vqgan-vae codebook dims should be 256 or smaller

README.md

@@ -499,10 +499,12 @@ loss.backward()
 
 ### DALL-E2 with Latent Diffusion
 
-This repository decides to take the next step and offer DALL-E2 combined with <a href="https://huggingface.co/spaces/multimodalart/latentdiffusion">latent diffusion</a>, from Rombach et al.
+This repository decides to take the next step and offer DALL-E v2 combined with <a href="https://huggingface.co/spaces/multimodalart/latentdiffusion">latent diffusion</a>, from Rombach et al.
 
 You can use it as follows. Latent diffusion can be limited to just the first U-Net in the cascade, or to any number you wish.
 
+The repository also comes equipped with all the necessary settings to recreate `ViT-VQGan` from the <a href="https://arxiv.org/abs/2110.04627">Improved VQGans</a> paper. Furthermore, the <a href="https://github.com/lucidrains/vector-quantize-pytorch">vector quantization</a> library also comes equipped to do <a href="https://arxiv.org/abs/2203.01941">residual or multi-headed quantization</a>, which I believe will give an even further boost in performance to the autoencoder.
+
 ```python
 import torch
 from dalle2_pytorch import Unet, Decoder, CLIP, VQGanVAE
@@ -644,14 +646,15 @@ Once built, images will be saved to the same directory the command is invoked
 - [x] for decoder, allow ability to customize objective (predict epsilon vs x0), in case latent diffusion does better with prediction of x0
 - [x] use attention-based upsampling https://arxiv.org/abs/2112.11435
 - [x] use inheritance just this once for sharing logic between decoder and prior network ddpms
-- [ ] abstract interface for CLIP adapter class, so other CLIPs can be brought in
+- [x] bring in vit-vqgan https://arxiv.org/abs/2110.04627 for the latent diffusion
+- [x] abstract interface for CLIP adapter class, so other CLIPs can be brought in
 - [ ] 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
 - [ ] transcribe code to Jax, which lowers the activation energy for distributed training, given access to TPUs
 - [ ] train on a toy task, offer in colab
+- [ ] think about how best to design a declarative training config that handles preencoding for prior and training of multiple networks in decoder
 - [ ] extend diffusion head to use diffusion-gan (potentially using lightweight-gan) to speed up inference
 - [ ] bring in tools to train vqgan-vae
-- [ ] bring in vit-vqgan https://arxiv.org/abs/2110.04627 for the latent diffusion
 
 ## Citations
 
@@ -697,4 +700,14 @@ Once built, images will be saved to the same directory the command is invoked
 }
 ```
 
+```bibtex
+@article{Yu2021VectorquantizedIM,
+    title   = {Vector-quantized Image Modeling with Improved VQGAN},
+    author  = {Jiahui Yu and Xin Li and Jing Yu Koh and Han Zhang and Ruoming Pang and James Qin and Alexander Ku and Yuanzhong Xu and Jason Baldridge and Yonghui Wu},
+    journal = {ArXiv},
+    year    = {2021},
+    volume  = {abs/2110.04627}
+}
+```
+
 *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/attention.py

@@ -1,125 +0,0 @@
-import torch
-from torch import nn, einsum
-import torch.nn.functional as F
-
-from einops import rearrange, repeat
-
-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
-
-# attention-based upsampling
-# from https://arxiv.org/abs/2112.11435
-
-class QueryAndAttend(nn.Module):
-    def __init__(
-        self,
-        *,
-        dim,
-        num_queries = 1,
-        dim_head = 32,
-        heads = 8,
-        window_size = 3
-    ):
-        super().__init__()
-        self.scale = dim_head ** -0.5
-        inner_dim = dim_head * heads
-        self.heads = heads
-        self.dim_head = dim_head
-        self.window_size = window_size
-        self.num_queries = num_queries
-
-        self.rel_pos_bias = nn.Parameter(torch.randn(heads, num_queries, window_size * window_size, 1, 1))
-
-        self.queries = nn.Parameter(torch.randn(heads, num_queries, dim_head))
-        self.to_kv = nn.Conv2d(dim, dim_head * 2, 1, bias = False)
-        self.to_out = nn.Conv2d(inner_dim, dim, 1, bias = False)
-
-    def forward(self, x):
-        """
-        einstein notation
-        b - batch
-        h - heads
-        l - num queries
-        d - head dimension
-        x - height
-        y - width
-        j - source sequence for attending to (kernel size squared in this case)
-        """
-
-        wsz, heads, dim_head, num_queries = self.window_size, self.heads, self.dim_head, self.num_queries
-        batch, _, height, width = x.shape
-
-        is_one_query = self.num_queries == 1
-
-        # queries, keys, values
-
-        q = self.queries * self.scale
-        k, v = self.to_kv(x).chunk(2, dim = 1)
-
-        # similarities
-
-        sim = einsum('h l d, b d x y -> b h l x y', q, k)
-        sim = rearrange(sim, 'b ... x y -> b (...) x y')
-
-        # unfold the similarity scores, with float(-inf) as padding value
-
-        mask_value = -torch.finfo(sim.dtype).max
-        sim = F.pad(sim, ((wsz // 2,) * 4), value = mask_value)
-        sim = F.unfold(sim, kernel_size = wsz)
-        sim = rearrange(sim, 'b (h l j) (x y) -> b h l j x y', h = heads, l = num_queries, x = height, y = width)
-
-        # rel pos bias
-
-        sim = sim + self.rel_pos_bias
-
-        # numerically stable attention
-
-        sim = sim - sim.amax(dim = -3, keepdim = True).detach()
-        attn = sim.softmax(dim = -3)
-
-        # unfold values
-
-        v = F.pad(v, ((wsz // 2,) * 4), value = 0.)
-        v = F.unfold(v, kernel_size = wsz)
-        v = rearrange(v, 'b (d j) (x y) -> b d j x y', d = dim_head, x = height, y = width)
-
-        # aggregate values
-
-        out = einsum('b h l j x y, b d j x y -> b l h d x y', attn, v)
-
-        # combine heads
-
-        out = rearrange(out, 'b l h d x y -> (b l) (h d) x y')
-        out = self.to_out(out)
-        out = rearrange(out, '(b l) d x y -> b l d x y', b = batch)
-
-        # return original input if one query
-
-        if is_one_query:
-            out = rearrange(out, 'b 1 ... -> b ...')
-
-        return out
-
-class QueryAttnUpsample(nn.Module):
-    def __init__(self, dim, **kwargs):
-        super().__init__()
-        self.norm = LayerNormChan(dim)
-        self.qna = QueryAndAttend(dim = dim, num_queries = 4, **kwargs)
-
-    def forward(self, x):
-        x = self.norm(x)
-        out = self.qna(x)
-        out = rearrange(out, 'b (w1 w2) c h w -> b c (h w1) (w w2)', w1 = 2, w2 = 2)
-        return out

dalle2_pytorch/dalle2_pytorch.py

@@ -7,6 +7,7 @@ from contextlib import contextmanager
 import torch
 import torch.nn.functional as F
 from torch import nn, einsum
+import torchvision.transforms as T
 
 from einops import rearrange, repeat
 from einops.layers.torch import Rearrange
@@ -17,7 +18,6 @@ from kornia.filters import gaussian_blur2d
 
 from dalle2_pytorch.tokenizer import tokenizer
 from dalle2_pytorch.vqgan_vae import NullVQGanVAE, VQGanVAE
-from dalle2_pytorch.attention import QueryAttnUpsample
 
 # use x-clip
 
@@ -36,6 +36,10 @@ def default(val, d):
 def cast_tuple(val, length = 1):
     return val if isinstance(val, tuple) else ((val,) * length)
 
+@contextmanager
+def null_context(*args, **kwargs):
+    yield
+
 def eval_decorator(fn):
     def inner(model, *args, **kwargs):
         was_training = model.training
@@ -86,6 +90,59 @@ def resize_image_to(t, image_size, mode = 'bilinear'): # take a look at https://
 
     return F.interpolate(t, size = shape, mode = mode, align_corners = False)
 
+# clip related adapters
+
+class BaseClipAdapter(nn.Module):
+    def __init__(self, clip):
+        super().__init__()
+        self.clip = clip
+
+    @property
+    def dim_latent(self):
+        raise NotImplementedError
+
+    @property
+    def image_size(self):
+        raise NotImplementedError
+
+    @property
+    def image_channels(self):
+        raise NotImplementedError
+
+    def embed_text(self, text):
+        raise NotImplementedError
+
+    def embed_image(self, image):
+        raise NotImplementedError
+
+class XClipAdapter(BaseClipAdapter):
+    @property
+    def dim_latent(self):
+        return self.clip.dim_latent
+
+    @property
+    def image_size(self):
+        return self.clip.image_size
+
+    @property
+    def image_channels(self):
+        return self.clip.image_channels
+
+    @torch.no_grad()
+    def embed_text(self, text):
+        encoder_output = self.clip.text_transformer(text)
+        text_cls, text_encodings = encoder_output[:, 0], encoder_output[:, 1:]
+        text_embed = self.clip.to_text_latent(text_cls)
+        return l2norm(text_embed), text_encodings
+
+    @torch.no_grad()
+    def embed_image(self, image):
+        image = resize_image_to(image, self.image_size)
+        encoder_output = self.clip.visual_transformer(image)
+        image_cls, image_encodings = encoder_output[:, 0], encoder_output[:, 1:]
+        image_embed = self.clip.to_visual_latent(image_cls)
+        return l2norm(image_embed), image_encodings
+
 # classifier free guidance functions
 
 def prob_mask_like(shape, prob, device):
@@ -590,7 +647,10 @@ class DiffusionPrior(BaseGaussianDiffusion):
         )
 
         if exists(clip):
-            assert isinstance(clip, CLIP)
+            if isinstance(clip, CLIP):
+                clip = XClipAdapter(clip)
+
+            assert isinstance(clip, BaseClipAdapter)
             freeze_model_and_make_eval_(clip)
             self.clip = clip
         else:
@@ -607,29 +667,6 @@ class DiffusionPrior(BaseGaussianDiffusion):
         self.predict_x_start = predict_x_start
         # in paper, they do not predict the noise, but predict x0 directly for image embedding, claiming empirically better results. I'll just offer both.
 
-    @torch.no_grad()
-    def get_image_embed(self, image):
-        assert exists(self.clip)
-
-        image_encoding = self.clip.visual_transformer(image)
-        image_cls = image_encoding[:, 0]
-        image_embed = self.clip.to_visual_latent(image_cls)
-        return l2norm(image_embed)
-
-    @torch.no_grad()
-    def get_text_cond(self, text):
-        assert exists(self.clip)
-
-        text_encodings = self.clip.text_transformer(text)
-        text_cls, text_encodings = text_encodings[:, 0], text_encodings[:, 1:]
-        text_embed = self.clip.to_text_latent(text_cls)
-        text_embed = l2norm(text_embed)
-
-        if not self.condition_on_text_encodings:            
-            return dict(text_embed = text_embed)
-
-        return dict(text_encodings = text_encodings, text_embed = text_embed, mask = text != 0)
-
     def p_mean_variance(self, x, t, text_cond, clip_denoised: bool):
         pred = self.net(x, t, **text_cond)
 
@@ -701,7 +738,12 @@ class DiffusionPrior(BaseGaussianDiffusion):
         batch_size = text.shape[0]
         image_embed_dim = self.image_embed_dim
 
-        text_cond = self.get_text_cond(text)
+        text_embed, text_encodings = self.clip.embed_text(text)
+
+        text_cond = dict(text_embed = text_embed)
+
+        if self.condition_on_text_encodings:
+            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text != 0}
 
         image_embeds = self.p_sample_loop((batch_size, image_embed_dim), text_cond = text_cond)
         text_embeds = text_cond['text_embed']
@@ -733,18 +775,19 @@ class DiffusionPrior(BaseGaussianDiffusion):
         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):
-            image_embed = self.get_image_embed(image)
+            image_embed, _ = self.clip.embed_image(image)
 
         # calculate text conditionings, based on what is passed in
 
         if exists(text):
-            text_cond = self.get_text_cond(text)
-        else:
-            text_cond = dict(
-                text_embed = text_embed,
-                text_encodings = text_encodings,
-                mask = text_mask
-            )
+            text_embed, text_encodings = self.clip.embed_text(text)
+            text_mask = text != 0
+
+        text_cond = dict(text_embed = text_embed)
+
+        if self.condition_on_text_encodings:
+            assert exists(text_encodings), 'text encodings must be present for diffusion prior if specified'
+            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text_mask}
 
         # timestep conditioning from ddpm
 
@@ -753,8 +796,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
 
         # calculate forward loss
 
-        loss = self.p_losses(image_embed, times, text_cond = text_cond, *args, **kwargs)
-        return loss
+        return self.p_losses(image_embed, times, text_cond = text_cond, *args, **kwargs)
 
 # decoder
 
@@ -906,6 +948,7 @@ class Unet(nn.Module):
         dim,
         *,
         image_embed_dim,
+        text_embed_dim = None,
         cond_dim = None,
         num_image_tokens = 4,
         num_time_tokens = 2,
@@ -959,7 +1002,7 @@ class Unet(nn.Module):
             Rearrange('b (n d) -> b n d', n = num_image_tokens)
         ) if image_embed_dim != cond_dim else nn.Identity()
 
-        self.text_to_cond = nn.LazyLinear(cond_dim)
+        self.text_to_cond = nn.LazyLinear(cond_dim) if not exists(text_embed_dim) else nn.Linear(text_embed_dim, 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
@@ -1204,8 +1247,12 @@ class Decoder(BaseGaussianDiffusion):
             loss_type = loss_type
         )
 
-        assert isinstance(clip, CLIP)
+        if isinstance(clip, CLIP):
+            clip = XClipAdapter(clip)
+
         freeze_model_and_make_eval_(clip)
+        assert isinstance(clip, BaseClipAdapter)
+
         self.clip = clip
         self.clip_image_size = clip.image_size
         self.channels = clip.image_channels
@@ -1286,10 +1333,6 @@ class Decoder(BaseGaussianDiffusion):
         yield
         unet.cpu()
 
-    @torch.no_grad()
-    def get_text_encodings(self, text):
-        text_encodings = self.clip.text_transformer(text)
-        return text_encodings[:, 1:]
 
     @torch.no_grad()
     def get_image_embed(self, image):
@@ -1375,14 +1418,20 @@ class Decoder(BaseGaussianDiffusion):
     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
+        text_encodings = None
+        if exists(text):
+            _, text_encodings = self.clip.embed_text(text)
 
         assert not (self.condition_on_text_encodings and not exists(text_encodings)), 'text or text encodings must be passed into decoder if specified'
+        assert not (not self.condition_on_text_encodings and exists(text_encodings)), 'decoder specified not to be conditioned on text, yet it is presented'
 
         img = None
 
         for unet, vae, channel, image_size, predict_x_start in tqdm(zip(self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start)):
-            with self.one_unet_in_gpu(unet = unet):
+
+            context = self.one_unet_in_gpu(unet = unet) if image_embed.is_cuda else null_context()
+
+            with context:
                 lowres_cond_img = None
                 shape = (batch_size, channel, image_size, image_size)
 
@@ -1435,11 +1484,14 @@ class Decoder(BaseGaussianDiffusion):
         times = torch.randint(0, self.num_timesteps, (b,), device = device, dtype = torch.long)
 
         if not exists(image_embed):
-            image_embed = self.get_image_embed(image)
+            image_embed, _ = self.clip.embed_image(image)
 
-        text_encodings = self.get_text_encodings(text) if exists(text) and not exists(text_encodings) else None
+        text_encodings = None
+        if exists(text) and not exists(text_encodings):
+            _, text_encodings = self.clip.embed_text(text)
 
         assert not (self.condition_on_text_encodings and not exists(text_encodings)), 'text or text encodings must be passed into decoder if specified'
+        assert not (not self.condition_on_text_encodings and exists(text_encodings)), 'decoder specified not to be conditioned on text, yet it is presented'
 
         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
         image = resize_image_to(image, target_image_size)
@@ -1472,12 +1524,15 @@ class DALLE2(nn.Module):
         self.prior_num_samples = prior_num_samples
         self.decoder_need_text_cond = self.decoder.condition_on_text_encodings
 
+        self.to_pil = T.ToPILImage()
+
     @torch.no_grad()
     @eval_decorator
     def forward(
         self,
         text,
-        cond_scale = 1.
+        cond_scale = 1.,
+        return_pil_images = False
     ):
         device = next(self.parameters()).device
         one_text = isinstance(text, str) or (not is_list_str(text) and text.shape[0] == 1)
@@ -1491,7 +1546,11 @@ class DALLE2(nn.Module):
         text_cond = text if self.decoder_need_text_cond else None
         images = self.decoder.sample(image_embed, text = text_cond, cond_scale = cond_scale)
 
+        if return_pil_images:
+            images = list(map(self.to_pil, images.unbind(dim = 0)))
+
         if one_text:
             return images[0]
 
         return images
+

dalle2_pytorch/train.py

@@ -35,7 +35,7 @@ class EMA(nn.Module):
 
         self.update_moving_average(self.ema_model, self.online_model)
 
-    def update_moving_average(ma_model, current_model):
+    def update_moving_average(self, ma_model, current_model):
         def calculate_ema(beta, old, new):
             if not exists(old):
                 return new

dalle2_pytorch/vqgan_vae.py

@@ -12,8 +12,8 @@ from torch.autograd import grad as torch_grad
 import torchvision
 
 from einops import rearrange, reduce, repeat
-
-from dalle2_pytorch.attention import QueryAttnUpsample
+from einops_exts import rearrange_many
+from einops.layers.torch import Rearrange
 
 # constants
 
@@ -146,6 +146,8 @@ class LayerNormChan(nn.Module):
         mean = torch.mean(x, dim = 1, keepdim = True)
         return (x - mean) / (var + self.eps).sqrt() * self.gamma
 
+# discriminator
+
 class Discriminator(nn.Module):
     def __init__(
         self,
@@ -179,6 +181,8 @@ class Discriminator(nn.Module):
 
         return self.to_logits(x)
 
+# positional encoding
+
 class ContinuousPositionBias(nn.Module):
     """ from https://arxiv.org/abs/2111.09883 """
 
@@ -213,6 +217,84 @@ class ContinuousPositionBias(nn.Module):
         bias = rearrange(rel_pos, 'i j h -> h i j')
         return x + bias
 
+# resnet encoder / decoder
+
+class ResnetEncDec(nn.Module):
+    def __init__(
+        self,
+        dim,
+        *,
+        channels = 3,
+        layers = 4,
+        layer_mults = None,
+        num_resnet_blocks = 1,
+        resnet_groups = 16,
+        first_conv_kernel_size = 5,
+        use_attn = True,
+        attn_dim_head = 64,
+        attn_heads = 8,
+        attn_dropout = 0.,
+    ):
+        super().__init__()
+        assert dim % resnet_groups == 0, f'dimension {dim} must be divisible by {resnet_groups} (groups for the groupnorm)'
+
+        self.layers = layers
+
+        self.encoders = MList([])
+        self.decoders = MList([])
+
+        layer_mults = default(layer_mults, list(map(lambda t: 2 ** t, range(layers))))
+        assert len(layer_mults) == layers, 'layer multipliers must be equal to designated number of layers'
+
+        layer_dims = [dim * mult for mult in layer_mults]
+        dims = (dim, *layer_dims)
+
+        self.encoded_dim = 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,) * (layers - 1)), num_resnet_blocks)
+
+        if not isinstance(use_attn, tuple):
+            use_attn = (*((False,) * (layers - 1)), use_attn)
+
+        assert len(num_resnet_blocks) == layers, 'number of resnet blocks config must be equal to number of layers'
+        assert len(use_attn) == layers
+
+        for layer_index, (dim_in, dim_out), layer_num_resnet_blocks, layer_use_attn in zip(range(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.ConvTranspose2d(dim_out, dim_in, 4, 2, 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))
+
+    def get_encoded_fmap_size(self, image_size):
+        return image_size // (2 ** self.layers)
+
+    def encode(self, x):
+        for enc in self.encoders:
+            x = enc(x)
+        return x
+
+    def decode(self, x):
+        for dec in self.decoders:
+            x = dec(x)
+        return x
+
 class GLUResBlock(nn.Module):
     def __init__(self, chan, groups = 16):
         super().__init__()
@@ -246,6 +328,7 @@ class ResBlock(nn.Module):
         return self.net(x) + x
 
 # vqgan attention layer
+
 class VQGanAttention(nn.Module):
     def __init__(
         self,
@@ -290,6 +373,145 @@ class VQGanAttention(nn.Module):
 
         return out + residual
 
+# ViT encoder / decoder
+
+class RearrangeImage(nn.Module):
+    def forward(self, x):
+        n = x.shape[1]
+        w = h = int(sqrt(n))
+        return rearrange(x, 'b (h w) ... -> b h w ...', h = h, w = w)
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        *,
+        heads = 8,
+        dim_head = 32
+    ):
+        super().__init__()
+        self.norm = nn.LayerNorm(dim)
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+        inner_dim = dim_head * heads
+
+        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
+        self.to_out = nn.Linear(inner_dim, dim)
+
+    def forward(self, x):
+        h = self.heads
+
+        x = self.norm(x)
+
+        q, k, v = self.to_qkv(x).chunk(3, dim = -1)
+        q, k, v = rearrange_many((q, k, v), 'b n (h d) -> b h n d', h = h)
+
+        q = q * self.scale
+        sim = einsum('b h i d, b h j d -> b h i j', q, k)
+
+        sim = sim - sim.amax(dim = -1, keepdim = True).detach()
+        attn = sim.softmax(dim = -1)
+
+        out = einsum('b h i j, b h j d -> b h i d', attn, v)
+
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return self.to_out(out)
+
+def FeedForward(dim, mult = 4):
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, dim * mult, bias = False),
+        nn.GELU(),
+        nn.Linear(dim * mult, dim, bias = False)
+    )
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        dim,
+        *,
+        layers,
+        dim_head = 32,
+        heads = 8,
+        ff_mult = 4
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(layers):
+            self.layers.append(nn.ModuleList([
+                Attention(dim = dim, dim_head = dim_head, heads = heads),
+                FeedForward(dim = dim, mult = ff_mult)
+            ]))
+
+        self.norm = nn.LayerNorm(dim)
+
+    def forward(self, x):
+        for attn, ff in self.layers:
+            x = attn(x) + x
+            x = ff(x) + x
+
+        return self.norm(x)
+
+class ViTEncDec(nn.Module):
+    def __init__(
+        self,
+        dim,
+        channels = 3,
+        layers = 4,
+        patch_size = 8,
+        dim_head = 32,
+        heads = 8,
+        ff_mult = 4
+    ):
+        super().__init__()
+        self.encoded_dim = dim
+        self.patch_size = patch_size
+
+        input_dim = channels * (patch_size ** 2)
+
+        self.encoder = nn.Sequential(
+            Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_size, p2 = patch_size),
+            nn.Linear(input_dim, dim),
+            Transformer(
+                dim = dim,
+                dim_head = dim_head,
+                heads = heads,
+                ff_mult = ff_mult,
+                layers = layers
+            ),
+            RearrangeImage(),
+            Rearrange('b h w c -> b c h w')
+        )
+
+        self.decoder = nn.Sequential(
+            Rearrange('b c h w -> b (h w) c'),
+            Transformer(
+                dim = dim,
+                dim_head = dim_head,
+                heads = heads,
+                ff_mult = ff_mult,
+                layers = layers
+            ),
+            nn.Sequential(
+                nn.Linear(dim, dim * 4, bias = False),
+                nn.Tanh(),
+                nn.Linear(dim * 4, input_dim, bias = False),
+            ),
+            RearrangeImage(),
+            Rearrange('b h w (p1 p2 c) -> b c (h p1) (w p2)', p1 = patch_size, p2 = patch_size)
+        )
+
+    def get_encoded_fmap_size(self, image_size):
+        return image_size // self.patch_size
+
+    def encode(self, x):
+        return self.encoder(x)
+
+    def decode(self, x):
+        return self.decoder(x)
+
+# main vqgan-vae classes
+
 class NullVQGanVAE(nn.Module):
     def __init__(
         self,
@@ -320,81 +542,45 @@ class VQGanVAE(nn.Module):
         image_size,
         channels = 3,
         layers = 4,
-        layer_mults = None,
         l2_recon_loss = False,
         use_hinge_loss = True,
-        num_resnet_blocks = 1,
         vgg = None,
+        vq_codebook_dim = 256,
         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,
+        vae_type = 'resnet',
+        discr_layers = 4,
         **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)
+        encdec_kwargs, kwargs = groupby_prefix_and_trim('encdec_', kwargs)
 
         self.image_size = image_size
         self.channels = channels
-        self.layers = layers
-        self.fmap_size = image_size // (layers ** 2)
         self.codebook_size = vq_codebook_size
 
-        self.encoders = MList([])
-        self.decoders = MList([])
+        if vae_type == 'resnet':
+            enc_dec_klass = ResnetEncDec
+        elif vae_type == 'vit':
+            enc_dec_klass = ViTEncDec
+        else:
+            raise ValueError(f'{vae_type} not valid')
 
-        layer_mults = default(layer_mults, list(map(lambda t: 2 ** t, range(layers))))
-        assert len(layer_mults) == 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]
-
-        self.encoded_dim = 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,) * (layers - 1)), num_resnet_blocks)
-
-        if not isinstance(use_attn, tuple):
-            use_attn = (*((False,) * (layers - 1)), use_attn)
-
-        assert len(num_resnet_blocks) == layers, 'number of resnet blocks config must be equal to number of layers'
-        assert len(use_attn) == layers
-
-        for layer_index, (dim_in, dim_out), layer_num_resnet_blocks, layer_use_attn in zip(range(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.ConvTranspose2d(dim_out, dim_in, 4, 2, 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.enc_dec = enc_dec_klass(
+            dim = dim,
+            channels = channels,
+            layers = layers,
+            **encdec_kwargs
+        )
 
         self.vq = VQ(
-            dim = codebook_dim,
+            dim = self.enc_dec.encoded_dim,
+            codebook_dim = vq_codebook_dim,
             codebook_size = vq_codebook_size,
             decay = vq_decay,
             commitment_weight = vq_commitment_weight,
@@ -427,13 +613,21 @@ class VQGanVAE(nn.Module):
 
         # gan related losses
 
+        layer_mults = list(map(lambda t: 2 ** t, range(discr_layers)))
+        layer_dims = [dim * mult for mult in layer_mults]
+        dims = (dim, *layer_dims)
+
         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
 
+    @property
+    def encoded_dim(self):
+        return self.enc_dec.encoded_dim
+
     def get_encoded_fmap_size(self, image_size):
-        return image_size // (2 ** self.layers)
+        return self.enc_dec.get_encoded_fmap_size(image_size)
 
     def copy_for_eval(self):
         device = next(self.parameters()).device
@@ -459,16 +653,13 @@ class VQGanVAE(nn.Module):
         return self.vq.codebook
 
     def encode(self, fmap):
-        for enc in self.encoders:
-            fmap = enc(fmap)
-
+        fmap = self.enc_dec.encode(fmap)
         return fmap
 
     def decode(self, fmap, return_indices_and_loss = False):
         fmap, indices, commit_loss = self.vq(fmap)
 
-        for dec in self.decoders:
-            fmap = dec(fmap)
+        fmap = self.enc_dec.decode(fmap)
 
         if not return_indices_and_loss:
             return fmap

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.51',
+  version = '0.0.63',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',