unet so that it can also be conditioned on text encodings (although in paper they hinted this didn't make much a difference)
-- [ ] look into Jonathan Ho's cascading DDPM for the decoder, as that seems to be what they are using. get caught up on DDPM literature
-- [ ] figure out all the current bag of tricks needed to make DDPMs great (starting with the blur trick mentioned in paper)
+- [x] figure out all the current bag of tricks needed to make DDPMs great (starting with the blur trick mentioned in paper)
+- [x] build the cascading ddpm by having Decoder class manage multiple unets at different resolutions
+- [x] add efficient attention in unet
+- [x] be able to finely customize what to condition on (text, image embed) for specific unet in the cascade (super resolution ddpms near the end may not need too much conditioning)
+- [x] offload unets not being trained on to CPU for memory efficiency (for training each resolution unets separately)
+- [ ] build out latent diffusion architecture, with the vq-reg variant (vqgan-vae), make it completely optional
+- [ ] become an expert with unets, cleanup unet code, make it fully configurable, port all learnings over to https://github.com/lucidrains/x-unet
+- [ ] copy the cascading ddpm code to a separate repo (perhaps https://github.com/lucidrains/denoising-diffusion-pytorch) as the main contribution of dalle2 really is just the prior network
 - [ ] train on a toy task, offer in colab
-- [ ] add attention to unet - apply some personal tricks with efficient attention
-- [ ] figure out the big idea behind latent diffusion and what can be ported over
-- [ ] consider U2-net for decoder https://arxiv.org/abs/2005.09007
+- [ ] extend diffusion head to use diffusion-gan (potentially using lightweight-gan) to speed up inference
 
 ## Citations
 
@@ -374,4 +465,12 @@ Offer training wrappers
 }
 ```
 
+```bibtex
+@inproceedings{Tu2022MaxViTMV,
+    title   = {MaxViT: Multi-Axis Vision Transformer},
+    author  = {Zhe-Wei Tu and Hossein Talebi and Han Zhang and Feng Yang and Peyman Milanfar and Alan Conrad Bovik and Yinxiao Li},
+    year    = {2022}
+}
+```
+
 *Creating noise from data is easy; creating data from noise is generative modeling.* - Yang Song's <a href="https://arxiv.org/abs/2011.13456">paper</a>

dalle2_pytorch/__init__.py

@@ -1,2 +1,4 @@
 from dalle2_pytorch.dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder
+
+from dalle2_pytorch.vqgan_vae import VQGanVAE
 from x_clip import CLIP

dalle2_pytorch/cli.py

@@ -6,4 +6,4 @@ def main():
 @click.command()
 @click.argument('text')
 def dream(text):
-    return image
+    return 'not ready yet'

dalle2_pytorch/dalle2_pytorch.py

@@ -1,6 +1,8 @@
 import math
 from tqdm import tqdm
 from inspect import isfunction
+from functools import partial
+from contextlib import contextmanager
 
 import torch
 import torch.nn.functional as F
@@ -11,7 +13,7 @@ from einops.layers.torch import Rearrange
 from einops_exts import rearrange_many, repeat_many, check_shape
 from einops_exts.torch import EinopsToAndFrom
 
-from kornia.filters.gaussian import GaussianBlur2d
+from kornia.filters import gaussian_blur2d
 
 from dalle2_pytorch.tokenizer import tokenizer
 
@@ -29,6 +31,9 @@ def default(val, d):
         return val
     return d() if isfunction(d) else d
 
+def cast_tuple(val, length = 1):
+    return val if isinstance(val, tuple) else ((val,) * length)
+
 def eval_decorator(fn):
     def inner(model, *args, **kwargs):
         was_training = model.training
@@ -64,6 +69,15 @@ def freeze_model_and_make_eval_(model):
 def l2norm(t):
     return F.normalize(t, dim = -1)
 
+def resize_image_to(t, image_size, mode = 'bilinear'): # take a look at https://github.com/assafshocher/ResizeRight
+    shape = cast_tuple(image_size, 2)
+    orig_image_size = t.shape[-2:]
+
+    if orig_image_size == shape:
+        return t
+
+    return F.interpolate(t, size = shape, mode = mode)
+
 # classifier free guidance functions
 
 def prob_mask_like(shape, prob, device):
@@ -92,12 +106,35 @@ def cosine_beta_schedule(timesteps, s = 0.008):
     as proposed in https://openreview.net/forum?id=-NEXDKk8gZ
     """
     steps = timesteps + 1
-    x = torch.linspace(0, steps, steps)
-    alphas_cumprod = torch.cos(((x / steps) + s) / (1 + s) * torch.pi * 0.5) ** 2
+    x = torch.linspace(0, timesteps, steps)
+    alphas_cumprod = torch.cos(((x / timesteps) + s) / (1 + s) * torch.pi * 0.5) ** 2
     alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
     betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
     return torch.clip(betas, 0, 0.999)
 
+
+def linear_beta_schedule(timesteps):
+    scale = 1000 / timesteps
+    beta_start = scale * 0.0001
+    beta_end = scale * 0.02
+    return torch.linspace(beta_start, beta_end, timesteps)
+
+
+def quadratic_beta_schedule(timesteps):
+    scale = 1000 / timesteps
+    beta_start = scale * 0.0001
+    beta_end = scale * 0.02
+    return torch.linspace(beta_start**2, beta_end**2, timesteps) ** 2
+
+
+def sigmoid_beta_schedule(timesteps):
+    scale = 1000 / timesteps
+    beta_start = scale * 0.0001
+    beta_end = scale * 0.02
+    betas = torch.linspace(-6, 6, timesteps)
+    return torch.sigmoid(betas) * (beta_end - beta_start) + beta_start
+
+
 # diffusion prior
 
 class RMSNorm(nn.Module):
@@ -429,8 +466,9 @@ class DiffusionPrior(nn.Module):
         clip,
         timesteps = 1000,
         cond_drop_prob = 0.2,
-        loss_type = 'l1',
-        predict_x0 = True
+        loss_type = "l1",
+        predict_x0 = True,
+        beta_schedule = "cosine",
     ):
         super().__init__()
         assert isinstance(clip, CLIP)
@@ -446,10 +484,21 @@ class DiffusionPrior(nn.Module):
         self.predict_x0 = predict_x0
         # in paper, they do not predict the noise, but predict x0 directly for image embedding, claiming empirically better results. I'll just offer both.
 
-        betas = cosine_beta_schedule(timesteps)
+        if beta_schedule == "cosine":
+            betas = cosine_beta_schedule(timesteps)
+        elif beta_schedule == "linear":
+            betas = linear_beta_schedule(timesteps)
+        elif beta_schedule == "quadratic":
+            betas = quadratic_beta_schedule(timesteps)
+        elif beta_schedule == "jsd":
+            betas = 1.0 / torch.linspace(timesteps, 1, timesteps)
+        elif beta_schedule == "sigmoid":
+            betas = sigmoid_beta_schedule(timesteps)
+        else:
+            raise NotImplementedError()
 
         alphas = 1. - betas
-        alphas_cumprod = torch.cumprod(alphas, axis=0)
+        alphas_cumprod = torch.cumprod(alphas, axis = 0)
         alphas_cumprod_prev = F.pad(alphas_cumprod[:-1], (1, 0), value = 1.)
 
         timesteps, = betas.shape
@@ -550,31 +599,6 @@ class DiffusionPrior(nn.Module):
             img = self.p_sample(img, torch.full((b,), i, device = device, dtype = torch.long), text_cond = text_cond)
         return img
 
-    @torch.no_grad()
-    def sample(self, text, num_samples_per_batch = 2):
-        # in the paper, what they did was
-        # sample 2 image embeddings, choose the top 1 similarity, as judged by CLIP
-        text = repeat(text, 'b ... -> (b r) ...', r = num_samples_per_batch)
-
-        batch_size = text.shape[0]
-        image_embed_dim = self.image_embed_dim
-
-        text_cond = self.get_text_cond(text)
-
-        image_embeds = self.p_sample_loop((batch_size, image_embed_dim), text_cond = text_cond)
-        text_embeds = text_cond['text_embed']
-
-        text_embeds = rearrange(text_embeds, '(b r) d -> b r d', r = num_samples_per_batch)
-        image_embeds = rearrange(image_embeds, '(b r) d -> b r d', r = num_samples_per_batch)
-
-        text_image_sims = einsum('b r d, b r d -> b r', l2norm(text_embeds), l2norm(image_embeds))
-        top_sim_indices = text_image_sims.topk(k = 1).indices
-
-        top_sim_indices = repeat(top_sim_indices, 'b 1 -> b 1 d', d = image_embed_dim)
-
-        top_image_embeds = image_embeds.gather(1, top_sim_indices)
-        return rearrange(top_image_embeds, 'b 1 d -> b d')
-
     def q_sample(self, x_start, t, noise=None):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
@@ -601,11 +625,39 @@ class DiffusionPrior(nn.Module):
             loss = F.l1_loss(to_predict, x_recon)
         elif self.loss_type == 'l2':
             loss = F.mse_loss(to_predict, x_recon)
+        elif self.loss_type == "huber":
+            loss = F.smooth_l1_loss(to_predict, x_recon)
         else:
             raise NotImplementedError()
 
         return loss
 
+    @torch.no_grad()
+    @eval_decorator
+    def sample(self, text, num_samples_per_batch = 2):
+        # in the paper, what they did was
+        # sample 2 image embeddings, choose the top 1 similarity, as judged by CLIP
+        text = repeat(text, 'b ... -> (b r) ...', r = num_samples_per_batch)
+
+        batch_size = text.shape[0]
+        image_embed_dim = self.image_embed_dim
+
+        text_cond = self.get_text_cond(text)
+
+        image_embeds = self.p_sample_loop((batch_size, image_embed_dim), text_cond = text_cond)
+        text_embeds = text_cond['text_embed']
+
+        text_embeds = rearrange(text_embeds, '(b r) d -> b r d', r = num_samples_per_batch)
+        image_embeds = rearrange(image_embeds, '(b r) d -> b r d', r = num_samples_per_batch)
+
+        text_image_sims = einsum('b r d, b r d -> b r', l2norm(text_embeds), l2norm(image_embeds))
+        top_sim_indices = text_image_sims.topk(k = 1).indices
+
+        top_sim_indices = repeat(top_sim_indices, 'b 1 -> b 1 d', d = image_embed_dim)
+
+        top_image_embeds = image_embeds.gather(1, top_sim_indices)
+        return rearrange(top_image_embeds, 'b 1 d -> b d')
+
     def forward(self, text, image, *args, **kwargs):
         b, device, img_size, = image.shape[0], image.device, self.image_size
         check_shape(image, 'b c h w', h = img_size, w = img_size, c = self.channels)
@@ -747,6 +799,20 @@ class CrossAttention(nn.Module):
         out = rearrange(out, 'b h n d -> b n (h d)')
         return self.to_out(out)
 
+class GridAttention(nn.Module):
+    def __init__(self, *args, window_size = 8, **kwargs):
+        super().__init__()
+        self.window_size = window_size
+        self.attn = Attention(*args, **kwargs)
+
+    def forward(self, x):
+        h, w = x.shape[-2:]
+        wsz = self.window_size
+        x = rearrange(x, 'b c (w1 h) (w2 w) -> (b h w) (w1 w2) c', w1 = wsz, w2 = wsz)
+        out = self.attn(x)
+        out = rearrange(out, '(b h w) (w1 w2) c -> b c (w1 h) (w2 w)', w1 = wsz, w2 = wsz, h = h // wsz, w = w // wsz)
+        return out
+
 class Unet(nn.Module):
     def __init__(
         self,
@@ -755,20 +821,29 @@ class Unet(nn.Module):
         image_embed_dim,
         cond_dim = None,
         num_image_tokens = 4,
+        num_time_tokens = 2,
         out_dim = None,
         dim_mults=(1, 2, 4, 8),
         channels = 3,
+        attn_dim_head = 32,
+        attn_heads = 8,
         lowres_cond = False, # for cascading diffusion - https://cascaded-diffusion.github.io/
-        lowres_cond_upsample_mode = 'bilinear',
-        blur_sigma = 0.1
+        sparse_attn = False,
+        sparse_attn_window = 8,  # window size for sparse attention
+        attend_at_middle = True, # whether to have a layer of attention at the bottleneck (can turn off for higher resolution in cascading DDPM, before bringing in efficient attention)
+        cond_on_text_encodings = False,
+        cond_on_image_embeds = False,
     ):
         super().__init__()
+        # save locals to take care of some hyperparameters for cascading DDPM
+
+        self._locals = locals()
+        del self._locals['self']
+        del self._locals['__class__']
 
         # for eventual cascading diffusion
 
         self.lowres_cond = lowres_cond
-        self.lowres_cond_upsample_mode = lowres_cond_upsample_mode
-        self.lowres_cond_blur = GaussianBlur2d((3, 3), (blur_sigma, blur_sigma))
 
         # determine dimensions
 
@@ -787,8 +862,8 @@ class Unet(nn.Module):
             SinusoidalPosEmb(dim),
             nn.Linear(dim, dim * 4),
             nn.GELU(),
-            nn.Linear(dim * 4, cond_dim),
-            Rearrange('b d -> b 1 d')
+            nn.Linear(dim * 4, cond_dim * num_time_tokens),
+            Rearrange('b (r d) -> b r d', r = num_time_tokens)
         )
 
         self.image_to_cond = nn.Sequential(
@@ -798,11 +873,21 @@ class Unet(nn.Module):
 
         self.text_to_cond = nn.LazyLinear(cond_dim)
 
+        # finer control over whether to condition on image embeddings and text encodings
+        # so one can have the latter unets in the cascading DDPMs only focus on super-resoluting
+
+        self.cond_on_text_encodings = cond_on_text_encodings
+        self.cond_on_image_embeds = cond_on_image_embeds
+
         # for classifier free guidance
 
         self.null_image_embed = nn.Parameter(torch.randn(1, num_image_tokens, cond_dim))
         self.null_text_embed = nn.Parameter(torch.randn(1, 1, cond_dim))
 
+        # attention related params
+
+        attn_kwargs = dict(heads = attn_heads, dim_head = attn_dim_head)
+
         # layers
 
         self.downs = nn.ModuleList([])
@@ -810,27 +895,32 @@ class Unet(nn.Module):
         num_resolutions = len(in_out)
 
         for ind, (dim_in, dim_out) in enumerate(in_out):
+            is_first = ind == 0
             is_last = ind >= (num_resolutions - 1)
+            layer_cond_dim = cond_dim if not is_first else None
 
             self.downs.append(nn.ModuleList([
                 ConvNextBlock(dim_in, dim_out, norm = ind != 0),
-                ConvNextBlock(dim_out, dim_out, cond_dim = cond_dim),
+                Residual(GridAttention(dim_out, window_size = sparse_attn_window, **attn_kwargs)) if sparse_attn else nn.Identity(),
+                ConvNextBlock(dim_out, dim_out, cond_dim = layer_cond_dim),
                 Downsample(dim_out) if not is_last else nn.Identity()
             ]))
 
         mid_dim = dims[-1]
 
         self.mid_block1 = ConvNextBlock(mid_dim, mid_dim, cond_dim = cond_dim)
-        self.mid_attn = EinopsToAndFrom('b c h w', 'b (h w) c', Residual(Attention(mid_dim)))
+        self.mid_attn = EinopsToAndFrom('b c h w', 'b (h w) c', Residual(Attention(mid_dim, **attn_kwargs))) if attend_at_middle else None
         self.mid_block2 = ConvNextBlock(mid_dim, mid_dim, cond_dim = cond_dim)
 
         for ind, (dim_in, dim_out) in enumerate(reversed(in_out[1:])):
-            is_last = ind >= (num_resolutions - 1)
+            is_last = ind >= (num_resolutions - 2)
+            layer_cond_dim = cond_dim if not is_last else None
 
             self.ups.append(nn.ModuleList([
-                ConvNextBlock(dim_out * 2, dim_in, cond_dim = cond_dim),
-                ConvNextBlock(dim_in, dim_in, cond_dim = cond_dim),
-                Upsample(dim_in) if not is_last else nn.Identity()
+                ConvNextBlock(dim_out * 2, dim_in, cond_dim = layer_cond_dim),
+                Residual(GridAttention(dim_in, window_size = sparse_attn_window, **attn_kwargs)) if sparse_attn else nn.Identity(),
+                ConvNextBlock(dim_in, dim_in, cond_dim = layer_cond_dim),
+                Upsample(dim_in)
             ]))
 
         out_dim = default(out_dim, channels)
@@ -839,6 +929,15 @@ class Unet(nn.Module):
             nn.Conv2d(dim, out_dim, 1)
         )
 
+    # if the current settings for the unet are not correct
+    # for cascading DDPM, then reinit the unet with the right settings
+    def force_lowres_cond(self, lowres_cond):
+        if lowres_cond == self.lowres_cond:
+            return self
+
+        updated_kwargs = {**self._locals, 'lowres_cond': lowres_cond}
+        return self.__class__(**updated_kwargs)
+
     def forward_with_cond_scale(
         self,
         *args,
@@ -861,20 +960,17 @@ class Unet(nn.Module):
         image_embed,
         lowres_cond_img = None,
         text_encodings = None,
-        cond_drop_prob = 0.
+        cond_drop_prob = 0.,
+        blur_sigma = None,
+        blur_kernel_size = None
     ):
         batch_size, device = x.shape[0], x.device
 
         # add low resolution conditioning, if present
 
-        assert not self.lowres_cond and not exists(lowres_cond_img), 'low resolution conditioning image must be present'
+        assert not (self.lowres_cond and not exists(lowres_cond_img)), 'low resolution conditioning image must be present'
 
         if exists(lowres_cond_img):
-            if self.training:
-                # when training, blur the low resolution conditional image
-                lowres_cond_img = self.lowres_cond_blur(lowres_cond_img)
-
-            lowres_cond_img = F.interpolate(lowres_cond_img, size = x.shape[-2:], mode = self.lowres_cond_upsample_mode)
             x = torch.cat((x, lowres_cond_img), dim = 1)
 
         # time conditioning
@@ -889,17 +985,22 @@ class Unet(nn.Module):
         # mask out image embedding depending on condition dropout
         # for classifier free guidance
 
-        image_tokens = self.image_to_cond(image_embed)
+        image_tokens = None
 
-        image_tokens = torch.where(
-            cond_prob_mask,
-            image_tokens,
-            self.null_image_embed
-        )
+        if self.cond_on_image_embeds:
+            image_tokens = self.image_to_cond(image_embed)
+
+            image_tokens = torch.where(
+                cond_prob_mask,
+                image_tokens,
+                self.null_image_embed
+            )
 
         # take care of text encodings (optional)
 
-        if exists(text_encodings):
+        text_tokens = None
+
+        if exists(text_encodings) and self.cond_on_text_encodings:
             text_tokens = self.text_to_cond(text_encodings)
             text_tokens = torch.where(
                 cond_prob_mask,
@@ -909,59 +1010,155 @@ class Unet(nn.Module):
 
         # main conditioning tokens (c)
 
-        c = torch.cat((time_tokens, image_tokens), dim = -2)
+        c = time_tokens
+
+        if exists(image_tokens):
+            c = torch.cat((c, image_tokens), dim = -2)
 
         # text and image conditioning tokens (mid_c)
         # to save on compute, only do cross attention based conditioning on the inner most layers of the Unet
 
-        mid_c = c if not exists(text_encodings) else torch.cat((c, text_tokens), dim = -2)
+        mid_c = c if not exists(text_tokens) else torch.cat((c, text_tokens), dim = -2)
 
         # go through the layers of the unet, down and up
 
         hiddens = []
 
-        for convnext, convnext2, downsample in self.downs:
+        for convnext, sparse_attn, convnext2, downsample in self.downs:
             x = convnext(x, c)
+            x = sparse_attn(x)
             x = convnext2(x, c)
             hiddens.append(x)
             x = downsample(x)
 
         x = self.mid_block1(x, mid_c)
-        x = self.mid_attn(x)
+
+        if exists(self.mid_attn):
+            x = self.mid_attn(x)
+
         x = self.mid_block2(x, mid_c)
 
-        for convnext, convnext2, upsample in self.ups:
+        for convnext, sparse_attn, convnext2, upsample in self.ups:
             x = torch.cat((x, hiddens.pop()), dim=1)
             x = convnext(x, c)
+            x = sparse_attn(x)
             x = convnext2(x, c)
             x = upsample(x)
 
         return self.final_conv(x)
 
+class LowresConditioner(nn.Module):
+    def __init__(
+        self,
+        cond_upsample_mode = 'bilinear',
+        downsample_first = True,
+        blur_sigma = 0.1,
+        blur_kernel_size = 3,
+    ):
+        super().__init__()
+        self.cond_upsample_mode = cond_upsample_mode
+        self.downsample_first = downsample_first
+        self.blur_sigma = blur_sigma
+        self.blur_kernel_size = blur_kernel_size
+
+    def forward(
+        self,
+        cond_fmap,
+        *,
+        target_image_size,
+        downsample_image_size = None,
+        blur_sigma = None,
+        blur_kernel_size = None
+    ):
+        target_image_size = cast_tuple(target_image_size, 2)
+
+        if self.training and self.downsample_first and exists(downsample_image_size):
+            cond_fmap = resize_image_to(cond_fmap, target_image_size, mode = self.cond_upsample_mode)
+
+        if self.training:
+            # when training, blur the low resolution conditional image
+            blur_sigma = default(blur_sigma, self.blur_sigma)
+            blur_kernel_size = default(blur_kernel_size, self.blur_kernel_size)
+            cond_fmap = gaussian_blur2d(cond_fmap, cast_tuple(blur_kernel_size, 2), cast_tuple(blur_sigma, 2))
+
+        cond_fmap = resize_image_to(cond_fmap, target_image_size, mode = self.cond_upsample_mode)
+
+        return cond_fmap
+
 class Decoder(nn.Module):
     def __init__(
         self,
-        net,
+        unet,
         *,
         clip,
         timesteps = 1000,
         cond_drop_prob = 0.2,
-        loss_type = 'l1'
+        loss_type = 'l1',
+        beta_schedule = 'cosine',
+        image_sizes = None,                         # for cascading ddpm, image size at each stage
+        lowres_cond_upsample_mode = 'bilinear',     # cascading ddpm - low resolution upsample mode
+        lowres_downsample_first = True,             # cascading ddpm - resizes to lower resolution, then to next conditional resolution + blur
+        blur_sigma = 0.1,                           # cascading ddpm - blur sigma
+        blur_kernel_size = 3,                       # cascading ddpm - blur kernel size
     ):
         super().__init__()
         assert isinstance(clip, CLIP)
         freeze_model_and_make_eval_(clip)
         self.clip = clip
-
-        self.net = net
+        self.clip_image_size = clip.image_size
         self.channels = clip.image_channels
-        self.image_size = clip.image_size
+
+        # automatically take care of ensuring that first unet is unconditional
+        # while the rest of the unets are conditioned on the low resolution image produced by previous unet
+
+        self.unets = nn.ModuleList([])
+        for ind, one_unet in enumerate(cast_tuple(unet)):
+            is_first = ind == 0
+            one_unet = one_unet.force_lowres_cond(not is_first)
+            self.unets.append(one_unet)
+
+        # unet image sizes
+
+        image_sizes = default(image_sizes, (clip.image_size,))
+        image_sizes = tuple(sorted(set(image_sizes)))
+
+        assert len(self.unets) == len(image_sizes), f'you did not supply the correct number of u-nets ({len(self.unets)}) for resolutions {image_sizes}'
+        self.image_sizes = image_sizes
+        self.sample_channels = cast_tuple(self.channels, len(image_sizes))
+
+        # cascading ddpm related stuff
+
+        lowres_conditions = tuple(map(lambda t: t.lowres_cond, self.unets))
+        assert lowres_conditions == (False, *((True,) * (len(self.unets) - 1))), 'the first unet must be unconditioned (by low resolution image), and the rest of the unets must have `lowres_cond` set to True'
+
+        self.to_lowres_cond = LowresConditioner(
+            cond_upsample_mode = lowres_cond_upsample_mode,
+            downsample_first = lowres_downsample_first,
+            blur_sigma = blur_sigma,
+            blur_kernel_size = blur_kernel_size,
+        )
+
+        # classifier free guidance
+
         self.cond_drop_prob = cond_drop_prob
 
-        betas = cosine_beta_schedule(timesteps)
+        # noise schedule
+
+        if beta_schedule == "cosine":
+            betas = cosine_beta_schedule(timesteps)
+        elif beta_schedule == "linear":
+            betas = linear_beta_schedule(timesteps)
+        elif beta_schedule == "quadratic":
+            betas = quadratic_beta_schedule(timesteps)
+        elif beta_schedule == "jsd":
+            betas = 1.0 / torch.linspace(timesteps, 1, timesteps)
+        elif beta_schedule == "sigmoid":
+            betas = sigmoid_beta_schedule(timesteps)
+        else:
+            raise NotImplementedError()
 
         alphas = 1. - betas
-        alphas_cumprod = torch.cumprod(alphas, axis=0)
+        alphas_cumprod = torch.cumprod(alphas, axis = 0)
         alphas_cumprod_prev = F.pad(alphas_cumprod[:-1], (1, 0), value = 1.)
 
         timesteps, = betas.shape
@@ -994,11 +1191,31 @@ class Decoder(nn.Module):
         self.register_buffer('posterior_mean_coef1', betas * torch.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))
         self.register_buffer('posterior_mean_coef2', (1. - alphas_cumprod_prev) * torch.sqrt(alphas) / (1. - alphas_cumprod))
 
+    def get_unet(self, unet_number):
+        assert 0 < unet_number <= len(self.unets)
+        index = unet_number - 1
+        return self.unets[index]
+
+    @contextmanager
+    def one_unet_in_gpu(self, unet_number = None, unet = None):
+        assert exists(unet_number) ^ exists(unet)
+
+        if exists(unet_number):
+            unet = self.get_unet(unet_number)
+
+        self.cuda()
+        self.unets.cpu()
+
+        unet.cuda()
+        yield
+        unet.cpu()
+
     def get_text_encodings(self, text):
         text_encodings = self.clip.text_transformer(text)
         return text_encodings[:, 1:]
 
     def get_image_embed(self, image):
+        image = resize_image_to(image, self.clip_image_size)
         image_encoding = self.clip.visual_transformer(image)
         image_cls = image_encoding[:, 0]
         image_embed = self.clip.to_visual_latent(image_cls)
@@ -1025,8 +1242,9 @@ class Decoder(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 p_mean_variance(self, x, t, image_embed, text_encodings = None, clip_denoised = True, cond_scale = 1.):
-        x_recon = self.predict_start_from_noise(x, t = t, noise = self.net.forward_with_cond_scale(x, t, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale))
+    def p_mean_variance(self, unet, x, t, image_embed, text_encodings = None, lowres_cond_img = None, clip_denoised = True, cond_scale = 1.):
+        pred_noise = 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)
+        x_recon = self.predict_start_from_noise(x, t = t, noise = pred_noise)
 
         if clip_denoised:
             x_recon.clamp_(-1., 1.)
@@ -1035,33 +1253,26 @@ class Decoder(nn.Module):
         return model_mean, posterior_variance, posterior_log_variance
 
     @torch.no_grad()
-    def p_sample(self, x, t, image_embed, text_encodings = None, cond_scale = 1., clip_denoised = True, repeat_noise = False):
+    def p_sample(self, unet, x, t, image_embed, text_encodings = None, cond_scale = 1., lowres_cond_img = None, clip_denoised = True, repeat_noise = False):
         b, *_, device = *x.shape, x.device
-        model_mean, _, model_log_variance = self.p_mean_variance(x = x, t = t, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, clip_denoised = clip_denoised)
+        model_mean, _, model_log_variance = 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, clip_denoised = clip_denoised)
         noise = noise_like(x.shape, device, repeat_noise)
         # no noise when t == 0
         nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
         return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
 
     @torch.no_grad()
-    def p_sample_loop(self, shape, image_embed, text_encodings = None, cond_scale = 1):
+    def p_sample_loop(self, unet, shape, image_embed, lowres_cond_img = None, text_encodings = None, cond_scale = 1):
         device = self.betas.device
 
         b = shape[0]
-        img = torch.randn(shape, device=device)
+        img = torch.randn(shape, device = device)
+
+        for i in tqdm(reversed(range(0, self.num_timesteps)), desc = 'sampling loop time step', total = self.num_timesteps):
+            img = self.p_sample(unet, img, torch.full((b,), i, device = device, dtype = torch.long), image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img)
 
-        for i in tqdm(reversed(range(0, self.num_timesteps)), desc='sampling loop time step', total=self.num_timesteps):
-            img = self.p_sample(img, torch.full((b,), i, device = device, dtype = torch.long), image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale)
         return img
 
-    @torch.no_grad()
-    def sample(self, image_embed, text = None, cond_scale = 1.):
-        batch_size = image_embed.shape[0]
-        image_size = self.image_size
-        channels = self.channels
-        text_encodings = self.get_text_encodings(text) if exists(text) else None
-        return self.p_sample_loop((batch_size, channels, image_size, image_size), image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale)
-
     def q_sample(self, x_start, t, noise=None):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
@@ -1070,16 +1281,17 @@ class Decoder(nn.Module):
             extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
         )
 
-    def p_losses(self, x_start, t, *, image_embed, text_encodings = None, noise = None):
+    def p_losses(self, unet, x_start, t, *, image_embed, lowres_cond_img = None, text_encodings = None, noise = None):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
         x_noisy = self.q_sample(x_start = x_start, t = t, noise = noise)
 
-        x_recon = self.net(
+        x_recon = unet(
             x_noisy,
             t,
             image_embed = image_embed,
             text_encodings = text_encodings,
+            lowres_cond_img = lowres_cond_img,
             cond_drop_prob = self.cond_drop_prob
         )
 
@@ -1087,22 +1299,64 @@ class Decoder(nn.Module):
             loss = F.l1_loss(noise, x_recon)
         elif self.loss_type == 'l2':
             loss = F.mse_loss(noise, x_recon)
+        elif self.loss_type == "huber":
+            loss = F.smooth_l1_loss(noise, x_recon)
         else:
             raise NotImplementedError()
 
         return loss
 
-    def forward(self, image, text = None):
-        b, device, img_size, = image.shape[0], image.device, self.image_size
-        check_shape(image, 'b c h w', h = img_size, w = img_size, c = self.channels)
+    @torch.no_grad()
+    @eval_decorator
+    def sample(self, image_embed, text = None, cond_scale = 1.):
+        batch_size = image_embed.shape[0]
+
+        text_encodings = self.get_text_encodings(text) if exists(text) else None
+
+        img = None
+
+        for unet, channel, image_size in tqdm(zip(self.unets, self.sample_channels, self.image_sizes)):
+            with self.one_unet_in_gpu(unet = unet):
+                lowres_cond_img = self.to_lowres_cond(
+                    img,
+                    target_image_size = image_size
+                ) if unet.lowres_cond else None
+
+                img = self.p_sample_loop(
+                    unet,
+                    (batch_size, channel, image_size, image_size),
+                    image_embed = image_embed,
+                    text_encodings = text_encodings,
+                    cond_scale = cond_scale,
+                    lowres_cond_img = lowres_cond_img
+                )
+
+        return img
+
+    def forward(self, image, text = None, image_embed = None, text_encodings = None, unet_number = None):
+        assert not (len(self.unets) > 1 and not exists(unet_number)), f'you must specify which unet you want trained, from a range of 1 to {len(self.unets)}, if you are training cascading DDPM (multiple unets)'
+        unet_number = default(unet_number, 1)
+        unet_index = unet_number - 1
+
+        unet = self.get_unet(unet_number)
+
+        target_image_size = self.image_sizes[unet_index]
+
+        b, c, h, w, device, = *image.shape, image.device
+
+        check_shape(image, 'b c h w', c = self.channels)
+        assert h >= target_image_size and w >= target_image_size
 
         times = torch.randint(0, self.num_timesteps, (b,), device = device, dtype = torch.long)
 
-        image_embed = self.get_image_embed(image)
-        text_encodings = self.get_text_encodings(text) if exists(text) else None
+        if not exists(image_embed):
+            image_embed = self.get_image_embed(image)
 
-        loss = self.p_losses(image, times, image_embed = image_embed, text_encodings = text_encodings)
-        return loss
+        text_encodings = self.get_text_encodings(text) if exists(text) and not exists(text_encodings) else None
+
+        lowres_cond_img = self.to_lowres_cond(image, target_image_size = target_image_size, downsample_image_size = self.image_sizes[unet_index - 1]) if unet_number > 1 else None
+        ddpm_image = resize_image_to(image, target_image_size)
+        return self.p_losses(unet, ddpm_image, times, image_embed = image_embed, text_encodings = text_encodings, lowres_cond_img = lowres_cond_img)
 
 # main class
 

dalle2_pytorch/vqgan_vae.py

@@ -0,0 +1,532 @@
+import copy
+import math
+from math import sqrt
+from functools import partial, wraps
+
+from vector_quantize_pytorch import VectorQuantize as VQ
+
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+from torch.autograd import grad as torch_grad
+import torchvision
+
+from einops import rearrange, reduce, repeat
+
+# constants
+
+MList = nn.ModuleList
+
+# helper functions
+
+def exists(val):
+    return val is not None
+
+def default(val, d):
+    return val if exists(val) else d
+
+# decorators
+
+def eval_decorator(fn):
+    def inner(model, *args, **kwargs):
+        was_training = model.training
+        model.eval()
+        out = fn(model, *args, **kwargs)
+        model.train(was_training)
+        return out
+    return inner
+
+def remove_vgg(fn):
+    @wraps(fn)
+    def inner(self, *args, **kwargs):
+        has_vgg = hasattr(self, 'vgg')
+        if has_vgg:
+            vgg = self.vgg
+            delattr(self, 'vgg')
+
+        out = fn(self, *args, **kwargs)
+
+        if has_vgg:
+            self.vgg = vgg
+
+        return out
+    return inner
+
+# keyword argument helpers
+
+def pick_and_pop(keys, d):
+    values = list(map(lambda key: d.pop(key), keys))
+    return dict(zip(keys, values))
+
+def group_dict_by_key(cond, d):
+    return_val = [dict(),dict()]
+    for key in d.keys():
+        match = bool(cond(key))
+        ind = int(not match)
+        return_val[ind][key] = d[key]
+    return (*return_val,)
+
+def string_begins_with(prefix, str):
+    return str.startswith(prefix)
+
+def group_by_key_prefix(prefix, d):
+    return group_dict_by_key(partial(string_begins_with, prefix), d)
+
+def groupby_prefix_and_trim(prefix, d):
+    kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d)
+    kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
+    return kwargs_without_prefix, kwargs
+
+# tensor helper functions
+
+def log(t, eps = 1e-10):
+    return torch.log(t + eps)
+
+def gradient_penalty(images, output, weight = 10):
+    batch_size = images.shape[0]
+    gradients = torch_grad(outputs = output, inputs = images,
+                           grad_outputs = torch.ones(output.size(), device = images.device),
+                           create_graph = True, retain_graph = True, only_inputs = True)[0]
+
+    gradients = rearrange(gradients, 'b ... -> b (...)')
+    return weight * ((gradients.norm(2, dim = 1) - 1) ** 2).mean()
+
+def l2norm(t):
+    return F.normalize(t, dim = -1)
+
+def leaky_relu(p = 0.1):
+    return nn.LeakyReLU(0.1)
+
+def stable_softmax(t, dim = -1, alpha = 32 ** 2):
+    t = t / alpha
+    t = t - torch.amax(t, dim = dim, keepdim = True).detach()
+    return (t * alpha).softmax(dim = dim)
+
+def safe_div(numer, denom, eps = 1e-8):
+    return numer / (denom + eps)
+
+# gan losses
+
+def hinge_discr_loss(fake, real):
+    return (F.relu(1 + fake) + F.relu(1 - real)).mean()
+
+def hinge_gen_loss(fake):
+    return -fake.mean()
+
+def bce_discr_loss(fake, real):
+    return (-log(1 - torch.sigmoid(fake)) - log(torch.sigmoid(real))).mean()
+
+def bce_gen_loss(fake):
+    return -log(torch.sigmoid(fake)).mean()
+
+def grad_layer_wrt_loss(loss, layer):
+    return torch_grad(
+        outputs = loss,
+        inputs = layer,
+        grad_outputs = torch.ones_like(loss),
+        retain_graph = True
+    )[0].detach()
+
+# vqgan vae
+
+class LayerNormChan(nn.Module):
+    def __init__(
+        self,
+        dim,
+        eps = 1e-5
+    ):
+        super().__init__()
+        self.eps = eps
+        self.gamma = nn.Parameter(torch.ones(1, dim, 1, 1))
+
+    def forward(self, x):
+        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
+        mean = torch.mean(x, dim = 1, keepdim = True)
+        return (x - mean) / (var + self.eps).sqrt() * self.gamma
+
+class Discriminator(nn.Module):
+    def __init__(
+        self,
+        dims,
+        channels = 3,
+        groups = 16,
+        init_kernel_size = 5
+    ):
+        super().__init__()
+        dim_pairs = zip(dims[:-1], dims[1:])
+
+        self.layers = MList([nn.Sequential(nn.Conv2d(channels, dims[0], init_kernel_size, padding = init_kernel_size // 2), leaky_relu())])
+
+        for dim_in, dim_out in dim_pairs:
+            self.layers.append(nn.Sequential(
+                nn.Conv2d(dim_in, dim_out, 4, stride = 2, padding = 1),
+                nn.GroupNorm(groups, dim_out),
+                leaky_relu()
+            ))
+
+        dim = dims[-1]
+        self.to_logits = nn.Sequential( # return 5 x 5, for PatchGAN-esque training
+            nn.Conv2d(dim, dim, 1),
+            leaky_relu(),
+            nn.Conv2d(dim, 1, 4)
+        )
+
+    def forward(self, x):
+        for net in self.layers:
+            x = net(x)
+
+        return self.to_logits(x)
+
+class ContinuousPositionBias(nn.Module):
+    """ from https://arxiv.org/abs/2111.09883 """
+
+    def __init__(self, *, dim, heads, layers = 2):
+        super().__init__()
+        self.net = MList([])
+        self.net.append(nn.Sequential(nn.Linear(2, dim), leaky_relu()))
+
+        for _ in range(layers - 1):
+            self.net.append(nn.Sequential(nn.Linear(dim, dim), leaky_relu()))
+
+        self.net.append(nn.Linear(dim, heads))
+        self.register_buffer('rel_pos', None, persistent = False)
+
+    def forward(self, x):
+        n, device = x.shape[-1], x.device
+        fmap_size = int(sqrt(n))
+
+        if not exists(self.rel_pos):
+            pos = torch.arange(fmap_size, device = device)
+            grid = torch.stack(torch.meshgrid(pos, pos, indexing = 'ij'))
+            grid = rearrange(grid, 'c i j -> (i j) c')
+            rel_pos = rearrange(grid, 'i c -> i 1 c') - rearrange(grid, 'j c -> 1 j c')
+            rel_pos = torch.sign(rel_pos) * torch.log(rel_pos.abs() + 1)
+            self.register_buffer('rel_pos', rel_pos, persistent = False)
+
+        rel_pos = self.rel_pos.float()
+
+        for layer in self.net:
+            rel_pos = layer(rel_pos)
+
+        bias = rearrange(rel_pos, 'i j h -> h i j')
+        return x + bias
+
+class GLUResBlock(nn.Module):
+    def __init__(self, chan, groups = 16):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Conv2d(chan, chan * 2, 3, padding = 1),
+            nn.GLU(dim = 1),
+            nn.GroupNorm(groups, chan),
+            nn.Conv2d(chan, chan * 2, 3, padding = 1),
+            nn.GLU(dim = 1),
+            nn.GroupNorm(groups, chan),
+            nn.Conv2d(chan, chan, 1)
+        )
+
+    def forward(self, x):
+        return self.net(x) + x
+
+class ResBlock(nn.Module):
+    def __init__(self, chan, groups = 16):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Conv2d(chan, chan, 3, padding = 1),
+            nn.GroupNorm(groups, chan),
+            leaky_relu(),
+            nn.Conv2d(chan, chan, 3, padding = 1),
+            nn.GroupNorm(groups, chan),
+            leaky_relu(),
+            nn.Conv2d(chan, chan, 1)
+        )
+
+    def forward(self, x):
+        return self.net(x) + x
+
+class VQGanAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        dim,
+        dim_head = 64,
+        heads = 8,
+        dropout = 0.
+    ):
+        super().__init__()
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+        inner_dim = heads * dim_head
+
+        self.dropout = nn.Dropout(dropout)
+        self.pre_norm = LayerNormChan(dim)
+
+        self.cpb = ContinuousPositionBias(dim = dim // 4, heads = heads)
+        self.to_qkv = nn.Conv2d(dim, inner_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(inner_dim, dim, 1, bias = False)
+
+    def forward(self, x):
+        h = self.heads
+        height, width, residual = *x.shape[-2:], x.clone()
+
+        x = self.pre_norm(x)
+
+        q, k, v = self.to_qkv(x).chunk(3, dim = 1)
+
+        q, k, v = map(lambda t: rearrange(t, 'b (h c) x y -> b h c (x y)', h = h), (q, k, v))
+
+        sim = einsum('b h c i, b h c j -> b h i j', q, k) * self.scale
+
+        sim = self.cpb(sim)
+
+        attn = stable_softmax(sim, dim = -1)
+        attn = self.dropout(attn)
+
+        out = einsum('b h i j, b h c j -> b h c i', attn, v)
+        out = rearrange(out, 'b h c (x y) -> b (h c) x y', x = height, y = width)
+        out = self.to_out(out)
+
+        return out + residual
+
+class VQGanVAE(nn.Module):
+    def __init__(
+        self,
+        *,
+        dim,
+        image_size,
+        channels = 3,
+        num_layers = 4,
+        layer_mults = None,
+        l2_recon_loss = False,
+        use_hinge_loss = True,
+        num_resnet_blocks = 1,
+        vgg = None,
+        vq_codebook_size = 512,
+        vq_decay = 0.8,
+        vq_commitment_weight = 1.,
+        vq_kmeans_init = True,
+        vq_use_cosine_sim = True,
+        use_attn = True,
+        attn_dim_head = 64,
+        attn_heads = 8,
+        resnet_groups = 16,
+        attn_dropout = 0.,
+        first_conv_kernel_size = 5,
+        use_vgg_and_gan = True,
+        **kwargs
+    ):
+        super().__init__()
+        assert dim % resnet_groups == 0, f'dimension {dim} must be divisible by {resnet_groups} (groups for the groupnorm)'
+
+        vq_kwargs, kwargs = groupby_prefix_and_trim('vq_', kwargs)
+
+        self.image_size = image_size
+        self.channels = channels
+        self.num_layers = num_layers
+        self.fmap_size = image_size // (num_layers ** 2)
+        self.codebook_size = vq_codebook_size
+
+        self.encoders = MList([])
+        self.decoders = MList([])
+
+        layer_mults = default(layer_mults, list(map(lambda t: 2 ** t, range(num_layers))))
+        assert len(layer_mults) == num_layers, 'layer multipliers must be equal to designated number of layers'
+
+        layer_dims = [dim * mult for mult in layer_mults]
+        dims = (dim, *layer_dims)
+        codebook_dim = layer_dims[-1]
+
+        dim_pairs = zip(dims[:-1], dims[1:])
+
+        append = lambda arr, t: arr.append(t)
+        prepend = lambda arr, t: arr.insert(0, t)
+
+        if not isinstance(num_resnet_blocks, tuple):
+            num_resnet_blocks = (*((0,) * (num_layers - 1)), num_resnet_blocks)
+
+        if not isinstance(use_attn, tuple):
+            use_attn = (*((False,) * (num_layers - 1)), use_attn)
+
+        assert len(num_resnet_blocks) == num_layers, 'number of resnet blocks config must be equal to number of layers'
+        assert len(use_attn) == num_layers
+
+        for layer_index, (dim_in, dim_out), layer_num_resnet_blocks, layer_use_attn in zip(range(num_layers), dim_pairs, num_resnet_blocks, use_attn):
+            append(self.encoders, nn.Sequential(nn.Conv2d(dim_in, dim_out, 4, stride = 2, padding = 1), leaky_relu()))
+            prepend(self.decoders, nn.Sequential(nn.Upsample(scale_factor = 2, mode = 'bilinear', align_corners = False), nn.Conv2d(dim_out, dim_in, 3, padding = 1), leaky_relu()))
+
+            if layer_use_attn:
+                prepend(self.decoders, VQGanAttention(dim = dim_out, heads = attn_heads, dim_head = attn_dim_head, dropout = attn_dropout))
+
+            for _ in range(layer_num_resnet_blocks):
+                append(self.encoders, ResBlock(dim_out, groups = resnet_groups))
+                prepend(self.decoders, GLUResBlock(dim_out, groups = resnet_groups))
+
+            if layer_use_attn:
+                append(self.encoders, VQGanAttention(dim = dim_out, heads = attn_heads, dim_head = attn_dim_head, dropout = attn_dropout))
+
+        prepend(self.encoders, nn.Conv2d(channels, dim, first_conv_kernel_size, padding = first_conv_kernel_size // 2))
+        append(self.decoders, nn.Conv2d(dim, channels, 1))
+
+        self.vq = VQ(
+            dim = codebook_dim,
+            codebook_size = vq_codebook_size,
+            decay = vq_decay,
+            commitment_weight = vq_commitment_weight,
+            accept_image_fmap = True,
+            kmeans_init = vq_kmeans_init,
+            use_cosine_sim = vq_use_cosine_sim,
+            **vq_kwargs
+        )
+
+        # reconstruction loss
+
+        self.recon_loss_fn = F.mse_loss if l2_recon_loss else F.l1_loss
+
+        # turn off GAN and perceptual loss if grayscale
+
+        self.vgg = None
+        self.discr = None
+        self.use_vgg_and_gan = use_vgg_and_gan
+
+        if not use_vgg_and_gan:
+            return
+
+        # preceptual loss
+
+        if exists(vgg):
+            self.vgg = vgg
+        else:
+            self.vgg = torchvision.models.vgg16(pretrained = True)
+            self.vgg.classifier = nn.Sequential(*self.vgg.classifier[:-2])
+
+        # gan related losses
+
+        self.discr = Discriminator(dims = dims, channels = channels)
+
+        self.discr_loss = hinge_discr_loss if use_hinge_loss else bce_discr_loss
+        self.gen_loss = hinge_gen_loss if use_hinge_loss else bce_gen_loss
+
+    def copy_for_eval(self):
+        device = next(self.parameters()).device
+        vae_copy = copy.deepcopy(self.cpu())
+
+        if vae_copy.use_vgg_and_gan:
+            del vae_copy.discr
+            del vae_copy.vgg
+
+        vae_copy.eval()
+        return vae_copy.to(device)
+
+    @remove_vgg
+    def state_dict(self, *args, **kwargs):
+        return super().state_dict(*args, **kwargs)
+
+    @remove_vgg
+    def load_state_dict(self, *args, **kwargs):
+        return super().load_state_dict(*args, **kwargs)
+
+    @property
+    def codebook(self):
+        return self.vq.codebook
+
+    def encode(self, fmap):
+        for enc in self.encoders:
+            fmap = enc(fmap)
+
+        return fmap
+
+    def decode(self, fmap):
+        fmap = self.vq(fmap)
+
+        for dec in self.decoders:
+            fmap = dec(fmap)
+
+        return fmap
+
+    def forward(
+        self,
+        img,
+        return_loss = False,
+        return_discr_loss = False,
+        return_recons = False
+    ):
+        batch, channels, height, width, device = *img.shape, img.device
+        assert height == self.image_size and width == self.image_size, 'height and width of input image must be equal to {self.image_size}'
+        assert channels == self.channels, 'number of channels on image or sketch is not equal to the channels set on this VQGanVAE'
+
+        fmap, indices, commit_loss = self.encode(img)
+
+        fmap = self.decode(fmap)
+
+        if not return_loss and not return_discr_loss:
+            return fmap
+
+        assert return_loss ^ return_discr_loss, 'you should either return autoencoder loss or discriminator loss, but not both'
+
+        # whether to return discriminator loss
+
+        if return_discr_loss:
+            assert exists(self.discr), 'discriminator must exist to train it'
+
+            fmap.detach_()
+            img.requires_grad_()
+
+            fmap_discr_logits, img_discr_logits = map(self.discr, (fmap, img))
+
+            gp = gradient_penalty(img, img_discr_logits)
+
+            discr_loss = self.discr_loss(fmap_discr_logits, img_discr_logits)
+
+            loss = discr_loss + gp
+
+            if return_recons:
+                return loss, fmap
+
+            return loss
+
+        # reconstruction loss
+
+        recon_loss = self.recon_loss_fn(fmap, img)
+
+        # early return if training on grayscale
+
+        if not self.use_vgg_and_gan:
+            if return_recons:
+                return recon_loss, fmap
+
+            return recon_loss
+
+        # perceptual loss
+
+        img_vgg_input = img
+        fmap_vgg_input = fmap
+
+        if img.shape[1] == 1:
+            # handle grayscale for vgg
+            img_vgg_input, fmap_vgg_input = map(lambda t: repeat(t, 'b 1 ... -> b c ...', c = 3), (img_vgg_input, fmap_vgg_input))
+
+        img_vgg_feats = self.vgg(img_vgg_input)
+        recon_vgg_feats = self.vgg(fmap_vgg_input)
+        perceptual_loss = F.mse_loss(img_vgg_feats, recon_vgg_feats)
+
+        # generator loss
+
+        gen_loss = self.gen_loss(self.discr(fmap))
+
+        # calculate adaptive weight
+
+        last_dec_layer = self.decoders[-1].weight
+
+        norm_grad_wrt_gen_loss = grad_layer_wrt_loss(gen_loss, last_dec_layer).norm(p = 2)
+        norm_grad_wrt_perceptual_loss = grad_layer_wrt_loss(perceptual_loss, last_dec_layer).norm(p = 2)
+
+        adaptive_weight = safe_div(norm_grad_wrt_perceptual_loss, norm_grad_wrt_gen_loss)
+        adaptive_weight.clamp_(max = 1e4)
+
+        # combine losses
+
+        loss = recon_loss + perceptual_loss + commit_loss + adaptive_weight * gen_loss
+
+        if return_recons:
+            return loss, fmap
+
+        return loss

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.17',
+  version = '0.0.33',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',
@@ -30,6 +30,7 @@ setup(
     'torch>=1.10',
     'torchvision',
     'tqdm',
+    'vector-quantize-pytorch',
     'x-clip>=0.4.4',
     'youtokentome'
   ],