Commit
•
c09faa6
1
Parent(s):
6a89764
up
Browse files- create_model.py +29 -0
- run_flax_speech_recognition_seq2seq.py +1 -0
- run_librispeech.sh +34 -0
create_model.py
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import jax.numpy as jnp
|
2 |
+
from transformers import AutoFeatureExtractor, AutoTokenizer, FlaxSpeechEncoderDecoderModel
|
3 |
+
|
4 |
+
encoder_id = "facebook/wav2vec2-large-lv60"
|
5 |
+
decoder_id = "facebook/bart-large-cnn"
|
6 |
+
|
7 |
+
model = FlaxSpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True, decoder_from_pt=True)
|
8 |
+
|
9 |
+
model.config.encoder.feat_proj_dropout = 0.0
|
10 |
+
model.config.encoder.final_dropout = 0.0
|
11 |
+
model.config.encoder.mask_time_prob = 0.1
|
12 |
+
model.config.decoder_start_token_id = model.config.decoder.bos_token_id
|
13 |
+
model.config.pad_token_id = model.config.decoder.pad_token_id
|
14 |
+
model.config.eos_token_id = model.config.decoder.eos_token_id
|
15 |
+
model.config.max_length = 40
|
16 |
+
model.config.num_beams = 1
|
17 |
+
model.config.encoder.layerdrop = 0.0
|
18 |
+
model.config.use_cache = False
|
19 |
+
model.config.processor_class = "Wav2Vec2Processor"
|
20 |
+
|
21 |
+
# check if generation works
|
22 |
+
out = model.generate(jnp.ones((1, 2000)))
|
23 |
+
|
24 |
+
model.save_pretrained("./")
|
25 |
+
|
26 |
+
feature_extractor = AutoFeatureExtractor.from_pretrained(encoder_id)
|
27 |
+
feature_extractor.save_pretrained("./")
|
28 |
+
tokenizer = AutoTokenizer.from_pretrained(decoder_id)
|
29 |
+
tokenizer.save_pretrained("./")
|
run_flax_speech_recognition_seq2seq.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
/home/sanchitgandhi/transformers/examples/flax/speech-recognition/run_flax_speech_recognition_seq2seq.py
|
run_librispeech.sh
ADDED
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env bash
|
2 |
+
python run_flax_speech_recognition_seq2seq.py \
|
3 |
+
--dataset_name="librispeech_asr" \
|
4 |
+
--model_name_or_path="./" \
|
5 |
+
--dataset_config_name="clean" \
|
6 |
+
--train_split_name="train.100" \
|
7 |
+
--eval_split_name="validation" \
|
8 |
+
--dataset_cache_dir="~/cache/huggingface/datasets" \
|
9 |
+
--output_dir="./" \
|
10 |
+
--preprocessing_num_workers="16" \
|
11 |
+
--length_column_name="input_length" \
|
12 |
+
--overwrite_output_dir \
|
13 |
+
--num_train_epochs="15" \
|
14 |
+
--per_device_train_batch_size="2" \
|
15 |
+
--per_device_eval_batch_size="2" \
|
16 |
+
--gradient_accumulation_steps="1" \
|
17 |
+
--logging_steps="25" \
|
18 |
+
--max_duration_in_seconds="15" \
|
19 |
+
--max_target_length="64" \
|
20 |
+
--generation_max_length="40" \
|
21 |
+
--generation_num_beams="1" \
|
22 |
+
--learning_rate="3e-4" \
|
23 |
+
--warmup_steps="500" \
|
24 |
+
--text_column_name="text" \
|
25 |
+
--save_total_limit="1" \
|
26 |
+
--freeze_feature_encoder \
|
27 |
+
--predict_with_generate \
|
28 |
+
--do_lower_case \
|
29 |
+
--do_eval \
|
30 |
+
--do_train \
|
31 |
+
--push_to_hub \
|
32 |
+
--use_auth_token \
|
33 |
+
--wandb_project="flax-wav2vec2-2-bart-large-cnn"
|
34 |
+
|