enforce clip anytorch version

README.md

@@ -627,6 +627,18 @@ images = dalle2(
 # save your image (in this example, of size 256x256)
 ```
 
+Alternatively, you can also use <a href="https://github.com/mlfoundations/open_clip">Open Clip</a>
+
+```bash
+$ pip install open-clip-torch
+```
+
+```python
+from dalle2_pytorch import OpenClipAdapter
+
+clip = OpenClipAdapter()
+```
+
 Now you'll just have to worry about training the Prior and the Decoder!
 
 ## Inpainting

dalle2_pytorch/dalle2_pytorch.py

@@ -339,6 +339,75 @@ class OpenAIClipAdapter(BaseClipAdapter):
         image_embed = self.clip.encode_image(image)
         return EmbeddedImage(l2norm(image_embed.float()), None)
 
+class OpenClipAdapter(BaseClipAdapter):
+    def __init__(
+        self,
+        name = 'ViT-B/32',
+        pretrained = 'laion400m_e32'
+    ):
+        import open_clip
+        clip, _, preprocess = open_clip.create_model_and_transforms(name, pretrained = pretrained)
+
+        super().__init__(clip)
+        self.eos_id = 49407
+
+        text_attention_final = self.find_layer('ln_final')
+        self.handle = text_attention_final.register_forward_hook(self._hook)
+        self.clip_normalize = preprocess.transforms[-1]
+        self.cleared = False
+
+    def find_layer(self,  layer):
+        modules = dict([*self.clip.named_modules()])
+        return modules.get(layer, None)
+
+    def clear(self):
+        if self.cleared:
+            return
+
+        self.handle()
+
+    def _hook(self, _, inputs, outputs):
+        self.text_encodings = outputs
+
+    @property
+    def dim_latent(self):
+        return 512
+
+    @property
+    def image_size(self):
+        return self.clip.visual.image_size
+
+    @property
+    def image_channels(self):
+        return 3
+
+    @property
+    def max_text_len(self):
+        return self.clip.context_length
+
+    @torch.no_grad()
+    def embed_text(self, text):
+        text = text[..., :self.max_text_len]
+
+        is_eos_id = (text == self.eos_id)
+        text_mask_excluding_eos = is_eos_id.cumsum(dim = -1) == 0
+        text_mask = F.pad(text_mask_excluding_eos, (1, -1), value = True)
+        assert not self.cleared
+
+        text_embed = self.clip.encode_text(text)
+        text_encodings = self.text_encodings
+        text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
+        del self.text_encodings
+        return EmbeddedText(l2norm(text_embed.float()), text_encodings.float())
+
+    @torch.no_grad()
+    def embed_image(self, image):
+        assert not self.cleared
+        image = self.validate_and_resize_image(image)
+        image = self.clip_normalize(image)
+        image_embed = self.clip.encode_image(image)
+        return EmbeddedImage(l2norm(image_embed.float()), None)
+
 # classifier free guidance functions
 
 def prob_mask_like(shape, prob, device):
@@ -701,11 +770,12 @@ class Attention(nn.Module):
         dropout = 0.,
         causal = False,
         rotary_emb = None,
-        pb_relax_alpha = 128
+        cosine_sim = True,
+        cosine_sim_scale = 16
     ):
         super().__init__()
-        self.pb_relax_alpha = pb_relax_alpha
+        self.scale = cosine_sim_scale if cosine_sim else (dim_head ** -0.5)
-        self.scale = dim_head ** -0.5 * (pb_relax_alpha ** -1)
+        self.cosine_sim = cosine_sim
 
         self.heads = heads
         inner_dim = dim_head * heads
@@ -745,6 +815,13 @@ class Attention(nn.Module):
         k = torch.cat((nk, k), dim = -2)
         v = torch.cat((nv, v), dim = -2)
 
+        # whether to use cosine sim
+
+        if self.cosine_sim:
+            q, k = map(l2norm, (q, k))
+
+        q, k = map(lambda t: t * math.sqrt(self.scale), (q, k))
+
         # calculate query / key similarities
 
         sim = einsum('b h i d, b j d -> b h i j', q, k)
@@ -770,9 +847,6 @@ class Attention(nn.Module):
 
         # attention
 
-        sim = sim - sim.amax(dim = -1, keepdim = True).detach()
-        sim = sim * self.pb_relax_alpha
-
         attn = sim.softmax(dim = -1)
         attn = self.dropout(attn)
 
@@ -1485,7 +1559,8 @@ class LinearAttention(nn.Module):
         self,
         dim,
         dim_head = 32,
-        heads = 8
+        heads = 8,
+        **kwargs
     ):
         super().__init__()
         self.scale = dim_head ** -0.5
@@ -1604,6 +1679,7 @@ class Unet(nn.Module):
         lowres_noise_cond = False,       # for conditioning on low resolution noising, based on Imagen
         sparse_attn = False,
         cosine_sim_cross_attn = False,
+        cosine_sim_self_attn = False,
         attend_at_middle = True,         # whether to have a layer of attention at the bottleneck (can turn off for higher resolution in cascading DDPM, before bringing in efficient attention)
         cond_on_text_encodings = False,
         max_text_len = 256,
@@ -1724,7 +1800,7 @@ class Unet(nn.Module):
 
         # attention related params
 
-        attn_kwargs = dict(heads = attn_heads, dim_head = attn_dim_head)
+        attn_kwargs = dict(heads = attn_heads, dim_head = attn_dim_head, cosine_sim = cosine_sim_self_attn)
 
         self_attn = cast_tuple(self_attn, num_stages)
 

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.4.2'
+__version__ = '1.4.6'

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',