Create create_model.py
Browse files- create_model.py +27 -0
create_model.py
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import SpeechEncoderDecoderModel, AutoFeatureExtractor, AutoTokenizer
|
2 |
+
import torch
|
3 |
+
|
4 |
+
|
5 |
+
encoder_id = "facebook/wav2vec2-xls-r-300m"
|
6 |
+
decoder_id = "facebook/bart-base"
|
7 |
+
|
8 |
+
model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True)
|
9 |
+
model.config.encoder.feat_proj_dropout = 0.0
|
10 |
+
model.config.encoder.mask_time_prob = 0.0
|
11 |
+
model.config.decoder_start_token_id = model.decoder.config.bos_token_id
|
12 |
+
model.config.pad_token_id = model.decoder.config.pad_token_id
|
13 |
+
model.config.eos_token_id = model.decoder.config.eos_token_id
|
14 |
+
model.config.max_length = 40
|
15 |
+
model.config.encoder.layerdrop = 0.0
|
16 |
+
model.config.use_cache = False
|
17 |
+
model.config.processor_class = "Wav2Vec2Processor"
|
18 |
+
|
19 |
+
# check if generation works
|
20 |
+
out = model.generate(torch.ones((1, 2000)))
|
21 |
+
|
22 |
+
model.save_pretrained("./")
|
23 |
+
|
24 |
+
feature_etxractor = AutoFeatureExtractor.from_pretrained(encoder_id)
|
25 |
+
feature_etxractor.save_pretrained("./")
|
26 |
+
tokenizer = AutoTokenizer.from_pretrained(decoder_id)
|
27 |
+
tokenizer.save_pretrained("./")
|