Load the Pretrained Model and the dataset¶

We use bert-base-uncased as the model and SST-2 as the dataset for example. More models can be found in PaddleNLP Model Zoo.

Obviously, PaddleNLP is needed to run this notebook, which is easy to install:

pip install setuptools_scm 
pip install --upgrade paddlenlp

In [3]:

import paddle
import paddlenlp
from paddlenlp.transformers import BertForSequenceClassification, BertTokenizer

MODEL_NAME = "bert-base-uncased"
model = BertForSequenceClassification.from_pretrained(MODEL_NAME, num_classes=2)
tokenizer = BertTokenizer.from_pretrained(MODEL_NAME)

from paddlenlp.datasets import load_dataset
train_ds, dev_ds, test_ds = load_dataset(
    "glue", name='sst-2', splits=["train", "dev", "test"]
)

[2021-11-04 16:50:48,431] [    INFO] - Already cached /root/.paddlenlp/models/bert-base-uncased/bert-base-uncased.pdparams
W1104 16:50:48.433992 22865 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 10.2
W1104 16:50:48.439213 22865 device_context.cc:465] device: 0, cuDNN Version: 7.6.
[2021-11-04 16:50:58,691] [    INFO] - Already cached /root/.paddlenlp/models/bert-base-uncased/bert-base-uncased-vocab.txt
INFO:paddle.utils.download:unique_endpoints {'10.255.126.17:35174'}

Prepare the Model¶

Train the model¶

In [ ]:

# training the model and save to save_dir
# only needs to run once.
# total steps ~2100 (1 epoch)

from assets.utils import training_model
training_model(model, tokenizer, train_ds, dev_ds, save_dir=f'assets/sst-2-{MODEL_NAME}')

# global step 2100, epoch: 1, batch: 2100, loss: 0.22977, acc: 0.91710
# eval loss: 0.20062, accu: 0.91972

Or Load the trained model¶

In [4]:

# Load the trained model.
state_dict = paddle.load(f'assets/sst-2-{MODEL_NAME}/model_state.pdparams')
model.set_dict(state_dict)

See the prediction results¶

In [5]:

from assets.utils import predict

reviews = [
    "it 's a charming and often affecting journey . ",
    'the movie achieves as great an impact by keeping these thoughts hidden as ... ( quills ) did by showing them . ',
    'this one is definitely one to skip , even for horror movie fanatics . ',
    'in its best moments , resembles a bad high school production of grease , without benefit of song . '
]

data = [ {"text": r} for r in reviews]

label_map = {0: 'negative', 1: 'positive'}
batch_size = 32

results = predict(
    model, data, tokenizer, label_map, batch_size=batch_size)

for idx, text in enumerate(data):
    print('Data: {} \t Lable: {}'.format(text, results[idx]))

Data: {'text': "it 's a charming and often affecting journey . "} 	 Lable: positive
Data: {'text': 'the movie achieves as great an impact by keeping these thoughts hidden as ... ( quills ) did by showing them . '} 	 Lable: positive
Data: {'text': 'this one is definitely one to skip , even for horror movie fanatics . '} 	 Lable: positive
Data: {'text': 'in its best moments , resembles a bad high school production of grease , without benefit of song . '} 	 Lable: negative

Prepare for Interpretations¶

In [ ]:

import interpretdl as it
import numpy as np
from assets.utils import convert_example, aggregate_subwords_and_importances
from paddlenlp.data import Stack, Tuple, Pad
from interpretdl.data_processor.visualizer import VisualizationTextRecord, visualize_text

def preprocess_fn(data):
    examples = []
    
    if not isinstance(data, list):
        data = [data]
    
    for text in data:
        input_ids, segment_ids = convert_example(
            text,
            tokenizer,
            max_seq_length=128,
            is_test=True
        )
        examples.append((input_ids, segment_ids))

    batchify_fn = lambda samples, fn=Tuple(
        Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input id
        Pad(axis=0, pad_val=tokenizer.pad_token_id),  # segment id
    ): fn(samples)
    
    input_ids, segment_ids = batchify_fn(examples)
    return paddle.to_tensor(input_ids, stop_gradient=False), paddle.to_tensor(segment_ids, stop_gradient=False)

IG Interpreter¶

In [9]:

ig = it.IntGradNLPInterpreter(model, device='gpu:0')

pred_labels, pred_probs, avg_gradients = ig.interpret(
    preprocess_fn(data),
    steps=50,
    return_pred=True)

true_labels = [1, 1, 0, 0] * 5
recs = []
for i in range(avg_gradients.shape[0]):
    subwords = " ".join(tokenizer._tokenize(data[i]['text'])).split(' ')
    subword_importances = avg_gradients[i]
    words, word_importances = aggregate_subwords_and_importances(subwords, subword_importances)
    word_importances = np.array(word_importances) / np.linalg.norm(
        word_importances)
    
    pred_label = pred_labels[i]
    pred_prob = pred_probs[i, pred_label]
    true_label = true_labels[i]
    interp_class = pred_label
    
    if interp_class == 0:
        word_importances = -word_importances
    recs.append(
        VisualizationTextRecord(words, word_importances, true_label,
                                pred_label, pred_prob, interp_class)
    )

visualize_text(recs)
# The visualization is not available at github

True Label	Predicted Label (Prob)	Target Label	Word Importance
1	1 (1.00)	1	it ' s a charming and often affecting journey .
1	1 (0.99)	1	the movie achieves as great an impact by keeping these thoughts hidden as . . . ( quills ) did by showing them .
0	1 (0.76)	1	this one is definitely one to skip , even for horror movie fanatics .
0	0 (1.00)	0	in its best moments , resembles a bad high school production of grease , without benefit of song .

LIME Interpreter¶

In [8]:

true_labels = [1, 1, 0, 0] * 5
recs = []

lime = it.LIMENLPInterpreter(model, device='gpu:0')
for i, review in enumerate(data):
    pred_class, pred_prob, lime_weights = lime.interpret(
        review,
        preprocess_fn,
        num_samples=1000,
        batch_size=32,
        unk_id=tokenizer.convert_tokens_to_ids('[UNK]'),
        pad_id=tokenizer.convert_tokens_to_ids('[PAD]'),
        return_pred=True)

    # subwords
    subwords = " ".join(tokenizer._tokenize(review['text'])).split(' ')
    interp_class = list(lime_weights.keys())[0]
    weights = lime_weights[interp_class][1 : -1]
    subword_importances = [t[1] for t in lime_weights[interp_class][1 : -1]]
    
    words, word_importances = aggregate_subwords_and_importances(subwords, subword_importances)
    word_importances = np.array(word_importances) / np.linalg.norm(
        word_importances)
    
    true_label = true_labels[i]
    
    if interp_class == 0:
        word_importances = -word_importances
        
    rec = VisualizationTextRecord(
        words, 
        word_importances, 
        true_label,                   
        pred_class[0], 
        pred_prob[0],
        interp_class
    )
    
    recs.append(rec)

visualize_text(recs)
# The visualization is not available at github

True Label	Predicted Label (Prob)	Target Label	Word Importance
1	1 (1.00)	1	it ' s a charming and often affecting journey .
1	1 (0.99)	1	the movie achieves as great an impact by keeping these thoughts hidden as . . . ( quills ) did by showing them .
0	1 (0.82)	1	this one is definitely one to skip , even for horror movie fanatics .
0	0 (1.00)	0	in its best moments , resembles a bad high school production of grease , without benefit of song .

GradShapNLPInterpreter¶

In [7]:

ig = it.GradShapNLPInterpreter(model, device='gpu:0')

pred_labels, pred_probs, avg_gradients = ig.interpret(
    preprocess_fn(data),
    n_samples=10,
    noise_amount=0.1,
    return_pred=True)

true_labels = [1, 1, 0, 0] * 5
recs = []
for i in range(avg_gradients.shape[0]):
    subwords = " ".join(tokenizer._tokenize(data[i]['text'])).split(' ')
    subword_importances = avg_gradients[i]
    words, word_importances = aggregate_subwords_and_importances(subwords, subword_importances)
    word_importances = np.array(word_importances) / np.linalg.norm(
        word_importances)
    
    pred_label = pred_labels[i]
    pred_prob = pred_probs[i, pred_label]
    true_label = true_labels[i]
    interp_class = pred_label
    
    if interp_class == 0:
        word_importances = -word_importances
    recs.append(
        VisualizationTextRecord(words, word_importances, true_label,
                                pred_label, pred_prob, interp_class)
    )

visualize_text(recs)
# The visualization is not available at github

True Label	Predicted Label (Prob)	Target Label	Word Importance
1	1 (1.00)	1	it ' s a charming and often affecting journey .
1	1 (1.00)	1	the movie achieves as great an impact by keeping these thoughts hidden as . . . ( quills ) did by showing them .
0	1 (0.76)	1	this one is definitely one to skip , even for horror movie fanatics .
0	0 (1.00)	0	in its best moments , resembles a bad high school production of grease , without benefit of song .

In [ ]: