fix a potential issue in the low resolution conditioner, when downsampling and then upsampling using resize right, thanks to @marunine

README.md

@@ -44,6 +44,7 @@ This library would not have gotten to this working state without the help of
 - <a href="https://github.com/krish240574">Kumar</a> for working on the initial diffusion training script
 - <a href="https://github.com/rom1504">Romain</a> for the pull request reviews and project management
 - <a href="https://github.com/Ciaohe">He Cao</a> and <a href="https://github.com/xiankgx">xiankgx</a> for the Q&A and for identifying of critical bugs
+- <a href="https://github.com/marunine">Marunine</a> for identifying issues with resizing of the low resolution conditioner, when training the upsampler, in addition to various other bug fixes
 - <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
@@ -581,7 +582,8 @@ unet1 = Unet(
     image_embed_dim = 512,
     cond_dim = 128,
     channels = 3,
-    dim_mults=(1, 2, 4, 8)
+    dim_mults=(1, 2, 4, 8),
+    cond_on_text_encodings = True  # set to True for any unets that need to be conditioned on text encodings (ex. first unet in cascade)
 ).cuda()
 
 unet2 = Unet(
@@ -598,12 +600,11 @@ decoder = Decoder(
     clip = clip,
     timesteps = 100,
     image_cond_drop_prob = 0.1,
-    text_cond_drop_prob = 0.5,
-    condition_on_text_encodings = False  # set this to True if you wish to condition on text during training and sampling
+    text_cond_drop_prob = 0.5
 ).cuda()
 
 for unet_number in (1, 2):
-    loss = decoder(images, unet_number = unet_number) # this can optionally be decoder(images, text) if you wish to condition on the text encodings as well, though it was hinted in the paper it didn't do much
+    loss = decoder(images, text = text, unet_number = unet_number) # this can optionally be decoder(images, text) if you wish to condition on the text encodings as well, though it was hinted in the paper it didn't do much
     loss.backward()
 
 # do above for many steps

dalle2_pytorch/dalle2_pytorch.py

@@ -125,14 +125,19 @@ def log(t, eps = 1e-12):
 def l2norm(t):
     return F.normalize(t, dim = -1)
 
-def resize_image_to(image, target_image_size):
+def resize_image_to(image, target_image_size, clamp_range = None):
     orig_image_size = image.shape[-1]
 
     if orig_image_size == target_image_size:
         return image
 
     scale_factors = target_image_size / orig_image_size
-    return resize(image, scale_factors = scale_factors)
+    out = resize(image, scale_factors = scale_factors)
+
+    if exists(clamp_range):
+        out = out.clamp(*clamp_range)
+
+    return out
 
 # image normalization functions
 # ddpms expect images to be in the range of -1 to 1
@@ -1778,9 +1783,12 @@ class LowresConditioner(nn.Module):
         downsample_first = True,
         blur_sigma = 0.6,
         blur_kernel_size = 3,
+        input_image_range = None
     ):
         super().__init__()
         self.downsample_first = downsample_first
+        self.input_image_range = input_image_range
+
         self.blur_sigma = blur_sigma
         self.blur_kernel_size = blur_kernel_size
 
@@ -1794,7 +1802,7 @@ class LowresConditioner(nn.Module):
         blur_kernel_size = None
     ):
         if self.training and self.downsample_first and exists(downsample_image_size):
-            cond_fmap = resize_image_to(cond_fmap, downsample_image_size)
+            cond_fmap = resize_image_to(cond_fmap, downsample_image_size, clamp_range = self.input_image_range)
 
         if self.training:
             # when training, blur the low resolution conditional image
@@ -1814,7 +1822,7 @@ class LowresConditioner(nn.Module):
 
             cond_fmap = gaussian_blur2d(cond_fmap, cast_tuple(blur_kernel_size, 2), cast_tuple(blur_sigma, 2))
 
-        cond_fmap = resize_image_to(cond_fmap, target_image_size)
+        cond_fmap = resize_image_to(cond_fmap, target_image_size, clamp_range = self.input_image_range)
 
         return cond_fmap
 
@@ -1930,10 +1938,6 @@ class Decoder(nn.Module):
             self.unets.append(one_unet)
             self.vaes.append(one_vae.copy_for_eval())
 
-        # determine from unets whether conditioning on text encoding is needed
-
-        self.condition_on_text_encodings = any([unet.cond_on_text_encodings for unet in self.unets])
-
         # create noise schedulers per unet
 
         if not exists(beta_schedule):
@@ -1972,6 +1976,10 @@ class Decoder(nn.Module):
 
         self.predict_x_start = cast_tuple(predict_x_start, len(unets)) if not predict_x_start_for_latent_diffusion else tuple(map(lambda t: isinstance(t, VQGanVAE), self.vaes))
 
+        # input image range
+
+        self.input_image_range = (-1. if not auto_normalize_img else 0., 1.)
+
         # cascading ddpm related stuff
 
         lowres_conditions = tuple(map(lambda t: t.lowres_cond, self.unets))
@@ -1981,6 +1989,7 @@ class Decoder(nn.Module):
             downsample_first = lowres_downsample_first,
             blur_sigma = blur_sigma,
             blur_kernel_size = blur_kernel_size,
+            input_image_range = self.input_image_range
         )
 
         # classifier free guidance
@@ -2012,6 +2021,10 @@ class Decoder(nn.Module):
     def device(self):
         return self._dummy.device
 
+    @property
+    def condition_on_text_encodings(self):
+        return any([unet.cond_on_text_encodings for unet in self.unets])
+
     def get_unet(self, unet_number):
         assert 0 < unet_number <= len(self.unets)
         index = unet_number - 1

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.16.14'
+__version__ = '0.16.16'