more decoding method (#618)

* more decoding method

* all decode method test scripts; result readme

* exp libri confi

* parallel data scripts; more mask test; need pybind11 repo

* speed perturb config

* libri conf test set

.gitignore

@@ -1,7 +1,7 @@
 .DS_Store
 *.pyc
 .vscode
-*.log
+*log
 *.pdmodel
 *.pdiparams*
 *.zip

deepspeech/exps/deepspeech2/model.py

@@ -168,7 +168,7 @@ class DeepSpeech2Trainer(Trainer):
             train_dataset,
             batch_sampler=batch_sampler,
             collate_fn=collate_fn,
-            num_workers=config.data.num_workers, )
+            num_workers=config.data.num_workers)
         self.valid_loader = DataLoader(
             dev_dataset,
             batch_size=config.data.batch_size,

deepspeech/exps/u2/model.py

@@ -450,7 +450,7 @@ class U2Tester(U2Trainer):
         logger.info(msg)
 
         # test meta results
-        err_meta_path = os.path.splitext(self.args.checkpoint_path)[0] + '.err'
+        err_meta_path = os.path.splitext(self.args.result_file)[0] + '.err'
         err_type_str = "{}".format(error_rate_type)
         with open(err_meta_path, 'w') as f:
             data = json.dumps({
@@ -471,6 +471,8 @@ class U2Tester(U2Trainer):
                 errors_sum,
                 "ref_len":
                 len_refs,
+                "decode_method":
+                self.config.decoding.decoding_method,
             })
             f.write(data + '\n')
 

examples/aishell/s0/local/data.sh

@@ -66,19 +66,22 @@ fi
 if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     # format manifest with tokenids, vocab size
     for dataset in train dev test; do
+    {
         python3 ${MAIN_ROOT}/utils/format_data.py \
-        --feat_type "raw" \
-        --cmvn_path "data/mean_std.json" \
-        --unit_type "char" \
-        --vocab_path="data/vocab.txt" \
-        --manifest_path="data/manifest.${dataset}.raw" \
-        --output_path="data/manifest.${dataset}"
-    done
+                --feat_type "raw" \
+                --cmvn_path "data/mean_std.json" \
+                --unit_type "char" \
+                --vocab_path="data/vocab.txt" \
+                --manifest_path="data/manifest.${dataset}.raw" \
+                --output_path="data/manifest.${dataset}"
 
-    if [ $? -ne 0 ]; then
-        echo "Formt mnaifest failed. Terminated."
-        exit 1
-    fi
+        if [ $? -ne 0 ]; then
+                echo "Formt mnaifest failed. Terminated."
+                exit 1
+        fi
+    } &
+    done
+    wait
 fi
 
 echo "Aishell data preparation done."

examples/aishell/s1/README.md

@@ -0,0 +1,14 @@
+# Aishell
+
+## Conformer
+| Model | Config | Augmentation| Test set | Decode method | Loss | WER |
+| --- | --- | --- | --- | --- | --- |
+| conformer | conf/conformer.yaml | spec_aug + shift | test | attention | - | 0.059858 |
+| conformer | conf/conformer.yaml | spec_aug + shift | test | ctc_greedy_search | - | 0.062311 |
+| conformer | conf/conformer.yaml | spec_aug + shift | test | ctc_prefix_beam_search | - | 0.062196 |
+| conformer | conf/conformer.yaml | spec_aug + shift | test | attention_rescoring | - | 0.054694 |
+
+## Transformer
+| Model | Config | Augmentation| Test set | Decode method | Loss | WER |
+| --- | --- | --- | --- | --- | --- |
+| transformer | conf/transformer.yaml | spec_aug + shift | test | attention | - | - |

examples/aishell/s1/local/data.sh

@@ -14,7 +14,7 @@ if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
     python3 ${TARGET_DIR}/aishell/aishell.py \
     --manifest_prefix="data/manifest" \
     --target_dir="${TARGET_DIR}/aishell"
-    
+
     if [ $? -ne 0 ]; then
         echo "Prepare Aishell failed. Terminated."
         exit 1
@@ -33,7 +33,7 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     --count_threshold=0 \
     --vocab_path="data/vocab.txt" \
     --manifest_paths "data/manifest.train.raw"
-    
+
     if [ $? -ne 0 ]; then
         echo "Build vocabulary failed. Terminated."
         exit 1
@@ -56,7 +56,7 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
     --num_samples=-1 \
     --num_workers=${num_workers} \
     --output_path="data/mean_std.json"
-    
+
     if [ $? -ne 0 ]; then
         echo "Compute mean and stddev failed. Terminated."
         exit 1
@@ -67,19 +67,22 @@ fi
 if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
     # format manifest with tokenids, vocab size
     for dataset in train dev test; do
+    {
         python3 ${MAIN_ROOT}/utils/format_data.py \
-        --feat_type "raw" \
-        --cmvn_path "data/mean_std.json" \
-        --unit_type "char" \
-        --vocab_path="data/vocab.txt" \
-        --manifest_path="data/manifest.${dataset}.raw" \
-        --output_path="data/manifest.${dataset}"
+            --feat_type "raw" \
+            --cmvn_path "data/mean_std.json" \
+            --unit_type "char" \
+            --vocab_path="data/vocab.txt" \
+            --manifest_path="data/manifest.${dataset}.raw" \
+            --output_path="data/manifest.${dataset}"
+
+        if [ $? -ne 0 ]; then
+            echo "Formt mnaifest failed. Terminated."
+            exit 1
+        fi
+    } &
     done
-    
-    if [ $? -ne 0 ]; then
-        echo "Formt mnaifest failed. Terminated."
-        exit 1
-    fi
+    wait
 fi
 
 echo "Aishell data preparation done."

examples/aishell/s1/local/test.sh

@@ -21,17 +21,39 @@ ckpt_prefix=$2
 #    exit 1
 #fi
 
-python3 -u ${BIN_DIR}/test.py \
---device ${device} \
---nproc 1 \
---config ${config_path} \
---result_file ${ckpt_prefix}.rsl \
---checkpoint_path ${ckpt_prefix}
-
-if [ $? -ne 0 ]; then
-    echo "Failed in evaluation!"
-    exit 1
-fi
 
+for type in attention ctc_greedy_search; do
+    echo "decoding ${type}"
+    batch_size=64
+    python3 -u ${BIN_DIR}/test.py \
+    --device ${device} \
+    --nproc 1 \
+    --config ${config_path} \
+    --result_file ${ckpt_prefix}.${type}.rsl \
+    --checkpoint_path ${ckpt_prefix} \
+    --opts decoding.decoding_method ${type} decoding.batch_size ${batch_size}
+
+    if [ $? -ne 0 ]; then
+        echo "Failed in evaluation!"
+        exit 1
+    fi
+done
+
+for type in ctc_prefix_beam_search attention_rescoring; do
+    echo "decoding ${type}"
+    batch_size=1
+    python3 -u ${BIN_DIR}/test.py \
+    --device ${device} \
+    --nproc 1 \
+    --config ${config_path} \
+    --result_file ${ckpt_prefix}.${type}.rsl \
+    --checkpoint_path ${ckpt_prefix} \
+    --opts decoding.decoding_method ${type} decoding.batch_size ${batch_size}
+
+    if [ $? -ne 0 ]; then
+        echo "Failed in evaluation!"
+        exit 1
+    fi
+done
 
 exit 0

examples/librispeech/README.md

@@ -1,3 +1,3 @@
 # ASR
 * s0 is for deepspeech2
-* s1 is for U2
+* s1 is for transformer/conformer/U2

examples/librispeech/s1/README.md

@@ -0,0 +1,16 @@
+# LibriSpeech
+
+## Conformer
+| Model | Config | Augmentation| Test set | Decode method | Loss | WER |
+| --- | --- | --- | --- | --- | --- |
+| conformer | conf/conformer.yaml | spec_aug + shift | test-all | attention | test-all 6.35 | 0.057117 |
+| conformer | conf/conformer.yaml | spec_aug + shift | test-clean | attention | test-all 6.35 | 0.030162 |
+| conformer | conf/conformer.yaml | spec_aug + shift | test-clean | ctc_greedy_search | test-all 6.35 | 0.037910 |
+| conformer | conf/conformer.yaml | spec_aug + shift | test-clean | ctc_prefix_beam_search | test-all 6.35 | 0.037761 |
+| conformer | conf/conformer.yaml | spec_aug + shift | test-clean | attention_rescoring | test-all 6.35 | 0.032115 |
+
+## Transformer
+| Model | Config | Augmentation| Test set | Decode method | Loss | WER |
+| --- | --- | --- | --- | --- | --- |
+| transformer | conf/transformer.yaml | spec_aug + shift | test-all | attention | test-all 6.98 | 0.066500 |
+| transformer | conf/transformer.yaml | spec_aug + shift | test-clean | attention | test-all 6.98 | 0.036 |

examples/librispeech/s1/conf/conformer.yaml

@@ -2,7 +2,7 @@
 data:
   train_manifest: data/manifest.train
   dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
+  test_manifest: data/manifest.test-clean
   vocab_filepath: data/vocab.txt 
   unit_type: 'spm'
   spm_model_prefix: 'data/bpe_unigram_5000'
@@ -14,7 +14,7 @@ data:
   min_output_len: 0.0 # tokens
   max_output_len: 400.0 # tokens
   min_output_input_ratio: 0.05
-  max_output_input_ratio: 10.0
+  max_output_input_ratio: 10.0 
   raw_wav: True  # use raw_wav or kaldi feature
   specgram_type: fbank #linear, mfcc, fbank
   feat_dim: 80

examples/librispeech/s1/conf/transformer.yaml

@@ -2,7 +2,7 @@
 data:
   train_manifest: data/manifest.train
   dev_manifest: data/manifest.dev
-  test_manifest: data/manifest.test
+  test_manifest: data/manifest.test-clean
   vocab_filepath: data/vocab.txt
   unit_type: 'spm'
   spm_model_prefix: 'data/bpe_unigram_5000'

examples/librispeech/s1/local/test.sh

@@ -21,17 +21,39 @@ ckpt_prefix=$2
 #    exit 1
 #fi
 
-python3 -u ${BIN_DIR}/test.py \
---device ${device} \
---nproc 1 \
---config ${config_path} \
---result_file ${ckpt_prefix}.rsl \
---checkpoint_path ${ckpt_prefix}
-
-if [ $? -ne 0 ]; then
-    echo "Failed in evaluation!"
-    exit 1
-fi
+for type in attention ctc_greedy_search; do
+    echo "decoding ${type}"
+    batch_size=64
+    python3 -u ${BIN_DIR}/test.py \
+    --device ${device} \
+    --nproc 1 \
+    --config ${config_path} \
+    --result_file ${ckpt_prefix}.${type}.rsl \
+    --checkpoint_path ${ckpt_prefix} \
+    --opts decoding.decoding_method ${type} decoding.batch_size ${batch_size}
+
+    if [ $? -ne 0 ]; then
+        echo "Failed in evaluation!"
+        exit 1
+    fi
+done
+
+for type in ctc_prefix_beam_search attention_rescoring; do
+    echo "decoding ${type}"
+    batch_size=1
+    python3 -u ${BIN_DIR}/test.py \
+    --device ${device} \
+    --nproc 1 \
+    --config ${config_path} \
+    --result_file ${ckpt_prefix}.${type}.rsl \
+    --checkpoint_path ${ckpt_prefix} \
+    --opts decoding.decoding_method ${type} decoding.batch_size ${batch_size}
+
+    if [ $? -ne 0 ]; then
+        echo "Failed in evaluation!"
+        exit 1
+    fi
+done
 
 
 exit 0

requirements.txt

@@ -8,4 +8,5 @@ SoundFile==0.9.0.post1
 sox
 tensorboardX
 typeguard
-yacs
+yacs
+pybind11

tests/mask_test.py

@@ -50,7 +50,9 @@ class TestU2Model(unittest.TestCase):
 
     def test_make_pad_mask(self):
         res = make_pad_mask(self.lengths)
+        res1 = make_non_pad_mask(self.lengths).logical_not()
         self.assertSequenceEqual(res.numpy().tolist(), self.pad_masks.tolist())
+        self.assertSequenceEqual(res.numpy().tolist(), res1.tolist())
 
 
 if __name__ == '__main__':