Commit
•
5685ae1
1
Parent(s):
88c729c
Update README.md
Browse files
README.md
CHANGED
|
@@ -7,19 +7,33 @@ tags:
|
|
| 7 |
- he
|
| 8 |
- generated_from_trainer
|
| 9 |
model-index:
|
| 10 |
-
- name: wav2vec2-xls-r-300m-
|
| 11 |
results: []
|
| 12 |
---
|
| 13 |
|
| 14 |
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
| 15 |
should probably proofread and complete it, then remove this comment. -->
|
| 16 |
|
| 17 |
-
# wav2vec2-xls-r-300m-
|
| 18 |
|
| 19 |
-
This model is a fine-tuned version of [
|
| 20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
- Loss: 0.4502
|
| 22 |
-
-
|
| 23 |
|
| 24 |
## Model description
|
| 25 |
|
|
@@ -37,6 +51,63 @@ More information needed
|
|
| 37 |
|
| 38 |
### Training hyperparameters
|
| 39 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
The following hyperparameters were used during training:
|
| 41 |
- learning_rate: 0.0003
|
| 42 |
- train_batch_size: 8
|
|
|
|
| 7 |
- he
|
| 8 |
- generated_from_trainer
|
| 9 |
model-index:
|
| 10 |
+
- name: wav2vec2-xls-r-300m-hebrew
|
| 11 |
results: []
|
| 12 |
---
|
| 13 |
|
| 14 |
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
| 15 |
should probably proofread and complete it, then remove this comment. -->
|
| 16 |
|
| 17 |
+
# wav2vec2-xls-r-300m-hebrew
|
| 18 |
|
| 19 |
+
This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the private datasets in 2 stages - firstly was fine-tuned on a small dataset with good samples and it achieves the following results on the evaluation set with the dataset:
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| split |size(gb) | n_samples | duration(hrs)| |
|
| 23 |
+
|---|---|---|---|---|
|
| 24 |
+
|train|4.19| 20306 | 28 | |
|
| 25 |
+
|dev |1.05| 5076 | 7 | |
|
| 26 |
+
|
| 27 |
+
- Loss: 0.5438
|
| 28 |
+
- Wer: 0.1773
|
| 29 |
+
|
| 30 |
+
Then the obtained model was fine-tuned on a large dataset with the small good dataset, with various samples from different sources, and with an unlabeled dataset that was weakly labeled using a previously trained model.
|
| 31 |
+
on a small dataset from previous step achieves
|
| 32 |
+
- WER: 0.1697
|
| 33 |
+
|
| 34 |
+
on a whole dataset
|
| 35 |
- Loss: 0.4502
|
| 36 |
+
- WER: 0.2318
|
| 37 |
|
| 38 |
## Model description
|
| 39 |
|
|
|
|
| 51 |
|
| 52 |
### Training hyperparameters
|
| 53 |
|
| 54 |
+
|
| 55 |
+
#### First training
|
| 56 |
+
|
| 57 |
+
The following hyperparameters were used during training:
|
| 58 |
+
- learning_rate: 0.0003
|
| 59 |
+
- train_batch_size: 8
|
| 60 |
+
- eval_batch_size: 8
|
| 61 |
+
- seed: 42
|
| 62 |
+
- distributed_type: multi-GPU
|
| 63 |
+
- num_devices: 2
|
| 64 |
+
- gradient_accumulation_steps: 4
|
| 65 |
+
- total_train_batch_size: 64
|
| 66 |
+
- total_eval_batch_size: 16
|
| 67 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
| 68 |
+
- lr_scheduler_type: linear
|
| 69 |
+
- lr_scheduler_warmup_steps: 1000
|
| 70 |
+
- num_epochs: 100.0
|
| 71 |
+
- mixed_precision_training: Native AMP
|
| 72 |
+
|
| 73 |
+
Training results
|
| 74 |
+
|
| 75 |
+
| Training Loss | Epoch | Step | Validation Loss | Wer |
|
| 76 |
+
|:-------------:|:-----:|:-----:|:---------------:|:------:|
|
| 77 |
+
| No log | 3.15 | 1000 | 0.5203 | 0.4333 |
|
| 78 |
+
| 1.4284 | 6.31 | 2000 | 0.4816 | 0.3951 |
|
| 79 |
+
| 1.4284 | 9.46 | 3000 | 0.4315 | 0.3546 |
|
| 80 |
+
| 1.283 | 12.62 | 4000 | 0.4278 | 0.3404 |
|
| 81 |
+
| 1.283 | 15.77 | 5000 | 0.4090 | 0.3054 |
|
| 82 |
+
| 1.1777 | 18.93 | 6000 | 0.3893 | 0.3006 |
|
| 83 |
+
| 1.1777 | 22.08 | 7000 | 0.3968 | 0.2857 |
|
| 84 |
+
| 1.0994 | 25.24 | 8000 | 0.3892 | 0.2751 |
|
| 85 |
+
| 1.0994 | 28.39 | 9000 | 0.4061 | 0.2690 |
|
| 86 |
+
| 1.0323 | 31.54 | 10000 | 0.4114 | 0.2507 |
|
| 87 |
+
| 1.0323 | 34.7 | 11000 | 0.4021 | 0.2508 |
|
| 88 |
+
| 0.9623 | 37.85 | 12000 | 0.4032 | 0.2378 |
|
| 89 |
+
| 0.9623 | 41.01 | 13000 | 0.4148 | 0.2374 |
|
| 90 |
+
| 0.9077 | 44.16 | 14000 | 0.4350 | 0.2323 |
|
| 91 |
+
| 0.9077 | 47.32 | 15000 | 0.4515 | 0.2246 |
|
| 92 |
+
| 0.8573 | 50.47 | 16000 | 0.4474 | 0.2180 |
|
| 93 |
+
| 0.8573 | 53.63 | 17000 | 0.4649 | 0.2171 |
|
| 94 |
+
| 0.8083 | 56.78 | 18000 | 0.4455 | 0.2102 |
|
| 95 |
+
| 0.8083 | 59.94 | 19000 | 0.4587 | 0.2092 |
|
| 96 |
+
| 0.769 | 63.09 | 20000 | 0.4794 | 0.2012 |
|
| 97 |
+
| 0.769 | 66.25 | 21000 | 0.4845 | 0.2007 |
|
| 98 |
+
| 0.7308 | 69.4 | 22000 | 0.4937 | 0.2008 |
|
| 99 |
+
| 0.7308 | 72.55 | 23000 | 0.4920 | 0.1895 |
|
| 100 |
+
| 0.6927 | 75.71 | 24000 | 0.5179 | 0.1911 |
|
| 101 |
+
| 0.6927 | 78.86 | 25000 | 0.5202 | 0.1877 |
|
| 102 |
+
| 0.6622 | 82.02 | 26000 | 0.5266 | 0.1840 |
|
| 103 |
+
| 0.6622 | 85.17 | 27000 | 0.5351 | 0.1854 |
|
| 104 |
+
| 0.6315 | 88.33 | 28000 | 0.5373 | 0.1811 |
|
| 105 |
+
| 0.6315 | 91.48 | 29000 | 0.5331 | 0.1792 |
|
| 106 |
+
| 0.6075 | 94.64 | 30000 | 0.5390 | 0.1779 |
|
| 107 |
+
| 0.6075 | 97.79 | 31000 | 0.5459 | 0.1773 |
|
| 108 |
+
|
| 109 |
+
#### Second training
|
| 110 |
+
|
| 111 |
The following hyperparameters were used during training:
|
| 112 |
- learning_rate: 0.0003
|
| 113 |
- train_batch_size: 8
|