Help protect the Great Barrier Reef with TensorFlow on Kaggle Join Challenge

Bangla Article Classification With TF-Hub

View on TensorFlow.org View on GitHub Download notebook

This Colab is a demonstration of using Tensorflow Hub for text classification in non-English/local languages. Here we choose Bangla as the local language and use pretrained word embeddings to solve a multiclass classification task where we classify Bangla news articles in 5 categories. The pretrained embeddings for Bangla comes from fastText which is a library by Facebook with released pretrained word vectors for 157 languages.

We'll use TF-Hub's pretrained embedding exporter for converting the word embeddings to a text embedding module first and then use the module to train a classifier with tf.keras, Tensorflow's high level user friendly API to build deep learning models. Even if we are using fastText embeddings here, it's possible to export any other embeddings pretrained from other tasks and quickly get results with Tensorflow hub.

Setup

# https://github.com/pypa/setuptools/issues/1694#issuecomment-466010982
pip install gdown --no-use-pep517
sudo apt-get install -y unzip
Reading package lists...
Building dependency tree...
Reading state information...
unzip is already the newest version (6.0-21ubuntu1.1).
The following packages were automatically installed and are no longer required:
  linux-gcp-5.4-headers-5.4.0-1040 linux-gcp-5.4-headers-5.4.0-1043
  linux-gcp-5.4-headers-5.4.0-1044 linux-gcp-5.4-headers-5.4.0-1049
  linux-headers-5.4.0-1049-gcp linux-image-5.4.0-1049-gcp
  linux-modules-5.4.0-1049-gcp linux-modules-extra-5.4.0-1049-gcp
Use 'sudo apt autoremove' to remove them.
0 upgraded, 0 newly installed, 0 to remove and 143 not upgraded.
import os

import tensorflow as tf
import tensorflow_hub as hub

import gdown
import numpy as np
from sklearn.metrics import classification_report
import matplotlib.pyplot as plt
import seaborn as sns
/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/pkg_resources/__init__.py:119: PkgResourcesDeprecationWarning: 0.18ubuntu0.18.04.1 is an invalid version and will not be supported in a future release
  PkgResourcesDeprecationWarning,

Dataset

We will use BARD (Bangla Article Dataset) which has around 376,226 articles collected from different Bangla news portals and labelled with 5 categories: economy, state, international, sports, and entertainment. We download the file from Google Drive this (bit.ly/BARD_DATASET) link is referring to from this GitHub repository.

gdown.download(
    url='https://drive.google.com/uc?id=1Ag0jd21oRwJhVFIBohmX_ogeojVtapLy',
    output='bard.zip',
    quiet=True
)
'bard.zip'
unzip -qo bard.zip

Export pretrained word vectors to TF-Hub module

TF-Hub provides some useful scripts for converting word embeddings to TF-hub text embedding modules here. To make the module for Bangla or any other languages, we simply have to download the word embedding .txt or .vec file to the same directory as export_v2.py and run the script.

The exporter reads the embedding vectors and exports it to a Tensorflow SavedModel. A SavedModel contains a complete TensorFlow program including weights and graph. TF-Hub can load the SavedModel as a module, which we will use to build the model for text classification. Since we are using tf.keras to build the model, we will use hub.KerasLayer, which provides a wrapper for a TF-Hub module to use as a Keras Layer.

First we will get our word embeddings from fastText and embedding exporter from TF-Hub repo.

curl -O https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.bn.300.vec.gz
curl -O https://raw.githubusercontent.com/tensorflow/hub/master/examples/text_embeddings_v2/export_v2.py
gunzip -qf cc.bn.300.vec.gz --k
% Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  840M  100  840M    0     0  11.6M      0  0:01:12  0:01:12 --:--:-- 12.0M
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  7469  100  7469    0     0  19053      0 --:--:-- --:--:-- --:--:-- 19005

Then, we will run the exporter script on our embedding file. Since fastText embeddings have a header line and are pretty large (around 3.3 GB for Bangla after converting to a module) we ignore the first line and export only the first 100, 000 tokens to the text embedding module.

python export_v2.py --embedding_file=cc.bn.300.vec --export_path=text_module --num_lines_to_ignore=1 --num_lines_to_use=100000
INFO:tensorflow:Assets written to: text_module/assets
I1105 11:55:29.817717 140238757988160 builder_impl.py:784] Assets written to: text_module/assets
module_path = "text_module"
embedding_layer = hub.KerasLayer(module_path, trainable=False)

The text embedding module takes a batch of sentences in a 1D tensor of strings as input and outputs the embedding vectors of shape (batch_size, embedding_dim) corresponding to the sentences. It preprocesses the input by splitting on spaces. Word embeddings are combined to sentence embeddings with the sqrtn combiner(See here). For demonstration we pass a list of Bangla words as input and get the corresponding embedding vectors.

embedding_layer(['বাস', 'বসবাস', 'ট্রেন', 'যাত্রী', 'ট্রাক'])
<tf.Tensor: shape=(5, 300), dtype=float64, numpy=
array([[ 0.0462, -0.0355,  0.0129, ...,  0.0025, -0.0966,  0.0216],
       [-0.0631, -0.0051,  0.085 , ...,  0.0249, -0.0149,  0.0203],
       [ 0.1371, -0.069 , -0.1176, ...,  0.029 ,  0.0508, -0.026 ],
       [ 0.0532, -0.0465, -0.0504, ...,  0.02  , -0.0023,  0.0011],
       [ 0.0908, -0.0404, -0.0536, ..., -0.0275,  0.0528,  0.0253]])>

Convert to Tensorflow Dataset

Since the dataset is really large instead of loading the entire dataset in memory we will use a generator to yield samples in run-time in batches using Tensorflow Dataset functions. The dataset is also very imbalanced, so, before using the generator, we will shuffle the dataset.

dir_names = ['economy', 'sports', 'entertainment', 'state', 'international']

file_paths = []
labels = []
for i, dir in enumerate(dir_names):
  file_names = ["/".join([dir, name]) for name in os.listdir(dir)]
  file_paths += file_names
  labels += [i] * len(os.listdir(dir))

np.random.seed(42)
permutation = np.random.permutation(len(file_paths))

file_paths = np.array(file_paths)[permutation]
labels = np.array(labels)[permutation]

We can check the distribution of labels in the training and validation examples after shuffling.

train_frac = 0.8
train_size = int(len(file_paths) * train_frac)
# plot training vs validation distribution
plt.subplot(1, 2, 1)
plt.hist(labels[0:train_size])
plt.title("Train labels")
plt.subplot(1, 2, 2)
plt.hist(labels[train_size:])
plt.title("Validation labels")
plt.tight_layout()

png

To create a Dataset using a generator, we first write a generator function which reads each of the articles from file_paths and the labels from the label array, and yields one training example at each step. We pass this generator function to the tf.data.Dataset.from_generator method and specify the output types. Each training example is a tuple containing an article of tf.string data type and one-hot encoded label. We split the dataset with a train-validation split of 80-20 using tf.data.Dataset.skip and tf.data.Dataset.take methods.

def load_file(path, label):
    return tf.io.read_file(path), label
def make_datasets(train_size):
  batch_size = 256

  train_files = file_paths[:train_size]
  train_labels = labels[:train_size]
  train_ds = tf.data.Dataset.from_tensor_slices((train_files, train_labels))
  train_ds = train_ds.map(load_file).shuffle(5000)
  train_ds = train_ds.batch(batch_size).prefetch(tf.data.AUTOTUNE)

  test_files = file_paths[train_size:]
  test_labels = labels[train_size:]
  test_ds = tf.data.Dataset.from_tensor_slices((test_files, test_labels))
  test_ds = test_ds.map(load_file)
  test_ds = test_ds.batch(batch_size).prefetch(tf.data.AUTOTUNE)


  return train_ds, test_ds
train_data, validation_data = make_datasets(train_size)

Model Training and Evaluation

Since we have already added a wrapper around our module to use it as any other layer in Keras, we can create a small Sequential model which is a linear stack of layers. We can add our text embedding module with model.add just like any other layer. We compile the model by specifying the loss and optimizer and train it for 10 epochs. The tf.keras API can handle Tensorflow Datasets as input, so we can pass a Dataset instance to the fit method for model training. Since we are using the generator function, tf.data will handle generating the samples, batching them and feeding them to the model.

Model

def create_model():
  model = tf.keras.Sequential([
    tf.keras.layers.Input(shape=[], dtype=tf.string),
    embedding_layer,
    tf.keras.layers.Dense(64, activation="relu"),
    tf.keras.layers.Dense(16, activation="relu"),
    tf.keras.layers.Dense(5),
  ])
  model.compile(loss=tf.losses.SparseCategoricalCrossentropy(from_logits=True),
      optimizer="adam", metrics=['accuracy'])
  return model
model = create_model()
# Create earlystopping callback
early_stopping_callback = tf.keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=0, patience=3)

Training

history = model.fit(train_data, 
                    validation_data=validation_data, 
                    epochs=5, 
                    callbacks=[early_stopping_callback])
Epoch 1/5
1176/1176 [==============================] - 34s 28ms/step - loss: 0.2181 - accuracy: 0.9279 - val_loss: 0.1580 - val_accuracy: 0.9449
Epoch 2/5
1176/1176 [==============================] - 32s 27ms/step - loss: 0.1411 - accuracy: 0.9505 - val_loss: 0.1411 - val_accuracy: 0.9503
Epoch 3/5
1176/1176 [==============================] - 32s 27ms/step - loss: 0.1307 - accuracy: 0.9534 - val_loss: 0.1359 - val_accuracy: 0.9524
Epoch 4/5
1176/1176 [==============================] - 32s 27ms/step - loss: 0.1248 - accuracy: 0.9555 - val_loss: 0.1318 - val_accuracy: 0.9527
Epoch 5/5
1176/1176 [==============================] - 32s 27ms/step - loss: 0.1196 - accuracy: 0.9567 - val_loss: 0.1247 - val_accuracy: 0.9555

Evaluation

We can visualize the accuracy and loss curves for training and validation data using the tf.keras.callbacks.History object returned by the tf.keras.Model.fit method, which contains the loss and accuracy value for each epoch.

# Plot training & validation accuracy values
plt.plot(history.history['accuracy'])
plt.plot(history.history['val_accuracy'])
plt.title('Model accuracy')
plt.ylabel('Accuracy')
plt.xlabel('Epoch')
plt.legend(['Train', 'Test'], loc='upper left')
plt.show()

# Plot training & validation loss values
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('Model loss')
plt.ylabel('Loss')
plt.xlabel('Epoch')
plt.legend(['Train', 'Test'], loc='upper left')
plt.show()

png

png

Prediction

We can get the predictions for the validation data and check the confusion matrix to see the model's performance for each of the 5 classes. Because tf.keras.Model.predict method returns an n-d array for probabilities for each class, they can be converted to class labels using np.argmax.

y_pred = model.predict(validation_data)
y_pred = np.argmax(y_pred, axis=1)
samples = file_paths[0:3]
for i, sample in enumerate(samples):
  f = open(sample)
  text = f.read()
  print(text[0:100])
  print("True Class: ", sample.split("/")[0])
  print("Predicted Class: ", dir_names[y_pred[i]])
  f.close()
রবিন উইলিয়ামস তাঁর হ্যাপি ফিট ছবিতে মজা করে বলেছিলেন, ‘আমি শুনতে পাচ্ছি, মানুষ কিছু একটা চাইছে...সে
True Class:  entertainment
Predicted Class:  state

নির্মাণ শেষে ফিতা কেটে মন্ত্রী ভবন উদ্বোধন করেছেন বহু আগেই। তবে এখনো চালু করা যায়নি খাগড়াছড়ি জেল
True Class:  state
Predicted Class:  state

কমলাপুর বীরশ্রেষ্ঠ মোস্তফা কামাল স্টেডিয়ামে কাল ফকিরেরপুল ইয়ংমেন্স ক্লাব ৩-০ গোলে হারিয়েছে স্বাধ
True Class:  sports
Predicted Class:  state

Compare Performance

Now we can take the correct labels for the validation data from labels and compare them with our predictions to get a classification_report.

y_true = np.array(labels[train_size:])
print(classification_report(y_true, y_pred, target_names=dir_names))
precision    recall  f1-score   support

      economy       0.83      0.76      0.79      3897
       sports       0.99      0.98      0.98     10204
entertainment       0.90      0.94      0.92      6256
        state       0.97      0.97      0.97     48512
international       0.92      0.93      0.93      6377

     accuracy                           0.96     75246
    macro avg       0.92      0.92      0.92     75246
 weighted avg       0.96      0.96      0.96     75246

We can also compare our model's performance with the published results obtained in the original paper, which had a 0.96 precision .The original authors described many preprocessing steps performed on the dataset, such as dropping punctuations and digits, removing top 25 most frequest stop words. As we can see in the classification_report, we also manage to obtain a 0.96 precision and accuracy after training for only 5 epochs without any preprocessing!

In this example, when we created the Keras layer from our embedding module, we set the parametertrainable=False, which means the embedding weights will not be updated during training. Try setting it to True to reach around 97% accuracy using this dataset after only 2 epochs.