control super condition factor

min_dalle/min_dalle.py

@@ -164,7 +164,8 @@ class MinDalle:
         text: str, 
         seed: int,
         grid_size: int,
-        log2_mid_count: int = 0,
+        log2_mid_count: int,
+        log2_supercondition_factor: int = 3,
         is_verbose: bool = False
     ) -> Iterator[Image.Image]:
         if is_verbose: print("tokenizing text")
@@ -200,6 +201,7 @@ class MinDalle:
                 print('sampling row {} of {}'.format(row_index + 1, row_count))
             attention_state, image_tokens = self.decoder.decode_row(
                 row_index,
+                log2_supercondition_factor,
                 encoder_state,
                 attention_mask,
                 attention_state,
@@ -216,6 +218,7 @@ class MinDalle:
         text: str,
         seed: int = -1,
         grid_size: int = 1,
+        log2_supercondition_factor: int = 3,
         is_verbose: bool = False
     ) -> Image.Image:
         log2_mid_count = 0
@@ -224,6 +227,7 @@ class MinDalle:
             seed,
             grid_size,
             log2_mid_count,
+            log2_supercondition_factor,
             is_verbose
         )
         return next(image_stream)

min_dalle/models/dalle_bart_decoder.py

@@ -116,7 +116,6 @@ class DalleBartDecoder(nn.Module):
         super().__init__()
         self.layer_count = layer_count
         self.embed_count = embed_count
-        self.condition_factor = 10.0
         self.embed_tokens = nn.Embedding(image_vocab_count + 1, embed_count)
         self.embed_positions = nn.Embedding(IMAGE_TOKEN_COUNT, embed_count)
         self.layers: List[DecoderLayer] = nn.ModuleList([
@@ -141,6 +140,7 @@ class DalleBartDecoder(nn.Module):
 
     def decode_step(
         self,
+        log2_supercondition_factor: int,
         attention_mask: BoolTensor,
         encoder_state: FloatTensor,
         attention_state: FloatTensor,
@@ -164,7 +164,7 @@ class DalleBartDecoder(nn.Module):
             )
         decoder_state = self.final_ln(decoder_state)
         logits = self.lm_head(decoder_state)
-        a = self.condition_factor
+        a = log2_supercondition_factor
         logits: FloatTensor = (
             logits[:image_count, -1] * (1 - a) + 
             logits[image_count:, -1] * a
@@ -182,6 +182,7 @@ class DalleBartDecoder(nn.Module):
     def decode_row(
         self,
         row_index: int,
+        log2_supercondition_factor: int,
         encoder_state: FloatTensor,
         attention_mask: BoolTensor,
         attention_state: FloatTensor,
@@ -190,6 +191,7 @@ class DalleBartDecoder(nn.Module):
         for col_index in range(16):
             i = 16 * row_index + col_index
             probs, attention_state = self.decode_step(
+                log2_supercondition_factor = log2_supercondition_factor,
                 attention_mask = attention_mask,
                 encoder_state = encoder_state,
                 attention_state = attention_state,