cast attention matrix back to original dtype pre-softmax in attention

README.md

@@ -49,6 +49,7 @@ This library would not have gotten to this working state without the help of
 - <a href="https://github.com/crowsonkb">Katherine</a> for her advice
 - <a href="https://stability.ai/">Stability AI</a> for the generous sponsorship
 - <a href="https://huggingface.co">🤗 Huggingface</a> and in particular <a href="https://github.com/sgugger">Sylvain</a> for the <a href="https://github.com/huggingface/accelerate">Accelerate</a> library
+- <a href="https://github.com/arogozhnikov">Alex</a> for <a href="https://github.com/arogozhnikov/einops">einops</a>, indispensable tool for tensor manipulation
 
 ... and many others. Thank you! 🙏
 
@@ -1253,4 +1254,35 @@ For detailed information on training the diffusion prior, please refer to the [d
 }
 ```
 
+```bibtex
+@misc{chen2022analog,
+    title   = {Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning},
+    author  = {Ting Chen and Ruixiang Zhang and Geoffrey Hinton},
+    year    = {2022},
+    eprint  = {2208.04202},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```
+
+```bibtex
+@article{Qiao2019WeightS,
+    title   = {Weight Standardization},
+    author  = {Siyuan Qiao and Huiyu Wang and Chenxi Liu and Wei Shen and Alan Loddon Yuille},
+    journal = {ArXiv},
+    year    = {2019},
+    volume  = {abs/1903.10520}
+}
+```
+
+```bibtex
+@inproceedings{rogozhnikov2022einops,
+    title   = {Einops: Clear and Reliable Tensor Manipulations with Einstein-like Notation},
+    author  = {Alex Rogozhnikov},
+    booktitle = {International Conference on Learning Representations},
+    year    = {2022},
+    url     = {https://openreview.net/forum?id=oapKSVM2bcj}
+}
+```
+
 *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

@@ -8,6 +8,7 @@ from pathlib import Path
 
 import torch
 import torch.nn.functional as F
+from torch.utils.checkpoint import checkpoint
 from torch import nn, einsum
 import torchvision.transforms as T
 
@@ -37,6 +38,8 @@ from coca_pytorch import CoCa
 
 NAT = 1. / math.log(2.)
 
+UnetOutput = namedtuple('UnetOutput', ['pred', 'var_interp_frac_unnormalized'])
+
 # helper functions
 
 def exists(val):
@@ -108,6 +111,28 @@ def pad_tuple_to_length(t, length, fillvalue = None):
         return t
     return (*t, *((fillvalue,) * remain_length))
 
+# checkpointing helper function
+
+def make_checkpointable(fn, **kwargs):
+    if isinstance(fn, nn.ModuleList):
+        return [maybe(make_checkpointable)(el, **kwargs) for el in fn]
+
+    condition = kwargs.pop('condition', None)
+
+    if exists(condition) and not condition(fn):
+        return fn
+
+    @wraps(fn)
+    def inner(*args):
+        input_needs_grad = any([isinstance(el, torch.Tensor) and el.requires_grad for el in args])
+
+        if not input_needs_grad:
+            return fn(*args)
+
+        return checkpoint(fn, *args)
+
+    return inner
+
 # for controlling freezing of CLIP
 
 def set_module_requires_grad_(module, requires_grad):
@@ -225,9 +250,15 @@ class XClipAdapter(BaseClipAdapter):
         text = text[..., :self.max_text_len]
         text_mask = text != 0
         encoder_output = self.clip.text_transformer(text)
-        text_cls, text_encodings = encoder_output[:, 0], encoder_output[:, 1:]
+
+        encoder_output_is_cls = encoder_output.ndim == 3
+
+        text_cls, text_encodings = (encoder_output[:, 0], encoder_output[:, 1:]) if encoder_output_is_cls else (encoder_output, None)
         text_embed = self.clip.to_text_latent(text_cls)
-        text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
+
+        if exists(text_encodings):
+            text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
+
         return EmbeddedText(l2norm(text_embed), text_encodings)
 
     @torch.no_grad()
@@ -847,7 +878,9 @@ class Attention(nn.Module):
 
         # attention
 
-        attn = sim.softmax(dim = -1)
+        attn = sim.softmax(dim = -1, dtype = torch.float32)
+        attn = attn.type(sim.dtype)
+
         attn = self.dropout(attn)
 
         # aggregate values
@@ -914,9 +947,12 @@ class DiffusionPriorNetwork(nn.Module):
         num_image_embeds = 1,
         num_text_embeds = 1,
         max_text_len = 256,
+        self_cond = False,
         **kwargs
     ):
         super().__init__()
+        self.dim = dim
+
         self.num_time_embeds = num_time_embeds
         self.num_image_embeds = num_image_embeds
         self.num_text_embeds = num_text_embeds
@@ -944,6 +980,10 @@ class DiffusionPriorNetwork(nn.Module):
         self.max_text_len = max_text_len
         self.null_text_embed = nn.Parameter(torch.randn(1, max_text_len, dim))
 
+        # whether to use self conditioning, Hinton's group's new ddpm technique
+
+        self.self_cond = self_cond
+
     def forward_with_cond_scale(
         self,
         *args,
@@ -965,12 +1005,19 @@ class DiffusionPriorNetwork(nn.Module):
         *,
         text_embed,
         text_encodings = None,
+        self_cond = None,
         cond_drop_prob = 0.
     ):
         batch, dim, device, dtype = *image_embed.shape, image_embed.device, image_embed.dtype
 
         num_time_embeds, num_image_embeds, num_text_embeds = self.num_time_embeds, self.num_image_embeds, self.num_text_embeds
 
+        # setup self conditioning
+
+        if self.self_cond:
+            self_cond = default(self_cond, lambda: torch.zeros(batch, self.dim, device = device, dtype = dtype))
+            self_cond = rearrange(self_cond, 'b d -> b 1 d')
+
         # in section 2.2, last paragraph
         # "... consisting of encoded text, CLIP text embedding, diffusion timestep embedding, noised CLIP image embedding, final embedding for prediction"
 
@@ -1020,13 +1067,16 @@ class DiffusionPriorNetwork(nn.Module):
         # whether text embedding is used for conditioning depends on whether text encodings are available for attention (for classifier free guidance, even though it seems from the paper it was not used in the prior ddpm, as the objective is different)
         # but let's just do it right
 
-        attend_padding = 1 + num_time_embeds + num_image_embeds # 1 for learned queries + number of image embeds + time embeds
+        attend_padding = 1 + num_time_embeds + num_image_embeds + int(self.self_cond) # 1 for learned queries + number of image embeds + time embeds
         mask = F.pad(mask, (0, attend_padding), value = True) # extend mask for text embedding, noised image embedding, time step embedding, and learned query
 
         time_embed = self.to_time_embeds(diffusion_timesteps)
 
         learned_queries = repeat(self.learned_query, 'd -> b 1 d', b = batch)
 
+        if self.self_cond:
+            learned_queries = torch.cat((image_embed, self_cond), dim = -2)
+
         tokens = torch.cat((
             text_encodings,
             text_embed,
@@ -1128,45 +1178,50 @@ class DiffusionPrior(nn.Module):
     def l2norm_clamp_embed(self, image_embed):
         return l2norm(image_embed) * self.image_embed_scale
 
-    def p_mean_variance(self, x, t, text_cond, clip_denoised = False, cond_scale = 1.):
+    def p_mean_variance(self, x, t, text_cond, self_cond = None, clip_denoised = False, cond_scale = 1.):
         assert not (cond_scale != 1. and not self.can_classifier_guidance), 'the model was not trained with conditional dropout, and thus one cannot use classifier free guidance (cond_scale anything other than 1)'
 
-        pred = self.net.forward_with_cond_scale(x, t, cond_scale = cond_scale, **text_cond)
+        pred = self.net.forward_with_cond_scale(x, t, cond_scale = cond_scale, self_cond = self_cond, **text_cond)
 
         if self.predict_x_start:
-            x_recon = pred
+            x_start = pred
         else:
-            x_recon = self.noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
+            x_start = self.noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
 
         if clip_denoised and not self.predict_x_start:
-            x_recon.clamp_(-1., 1.)
+            x_start.clamp_(-1., 1.)
 
         if self.predict_x_start and self.sampling_clamp_l2norm:
-            x_recon = l2norm(x_recon) * self.image_embed_scale
+            x_start = l2norm(x_start) * self.image_embed_scale
 
-        model_mean, posterior_variance, posterior_log_variance = self.noise_scheduler.q_posterior(x_start=x_recon, x_t=x, t=t)
-        return model_mean, posterior_variance, posterior_log_variance
+        model_mean, posterior_variance, posterior_log_variance = self.noise_scheduler.q_posterior(x_start=x_start, x_t=x, t=t)
+        return model_mean, posterior_variance, posterior_log_variance, x_start
 
     @torch.no_grad()
-    def p_sample(self, x, t, text_cond = None, clip_denoised = True, cond_scale = 1.):
+    def p_sample(self, x, t, text_cond = None, self_cond = None, clip_denoised = True, cond_scale = 1.):
         b, *_, device = *x.shape, x.device
-        model_mean, _, model_log_variance = self.p_mean_variance(x = x, t = t, text_cond = text_cond, clip_denoised = clip_denoised, cond_scale = cond_scale)
+        model_mean, _, model_log_variance, x_start = self.p_mean_variance(x = x, t = t, text_cond = text_cond, self_cond = self_cond, clip_denoised = clip_denoised, cond_scale = cond_scale)
         noise = torch.randn_like(x)
         # 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
+        pred = model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+        return pred, x_start
 
     @torch.no_grad()
     def p_sample_loop_ddpm(self, shape, text_cond, cond_scale = 1.):
         batch, device = shape[0], self.device
+
         image_embed = torch.randn(shape, device = device)
+        x_start = None # for self-conditioning
 
         if self.init_image_embed_l2norm:
             image_embed = l2norm(image_embed) * self.image_embed_scale
 
         for i in tqdm(reversed(range(0, self.noise_scheduler.num_timesteps)), desc='sampling loop time step', total=self.noise_scheduler.num_timesteps):
             times = torch.full((batch,), i, device = device, dtype = torch.long)
-            image_embed = self.p_sample(image_embed, times, text_cond = text_cond, cond_scale = cond_scale)
+
+            self_cond = x_start if self.net.self_cond else None
+            image_embed, x_start = self.p_sample(image_embed, times, text_cond = text_cond, self_cond = self_cond, cond_scale = cond_scale)
 
         if self.sampling_final_clamp_l2norm and self.predict_x_start:
             image_embed = self.l2norm_clamp_embed(image_embed)
@@ -1184,6 +1239,8 @@ class DiffusionPrior(nn.Module):
 
         image_embed = torch.randn(shape, device = device)
 
+        x_start = None # for self-conditioning
+
         if self.init_image_embed_l2norm:
             image_embed = l2norm(image_embed) * self.image_embed_scale
 
@@ -1193,7 +1250,9 @@ class DiffusionPrior(nn.Module):
 
             time_cond = torch.full((batch,), time, device = device, dtype = torch.long)
 
-            pred = self.net.forward_with_cond_scale(image_embed, time_cond, cond_scale = cond_scale, **text_cond)
+            self_cond = x_start if self.net.self_cond else None
+
+            pred = self.net.forward_with_cond_scale(image_embed, time_cond, self_cond = self_cond, cond_scale = cond_scale, **text_cond)
 
             if self.predict_x_start:
                 x_start = pred
@@ -1228,18 +1287,27 @@ class DiffusionPrior(nn.Module):
         is_ddim = timesteps < self.noise_scheduler.num_timesteps
 
         if not is_ddim:
-            return self.p_sample_loop_ddpm(*args, **kwargs)
+            normalized_image_embed = self.p_sample_loop_ddpm(*args, **kwargs)
+        else:
+            normalized_image_embed = self.p_sample_loop_ddim(*args, **kwargs, timesteps = timesteps)
 
-        return self.p_sample_loop_ddim(*args, **kwargs, timesteps = timesteps)
+        image_embed = normalized_image_embed / self.image_embed_scale
+        return image_embed
 
     def p_losses(self, image_embed, times, text_cond, noise = None):
         noise = default(noise, lambda: torch.randn_like(image_embed))
 
         image_embed_noisy = self.noise_scheduler.q_sample(x_start = image_embed, t = times, noise = noise)
 
+        self_cond = None
+        if self.net.self_cond and random.random() < 0.5:
+            with torch.no_grad():
+                self_cond = self.net(image_embed_noisy, times, **text_cond).detach()
+
         pred = self.net(
             image_embed_noisy,
             times,
+            self_cond = self_cond,
             cond_drop_prob = self.cond_drop_prob,
             **text_cond
         )
@@ -1293,8 +1361,6 @@ class DiffusionPrior(nn.Module):
 
         # retrieve original unscaled image embed
 
-        image_embeds /= self.image_embed_scale
-
         text_embeds = text_cond['text_embed']
 
         text_embeds = rearrange(text_embeds, '(b r) d -> b r d', r = num_samples_per_batch)
@@ -1393,6 +1459,26 @@ def Downsample(dim, *, dim_out = None):
     dim_out = default(dim_out, dim)
     return nn.Conv2d(dim, dim_out, 4, 2, 1)
 
+class WeightStandardizedConv2d(nn.Conv2d):
+    """
+    https://arxiv.org/abs/1903.10520
+    weight standardization purportedly works synergistically with group normalization
+    """
+    def forward(self, x):
+        eps = 1e-5 if x.dtype == torch.float32 else 1e-3
+
+        weight = self.weight
+        flattened_weights = rearrange(weight, 'o ... -> o (...)')
+
+        mean = reduce(weight, 'o ... -> o 1 1 1', 'mean')
+
+        var = torch.var(flattened_weights, dim = -1, unbiased = False)
+        var = rearrange(var, 'o -> o 1 1 1')
+
+        weight = (weight - mean) * (var + eps).rsqrt()
+
+        return F.conv2d(x, weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
+
 class SinusoidalPosEmb(nn.Module):
     def __init__(self, dim):
         super().__init__()
@@ -1411,10 +1497,13 @@ class Block(nn.Module):
         self,
         dim,
         dim_out,
-        groups = 8
+        groups = 8,
+        weight_standardization = False
     ):
         super().__init__()
-        self.project = nn.Conv2d(dim, dim_out, 3, padding = 1)
+        conv_klass = nn.Conv2d if not weight_standardization else WeightStandardizedConv2d
+
+        self.project = conv_klass(dim, dim_out, 3, padding = 1)
         self.norm = nn.GroupNorm(groups, dim_out)
         self.act = nn.SiLU()
 
@@ -1438,6 +1527,7 @@ class ResnetBlock(nn.Module):
         cond_dim = None,
         time_cond_dim = None,
         groups = 8,
+        weight_standardization = False,
         cosine_sim_cross_attn = False
     ):
         super().__init__()
@@ -1463,8 +1553,8 @@ class ResnetBlock(nn.Module):
                 )
             )
 
-        self.block1 = Block(dim, dim_out, groups = groups)
-        self.block2 = Block(dim_out, dim_out, groups = groups)
+        self.block1 = Block(dim, dim_out, groups = groups, weight_standardization = weight_standardization)
+        self.block2 = Block(dim_out, dim_out, groups = groups, weight_standardization = weight_standardization)
         self.res_conv = nn.Conv2d(dim, dim_out, 1) if dim != dim_out else nn.Identity()
 
     def forward(self, x, time_emb = None, cond = None):
@@ -1548,7 +1638,8 @@ class CrossAttention(nn.Module):
             mask = rearrange(mask, 'b j -> b 1 1 j')
             sim = sim.masked_fill(~mask, max_neg_value)
 
-        attn = sim.softmax(dim = -1)
+        attn = sim.softmax(dim = -1, dtype = torch.float32)
+        attn = attn.type(sim.dtype)
 
         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)')
@@ -1677,6 +1768,7 @@ class Unet(nn.Module):
         attn_heads = 16,
         lowres_cond = False,             # for cascading diffusion - https://cascaded-diffusion.github.io/
         lowres_noise_cond = False,       # for conditioning on low resolution noising, based on Imagen
+        self_cond = False,               # set this to True to use the self-conditioning technique from - https://arxiv.org/abs/2208.04202
         sparse_attn = False,
         cosine_sim_cross_attn = False,
         cosine_sim_self_attn = False,
@@ -1688,6 +1780,7 @@ class Unet(nn.Module):
         init_dim = None,
         init_conv_kernel_size = 7,
         resnet_groups = 8,
+        resnet_weight_standardization = False,
         num_resnet_blocks = 2,
         init_cross_embed = True,
         init_cross_embed_kernel_sizes = (3, 7, 15),
@@ -1698,6 +1791,7 @@ class Unet(nn.Module):
         pixel_shuffle_upsample = True,
         final_conv_kernel_size = 1,
         combine_upsample_fmaps = False, # whether to combine the outputs of all upsample blocks, as in unet squared paper
+        checkpoint_during_training = False,
         **kwargs
     ):
         super().__init__()
@@ -1711,12 +1805,21 @@ class Unet(nn.Module):
 
         self.lowres_cond = lowres_cond
 
+        # whether to do self conditioning
+
+        self.self_cond = self_cond
+
         # determine dimensions
 
         self.channels = channels
         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
+         # initial number of channels depends on
+         # (1) low resolution conditioning from cascading ddpm paper, conditioned on previous unet output in the cascade
+         # (2) self conditioning (bit diffusion paper)
+
+        init_channels = channels * (1 + int(lowres_cond) + int(self_cond))
+
         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) if init_cross_embed else nn.Conv2d(init_channels, init_dim, init_conv_kernel_size, padding = init_conv_kernel_size // 2)
@@ -1825,7 +1928,7 @@ class Unet(nn.Module):
 
         # prepare resnet klass
 
-        resnet_block = partial(ResnetBlock, cosine_sim_cross_attn = cosine_sim_cross_attn)
+        resnet_block = partial(ResnetBlock, cosine_sim_cross_attn = cosine_sim_cross_attn, weight_standardization = resnet_weight_standardization)
 
         # give memory efficient unet an initial resnet block
 
@@ -1908,6 +2011,10 @@ class Unet(nn.Module):
 
         zero_init_(self.to_out) # since both OpenAI and @crowsonkb are doing it
 
+        # whether to checkpoint during training
+
+        self.checkpoint_during_training = checkpoint_during_training
+
     # if the current settings for the unet are not correct
     # for cascading DDPM, then reinit the unet with the right settings
     def cast_model_parameters(
@@ -1965,7 +2072,9 @@ class Unet(nn.Module):
         image_cond_drop_prob = 0.,
         text_cond_drop_prob = 0.,
         blur_sigma = None,
-        blur_kernel_size = None
+        blur_kernel_size = None,
+        disable_checkpoint = False,
+        self_cond = None
     ):
         batch_size, device = x.shape[0], x.device
 
@@ -1973,6 +2082,14 @@ class Unet(nn.Module):
 
         assert not (self.lowres_cond and not exists(lowres_cond_img)), 'low resolution conditioning image must be present'
 
+        # concat self conditioning, if needed
+
+        if self.self_cond:
+            self_cond = default(self_cond, lambda: torch.zeros_like(x))
+            x = torch.cat((x, self_cond), dim = 1)
+
+        # concat low resolution conditioning
+
         if exists(lowres_cond_img):
             x = torch.cat((x, lowres_cond_img), dim = 1)
 
@@ -2087,17 +2204,29 @@ class Unet(nn.Module):
         c = self.norm_cond(c)
         mid_c = self.norm_mid_cond(mid_c)
 
+        # gradient checkpointing
+
+        can_checkpoint = self.training and self.checkpoint_during_training and not disable_checkpoint
+        apply_checkpoint_fn = make_checkpointable if can_checkpoint else identity
+
+        # make checkpointable modules
+
+        init_resnet_block, mid_block1, mid_attn, mid_block2, final_resnet_block = [maybe(apply_checkpoint_fn)(module) for module in (self.init_resnet_block, self.mid_block1, self.mid_attn, self.mid_block2, self.final_resnet_block)]
+
+        can_checkpoint_cond = lambda m: isinstance(m, ResnetBlock)
+        downs, ups = [maybe(apply_checkpoint_fn)(m, condition = can_checkpoint_cond) for m in (self.downs, self.ups)]
+
         # initial resnet block
 
-        if exists(self.init_resnet_block):
-            x = self.init_resnet_block(x, t)
+        if exists(init_resnet_block):
+            x = init_resnet_block(x, t)
 
         # go through the layers of the unet, down and up
 
         down_hiddens = []
         up_hiddens = []
 
-        for pre_downsample, init_block, resnet_blocks, attn, post_downsample in self.downs:
+        for pre_downsample, init_block, resnet_blocks, attn, post_downsample in downs:
             if exists(pre_downsample):
                 x = pre_downsample(x)
 
@@ -2113,16 +2242,16 @@ class Unet(nn.Module):
             if exists(post_downsample):
                 x = post_downsample(x)
 
-        x = self.mid_block1(x, t, mid_c)
+        x = mid_block1(x, t, mid_c)
 
-        if exists(self.mid_attn):
-            x = self.mid_attn(x)
+        if exists(mid_attn):
+            x = mid_attn(x)
 
-        x = self.mid_block2(x, t, mid_c)
+        x = mid_block2(x, t, mid_c)
 
         connect_skip = lambda fmap: torch.cat((fmap, down_hiddens.pop() * self.skip_connect_scale), dim = 1)
 
-        for init_block, resnet_blocks, attn, upsample in self.ups:
+        for init_block, resnet_blocks, attn, upsample in ups:
             x = connect_skip(x)
             x = init_block(x, t, c)
 
@@ -2139,7 +2268,7 @@ class Unet(nn.Module):
 
         x = torch.cat((x, r), dim = 1)
 
-        x = self.final_resnet_block(x, t)
+        x = final_resnet_block(x, t)
 
         if exists(lowres_cond_img):
             x = torch.cat((x, lowres_cond_img), dim = 1)
@@ -2492,6 +2621,14 @@ class Decoder(nn.Module):
         index = unet_number - 1
         return self.unets[index]
 
+    def parse_unet_output(self, learned_variance, output):
+        var_interp_frac_unnormalized = None
+
+        if learned_variance:
+            output, var_interp_frac_unnormalized = output.chunk(2, dim = 1)
+
+        return UnetOutput(output, var_interp_frac_unnormalized)
+
     @contextmanager
     def one_unet_in_gpu(self, unet_number = None, unet = None):
         assert exists(unet_number) ^ exists(unet)
@@ -2530,23 +2667,22 @@ class Decoder(nn.Module):
         x = x.clamp(-s, s) / s
         return x
 
-    def p_mean_variance(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, lowres_cond_img = None, clip_denoised = True, predict_x_start = False, learned_variance = False, cond_scale = 1., model_output = None, lowres_noise_level = None):
+    def p_mean_variance(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, lowres_cond_img = None, self_cond = None, clip_denoised = True, predict_x_start = False, learned_variance = False, cond_scale = 1., model_output = None, lowres_noise_level = None):
         assert not (cond_scale != 1. and not self.can_classifier_guidance), 'the decoder was not trained with conditional dropout, and thus one cannot use classifier free guidance (cond_scale anything other than 1)'
 
-        pred = default(model_output, lambda: unet.forward_with_cond_scale(x, t, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, lowres_noise_level = lowres_noise_level))
+        model_output = default(model_output, lambda: unet.forward_with_cond_scale(x, t, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, self_cond = self_cond, lowres_noise_level = lowres_noise_level))
 
-        if learned_variance:
-            pred, var_interp_frac_unnormalized = pred.chunk(2, dim = 1)
+        pred, var_interp_frac_unnormalized = self.parse_unet_output(learned_variance, model_output)
 
         if predict_x_start:
-            x_recon = pred
+            x_start = pred
         else:
-            x_recon = noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
+            x_start = noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
 
         if clip_denoised:
-            x_recon = self.dynamic_threshold(x_recon)
+            x_start = self.dynamic_threshold(x_start)
 
-        model_mean, posterior_variance, posterior_log_variance = noise_scheduler.q_posterior(x_start=x_recon, x_t=x, t=t)
+        model_mean, posterior_variance, posterior_log_variance = noise_scheduler.q_posterior(x_start=x_start, x_t=x, t=t)
 
         if learned_variance:
             # if learned variance, posterio variance and posterior log variance are predicted by the network
@@ -2562,16 +2698,17 @@ class Decoder(nn.Module):
             posterior_log_variance = var_interp_frac * max_log + (1 - var_interp_frac) * min_log
             posterior_variance = posterior_log_variance.exp()
 
-        return model_mean, posterior_variance, posterior_log_variance
+        return model_mean, posterior_variance, posterior_log_variance, x_start
 
     @torch.no_grad()
-    def p_sample(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, cond_scale = 1., lowres_cond_img = None, predict_x_start = False, learned_variance = False, clip_denoised = True, lowres_noise_level = None):
+    def p_sample(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, cond_scale = 1., lowres_cond_img = None, self_cond = None, predict_x_start = False, learned_variance = False, clip_denoised = True, lowres_noise_level = None):
         b, *_, device = *x.shape, x.device
-        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, predict_x_start = predict_x_start, noise_scheduler = noise_scheduler, learned_variance = learned_variance, lowres_noise_level = lowres_noise_level)
+        model_mean, _, model_log_variance, x_start = self.p_mean_variance(unet, x = x, t = t, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, self_cond = self_cond, clip_denoised = clip_denoised, predict_x_start = predict_x_start, noise_scheduler = noise_scheduler, learned_variance = learned_variance, lowres_noise_level = lowres_noise_level)
         noise = torch.randn_like(x)
         # no noise when t == 0
         nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
-        return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+        pred = model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+        return pred, x_start
 
     @torch.no_grad()
     def p_sample_loop_ddpm(
@@ -2597,6 +2734,8 @@ class Decoder(nn.Module):
         b = shape[0]
         img = torch.randn(shape, device = device)
 
+        x_start = None # for self-conditioning
+
         is_inpaint = exists(inpaint_image)
         resample_times = inpaint_resample_times if is_inpaint else 1
 
@@ -2624,13 +2763,16 @@ class Decoder(nn.Module):
                     noised_inpaint_image = noise_scheduler.q_sample(inpaint_image, t = times)
                     img = (img * ~inpaint_mask) + (noised_inpaint_image * inpaint_mask)
 
-                img = self.p_sample(
+                self_cond = x_start if unet.self_cond else None
+
+                img, x_start = self.p_sample(
                     unet,
                     img,
                     times,
                     image_embed = image_embed,
                     text_encodings = text_encodings,
                     cond_scale = cond_scale,
+                    self_cond = self_cond,
                     lowres_cond_img = lowres_cond_img,
                     lowres_noise_level = lowres_noise_level,
                     predict_x_start = predict_x_start,
@@ -2689,6 +2831,8 @@ class Decoder(nn.Module):
 
         img = torch.randn(shape, device = device)
 
+        x_start = None # for self-conditioning
+
         if not is_latent_diffusion:
             lowres_cond_img = maybe(self.normalize_img)(lowres_cond_img)
 
@@ -2709,10 +2853,11 @@ class Decoder(nn.Module):
                     noised_inpaint_image = noise_scheduler.q_sample(inpaint_image, t = time_cond)
                     img = (img * ~inpaint_mask) + (noised_inpaint_image * inpaint_mask)
 
-                pred = unet.forward_with_cond_scale(img, time_cond, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, lowres_noise_level = lowres_noise_level)
+                self_cond = x_start if unet.self_cond else None
 
-                if learned_variance:
-                    pred, _ = pred.chunk(2, dim = 1)
+                unet_output = unet.forward_with_cond_scale(img, time_cond, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, self_cond = self_cond, lowres_cond_img = lowres_cond_img, lowres_noise_level = lowres_noise_level)
+
+                pred, _ = self.parse_unet_output(learned_variance, unet_output)
 
                 if predict_x_start:
                     x_start = pred
@@ -2769,21 +2914,37 @@ class Decoder(nn.Module):
 
         x_noisy = noise_scheduler.q_sample(x_start = x_start, t = times, noise = noise)
 
-        model_output = unet(
-            x_noisy,
-            times,
+        # unet kwargs
+
+        unet_kwargs = dict(
             image_embed = image_embed,
             text_encodings = text_encodings,
             lowres_cond_img = lowres_cond_img,
             lowres_noise_level = lowres_noise_level,
+        )
+
+        # self conditioning
+
+        self_cond = None
+
+        if unet.self_cond and random.random() < 0.5:
+            with torch.no_grad():
+                unet_output = unet(x_noisy, times, **unet_kwargs)
+                self_cond, _ = self.parse_unet_output(learned_variance, unet_output)
+                self_cond = self_cond.detach()
+
+        # forward to get model prediction
+
+        unet_output = unet(
+            x_noisy,
+            times,
+            **unet_kwargs,
+            self_cond = self_cond,
             image_cond_drop_prob = self.image_cond_drop_prob,
             text_cond_drop_prob = self.text_cond_drop_prob,
         )
 
-        if learned_variance:
-            pred, _ = model_output.chunk(2, dim = 1)
-        else:
-            pred = model_output
+        pred, _ = self.parse_unet_output(learned_variance, unet_output)
 
         target = noise if not predict_x_start else x_start
 
@@ -2806,7 +2967,7 @@ class Decoder(nn.Module):
         # if learning the variance, also include the extra weight kl loss
 
         true_mean, _, true_log_variance_clipped = noise_scheduler.q_posterior(x_start = x_start, x_t = x_noisy, t = times)
-        model_mean, _, model_log_variance = self.p_mean_variance(unet, x = x_noisy, t = times, image_embed = image_embed, noise_scheduler = noise_scheduler, clip_denoised = clip_denoised, learned_variance = True, model_output = model_output)
+        model_mean, _, model_log_variance, _ = self.p_mean_variance(unet, x = x_noisy, t = times, image_embed = image_embed, noise_scheduler = noise_scheduler, clip_denoised = clip_denoised, learned_variance = True, model_output = unet_output)
 
         # kl loss with detached model predicted mean, for stability reasons as in paper
 

dalle2_pytorch/trainer.py

@@ -9,7 +9,7 @@ from collections.abc import Iterable
 import torch
 import torch.nn.functional as F
 from torch import nn
-from torch.optim.lr_scheduler import LambdaLR
+from torch.optim.lr_scheduler import LambdaLR, CosineAnnealingLR
 from torch.cuda.amp import autocast, GradScaler
 
 from dalle2_pytorch.dalle2_pytorch import Decoder, DiffusionPrior
@@ -181,7 +181,8 @@ class DiffusionPriorTrainer(nn.Module):
         eps = 1e-6,
         max_grad_norm = None,
         group_wd_params = True,
-        warmup_steps = 1,
+        warmup_steps = None,
+        cosine_decay_max_steps = None,
         **kwargs
     ):
         super().__init__()
@@ -233,8 +234,11 @@ class DiffusionPriorTrainer(nn.Module):
             **self.optim_kwargs,
             **kwargs
         )
-        
-        self.scheduler = LambdaLR(self.optimizer, lr_lambda = lambda _: 1.0)
+
+        if exists(cosine_decay_max_steps):
+            self.scheduler = CosineAnnealingLR(optimizer, T_max = cosine_decay_max_steps)
+        else:
+            self.scheduler = LambdaLR(self.optimizer, lr_lambda = lambda _: 1.0)
         
         self.warmup_scheduler = warmup.LinearWarmup(self.optimizer, warmup_period = warmup_steps) if exists(warmup_steps) else None
 
@@ -271,6 +275,7 @@ class DiffusionPriorTrainer(nn.Module):
             # FIXME: LambdaLR can't be saved due to pickling issues
             save_obj = dict(
                 optimizer = self.optimizer.state_dict(),
+                scheduler = self.scheduler.state_dict(),
                 warmup_scheduler = self.warmup_scheduler,
                 model = self.accelerator.unwrap_model(self.diffusion_prior).state_dict(),
                 version = version.parse(__version__),
@@ -317,7 +322,9 @@ class DiffusionPriorTrainer(nn.Module):
         # unwrap the model when loading from checkpoint
         self.accelerator.unwrap_model(self.diffusion_prior).load_state_dict(loaded_obj['model'], strict = strict)
         self.step.copy_(torch.ones_like(self.step, device=self.device) * loaded_obj['step'].to(self.device))
+
         self.optimizer.load_state_dict(loaded_obj['optimizer'])
+        self.scheduler.load_state_dict(loaded_obj['scheduler'])
 
         # set warmupstep
         if exists(self.warmup_scheduler):
@@ -350,7 +357,8 @@ class DiffusionPriorTrainer(nn.Module):
 
         # accelerator will ocassionally skip optimizer steps in a "dynamic loss scaling strategy"
         if not self.accelerator.optimizer_step_was_skipped:
-            with self.warmup_scheduler.dampening():
+            sched_context = self.warmup_scheduler.dampening if exists(self.warmup_scheduler) else nullcontext
+            with sched_context():
                 self.scheduler.step()
 
         if self.use_ema:
@@ -433,6 +441,7 @@ class DecoderTrainer(nn.Module):
         wd = 1e-2,
         eps = 1e-8,
         warmup_steps = None,
+        cosine_decay_max_steps = None,
         max_grad_norm = 0.5,
         amp = False,
         group_wd_params = True,
@@ -454,7 +463,7 @@ class DecoderTrainer(nn.Module):
         # be able to finely customize learning rate, weight decay
         # per unet
 
-        lr, wd, eps, warmup_steps = map(partial(cast_tuple, length = self.num_unets), (lr, wd, eps, warmup_steps))
+        lr, wd, eps, warmup_steps, cosine_decay_max_steps = map(partial(cast_tuple, length = self.num_unets), (lr, wd, eps, warmup_steps, cosine_decay_max_steps))
 
         assert all([unet_lr <= 1e-2 for unet_lr in lr]), 'your learning rate is too high, recommend sticking with 1e-4, at most 5e-4'
 
@@ -462,7 +471,7 @@ class DecoderTrainer(nn.Module):
         schedulers = []
         warmup_schedulers = []
 
-        for unet, unet_lr, unet_wd, unet_eps, unet_warmup_steps in zip(decoder.unets, lr, wd, eps, warmup_steps):
+        for unet, unet_lr, unet_wd, unet_eps, unet_warmup_steps, unet_cosine_decay_max_steps in zip(decoder.unets, lr, wd, eps, warmup_steps, cosine_decay_max_steps):
             if isinstance(unet, nn.Identity):
                 optimizers.append(None)
                 schedulers.append(None)
@@ -478,7 +487,11 @@ class DecoderTrainer(nn.Module):
                 )
 
                 optimizers.append(optimizer)
-                scheduler = LambdaLR(optimizer, lr_lambda = lambda step: 1.0)
+
+                if exists(unet_cosine_decay_max_steps):
+                    scheduler = CosineAnnealingLR(optimizer, T_max = unet_cosine_decay_max_steps)
+                else:
+                    scheduler = LambdaLR(optimizer, lr_lambda = lambda step: 1.0)
 
                 warmup_scheduler = warmup.LinearWarmup(optimizer, warmup_period = unet_warmup_steps) if exists(unet_warmup_steps) else None
                 warmup_schedulers.append(warmup_scheduler)
@@ -558,9 +571,15 @@ class DecoderTrainer(nn.Module):
 
         for ind in range(0, self.num_unets):
             optimizer_key = f'optim{ind}'
+            scheduler_key = f'sched{ind}'
+
             optimizer = getattr(self, optimizer_key)
-            state_dict = optimizer.state_dict() if optimizer is not None else None
-            save_obj = {**save_obj, optimizer_key: state_dict}
+            scheduler = getattr(self, scheduler_key)
+
+            optimizer_state_dict = optimizer.state_dict() if exists(optimizer) else None
+            scheduler_state_dict = scheduler.state_dict() if exists(scheduler) else None
+
+            save_obj = {**save_obj, optimizer_key: optimizer_state_dict, scheduler_key: scheduler_state_dict}
 
         if self.use_ema:
             save_obj = {**save_obj, 'ema': self.ema_unets.state_dict()}
@@ -581,10 +600,18 @@ class DecoderTrainer(nn.Module):
 
             optimizer_key = f'optim{ind}'
             optimizer = getattr(self, optimizer_key)
+
+            scheduler_key = f'sched{ind}'
+            scheduler = getattr(self, scheduler_key)
+
             warmup_scheduler = self.warmup_schedulers[ind]
-            if optimizer is not None:
+
+            if exists(optimizer):
                 optimizer.load_state_dict(loaded_obj[optimizer_key])
 
+            if exists(scheduler):
+                scheduler.load_state_dict(loaded_obj[scheduler_key])
+
             if exists(warmup_scheduler):
                 warmup_scheduler.last_step = last_step
 

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.4.5'
+__version__ = '1.8.4'

setup.py

@@ -26,7 +26,7 @@ setup(
   install_requires=[
     'accelerate',
     'click',
-    'clip-anytorch',
+    'clip-anytorch>=2.4.0',
     'coca-pytorch>=0.0.5',
     'ema-pytorch>=0.0.7',
     'einops>=0.4',