make memory efficient unet design from imagen toggle-able

README.md

@@ -1207,4 +1207,14 @@ This library would not have gotten to this working state without the help of
 }
 ```
 
+```bibtex
+@article{Choi2022PerceptionPT,
+    title   = {Perception Prioritized Training of Diffusion Models},
+    author  = {Jooyoung Choi and Jungbeom Lee and Chaehun Shin and Sungwon Kim and Hyunwoo J. Kim and Sung-Hoon Yoon},
+    journal = {ArXiv},
+    year    = {2022},
+    volume  = {abs/2204.00227}
+}
+```
+
 *Creating noise from data is easy; creating data from noise is generative modeling.* - <a href="https://arxiv.org/abs/2011.13456">Yang Song's paper</a>

dalle2_pytorch/dalle2_pytorch.py

@@ -11,7 +11,7 @@ import torch.nn.functional as F
 from torch import nn, einsum
 import torchvision.transforms as T
 
-from einops import rearrange, repeat
+from einops import rearrange, repeat, reduce
 from einops.layers.torch import Rearrange
 from einops_exts import rearrange_many, repeat_many, check_shape
 from einops_exts.torch import EinopsToAndFrom
@@ -379,7 +379,7 @@ def sigmoid_beta_schedule(timesteps):
 
 
 class BaseGaussianDiffusion(nn.Module):
-    def __init__(self, *, beta_schedule, timesteps, loss_type):
+    def __init__(self, *, beta_schedule, timesteps, loss_type, p2_loss_weight_gamma = 0., p2_loss_weight_k = 1):
         super().__init__()
 
         if beta_schedule == "cosine":
@@ -444,6 +444,11 @@ class BaseGaussianDiffusion(nn.Module):
         register_buffer('posterior_mean_coef1', betas * torch.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))
         register_buffer('posterior_mean_coef2', (1. - alphas_cumprod_prev) * torch.sqrt(alphas) / (1. - alphas_cumprod))
 
+        # p2 loss reweighting
+
+        self.has_p2_loss_reweighting = p2_loss_weight_gamma > 0.
+        register_buffer('p2_loss_weight', (p2_loss_weight_k + alphas_cumprod / (1 - alphas_cumprod)) ** -p2_loss_weight_gamma)
+
     def q_posterior(self, x_start, x_t, t):
         posterior_mean = (
             extract(self.posterior_mean_coef1, t, x_t.shape) * x_start +
@@ -1079,8 +1084,9 @@ class DiffusionPrior(BaseGaussianDiffusion):
 def Upsample(dim):
     return nn.ConvTranspose2d(dim, dim, 4, 2, 1)
 
-def Downsample(dim):
-    return nn.Conv2d(dim, dim, 4, 2, 1)
+def Downsample(dim, *, dim_out = None):
+    dim_out = default(dim_out, dim)
+    return nn.Conv2d(dim, dim_out, 4, 2, 1)
 
 class SinusoidalPosEmb(nn.Module):
     def __init__(self, dim):
@@ -1346,6 +1352,7 @@ class Unet(nn.Module):
         init_cross_embed_kernel_sizes = (3, 7, 15),
         cross_embed_downsample = False,
         cross_embed_downsample_kernel_sizes = (2, 4),
+        memory_efficient = False,
         **kwargs
     ):
         super().__init__()
@@ -1365,7 +1372,7 @@ class Unet(nn.Module):
         self.channels_out = default(channels_out, channels)
 
         init_channels = channels if not lowres_cond else channels * 2 # in cascading diffusion, one concats the low resolution image, blurred, for conditioning the higher resolution synthesis
-        init_dim = default(init_dim, dim // 3 * 2)
+        init_dim = default(init_dim, dim)
 
         self.init_conv = CrossEmbedLayer(init_channels, dim_out = init_dim, kernel_sizes = init_cross_embed_kernel_sizes, stride = 1)
 
@@ -1456,10 +1463,11 @@ class Unet(nn.Module):
             layer_cond_dim = cond_dim if not is_first else None
 
             self.downs.append(nn.ModuleList([
-                ResnetBlock(dim_in, dim_out, time_cond_dim = time_cond_dim, groups = groups),
+                downsample_klass(dim_in, dim_out = dim_out) if memory_efficient else None,
+                ResnetBlock(dim_out if memory_efficient else dim_in, dim_out, time_cond_dim = time_cond_dim, groups = groups),
                 Residual(LinearAttention(dim_out, **attn_kwargs)) if sparse_attn else nn.Identity(),
                 nn.ModuleList([ResnetBlock(dim_out, dim_out, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim, groups = groups) for _ in range(layer_num_resnet_blocks)]),
-                downsample_klass(dim_out) if not is_last else nn.Identity()
+                downsample_klass(dim_out) if not is_last and not memory_efficient else None
             ]))
 
         mid_dim = dims[-1]
@@ -1468,7 +1476,9 @@ class Unet(nn.Module):
         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 = ResnetBlock(mid_dim, mid_dim, cond_dim = cond_dim, time_cond_dim = time_cond_dim, groups = resnet_groups[-1])
 
-        for ind, ((dim_in, dim_out), groups, layer_num_resnet_blocks) in enumerate(zip(reversed(in_out[1:]), reversed(resnet_groups), reversed(num_resnet_blocks))):
+        up_in_out_slice = slice(1 if not memory_efficient else None, None)
+
+        for ind, ((dim_in, dim_out), groups, layer_num_resnet_blocks) in enumerate(zip(reversed(in_out[up_in_out_slice]), reversed(resnet_groups), reversed(num_resnet_blocks))):
             is_last = ind >= (num_resolutions - 2)
             layer_cond_dim = cond_dim if not is_last else None
 
@@ -1649,7 +1659,10 @@ class Unet(nn.Module):
 
         hiddens = []
 
-        for init_block, sparse_attn, resnet_blocks, downsample in self.downs:
+        for pre_downsample, init_block, sparse_attn, resnet_blocks, post_downsample in self.downs:
+            if exists(pre_downsample):
+                x = pre_downsample(x)
+
             x = init_block(x, c, t)
             x = sparse_attn(x)
 
@@ -1657,7 +1670,9 @@ class Unet(nn.Module):
                 x = resnet_block(x, c, t)
 
             hiddens.append(x)
-            x = downsample(x)
+
+            if exists(post_downsample):
+                x = post_downsample(x)
 
         x = self.mid_block1(x, mid_c, t)
 
@@ -1755,12 +1770,16 @@ class Decoder(BaseGaussianDiffusion):
         unconditional = False,
         auto_normalize_img = True,                  # whether to take care of normalizing the image from [0, 1] to [-1, 1] and back automatically - you can turn this off if you want to pass in the [-1, 1] ranged image yourself from the dataloader
         use_dynamic_thres = False,                  # from the Imagen paper
-        dynamic_thres_percentile = 0.9
+        dynamic_thres_percentile = 0.9,
+        p2_loss_weight_gamma = 0.,                  # p2 loss weight, from https://arxiv.org/abs/2204.00227 - 0 is equivalent to weight of 1 across time - 1. is recommended
+        p2_loss_weight_k = 1
     ):
         super().__init__(
             beta_schedule = beta_schedule,
             timesteps = timesteps,
-            loss_type = loss_type
+            loss_type = loss_type,
+            p2_loss_weight_gamma = p2_loss_weight_gamma,
+            p2_loss_weight_k = p2_loss_weight_k
         )
 
         self.unconditional = unconditional
@@ -2028,7 +2047,13 @@ class Decoder(BaseGaussianDiffusion):
 
         target = noise if not predict_x_start else x_start
 
-        loss = self.loss_fn(pred, target)
+        loss = self.loss_fn(pred, target, reduction = 'none')
+        loss = reduce(loss, 'b ... -> b (...)', 'mean')
+
+        if self.has_p2_loss_reweighting:
+            loss = loss * extract(self.p2_loss_weight, times, loss.shape)
+
+        loss = loss.mean()
 
         if not learned_variance:
             # return simple loss if not using learned variance

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.7.0'
+__version__ = '0.8.1'