cleanup

nuwa_pytorch/nuwa_pytorch.py

@@ -569,6 +569,19 @@ def __init__(
 
         self.to_logits = nn.Linear(dim, num_image_tokens)
 
+    def embed_text(self, text, mask = None):
+        batch, seq_len, device = *text.shape, text.device
+        assert seq_len <= self.text_max_seq_len, 'your input text has a greater length than what was designated on initialization'
+
+        tokens = self.text_embedding(text)
+        pos_emb = self.text_pos_embedding(torch.arange(seq_len, device = device))
+        tokens = tokens + rearrange(pos_emb, 'n d -> 1 n d')
+
+        return self.text_transformer(
+            tokens,
+            mask = mask
+        )
+
     @torch.no_grad()
     @eval_decorator
     def generate(
@@ -580,16 +593,7 @@ def generate(
         temperature = 1.
     ):
         batch, seq_len, device = *text.shape, text.device
-        assert seq_len <= self.text_max_seq_len, 'your input text has a greater length than what was designated on initialization'
-
-        tokens = self.text_embedding(text)
-        pos_emb = self.text_pos_embedding(torch.arange(seq_len, device = device))
-        tokens = tokens + rearrange(pos_emb, 'n d -> 1 n d')
-
-        text_embeds = self.text_transformer(
-            tokens,
-            mask = text_mask
-        )
+        text_embeds = self.embed_text(text, mask = text_mask)
 
         bos = repeat(self.video_bos, 'd -> b 1 d', b = batch)
 
@@ -632,16 +636,7 @@ def forward(
         return_loss = False
     ):
         batch, seq_len, device = *text.shape, text.device
-        assert seq_len <= self.text_max_seq_len, 'your input text has a greater length than what was designated on initialization'
-
-        tokens = self.text_embedding(text)
-        pos_emb = self.text_pos_embedding(torch.arange(seq_len, device = device))
-        tokens = tokens + rearrange(pos_emb, 'n d -> 1 n d')
-
-        text_embeds = self.text_transformer(
-            tokens,
-            mask = text_mask
-        )
+        text_embeds = self.embed_text(text, mask = text_mask)
 
         frame_indices = self.vae.get_video_indices(video)
         frame_indices = rearrange(frame_indices, 'b ... -> b (...)')