Use tensorflow learning-rate decay in a Keras-to-TPU model

Question

I'm following the "How to train Keras model x20 times faster with TPU for free" guide ( click here ) to run a keras model on google's colab TPU. It works perfectly. But...I like to use cosine restart learning rate decay when I fit my models. I've coded up my own as a keras callback, but it won't work within this framework because the tensorflow TFOptimizer class doesn't have a learning-rate variable that can be reset. I see that tensorflow itself has a bunch of decay function in tf.train , like tf.train.cosine_decay but I can't figure out how to embed it within my model.

Here's the basic code from that blog post. Anyone have a fix?

import tensorflow as tf
import os
from tensorflow.python.keras.layers import Input, LSTM, Bidirectional, Dense, Embedding

def make_model(batch_size=None):
    source = Input(shape=(maxlen,), batch_size=batch_size,
                   dtype=tf.int32, name='Input')
    embedding = Embedding(input_dim=max_features,
                          output_dim=128, name='Embedding')(source)
    lstm = LSTM(32, name='LSTM')(embedding)
    predicted_var = Dense(1, activation='sigmoid', name='Output')(lstm)
    model = tf.keras.Model(inputs=[source], outputs=[predicted_var])
    model.compile(
        optimizer=tf.train.RMSPropOptimizer(learning_rate=0.01),
        loss='binary_crossentropy',
        metrics=['acc'])
    return model

training_model = make_model(batch_size=128)

# This address identifies the TPU we'll use when configuring TensorFlow.
TPU_WORKER = 'grpc://' + os.environ['COLAB_TPU_ADDR']
tf.logging.set_verbosity(tf.logging.INFO)

tpu_model = tf.contrib.tpu.keras_to_tpu_model(
    training_model,
    strategy=tf.contrib.tpu.TPUDistributionStrategy(
        tf.contrib.cluster_resolver.TPUClusterResolver(TPU_WORKER)))

history = tpu_model.fit(x_train, y_train,
                    epochs=20,
                    batch_size=128 * 8,
                    validation_split=0.2)

Answer 1

一种选择是手动设置学习率-这里有一个Keras + TPU示例，其中带回调： https : //github.com/tensorflow/tpu/blob/master/models/experimental/resnet50_keras/resnet50.py#L197- L201

Answer 2

The following seems to work, where lr is the initial learning rate you choose and M is the number of initial steps over which you want to the cosine decay to work.

def make_model(batch_size=None,lr=1.e-3,n_steps=2000):
    source = Input(shape=(maxlen,), batch_size=batch_size,
                   dtype=tf.int32, name='Input')
    embedding = Embedding(input_dim=max_features,
                          output_dim=128, name='Embedding')(source)
    lstm = LSTM(32, name='LSTM')(embedding)
    predicted_var = Dense(1, activation='sigmoid', name='Output')(lstm)
    model = tf.keras.Model(inputs=[source], outputs=[predicted_var])
    # implement cosine decay or other learning rate decay here
    global_step = tf.Variable(0)    
    global_step=1
    learning_rate = tf.train.cosine_decay_restarts(
        learning_rate=lr,
        global_step=global_step,
        first_decay_steps=n_steps,
        t_mul= 1.5,
        m_mul= 1.,
        alpha=0.1
    )
    # now feed this into the optimizer as shown below
    model.compile(
        optimizer=tf.train.RMSPropOptimizer(learning_rate=learning_rate),
        loss='binary_crossentropy',
        metrics=['acc'])
    return model