A Blog About D4T4 & M47H

Classifying Movie Reviews with TensorFlow

29 August ’18

09-09-2018 Update: My initial deployment of this model had training-serving skew since I was simply splitting words by spaces and feeding into the model. To properly serve this model, I needed to replicate the preprocessing in the serving input receiver. There is a nifty tool for this in the TF ecosystem called tf.Transform. See this post.


Recently, I've been having a lot of fun with Tensorflow! Here I'm building a DNN bag-of-words classifier with TF for classifying movie reviews as positive or negative. The data source is the ACL 2011 IMDB dataset. I used a custom estimator so that I could implement cosine annealing for learning rate decay.

I used Cloud ML Engine to deploy the model on GCP. I also used a Cloud Function to make the model accessible via a simple HTTP function. Give it a try!

Here's a link to all the code for the model build and a link to the Cloud Function for serving. Cheers!


Build TF Estimator

# set up feature columns
terms_feature_column = tf.feature_column.categorical_column_with_vocabulary_list(key='terms', 

terms_embedding_column = tf.feature_column.embedding_column(terms_feature_column, dimension=10)
feature_columns = [terms_embedding_column]

# create estimator spec
def make_model(features, labels, mode):

    # build graph
    net = tf.feature_column.input_layer(features, feature_columns)
    net = tf.layers.dense(net, units=10, activation=tf.nn.leaky_relu)
    net = tf.layers.dropout(net, rate=0.3, training=(mode == tf.estimator.ModeKeys.TRAIN))
    net = tf.layers.dense(net, units=10)
    logits = tf.layers.dense(net, 2)

    # compute predictions
    predicted_classes = tf.argmax(logits, 1)
    predicted_probs = tf.nn.softmax(logits)

    # generate predictions
    if mode == tf.estimator.ModeKeys.PREDICT:
        predictions = {
            'class': predicted_classes,
            'prob': predicted_probs

        export_outputs = {
          'predict': tf.estimator.export.PredictOutput(outputs=predictions)

        return tf.estimator.EstimatorSpec(mode, predictions=predictions, export_outputs=export_outputs)

    # compute loss
    loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits)

    # create training op with cosine annealing for learning rate
    if mode == tf.estimator.ModeKeys.TRAIN:
        global_step = tf.train.get_global_step()

        learning_rate = tf.train.cosine_decay(learning_rate=0.2, global_step=global_step, alpha=0.05, decay_steps=10000)

        optimizer = tf.train.AdagradOptimizer(learning_rate=learning_rate)
        optimizer = tf.contrib.estimator.clip_gradients_by_norm(optimizer, 5.0)

        train_op = optimizer.minimize(loss, global_step=global_step)

        return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op)

    # compute evaluation metrics
    eval_metric_ops = {
        'accuracy': tf.metrics.accuracy(labels=labels, predictions=predicted_classes),
        'auc': tf.metrics.auc(labels=labels, predictions=predicted_probs[:, 1])
    return tf.estimator.EstimatorSpec(mode, loss=loss, eval_metric_ops=eval_metric_ops)

# create estimator
classifier = tf.estimator.Estimator(model_fn=make_model)

# train and evaluate
classifier.train(input_fn=lambda: input_fn([train_path], num_epochs=10))
test_stats = classifier.evaluate(input_fn=lambda: input_fn([test_path], num_epochs=1))

# export
def serving_input_receiver_fn():
    reviews = tf.placeholder(dtype=tf.string, shape=(None), name='reviews')
    terms = tf.sparse_tensor_to_dense(tf.string_split(reviews), default_value='')
    return tf.estimator.export.ServingInputReceiver({'terms': terms}, {'reviews': reviews})

export_path = classifier.export_savedmodel(export_dir_base='exports',

export_path = export_path.decode('utf-8')

Model Results

Train Test
Accuracy 0.918600 0.870160
AUC 0.971972 0.942018
Loss 0.214713 0.321892

Upload to Cloud ML Engine


MODEL_TIMESTAMP=$(ls -t exports/ | head -1)


gsutil rsync -c -d -r exports/$MODEL_TIMESTAMP $DEPLOYMENT_SOURCE

gcloud ml-engine models create $MODEL_NAME

gcloud ml-engine versions create $MODEL_VERSION --model $MODEL_NAME --origin $DEPLOYMENT_SOURCE \
    --python-version 3.5 --runtime-version 1.9

NOTE: Make sure the Python environment in which you build your model matches the serving environment in Cloud ML!

Expose Model with a Cloud Function

# gets predictions from cloud ml engine
def ml_predict(request):
    import flask
    import json
    import googleapiclient.discovery
    import google.auth

    headers = {
        'Access-Control-Allow-Origin': '*',
        'Access-Control-Allow-Methods': 'POST',
        'Access-Control-Allow-Headers': 'Content-Type'

    # handle pre-flight options request
    if request.method == 'OPTIONS':
        return flask.make_response(('', 204, headers))

    _, project = google.auth.default()

    request_json = request.get_json()

    model = request_json['model']
    version = request_json['version']
    instances = request_json['instances']

    service ='ml', 'v1')
    name = 'projects/{}/models/{}/versions/{}'.format(project, model, version)

    response = service.projects().predict(
        body={'instances': instances}

    if 'error' in response:
        raise RuntimeError(response['error'])

    return flask.make_response((