from my vision transformer experience, dimension of attention head of 32 is sufficient for image feature maps

README.md

@@ -14,7 +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
 
-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>
+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
 
@@ -409,9 +409,11 @@ Offer training wrappers
 - [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)
 - [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
-- [ ] offload unets not being trained on to CPU for memory efficiency (for training each resolution unets separately)
-- [ ] build out latent diffusion architecture in separate file, as it is not faithful to dalle-2 (but offer it as as setting)
-- [ ] become an expert with unets, cleanup unet code, make it fully configurable, add efficient attention (conditional on resolution), port all learnings over to https://github.com/lucidrains/x-unet
+- [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, make it completely optional (additional autoencoder + some regularizations [kl and vq regs]) (figure out if latent diffusion + cascading ddpm can be used in conjunction)
+- [ ] become an expert with unets, cleanup unet code, make it fully configurable, port all learnings over to https://github.com/lucidrains/x-unet
 - [ ] train on a toy task, offer in colab
 
 ## Citations
@@ -461,4 +463,12 @@ 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/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

@@ -2,6 +2,7 @@ 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
@@ -105,8 +106,8 @@ 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)
@@ -463,11 +464,11 @@ class DiffusionPrior(nn.Module):
         net,
         *,
         clip,
-        timesteps=1000,
-        cond_drop_prob=0.2,
-        loss_type="l1",
-        predict_x0=True,
-        beta_schedule="cosine",
+        timesteps = 1000,
+        cond_drop_prob = 0.2,
+        loss_type = "l1",
+        predict_x0 = True,
+        beta_schedule = "cosine",
     ):
         super().__init__()
         assert isinstance(clip, CLIP)
@@ -798,6 +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,
@@ -806,14 +821,21 @@ 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,
         blur_kernel_size = 3,
+        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
@@ -846,8 +868,8 @@ 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(
@@ -857,11 +879,21 @@ 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([])
@@ -875,6 +907,7 @@ class Unet(nn.Module):
 
             self.downs.append(nn.ModuleList([
                 ConvNextBlock(dim_in, dim_out, norm = ind != 0),
+                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()
             ]))
@@ -882,7 +915,7 @@ class Unet(nn.Module):
         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))) if attend_at_middle else None
+        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:])):
@@ -891,6 +924,7 @@ class Unet(nn.Module):
 
             self.ups.append(nn.ModuleList([
                 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)
             ]))
@@ -964,17 +998,22 @@ 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,
@@ -984,19 +1023,23 @@ 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)
@@ -1008,9 +1051,10 @@ class Unet(nn.Module):
 
         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)
 
@@ -1104,6 +1148,25 @@ 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:]
@@ -1208,20 +1271,21 @@ class Decoder(nn.Module):
         text_encodings = self.get_text_encodings(text) if exists(text) else None
 
         img = None
+
         for unet, image_size in tqdm(zip(self.unets, self.image_sizes)):
-            shape = (batch_size, channels, image_size, image_size)
-            img = self.p_sample_loop(unet, shape, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = img)
+            with self.one_unet_in_gpu(unet = unet):
+                shape = (batch_size, channels, image_size, image_size)
+                img = self.p_sample_loop(unet, shape, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = img)
 
         return img
 
-    def forward(self, image, text = None, image_embed = None, unet_number = None):
+    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)
-        assert 1 <= unet_number <= len(self.unets)
 
-        index = unet_number - 1
-        unet = self.unets[index]
-        target_image_size = self.image_sizes[index]
+        unet = self.get_unet(unet_number)
+
+        target_image_size = self.image_sizes[unet_number - 1]
 
         b, c, h, w, device, = *image.shape, image.device
 
@@ -1233,9 +1297,9 @@ class Decoder(nn.Module):
         if not exists(image_embed):
             image_embed = self.get_image_embed(image)
 
-        text_encodings = self.get_text_encodings(text) if exists(text) else None
+        text_encodings = self.get_text_encodings(text) if exists(text) and not exists(text_encodings) else None
 
-        lowres_cond_img = image if index > 0 else None
+        lowres_cond_img = image 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)
 

dalle2_pytorch/latent_diffusion.py

@@ -0,0 +1,12 @@
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+
+from einops import rearrange
+
+class LatentDiffusion(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.23',
+  version = '0.0.31',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',