update

egs/librispeech/ASR/decode.sh

@@ -0,0 +1,12 @@
+export CUDA_VISIBLE_DEVICES=2
+
+for epoch in {30..30}; do
+  for ((avg=1; avg<=$epoch-1; avg++)); do
+    ./zipformer_lstm/decode.py \
+      --epoch $epoch \
+      --avg $avg \
+      --exp-dir ./zipformer_lstm/exp_dropout0.2 \
+      --max-duration 2000 \
+      --decoding-method greedy_search
+  done
+done

egs/librispeech/ASR/decode_single.sh

@@ -0,0 +1,8 @@
+export CUDA_VISIBLE_DEVICES=$1
+
+./zipformer_lstm/decode.py \
+  --epoch $2 \
+  --avg $3 \
+  --exp-dir ./zipformer_lstm/exp \
+  --max-duration 2000 \
+  --decoding-method beam_search

egs/librispeech/ASR/sync.sh

@@ -0,0 +1,11 @@
+project=icefall-asr-librispeech-zipformer-2023-11-04
+run=4V10032G_lstm1_decoderdropout0.2_bpe500
+recipe=zipformer_lstm
+
+wandb sync ${recipe}/exp_dropout0.2/tensorboard/ --sync-tensorboard  -p $project  --id $run
+
+while true
+do
+  wandb sync ${recipe}/exp_dropout0.2/tensorboard/ --sync-tensorboard  -p $project  --id $run  --append
+  sleep 60
+done

egs/librispeech/ASR/zipformer_lstm/beam_search.py

@@ -36,7 +36,7 @@ def greedy_search(model: nn.Module, encoder_out: torch.Tensor) -> List[int]:
     # support only batch_size == 1 for now
     assert encoder_out.size(0) == 1, encoder_out.size(0)
     blank_id = model.decoder.blank_id
-    device = model.encoder_embed.device
+    device = next(model.parameters()).device
 
     sos = torch.tensor([blank_id], device=device, dtype=torch.int64).reshape(1, 1)
     decoder_out, (h, c) = model.decoder(sos)

egs/librispeech/ASR/zipformer_lstm/decoder.py

@@ -76,7 +76,7 @@ class Decoder(nn.Module):
 
         self.vocab_size = vocab_size
 
-        # self.embedding_dropout = nn.Dropout(embedding_dropout)
+        self.embedding_dropout = nn.Dropout(embedding_dropout)
 
         self.rnn = nn.LSTM(
             input_size=decoder_dim,
@@ -113,6 +113,8 @@ class Decoder(nn.Module):
         # at utterance start, we use negative ids in beam_search.py
         embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
 
+        embedding_out = self.embedding_dropout(embedding_out)
+
         embedding_out = self.balancer(embedding_out)
 
         rnn_out, (h, c) = self.rnn(embedding_out, states)