update the result of timit

examples/timit/s1/README.md

@@ -1,3 +1,11 @@
 # TIMIT
 
-Results will be organized and updated soon.
+
+
+
+### Transformer
+| Model | Params | Config | Decode method          | PER    |
+| --- | --- | --- | --- | --- |
+| transformer | 5.17M | conf/transformer.yaml | attention              | 0.5531 |
+| transformer | 5.17M | conf/transformer.yaml | ctc_greedy_search      | 0.3922 |
+| transformer | 5.17M | conf/transformer.yaml | ctc_prefix_beam_search | 0.3768 |

examples/timit/s1/conf/transformer.yaml

@@ -3,12 +3,12 @@ data:
   train_manifest: data/manifest.train
   dev_manifest: data/manifest.dev
   test_manifest: data/manifest.test
-  min_input_len: 0.5  # second
+  min_input_len: 0.0  # second
-  max_input_len: 30.0 # second
+  max_input_len: 10.0 # second
   min_output_len: 0.0 # tokens
-  max_output_len: 400.0 # tokens
+  max_output_len: 150.0 # tokens
-  min_output_input_ratio: 0.05
+  min_output_input_ratio: 0.005
-  max_output_input_ratio: 100.0
+  max_output_input_ratio: 1000.0
 
 collator:
   vocab_filepath: data/vocab.txt
@@ -42,10 +42,10 @@ model:
     # encoder related
     encoder: transformer
     encoder_conf:
-        output_size: 256    # dimension of attention
+        output_size: 128    # dimension of attention
         attention_heads: 4
-        linear_units: 2048  # the number of units of position-wise feed forward
+        linear_units: 1024  # the number of units of position-wise feed forward
-        num_blocks: 12      # the number of encoder blocks
+        num_blocks: 6      # the number of encoder blocks
         dropout_rate: 0.1
         positional_dropout_rate: 0.1
         attention_dropout_rate: 0.0
@@ -56,7 +56,7 @@ model:
     decoder: transformer
     decoder_conf:
         attention_heads: 4
-        linear_units: 2048
+        linear_units: 1024
         num_blocks: 6
         dropout_rate: 0.1
         positional_dropout_rate: 0.1
@@ -65,26 +65,26 @@ model:
 
     # hybrid CTC/attention
     model_conf:
-        ctc_weight: 0.3
+        ctc_weight: 0.5
         ctc_dropoutrate: 0.0
-        ctc_grad_norm_type: instance
+        ctc_grad_norm_type: batch
         lsm_weight: 0.1     # label smoothing option
         length_normalized_loss: false
 
 
 training:
-  n_epoch: 120
+  n_epoch: 200
   accum_grad: 2
   global_grad_clip: 5.0
   optim: adam
   optim_conf:
-    lr: 0.002
+    lr: 0.004
     weight_decay: 1e-06
   scheduler: warmuplr     # pytorch v1.1.0+ required
   scheduler_conf:
-    warmup_steps: 400
+    warmup_steps: 2000
     lr_decay: 1.0
-  log_interval: 100
+  log_interval: 10
   checkpoint:
     kbest_n: 50
     latest_n: 5

examples/timit/s1/local/test.sh

@@ -36,7 +36,8 @@ for type in attention ctc_greedy_search; do
     --config ${config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} decoding.batch_size ${batch_size}
+    --opts decoding.decoding_method ${type} \
+    --opts decoding.batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"
@@ -52,7 +53,8 @@ for type in ctc_prefix_beam_search attention_rescoring; do
     --config ${config_path} \
     --result_file ${ckpt_prefix}.${type}.rsl \
     --checkpoint_path ${ckpt_prefix} \
-    --opts decoding.decoding_method ${type} decoding.batch_size ${batch_size}
+    --opts decoding.decoding_method ${type} \
+    --opts decoding.batch_size ${batch_size}
 
     if [ $? -ne 0 ]; then
         echo "Failed in evaluation!"