moshew/distilbilstm-finetuned-sst-2-english

x100 smaller with less than 0.5 accuracy drop vs. distilbert-base-uncased-finetuned-sst-2-english

Model description

2 Layers Bilstm model finetuned on SST-2 and distlled from RoBERTa teacher

distilbert-base-uncased-finetuned-sst-2-english: 92.2 accuracy, 67M parameters
moshew/distilbilstm-finetuned-sst-2-english: 91.9 accuracy, 0.66M parameters

How to get started with the model

Example on SST-2 test dataset classification: ​​

!pip install datasets
from datasets import load_dataset  
import numpy as np  
from sklearn.metrics import accuracy_score  
from keras.preprocessing.text import Tokenizer
from keras.utils import pad_sequences
import tensorflow as tf
from huggingface_hub import from_pretrained_keras

from datasets import load_dataset
sst2 = load_dataset("SetFit/sst2")
augmented_sst2_dataset = load_dataset("jmamou/augmented-glue-sst2")

# Tokenize our training data
tokenizer = Tokenizer(num_words=10000)
tokenizer.fit_on_texts(augmented_sst2_dataset['train']['sentence'])

# Encode test data sentences into sequences
test_sequences = tokenizer.texts_to_sequences(sst2['test']['text'])

# Pad the test sequences
test_padded = pad_sequences(test_sequences, padding = 'post', truncating = 'post', maxlen=64)

reloaded_model = from_pretrained_keras('moshew/distilbilstm-finetuned-sst-2-english')

#Evaluate model on SST2 test data (GLUE)
pred=reloaded_model.predict(test_padded)
pred_bin = np.argmax(pred,1)
accuracy_score(pred_bin, sst2['test']['label'])

0.9187259747391543

reloaded_model.summary()

Model: "model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_1 (InputLayer)        [(None, 64)]              0         
                                                                 
 embedding (Embedding)       (None, 64, 50)            500000    
                                                                 
 bidirectional (Bidirectiona  (None, 64, 128)          58880     
 l)                                                              
                                                                 
 bidirectional_1 (Bidirectio  (None, 128)              98816     
 nal)                                                            
                                                                 
 dropout (Dropout)           (None, 128)               0         
                                                                 
 dense (Dense)               (None, 2)                 258       
                                                                 
=================================================================
Total params: 657,954
Trainable params: 657,954
Non-trainable params: 0
_________________________________________________________________

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

Hyperparameters	Value
name	Adam
learning_rate	0.0010000000474974513
decay	0.0
beta_1	0.8999999761581421
beta_2	0.9990000128746033
epsilon	1e-07
amsgrad	False
training_precision	float32

Model Plot

View Model Plot