Update README.md
Browse files
README.md
CHANGED
@@ -2,17 +2,58 @@
|
|
2 |
library_name: keras
|
3 |
---
|
4 |
|
|
|
|
|
5 |
## Model description
|
6 |
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
-
|
|
|
|
|
10 |
|
11 |
-
More information needed
|
12 |
|
13 |
-
|
14 |
|
15 |
-
|
16 |
|
17 |
## Training procedure
|
18 |
|
|
|
2 |
library_name: keras
|
3 |
---
|
4 |
|
5 |
+
x100 smaller with less than 0.5 accuracy drop vs. distilbert-base-uncased-finetuned-sst-2-english
|
6 |
+
|
7 |
## Model description
|
8 |
|
9 |
+
2 Layers Bilstm model finetuned on SST-2 and distlled from RoBERTa teacher
|
10 |
+
|
11 |
+
distilbert-base-uncased-finetuned-sst-2-english: 92.2 accuracy, 67M parameters
|
12 |
+
moshew/distilbilstm-finetuned-sst-2-english: 91.9 accuracy, 66K parameters
|
13 |
+
|
14 |
+
## How to get started with the model
|
15 |
+
|
16 |
+
Example on SST-2 test dataset classification:
|
17 |
+
|
18 |
+
```python
|
19 |
+
from datasets import load_dataset
|
20 |
+
import numpy as np
|
21 |
+
from sklearn.metrics import accuracy_score
|
22 |
+
|
23 |
+
from keras.preprocessing.text import Tokenizer
|
24 |
+
from keras.utils import pad_sequences
|
25 |
+
import tensorflow as tf
|
26 |
+
|
27 |
+
from huggingface_hub import from_pretrained_keras
|
28 |
+
|
29 |
+
from datasets import load_dataset
|
30 |
+
sst2 = load_dataset("SetFit/sst2")
|
31 |
+
augmented_sst2_dataset = load_dataset("jmamou/augmented-glue-sst2")
|
32 |
+
|
33 |
+
oov_token = '<UNK>' # Required only if test is not given
|
34 |
+
pad_type = 'post'
|
35 |
+
trunc_type = 'post'
|
36 |
+
|
37 |
+
# Tokenize our training data
|
38 |
+
tokenizer = Tokenizer(num_words=10000)
|
39 |
+
tokenizer.fit_on_texts(augmented_sst2_dataset['train']['sentence'])
|
40 |
+
|
41 |
+
# Encode training data sentences into sequences
|
42 |
+
test_sequences = tokenizer.texts_to_sequences(sst2['test']['text'])
|
43 |
+
|
44 |
+
# Pad the training sequences
|
45 |
+
test_padded = pad_sequences(test_sequences, padding=pad_type, truncating=trunc_type, maxlen=64)
|
46 |
+
|
47 |
+
reloaded_model = from_pretrained_keras('moshew/distilbilstm-finetuned-sst-2-english')
|
48 |
|
49 |
+
pred=reloaded_model.predict(test_padded)
|
50 |
+
pred_bin = np.argmax(pred,1)
|
51 |
+
accuracy_score(pred_bin, sst2['test']['label'])
|
52 |
|
|
|
53 |
|
54 |
+
```
|
55 |
|
56 |
+
0.9187259747391543
|
57 |
|
58 |
## Training procedure
|
59 |
|