misc minor fixes

facebookresearch · Apr 2, 2021 · fc34a97 · fc34a97
1 parent 0b8eb15
commit fc34a97
Show file tree

Hide file tree

Showing 3 changed files with 6 additions and 5 deletions.
diff --git a/conf/encoder/speech_mixed.yaml b/conf/encoder/speech_mixed.yaml
@@ -33,11 +33,11 @@ ctx_sequence_length: 256 # Max length of the encoder input sequence
 ctx_dropout: 0.1
 ctx_pretrained: True  # if False, the model won't load pre-trained BERT weights
 
+# whether to fix (don't update) context encoder during training or not
+fix_ctx_encoder: False
 
 # -------------- COMMON -------------------
 
-# whether to fix (don't update) context encoder during training or not
-fix_ctx_encoder: False
 
 #TODO: move to train config group?
 optimizer: hf-adam  # fairseq-adam

diff --git a/dpr/models/hf_models.py b/dpr/models/hf_models.py
@@ -28,7 +28,8 @@
     from transformers.optimization import AdamW
     from transformers.tokenization_bert import BertTokenizer
     from transformers.tokenization_roberta import RobertaTokenizer
-    from transformers import Wav2Vec2Model, Wav2Vec2Config  # will fail
+
+    # from transformers import Wav2Vec2Model, Wav2Vec2Config  # will fail
 
 from dpr.models.biencoder import BiEncoder
 from dpr.utils.data_utils import Tensorizer
@@ -205,7 +206,7 @@ def get_wav2vec_encoder(
     output_layer: int,
 ):
     encoder = Wav2Vec2HFEncoder.init_encoder(
-        pretrained_model, max_audio_t, extra_proj_d, final_drop, use_activation, output_layer
+        pretrained_model, max_audio_t, extra_proj_d, final_drop, use_activation, output_layer=output_layer
     )
     return encoder
 

diff --git a/dpr/models/mixed_models.py b/dpr/models/mixed_models.py
@@ -48,7 +48,7 @@ def get_audio_mixed_biencoder_components(cfg, inference_only: bool = False, **kw
             groups = get_hf_model_param_grouping(biencoder.ctx_model, weight_decay=cfg.train.weight_decay)
             q_groups = get_hf_model_param_grouping(biencoder.question_model, weight_decay=cfg.train.weight_decay)
             for g in q_groups:
-                g["lr"] = lr * cfg.encoder.audio_encoder_lr_factor
+                g["lr"] = lr * cfg.encoder.q_audio_encoder_lr_factor
                 logger.info("Setting lr=%s for wav2vec encoder param group", g["lr"])
                 groups.append(g)
         else: