Training ResNet50 in TensorFlow 2.0

resnet50-tf2-py3.py

import tensorflow as tf
import numpy as np

BATCH = 2
NUM_EPOCHS = 25


"""
ResNet50 is implemented here:
https://github.com/tensorflow/tensorflow/blob/bd754067dac90182d883f621b775d76ec7c6b87d/tensorflow/python/eager/benchmarks/resnet50/resnet50.py#L1
"""

class FakeData(object):
    def __init__(self):
        super(FakeData, self).__init__()
        self.length = 100
        self.X_train = np.random.random((224, 224, 3)).astype('float32')
        self.Y_train = np.array([np.random.randint(1000)]).astype('int32')

    def __iter__(self):
        for _ in range(self.length):
            yield self.X_train, self.Y_train

    def __len__(self):
        return self.length

    def output_shapes(self):
        return (self.X_train.shape, self.Y_train.shape)

    def output_types(self):
        return (tf.float32, tf.int32)


df = FakeData()


def get_data(df):
    tdf = tf.data.Dataset.from_generator(
        generator=df.__iter__,
        output_types=df.output_types(),
        output_shapes=df.output_shapes())
    tdf = tdf.batch(BATCH)
    tdf = tdf.prefetch(tf.data.experimental.AUTOTUNE)
    return tdf


# see https://www.tensorflow.org/guide/distributed_training#mirroredstrategy
mirrored_strategy = tf.distribute.MirroredStrategy()
with mirrored_strategy.scope():

    model = tf.keras.applications.resnet.ResNet50(
        input_shape=df.output_shapes()[0],
        include_top=True,
        weights=None)

    model.compile(
        optimizer=tf.keras.optimizers.SGD(lr=0.001),
        loss='binary_crossentropy',
        metrics=['accuracy'])


model.fit(get_data(df), epochs=NUM_EPOCHS)

Line 18 is not generating an one-hot label. I was able to run after changing it to one-hot variable.

zeros = np.zeros(1000)
zeros[np.random.randint(1000)] = 1
self.Y_train = zeros.astype('int32')