fix format of coding

Author: a2d8a4v

python/ctranslate2/converters/transformers.py

@@ -148,7 +148,7 @@ def _load(self):
                 tokenizer = self.load_tokenizer(
                     tokenizer_class, self._model_name_or_path, **tokenizer_kwargs
                 )
-            except:
+            except Exception:
                 tokenizer = None
                 print("Escape tokenizer, which does not exist.")
 
@@ -1081,7 +1081,6 @@ def set_common_layers(self, spec, module):
         self.set_layer_norm(spec.layer_norm, module.layer_norm)
 
 
-
 @register_loader("WavLMConfig")
 class WavLMLoader(BartLoader):
     @property
@@ -1119,7 +1118,7 @@ def set_vocabulary(self, spec, tokens):
 
     def set_feature_extractor(self, spec, feature_extractor):
         spec.feat_layer0.conv.weight = feature_extractor.conv_layers[0].conv.weight
-        # spec.feat_layer0.conv.bias = feature_extractor.conv_layers[0].conv.bias // wavlm has no bias
+        # wavlm has no bias in conv
         self.set_layer_norm(
             spec.feat_layer0.layer_norm, feature_extractor.conv_layers[0].layer_norm
         )
@@ -1161,7 +1160,7 @@ def set_wavlm_encoder_layer(self, spec, encoder):
                 layer_spec.self_attention,
                 layer.self_attn,
                 self_attention=True,
-                has_rel_attn_embed=(layer_index==0),
+                has_rel_attn_embed=(layer_index == 0),
             )
             self.set_layer_norm(
                 layer_spec.self_attention.layer_norm,
@@ -1187,7 +1186,8 @@ def set_attention(self, spec, attention, self_attention=False, has_rel_attn_embe
         self.set_linear(spec.linear[-1], attention.out_proj)
 
         self.set_linear(spec.gru_relative_position_linear, attention.gru_rel_pos_linear)
-        spec.gru_relative_position_const = attention.gru_rel_pos_const.data # is torch.nn.parameter.Parameter
+        # which is torch.nn.parameter.Parameter
+        spec.gru_relative_position_const = attention.gru_rel_pos_const.data
 
         if has_rel_attn_embed:
             spec.relative_attention_bias = attention.rel_attn_embed.weight

python/ctranslate2/specs/wavlm_spec.py

@@ -65,8 +65,9 @@ def __init__(self, feat_layers, num_layers, num_heads, return_hidden):
         self.pos_conv_embed = WavLMPosEmbedConvLayer()
         self.layer_norm = common_spec.LayerNormSpec()
         self.layer = [
-            transformer_spec.TransformerEncoderLayerSpec(gated_relative_attention_bias=True,
-                                                         relative_attention_bias=(i == 0)) for i in range(num_layers)
+            transformer_spec.TransformerEncoderLayerSpec(
+                gated_relative_attention_bias=True,
+                relative_attention_bias=(i == 0)) for i in range(num_layers)
         ]
         # if not return_hidden:
         #     self.lm_head = common_spec.LinearSpec()

python/tests/test_transformers.py

@@ -1116,7 +1116,10 @@ def test_transformers_wavlm(
 
         hg_output = hf_model(input_values.unsqueeze(0))
 
-        similarity = torch.nn.functional.cosine_similarity(last_hidden_state, hg_output.last_hidden_state.flatten(0, -1), dim=0)
+        similarity = torch.nn.functional.cosine_similarity(
+            last_hidden_state,
+            hg_output.last_hidden_state.flatten(0, -1),
+            dim=0)
 
         assert similarity == 1.0