Fix padding_idx (#942)

* fix padding_idx

* update RESULTS.md

egs/librispeech/ASR/RESULTS.md

@@ -540,6 +540,10 @@ for m in greedy_search fast_beam_search modified_beam_search ; do
 done
 ```
 
+Note that a small change is made to the `pruned_transducer_stateless7/decoder.py` in 
+this [PR](/ceph-data4/yangxiaoyu/softwares/icefall_development/icefall_random_padding/egs/librispeech/ASR/pruned_transducer_stateless7/exp_960h_no_paddingidx_ngpu4/tensorboard) to address the 
+problem of emitting the first symbol at the very beginning. If you need a 
+model without this issue, please download the model from here: <https://huggingface.co/marcoyang/icefall-asr-librispeech-pruned-transducer-stateless7-2023-03-10>
 
 ### LibriSpeech BPE training results (Pruned Stateless LSTM RNN-T + gradient filter)
 

egs/librispeech/ASR/pruned_transducer_stateless/decoder.py

@@ -58,7 +58,6 @@ class Decoder(nn.Module):
         self.embedding = nn.Embedding(
             num_embeddings=vocab_size,
             embedding_dim=embedding_dim,
-            padding_idx=blank_id,
         )
         self.blank_id = blank_id
         self.unk_id = unk_id

egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py

@@ -59,7 +59,6 @@ class Decoder(nn.Module):
         self.embedding = ScaledEmbedding(
             num_embeddings=vocab_size,
             embedding_dim=decoder_dim,
-            padding_idx=blank_id,
         )
         self.blank_id = blank_id
 

egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py

@@ -56,7 +56,6 @@ class Decoder(nn.Module):
         self.embedding = nn.Embedding(
             num_embeddings=vocab_size,
             embedding_dim=decoder_dim,
-            padding_idx=blank_id,
         )
         self.blank_id = blank_id