fix classifier free guidance for image hiddens summed to time hiddens, thanks to @xvjiarui for finding this bug

dalle2_pytorch/dalle2_pytorch.py

@@ -1,7 +1,6 @@
 import math
 import random
 from tqdm import tqdm
-from inspect import isfunction
 from functools import partial, wraps
 from contextlib import contextmanager
 from collections import namedtuple
@@ -57,7 +56,7 @@ def maybe(fn):
 def default(val, d):
     if exists(val):
         return val
-    return d() if isfunction(d) else d
+    return d() if callable(d) else d
 
 def cast_tuple(val, length = 1):
     if isinstance(val, list):
@@ -314,11 +313,6 @@ def extract(a, t, x_shape):
     out = a.gather(-1, t)
     return out.reshape(b, *((1,) * (len(x_shape) - 1)))
 
-def noise_like(shape, device, repeat=False):
-    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
-    noise = lambda: torch.randn(shape, device=device)
-    return repeat_noise() if repeat else noise()
-
 def meanflat(x):
     return x.mean(dim = tuple(range(1, len(x.shape))))
 
@@ -373,7 +367,7 @@ def quadratic_beta_schedule(timesteps):
     scale = 1000 / timesteps
     beta_start = scale * 0.0001
     beta_end = scale * 0.02
-    return torch.linspace(beta_start**2, beta_end**2, timesteps, dtype = torch.float64) ** 2
+    return torch.linspace(beta_start**0.5, beta_end**0.5, timesteps, dtype = torch.float64) ** 2
 
 
 def sigmoid_beta_schedule(timesteps):
@@ -946,10 +940,10 @@ class DiffusionPrior(BaseGaussianDiffusion):
         return model_mean, posterior_variance, posterior_log_variance
 
     @torch.no_grad()
-    def p_sample(self, x, t, text_cond = None, clip_denoised = True, repeat_noise = False, cond_scale = 1.):
+    def p_sample(self, x, t, text_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)
-        noise = noise_like(x.shape, device, repeat_noise)
+        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
@@ -1428,6 +1422,7 @@ class Unet(nn.Module):
         # for classifier free guidance
 
         self.null_image_embed = nn.Parameter(torch.randn(1, num_image_tokens, cond_dim))
+        self.null_image_hiddens = nn.Parameter(torch.randn(1, time_cond_dim))
 
         self.max_text_len = max_text_len
         self.null_text_embed = nn.Parameter(torch.randn(1, max_text_len, cond_dim))
@@ -1565,19 +1560,28 @@ class Unet(nn.Module):
         time_tokens = self.to_time_tokens(time_hiddens)
         t = self.to_time_cond(time_hiddens)
 
-        # image embedding to be summed to time embedding
-        # discovered by @mhh0318 in the paper
-
-        if exists(image_embed) and exists(self.to_image_hiddens):
-            image_hiddens = self.to_image_hiddens(image_embed)
-            t = t + image_hiddens
-
         # conditional dropout
 
         image_keep_mask = prob_mask_like((batch_size,), 1 - image_cond_drop_prob, device = device)
         text_keep_mask = prob_mask_like((batch_size,), 1 - text_cond_drop_prob, device = device)
 
-        image_keep_mask, text_keep_mask = rearrange_many((image_keep_mask, text_keep_mask), 'b -> b 1 1')
+        text_keep_mask = rearrange(text_keep_mask, 'b -> b 1 1')
+
+        # image embedding to be summed to time embedding
+        # discovered by @mhh0318 in the paper
+
+        if exists(image_embed) and exists(self.to_image_hiddens):
+            image_hiddens = self.to_image_hiddens(image_embed)
+            image_keep_mask_hidden = rearrange(image_keep_mask, 'b -> b 1')
+            null_image_hiddens = self.null_image_hiddens.to(image_hiddens.dtype)
+
+            image_hiddens = torch.where(
+                image_keep_mask_hidden,
+                image_hiddens,
+                null_image_hiddens
+            )
+
+            t = t + image_hiddens
 
         # mask out image embedding depending on condition dropout
         # for classifier free guidance
@@ -1585,11 +1589,12 @@ class Unet(nn.Module):
         image_tokens = None
 
         if self.cond_on_image_embeds:
+            image_keep_mask_embed = rearrange(image_keep_mask, 'b -> b 1 1')
             image_tokens = self.image_to_tokens(image_embed)
             null_image_embed = self.null_image_embed.to(image_tokens.dtype) # for some reason pytorch AMP not working
 
             image_tokens = torch.where(
-                image_keep_mask,
+                image_keep_mask_embed,
                 image_tokens,
                 null_image_embed
             )
@@ -1956,10 +1961,10 @@ class Decoder(BaseGaussianDiffusion):
         return model_mean, posterior_variance, posterior_log_variance
 
     @torch.no_grad()
-    def p_sample(self, unet, x, t, image_embed, text_encodings = None, text_mask = None, cond_scale = 1., lowres_cond_img = None, predict_x_start = False, learned_variance = False, clip_denoised = True, repeat_noise = False):
+    def p_sample(self, unet, x, t, image_embed, text_encodings = None, text_mask = None, cond_scale = 1., lowres_cond_img = None, predict_x_start = False, learned_variance = False, clip_denoised = True):
         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, text_mask = text_mask, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, clip_denoised = clip_denoised, predict_x_start = predict_x_start, learned_variance = learned_variance)
-        noise = noise_like(x.shape, device, repeat_noise)
+        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

dalle2_pytorch/optimizer.py

@@ -11,7 +11,7 @@ def get_optimizer(
     params,
     lr = 1e-4,
     wd = 1e-2,
-    betas = (0.9, 0.999),
+    betas = (0.9, 0.99),
     eps = 1e-8,
     filter_by_requires_grad = False,
     group_wd_params = True,

dalle2_pytorch/trainer.py

@@ -58,8 +58,15 @@ def num_to_groups(num, divisor):
         arr.append(remainder)
     return arr
 
-def get_pkg_version():
-    return __version__
+def clamp(value, min_value = None, max_value = None):
+    assert exists(min_value) or exists(max_value)
+    if exists(min_value):
+        value = max(value, min_value)
+
+    if exists(max_value):
+        value = min(value, max_value)
+
+    return value
 
 # decorators
 
@@ -175,12 +182,34 @@ def save_diffusion_model(save_path, model, optimizer, scaler, config, image_embe
 # exponential moving average wrapper
 
 class EMA(nn.Module):
+    """
+    Implements exponential moving average shadowing for your model.
+
+    Utilizes an inverse decay schedule to manage longer term training runs.
+    By adjusting the power, you can control how fast EMA will ramp up to your specified beta.
+
+    @crowsonkb's notes on EMA Warmup:
+    
+    If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are
+    good values for models you plan to train for a million or more steps (reaches decay
+    factor 0.999 at 31.6K steps, 0.9999 at 1M steps), gamma=1, power=3/4 for models
+    you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999 at
+    215.4k steps).
+    
+    Args:
+        inv_gamma (float): Inverse multiplicative factor of EMA warmup. Default: 1.
+        power (float): Exponential factor of EMA warmup. Default: 1.
+        min_value (float): The minimum EMA decay rate. Default: 0.
+    """
     def __init__(
         self,
         model,
-        beta = 0.99,
-        update_after_step = 1000,
+        beta = 0.9999,
+        update_after_step = 10000,
         update_every = 10,
+        inv_gamma = 1.0,
+        power = 2/3,
+        min_value = 0.0,
     ):
         super().__init__()
         self.beta = beta
@@ -190,6 +219,10 @@ class EMA(nn.Module):
         self.update_every = update_every
         self.update_after_step = update_after_step
 
+        self.inv_gamma = inv_gamma
+        self.power = power
+        self.min_value = min_value
+
         self.register_buffer('initted', torch.Tensor([False]))
         self.register_buffer('step', torch.tensor([0]))
 
@@ -198,7 +231,20 @@ class EMA(nn.Module):
         self.ema_model.to(device)
 
     def copy_params_from_model_to_ema(self):
-        self.ema_model.state_dict(self.online_model.state_dict())
+        for ma_param, current_param in zip(list(self.ema_model.parameters()), list(self.online_model.parameters())):
+            ma_param.data.copy_(current_param.data)
+
+        for ma_buffer, current_buffer in zip(list(self.ema_model.buffers()), list(self.online_model.buffers())):
+            ma_buffer.data.copy_(current_buffer.data)
+
+    def get_current_decay(self):
+        epoch = clamp(self.step.item() - self.update_after_step - 1, min_value = 0)
+        value = 1 - (1 + epoch / self.inv_gamma) ** - self.power
+
+        if epoch <= 0:
+            return 0.
+
+        return clamp(value, min_value = self.min_value, max_value = self.beta)
 
     def update(self):
         step = self.step.item()
@@ -217,19 +263,19 @@ class EMA(nn.Module):
 
         self.update_moving_average(self.ema_model, self.online_model)
 
+    @torch.no_grad()
     def update_moving_average(self, ma_model, current_model):
-        def calculate_ema(beta, old, new):
-            if not exists(old):
-                return new
-            return old * beta + (1 - beta) * new
+        current_decay = self.get_current_decay()
 
-        for current_params, ma_params in zip(current_model.parameters(), ma_model.parameters()):
-            old_weight, up_weight = ma_params.data, current_params.data
-            ma_params.data.copy_(calculate_ema(self.beta, old_weight, up_weight))
+        for current_params, ma_params in zip(list(current_model.parameters()), list(ma_model.parameters())):
+            difference = ma_params.data - current_params.data
+            difference.mul_(1.0 - current_decay)
+            ma_params.sub_(difference)
 
-        for current_buffer, ma_buffer in zip(current_model.buffers(), ma_model.buffers()):
-            new_buffer_value = calculate_ema(self.beta, ma_buffer, current_buffer)
-            ma_buffer.copy_(new_buffer_value)
+        for current_buffer, ma_buffer in zip(list(current_model.buffers()), list(ma_model.buffers())):
+            difference = ma_buffer - current_buffer
+            difference.mul_(1.0 - current_decay)
+            ma_buffer.sub_(difference)
 
     def __call__(self, *args, **kwargs):
         return self.ema_model(*args, **kwargs)
@@ -489,7 +535,7 @@ class DecoderTrainer(nn.Module):
         loaded_obj = torch.load(str(path))
 
         if version.parse(__version__) != loaded_obj['version']:
-            print(f'loading saved decoder at version {loaded_obj["version"]}, but current package version is {get_pkg_version()}')
+            print(f'loading saved decoder at version {loaded_obj["version"]}, but current package version is {__version__}')
 
         self.decoder.load_state_dict(loaded_obj['model'], strict = strict)
         self.step.copy_(torch.ones_like(self.step) * loaded_obj['step'])

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.6.10'
+__version__ = '0.7.0'

train_decoder.py

@@ -211,7 +211,7 @@ def recall_trainer(tracker, trainer, recall_source=None, **load_config):
     Loads the model with an appropriate method depending on the tracker
     """
     print(print_ribbon(f"Loading model from {recall_source}"))
-    state_dict = tracker.recall_state_dict(recall_source, **load_config)
+    state_dict = tracker.recall_state_dict(recall_source, **load_config.dict())
     trainer.load_state_dict(state_dict["trainer"])
     print("Model loaded")
     return state_dict["epoch"], state_dict["step"], state_dict["validation_losses"]