Inference is twice faster by calling CLIP just once and caching the r…

…esults

.gitignore

@@ -127,3 +127,5 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+
+save/

diffusion/gaussian_diffusion.py

@@ -637,6 +637,10 @@ def p_sample_loop(
         if dump_steps is not None:
             dump = []
 
+        if 'text' in model_kwargs['y'].keys():
+            # encoding once instead of each iteration saves lots of time
+            model_kwargs['y']['text_embed'] = model.encode_text(model_kwargs['y']['text'])
+
         for i, sample in enumerate(self.p_sample_loop_progressive(
             model,
             shape,

model/cfg_sampler.py

@@ -20,6 +20,7 @@ def __init__(self, model):
         self.nfeats = self.model.nfeats
         self.data_rep = self.model.data_rep
         self.cond_mode = self.model.cond_mode
+        self.encode_text = self.model.encode_text
 
     def forward(self, x, timesteps, y=None):
         cond_mode = self.model.cond_mode

model/mdm.py

@@ -148,7 +148,10 @@ def forward(self, x, timesteps, y=None):
 
         force_mask = y.get('uncond', False)
         if 'text' in self.cond_mode:
-            enc_text = self.encode_text(y['text'])
+            if 'text_embed' in y.keys():  # caching option
+                enc_text = y['text_embed']
+            else:
+                enc_text = self.encode_text(y['text'])
             emb += self.embed_text(self.mask_cond(enc_text, force_mask=force_mask))
         if 'action' in self.cond_mode:
             action_emb = self.embed_action(y['action'])