complete exp on zipformer-L

activate-icefall.sh

@@ -1 +1 @@
-export PYTHONPATH=/var/data/share20/qc/k2/Github/icefall:$PYTHONPATH
+export PYTHONPATH=/root/k2/Github/icefall:$PYTHONPATH

egs/reazonspeech/ASR/decode_greedy.sh

@@ -1,15 +1,20 @@
-num_epochs=30
+num_epochs=40
-for ((i=$num_epochs; i>=15; i--));
+for ((i=$num_epochs; i>=1; i--));
 do
     for ((j=1; j<=$i; j++));
     do
         python3 ./zipformer/decode.py \
             --epoch $i \
             --avg $j \
-            --exp-dir zipformer/exp \
+            --exp-dir zipformer/exp-large \
-            --max-duration 450 \
+            --max-duration 600 \
-            --lang data/lang_char \
+            --causal 0 \
             --decoding-method greedy_search \
-	    --blank-penalty 2
+            --num-encoder-layers 2,2,4,5,4,2 \
+            --feedforward-dim 512,768,1536,2048,1536,768 \
+            --encoder-dim 192,256,512,768,512,256 \
+            --encoder-unmasked-dim 192,192,256,320,256,192 \
+            --lang data/lang_char \
+            --blank-penalty 0
     done
 done

egs/reazonspeech/ASR/decode_modified_beam.sh

@@ -1,15 +1,20 @@
-num_epochs=30
+num_epochs=60
-for ((i=$num_epochs; i>=20; i--));
+for ((i=$num_epochs; i>=40; i--));
 do
-    for avg in 12 11 10 9 8 7 6 5;
+    for ((j=1; j<=$i; j++));
     do
         python3 ./zipformer/decode.py \
             --epoch $i \
-            --avg $avg \
+            --avg $j \
-            --exp-dir zipformer/exp \
+            --exp-dir zipformer/exp-large \
-            --max-duration 450 \
+            --max-duration 600 \
-            --lang data/lang_char \
+            --causal 0 \
             --decoding-method modified_beam_search \
-	    --blank-penalty 2.5
+            --num-encoder-layers 2,2,4,5,4,2 \
+            --feedforward-dim 512,768,1536,2048,1536,768 \
+            --encoder-dim 192,256,512,768,512,256 \
+            --encoder-unmasked-dim 192,192,256,320,256,192 \
+            --lang data/lang_char \
+            --blank-penalty 0
     done
 done

egs/reazonspeech/ASR/zipformer/train.py

@@ -328,7 +328,7 @@ def get_parser():
     )
 
     parser.add_argument(
-        "--base-lr", type=float, default=0.035, help="The base learning rate."
+        "--base-lr", type=float, default=0.015, help="The base learning rate."
     )
 
     parser.add_argument(
@@ -1069,6 +1069,9 @@ def train_one_epoch(
                     tb_writer, "train/valid_", params.batch_idx_train
                 )
 
+    # print('--------------------debug------------------')
+    # print(tot_loss)
+    # print(tot_loss["frames"])
     loss_value = tot_loss["loss"] / tot_loss["frames"]
     params.train_loss = loss_value
     if params.train_loss < params.best_train_loss:
@@ -1179,7 +1182,7 @@ def run(rank, world_size, args):
         # You should use ../local/display_manifest_statistics.py to get
         # an utterance duration distribution for your dataset to select
         # the threshold
-        if c.duration < 0.3 or c.duration > 30.0:
+        if c.duration < 1.0 or c.duration > 30.0:
             # logging.warning(
             #     f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
             # )