Commit
•
fb1fe48
1
Parent(s):
a5716eb
Update README.md
Browse files
README.md
CHANGED
@@ -87,7 +87,7 @@ processor = Wav2Vec2Processor.from_pretrained("gchhablani/wav2vec2-large-xlsr-pt
|
|
87 |
model = Wav2Vec2ForCTC.from_pretrained("gchhablani/wav2vec2-large-xlsr-pt")
|
88 |
model.to("cuda")
|
89 |
|
90 |
-
chars_to_ignore_regex = '[
|
91 |
resampler = torchaudio.transforms.Resample(48_000, 16_000)
|
92 |
|
93 |
# Preprocessing the datasets.
|
@@ -126,29 +126,29 @@ The Common Voice `train` and `validation` datasets were used for training. The s
|
|
126 |
|
127 |
```bash
|
128 |
#!/usr/bin/env bash
|
129 |
-
python run_common_voice.py
|
130 |
-
--model_name_or_path="facebook/wav2vec2-large-xlsr-53"
|
131 |
-
--dataset_config_name="pt"
|
132 |
-
--output_dir=/workspace/output_models/pt/wav2vec2-large-xlsr-pt
|
133 |
-
--cache_dir=/workspace/data
|
134 |
-
--overwrite_output_dir
|
135 |
-
--num_train_epochs="30"
|
136 |
-
--per_device_train_batch_size="32"
|
137 |
-
--per_device_eval_batch_size="32"
|
138 |
-
--evaluation_strategy="steps"
|
139 |
-
--learning_rate="3e-4"
|
140 |
-
--warmup_steps="500"
|
141 |
-
--fp16
|
142 |
-
--freeze_feature_extractor
|
143 |
-
--save_steps="500"
|
144 |
-
--eval_steps="500"
|
145 |
-
--save_total_limit="1"
|
146 |
-
--logging_steps="500"
|
147 |
-
--group_by_length
|
148 |
-
--feat_proj_dropout="0.0"
|
149 |
-
--layerdrop="0.1"
|
150 |
-
--gradient_checkpointing
|
151 |
-
--do_train --do_eval
|
152 |
```
|
153 |
|
154 |
Notebook containing the evaluation can be found [here](https://colab.research.google.com/drive/14e-zNK_5pm8EMY9EbeZerpHx7WsGycqG?usp=sharing).
|
|
|
87 |
model = Wav2Vec2ForCTC.from_pretrained("gchhablani/wav2vec2-large-xlsr-pt")
|
88 |
model.to("cuda")
|
89 |
|
90 |
+
chars_to_ignore_regex = '[\,\?\.\!\-\;\;\"\“\'\�]'
|
91 |
resampler = torchaudio.transforms.Resample(48_000, 16_000)
|
92 |
|
93 |
# Preprocessing the datasets.
|
|
|
126 |
|
127 |
```bash
|
128 |
#!/usr/bin/env bash
|
129 |
+
python run_common_voice.py \
|
130 |
+
--model_name_or_path="facebook/wav2vec2-large-xlsr-53" \
|
131 |
+
--dataset_config_name="pt" \
|
132 |
+
--output_dir=/workspace/output_models/pt/wav2vec2-large-xlsr-pt \
|
133 |
+
--cache_dir=/workspace/data \
|
134 |
+
--overwrite_output_dir \
|
135 |
+
--num_train_epochs="30" \
|
136 |
+
--per_device_train_batch_size="32" \
|
137 |
+
--per_device_eval_batch_size="32" \
|
138 |
+
--evaluation_strategy="steps" \
|
139 |
+
--learning_rate="3e-4" \
|
140 |
+
--warmup_steps="500" \
|
141 |
+
--fp16 \
|
142 |
+
--freeze_feature_extractor \
|
143 |
+
--save_steps="500" \
|
144 |
+
--eval_steps="500" \
|
145 |
+
--save_total_limit="1" \
|
146 |
+
--logging_steps="500" \
|
147 |
+
--group_by_length \
|
148 |
+
--feat_proj_dropout="0.0" \
|
149 |
+
--layerdrop="0.1" \
|
150 |
+
--gradient_checkpointing \
|
151 |
+
--do_train --do_eval \
|
152 |
```
|
153 |
|
154 |
Notebook containing the evaluation can be found [here](https://colab.research.google.com/drive/14e-zNK_5pm8EMY9EbeZerpHx7WsGycqG?usp=sharing).
|