Skip to content

GitLab

Unverified Commit 72f5834c authored by Asim Shankar's avatar Asim Shankar Committed by GitHub
Browse files

Merge pull request #3024 from asimshankar/mnist 

[mnist]: Updates
parents 0f5803bd 4d79fee3
Hide whitespace changes
Inline Side-by-side
Showing with 189 additions and 143 deletions
official/mnist/README.md
View file @ 72f5834c
......@@ -21,12 +21,19 @@ python mnist.py
The model will begin training and will automatically evaluate itself on the
validation data.
Illustrative unit tests and benchmarks can be run with:
```
python mnist_test.py
python mnist_test.py --benchmarks=.
```
## Exporting the model
You can export the model into Tensorflow [SavedModel](https://www.tensorflow.org/programmers_guide/saved_model) format by using the argument `--export_dir`:
```
python mnist.py --export_dir /tmp/mnist_saved_model
python mnist.py --export_dir /tmp/mnist_saved_model
```
The SavedModel will be saved in a timestamped directory under `/tmp/mnist_saved_model/` (e.g. `/tmp/mnist_saved_model/1513630966/`).
......@@ -35,7 +42,7 @@ The SavedModel will be saved in a timestamped directory under `/tmp/mnist_saved_
Use [`saved_model_cli`](https://www.tensorflow.org/programmers_guide/saved_model#cli_to_inspect_and_execute_savedmodel) to inspect and execute the SavedModel.
```
saved_model_cli run --dir /tmp/mnist_saved_model/TIMESTAMP --tag_set serve --signature_def classify --inputs image_raw=examples.npy
saved_model_cli run --dir /tmp/mnist_saved_model/TIMESTAMP --tag_set serve --signature_def classify --inputs image=examples.npy
```
`examples.npy` contains the data from `example5.png` and `example3.png` in a numpy array, in that order. The array values are normalized to values between 0 and 1.
......
official/mnist/mnist.py
View file @ 72f5834c
......@@ -63,6 +63,7 @@ parser.add_argument(
type=str,
help='The directory where the exported SavedModel will be stored.')
def train_dataset(data_dir):
"""Returns a tf.data.Dataset yielding (image, label) pairs for training."""
data = input_data.read_data_sets(data_dir, one_hot=True).train
......@@ -75,141 +76,115 @@ def eval_dataset(data_dir):
return tf.data.Dataset.from_tensors((data.images, data.labels))
def mnist_model(inputs, mode, data_format):
"""Takes the MNIST inputs and mode and outputs a tensor of logits."""
# Input Layer
# Reshape X to 4-D tensor: [batch_size, width, height, channels]
# MNIST images are 28x28 pixels, and have one color channel
inputs = tf.reshape(inputs, [-1, 28, 28, 1])
if data_format is None:
# When running on GPU, transpose the data from channels_last (NHWC) to
# channels_first (NCHW) to improve performance.
# See https://www.tensorflow.org/performance/performance_guide#data_formats
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
if data_format == 'channels_first':
inputs = tf.transpose(inputs, [0, 3, 1, 2])
# Convolutional Layer #1
# Computes 32 features using a 5x5 filter with ReLU activation.
# Padding is added to preserve width and height.
# Input Tensor Shape: [batch_size, 28, 28, 1]
# Output Tensor Shape: [batch_size, 28, 28, 32]
conv1 = tf.layers.conv2d(
inputs=inputs,
filters=32,
kernel_size=[5, 5],
padding='same',
activation=tf.nn.relu,
data_format=data_format)
# Pooling Layer #1
# First max pooling layer with a 2x2 filter and stride of 2
# Input Tensor Shape: [batch_size, 28, 28, 32]
# Output Tensor Shape: [batch_size, 14, 14, 32]
pool1 = tf.layers.max_pooling2d(
inputs=conv1, pool_size=[2, 2], strides=2, data_format=data_format)
# Convolutional Layer #2
# Computes 64 features using a 5x5 filter.
# Padding is added to preserve width and height.
# Input Tensor Shape: [batch_size, 14, 14, 32]
# Output Tensor Shape: [batch_size, 14, 14, 64]
conv2 = tf.layers.conv2d(
inputs=pool1,
filters=64,
kernel_size=[5, 5],
padding='same',
activation=tf.nn.relu,
data_format=data_format)
# Pooling Layer #2
# Second max pooling layer with a 2x2 filter and stride of 2
# Input Tensor Shape: [batch_size, 14, 14, 64]
# Output Tensor Shape: [batch_size, 7, 7, 64]
pool2 = tf.layers.max_pooling2d(
inputs=conv2, pool_size=[2, 2], strides=2, data_format=data_format)
# Flatten tensor into a batch of vectors
# Input Tensor Shape: [batch_size, 7, 7, 64]
# Output Tensor Shape: [batch_size, 7 * 7 * 64]
pool2_flat = tf.reshape(pool2, [-1, 7 * 7 * 64])
# Dense Layer
# Densely connected layer with 1024 neurons
# Input Tensor Shape: [batch_size, 7 * 7 * 64]
# Output Tensor Shape: [batch_size, 1024]
dense = tf.layers.dense(inputs=pool2_flat, units=1024, activation=tf.nn.relu)
# Add dropout operation; 0.6 probability that element will be kept
dropout = tf.layers.dropout(
inputs=dense, rate=0.4, training=(mode == tf.estimator.ModeKeys.TRAIN))
# Logits layer
# Input Tensor Shape: [batch_size, 1024]
# Output Tensor Shape: [batch_size, 10]
logits = tf.layers.dense(inputs=dropout, units=10)
return logits
def mnist_model_fn(features, labels, mode, params):
"""Model function for MNIST."""
if mode == tf.estimator.ModeKeys.PREDICT and isinstance(features,dict):
features = features['image_raw']
logits = mnist_model(features, mode, params['data_format'])
predictions = {
'classes': tf.argmax(input=logits, axis=1),
'probabilities': tf.nn.softmax(logits, name='softmax_tensor')
}
class Model(object):
"""Class that defines a graph to recognize digits in the MNIST dataset."""
def __init__(self, data_format):
"""Creates a model for classifying a hand-written digit.
Args:
data_format: Either 'channels_first' or 'channels_last'.
'channels_first' is typically faster on GPUs while 'channels_last' is
typically faster on CPUs. See
https://www.tensorflow.org/performance/performance_guide#data_formats
"""
if data_format == 'channels_first':
self._input_shape = [-1, 1, 28, 28]
else:
assert data_format == 'channels_last'
self._input_shape = [-1, 28, 28, 1]
self.conv1 = tf.layers.Conv2D(
32, 5, padding='same', data_format=data_format, activation=tf.nn.relu)
self.conv2 = tf.layers.Conv2D(
64, 5, padding='same', data_format=data_format, activation=tf.nn.relu)
self.fc1 = tf.layers.Dense(1024, activation=tf.nn.relu)
self.fc2 = tf.layers.Dense(10)
self.dropout = tf.layers.Dropout(0.4)
self.max_pool2d = tf.layers.MaxPooling2D(
(2, 2), (2, 2), padding='same', data_format=data_format)
def __call__(self, inputs, training):
"""Add operations to classify a batch of input images.
Args:
inputs: A Tensor representing a batch of input images.
training: A boolean. Set to True to add operations required only when
training the classifier.
Returns:
A logits Tensor with shape [<batch_size>, 10].
"""
y = tf.reshape(inputs, self._input_shape)
y = self.conv1(y)
y = self.max_pool2d(y)
y = self.conv2(y)
y = self.max_pool2d(y)
y = tf.layers.flatten(y)
y = self.fc1(y)
y = self.dropout(y, training=training)
return self.fc2(y)
def model_fn(features, labels, mode, params):
"""The model_fn argument for creating an Estimator."""
model = Model(params['data_format'])
image = features
if isinstance(image, dict):
image = features['image']
if mode == tf.estimator.ModeKeys.PREDICT:
export_outputs={'classify': tf.estimator.export.PredictOutput(predictions)}
return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions,
export_outputs=export_outputs)
loss = tf.losses.softmax_cross_entropy(onehot_labels=labels, logits=logits)
# Configure the training op
logits = model(image, training=False)
predictions = {
'classes': tf.argmax(logits, axis=1),
'probabilities': tf.nn.softmax(logits),
}
return tf.estimator.EstimatorSpec(
mode=tf.estimator.ModeKeys.PREDICT,
predictions=predictions,
export_outputs={
'classify': tf.estimator.export.PredictOutput(predictions)
})
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
train_op = optimizer.minimize(loss, tf.train.get_or_create_global_step())
else:
train_op = None
accuracy = tf.metrics.accuracy(
tf.argmax(labels, axis=1), predictions['classes'])
metrics = {'accuracy': accuracy}
# Create a tensor named train_accuracy for logging purposes
tf.identity(accuracy[1], name='train_accuracy')
tf.summary.scalar('train_accuracy', accuracy[1])
return tf.estimator.EstimatorSpec(
mode=mode,
predictions=predictions,
loss=loss,
train_op=train_op,
eval_metric_ops=metrics)
logits = model(image, training=True)
loss = tf.losses.softmax_cross_entropy(onehot_labels=labels, logits=logits)
accuracy = tf.metrics.accuracy(
labels=tf.argmax(labels, axis=1), predictions=tf.argmax(logits, axis=1))
# Name the accuracy tensor 'train_accuracy' to demonstrate the
# LoggingTensorHook.
tf.identity(accuracy[1], name='train_accuracy')
tf.summary.scalar('train_accuracy', accuracy[1])
return tf.estimator.EstimatorSpec(
mode=tf.estimator.ModeKeys.TRAIN,
loss=loss,
train_op=optimizer.minimize(loss, tf.train.get_or_create_global_step()))
if mode == tf.estimator.ModeKeys.EVAL:
logits = model(image, training=False)
loss = tf.losses.softmax_cross_entropy(onehot_labels=labels, logits=logits)
return tf.estimator.EstimatorSpec(
mode=tf.estimator.ModeKeys.EVAL,
loss=loss,
eval_metric_ops={
'accuracy':
tf.metrics.accuracy(
labels=tf.argmax(labels, axis=1),
predictions=tf.argmax(logits, axis=1)),
})
def main(unused_argv):
# Create the Estimator
data_format = FLAGS.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
mnist_classifier = tf.estimator.Estimator(
model_fn=mnist_model_fn,
model_fn=model_fn,
model_dir=FLAGS.model_dir,
params={
'data_format': FLAGS.data_format
'data_format': data_format
})
# Set up training hook that logs the training accuracy every 100 steps.
tensors_to_log = {'train_accuracy': 'train_accuracy'}
logging_hook = tf.train.LoggingTensorHook(
tensors=tensors_to_log, every_n_iter=100)
# Train the model
def train_input_fn():
# When choosing shuffle buffer sizes, larger sizes result in better
......@@ -221,6 +196,10 @@ def main(unused_argv):
(images, labels) = dataset.make_one_shot_iterator().get_next()
return (images, labels)
# Set up training hook that logs the training accuracy every 100 steps.
tensors_to_log = {'train_accuracy': 'train_accuracy'}
logging_hook = tf.train.LoggingTensorHook(
tensors=tensors_to_log, every_n_iter=100)
mnist_classifier.train(input_fn=train_input_fn, hooks=[logging_hook])
# Evaluate the model and print results
......@@ -233,10 +212,11 @@ def main(unused_argv):
# Export the model
if FLAGS.export_dir is not None:
image = tf.placeholder(tf.float32,[None, 28, 28])
serving_input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn(
{"image_raw":image})
mnist_classifier.export_savedmodel(FLAGS.export_dir, serving_input_fn)
image = tf.placeholder(tf.float32, [None, 28, 28])
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'image': image,
})
mnist_classifier.export_savedmodel(FLAGS.export_dir, input_fn)
if __name__ == '__main__':
......
official/mnist/mnist_test.py
View file @ 72f5834c
......@@ -18,30 +18,63 @@ from __future__ import division
from __future__ import print_function
import tensorflow as tf
import time
import mnist
tf.logging.set_verbosity(tf.logging.ERROR)
BATCH_SIZE = 100
class BaseTest(tf.test.TestCase):
def dummy_input_fn():
image = tf.random_uniform([BATCH_SIZE, 784])
labels = tf.random_uniform([BATCH_SIZE], maxval=9, dtype=tf.int32)
return image, tf.one_hot(labels, 10)
def input_fn(self):
features = tf.random_uniform([55000, 784])
labels = tf.random_uniform([55000], maxval=9, dtype=tf.int32)
return features, tf.one_hot(labels, 10)
def make_estimator():
data_format = 'channels_last'
if tf.test.is_built_with_cuda():
data_format = 'channels_first'
return tf.estimator.Estimator(
model_fn=mnist.model_fn, params={
'data_format': data_format
})
class Tests(tf.test.TestCase):
def test_mnist(self):
classifier = make_estimator()
classifier.train(input_fn=dummy_input_fn, steps=2)
eval_results = classifier.evaluate(input_fn=dummy_input_fn, steps=1)
loss = eval_results['loss']
global_step = eval_results['global_step']
accuracy = eval_results['accuracy']
self.assertEqual(loss.shape, ())
self.assertEqual(2, global_step)
self.assertEqual(accuracy.shape, ())
input_fn = lambda: tf.random_uniform([3, 784])
predictions_generator = classifier.predict(input_fn)
for i in range(3):
predictions = next(predictions_generator)
self.assertEqual(predictions['probabilities'].shape, (10,))
self.assertEqual(predictions['classes'].shape, ())
def mnist_model_fn_helper(self, mode):
features, labels = self.input_fn()
features, labels = dummy_input_fn()
image_count = features.shape[0]
spec = mnist.mnist_model_fn(
features, labels, mode, {'data_format': 'channels_last'})
spec = mnist.model_fn(features, labels, mode, {
'data_format': 'channels_last'
})
predictions = spec.predictions
self.assertAllEqual(predictions['probabilities'].shape, (image_count, 10))
self.assertEqual(predictions['probabilities'].dtype, tf.float32)
self.assertAllEqual(predictions['classes'].shape, (image_count,))
self.assertEqual(predictions['classes'].dtype, tf.int64)
if mode == tf.estimator.ModeKeys.PREDICT:
predictions = spec.predictions
self.assertAllEqual(predictions['probabilities'].shape, (image_count, 10))
self.assertEqual(predictions['probabilities'].dtype, tf.float32)
self.assertAllEqual(predictions['classes'].shape, (image_count,))
self.assertEqual(predictions['classes'].dtype, tf.int64)
if mode != tf.estimator.ModeKeys.PREDICT:
loss = spec.loss
......@@ -65,5 +98,31 @@ class BaseTest(tf.test.TestCase):
self.mnist_model_fn_helper(tf.estimator.ModeKeys.PREDICT)
class Benchmarks(tf.test.Benchmark):
def benchmark_train_step_time(self):
classifier = make_estimator()
# Run one step to warmup any use of the GPU.
classifier.train(input_fn=dummy_input_fn, steps=1)
have_gpu = tf.test.is_gpu_available()
num_steps = 1000 if have_gpu else 100
name = 'train_step_time_%s' % ('gpu' if have_gpu else 'cpu')
start = time.time()
classifier.train(input_fn=dummy_input_fn, steps=num_steps)
end = time.time()
wall_time = (end - start) / num_steps
self.report_benchmark(
iters=num_steps,
wall_time=wall_time,
name=name,
extras={
'examples_per_sec': BATCH_SIZE / wall_time
})
if __name__ == '__main__':
tf.logging.set_verbosity(tf.logging.ERROR)
tf.test.main()
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment