From 7961b6bf23b3d603343608995dcb6e0e070e867d Mon Sep 17 00:00:00 2001
From: yfyeung <yifanyeung@sjtu.edu.cn>
Date: Wed, 10 Jul 2024 00:16:24 -0700
Subject: [PATCH] update

---
 egs/librispeech/ASR/decode.sh                     | 12 ++++++++++++
 egs/librispeech/ASR/decode_single.sh              |  8 ++++++++
 egs/librispeech/ASR/sync.sh                       | 11 +++++++++++
 egs/librispeech/ASR/zipformer_lstm/beam_search.py |  2 +-
 egs/librispeech/ASR/zipformer_lstm/decoder.py     |  4 +++-
 5 files changed, 35 insertions(+), 2 deletions(-)
 create mode 100755 egs/librispeech/ASR/decode.sh
 create mode 100755 egs/librispeech/ASR/decode_single.sh
 create mode 100755 egs/librispeech/ASR/sync.sh

diff --git a/egs/librispeech/ASR/decode.sh b/egs/librispeech/ASR/decode.sh
new file mode 100755
index 000000000..e67c0188d
--- /dev/null
+++ b/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
diff --git a/egs/librispeech/ASR/decode_single.sh b/egs/librispeech/ASR/decode_single.sh
new file mode 100755
index 000000000..9a340c3ff
--- /dev/null
+++ b/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
diff --git a/egs/librispeech/ASR/sync.sh b/egs/librispeech/ASR/sync.sh
new file mode 100755
index 000000000..63f946ebf
--- /dev/null
+++ b/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
diff --git a/egs/librispeech/ASR/zipformer_lstm/beam_search.py b/egs/librispeech/ASR/zipformer_lstm/beam_search.py
index ae896274c..2c1e6781a 100644
--- a/egs/librispeech/ASR/zipformer_lstm/beam_search.py
+++ b/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)
diff --git a/egs/librispeech/ASR/zipformer_lstm/decoder.py b/egs/librispeech/ASR/zipformer_lstm/decoder.py
index 51c1952b9..c603e84ee 100644
--- a/egs/librispeech/ASR/zipformer_lstm/decoder.py
+++ b/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)