This sample trains an "MNIST" handwritten digit recognition model on a GPU or TPU backend using a Keras model. Data are handled using the tf.data.Datset API. This is a very simple sample provided for educational purposes. Do not expect outstanding TPU performance on a dataset as small as MNIST.
In [1]:
BATCH_SIZE = 64
LEARNING_RATE = 0.02
# GCS bucket for training logs and for saving the trained model
# You can leave this empty for local saving, unless you are using a TPU.
# TPUs do not have access to your local instance and can only write to GCS.
BUCKET="" # a valid bucket name must start with gs://
training_images_file = 'gs://mnist-public/train-images-idx3-ubyte'
training_labels_file = 'gs://mnist-public/train-labels-idx1-ubyte'
validation_images_file = 'gs://mnist-public/t10k-images-idx3-ubyte'
validation_labels_file = 'gs://mnist-public/t10k-labels-idx1-ubyte'
In [2]:
import os, re, math, json, time
import PIL.Image, PIL.ImageFont, PIL.ImageDraw
import numpy as np
import tensorflow as tf
from matplotlib import pyplot as plt
from tensorflow.python.platform import tf_logging
print("Tensorflow version " + tf.__version__)
In [3]:
tpu = None
try:
tpu = tf.distribute.cluster_resolver.TPUClusterResolver() # TPU detection relies on TPU_NAME env var
tf.tpu.experimental.initialize_tpu_system(tpu)
strategy = tf.distribute.experimental.TPUStrategy(tpu, steps_per_run=100)
print('Running on TPU ', tpu.cluster_spec().as_dict()['worker'])
except ValueError:
gpus = tf.config.experimental.list_logical_devices("GPU")
if len(gpus) > 1:
strategy = tf.distribute.MirroredStrategy([gpu.name for gpu in gpus])
print("running on multiple GPUs")
else:
strategy = tf.distribute.get_strategy() # the default strategy works on CPU and single GPU
print("Running on {}".format("a single GPU" if len(gpus)==1 else "CPU"))
# adjust batch size and learning rate for distributed computing
global_batch_size = BATCH_SIZE * strategy.num_replicas_in_sync # num replcas is 8 on a single TPU or N when runing on N GPUs.
learning_rate = LEARNING_RATE * strategy.num_replicas_in_sync
In [4]:
#@title visualization utilities [RUN ME]
"""
This cell contains helper functions used for visualization
and downloads only. You can skip reading it. There is very
little useful Keras/Tensorflow code here.
"""
# Matplotlib config
plt.rc('image', cmap='gray_r')
plt.rc('grid', linewidth=0)
plt.rc('xtick', top=False, bottom=False, labelsize='large')
plt.rc('ytick', left=False, right=False, labelsize='large')
plt.rc('axes', facecolor='F8F8F8', titlesize="large", edgecolor='white')
plt.rc('text', color='a8151a')
plt.rc('figure', facecolor='F0F0F0')# Matplotlib fonts
MATPLOTLIB_FONT_DIR = os.path.join(os.path.dirname(plt.__file__), "mpl-data/fonts/ttf")
# pull a batch from the datasets. This code is not very nice, it gets much better in eager mode (TODO)
def dataset_to_numpy_util(training_dataset, validation_dataset, N):
# get one batch from each: 10000 validation digits, N training digits
unbatched_train_ds = training_dataset.apply(tf.data.experimental.unbatch())
if tf.executing_eagerly():
# This is the TF 2.0 "eager execution" way of iterating through a tf.data.Dataset
for v_images, v_labels in validation_dataset:
break
for t_images, t_labels in unbatched_train_ds.batch(N):
break
validation_digits = v_images.numpy()
validation_labels = v_labels.numpy()
training_digits = t_images.numpy()
training_labels = t_labels.numpy()
else:
# This is the legacy TF 1.x way of iterating through a tf.data.Dataset
v_images, v_labels = validation_dataset.make_one_shot_iterator().get_next()
t_images, t_labels = unbatched_train_ds.batch(N).make_one_shot_iterator().get_next()
# Run once, get one batch. Session.run returns numpy results
with tf.Session() as ses:
(validation_digits, validation_labels,
training_digits, training_labels) = ses.run([v_images, v_labels, t_images, t_labels])
# these were one-hot encoded in the dataset
validation_labels = np.argmax(validation_labels, axis=1)
training_labels = np.argmax(training_labels, axis=1)
return (training_digits, training_labels,
validation_digits, validation_labels)
# create digits from local fonts for testing
def create_digits_from_local_fonts(n):
font_labels = []
img = PIL.Image.new('LA', (28*n, 28), color = (0,255)) # format 'LA': black in channel 0, alpha in channel 1
font1 = PIL.ImageFont.truetype(os.path.join(MATPLOTLIB_FONT_DIR, 'DejaVuSansMono-Oblique.ttf'), 25)
font2 = PIL.ImageFont.truetype(os.path.join(MATPLOTLIB_FONT_DIR, 'STIXGeneral.ttf'), 25)
d = PIL.ImageDraw.Draw(img)
for i in range(n):
font_labels.append(i%10)
d.text((7+i*28,0 if i<10 else -4), str(i%10), fill=(255,255), font=font1 if i<10 else font2)
font_digits = np.array(img.getdata(), np.float32)[:,0] / 255.0 # black in channel 0, alpha in channel 1 (discarded)
font_digits = np.reshape(np.stack(np.split(np.reshape(font_digits, [28, 28*n]), n, axis=1), axis=0), [n, 28*28])
return font_digits, font_labels
# utility to display a row of digits with their predictions
def display_digits(digits, predictions, labels, title, n):
plt.figure(figsize=(13,3))
digits = np.reshape(digits, [n, 28, 28])
digits = np.swapaxes(digits, 0, 1)
digits = np.reshape(digits, [28, 28*n])
plt.yticks([])
plt.xticks([28*x+14 for x in range(n)], predictions)
for i,t in enumerate(plt.gca().xaxis.get_ticklabels()):
if predictions[i] != labels[i]: t.set_color('red') # bad predictions in red
plt.imshow(digits)
plt.grid(None)
plt.title(title)
# utility to display multiple rows of digits, sorted by unrecognized/recognized status
def display_top_unrecognized(digits, predictions, labels, n, lines):
idx = np.argsort(predictions==labels) # sort order: unrecognized first
for i in range(lines):
display_digits(digits[idx][i*n:(i+1)*n], predictions[idx][i*n:(i+1)*n], labels[idx][i*n:(i+1)*n],
"{} sample validation digits out of {} with bad predictions in red and sorted first".format(n*lines, len(digits)) if i==0 else "", n)
In [5]:
#IS_COLAB_BACKEND = 'COLAB_GPU' in os.environ # this is always set on Colab, the value is 0 or 1 depending on GPU presence
#if IS_COLAB_BACKEND:
# from google.colab import auth
# auth.authenticate_user() # Authenticates the backend and also the TPU using your credentials so that they can access your private GCS buckets
Please read the best practices for building input pipelines with tf.data.Dataset
In [6]:
def read_label(tf_bytestring):
label = tf.io.decode_raw(tf_bytestring, tf.uint8)
label = tf.reshape(label, [])
label = tf.one_hot(label, 10)
return label
def read_image(tf_bytestring):
image = tf.io.decode_raw(tf_bytestring, tf.uint8)
image = tf.cast(image, tf.float32)/256.0
image = tf.reshape(image, [28*28])
return image
def load_dataset(image_file, label_file):
imagedataset = tf.data.FixedLengthRecordDataset(image_file, 28*28, header_bytes=16)
imagedataset = imagedataset.map(read_image, num_parallel_calls=16)
labelsdataset = tf.data.FixedLengthRecordDataset(label_file, 1, header_bytes=8)
labelsdataset = labelsdataset.map(read_label, num_parallel_calls=16)
dataset = tf.data.Dataset.zip((imagedataset, labelsdataset))
return dataset
def get_training_dataset(image_file, label_file, batch_size):
dataset = load_dataset(image_file, label_file)
dataset = dataset.cache() # this small dataset can be entirely cached in RAM, for TPU this is important to get good performance from such a small dataset
dataset = dataset.shuffle(5000, reshuffle_each_iteration=True)
dataset = dataset.repeat() # Mandatory for Keras for now
dataset = dataset.batch(batch_size, drop_remainder=True) # drop_remainder is important on TPU, batch size must be fixed
dataset = dataset.prefetch(-1) # fetch next batches while training on the current one (-1: autotune prefetch buffer size)
return dataset
def get_validation_dataset(image_file, label_file):
dataset = load_dataset(image_file, label_file)
dataset = dataset.cache() # this small dataset can be entirely cached in RAM, for TPU this is important to get good performance from such a small dataset
dataset = dataset.repeat() # Mandatory for Keras for now
dataset = dataset.batch(10000, drop_remainder=True) # 10000 items in eval dataset, all in one batch
return dataset
# instantiate the datasets
training_dataset = get_training_dataset(training_images_file, training_labels_file, global_batch_size)
validation_dataset = get_validation_dataset(validation_images_file, validation_labels_file)
In [8]:
N = 24
(training_digits, training_labels,
validation_digits, validation_labels) = dataset_to_numpy_util(training_dataset, validation_dataset, N)
display_digits(training_digits, training_labels, training_labels, "training digits and their labels", N)
display_digits(validation_digits[:N], validation_labels[:N], validation_labels[:N], "validation digits and their labels", N)
font_digits, font_labels = create_digits_from_local_fonts(N)
In [13]:
# This model trains to 99.4%— sometimes 99.5%— accuracy in 10 epochs (with a batch size of 64)
def make_model():
model = tf.keras.Sequential(
[
tf.keras.layers.Reshape(input_shape=(28*28,), target_shape=(28, 28, 1)),
tf.keras.layers.Conv2D(filters=6, kernel_size=3, padding='same', use_bias=False), # no bias necessary before batch norm
tf.keras.layers.BatchNormalization(scale=False, center=True), # no batch norm scaling necessary before "relu"
tf.keras.layers.Activation('relu'), # activation after batch norm
tf.keras.layers.Conv2D(filters=12, kernel_size=6, padding='same', use_bias=False, strides=2),
tf.keras.layers.BatchNormalization(scale=False, center=True),
tf.keras.layers.Activation('relu'),
tf.keras.layers.Conv2D(filters=24, kernel_size=6, padding='same', use_bias=False, strides=2),
tf.keras.layers.BatchNormalization(scale=False, center=True),
tf.keras.layers.Activation('relu'),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(200, use_bias=False),
tf.keras.layers.BatchNormalization(scale=False, center=True),
tf.keras.layers.Activation('relu'),
tf.keras.layers.Dropout(0.5), # Dropout on dense layer only
tf.keras.layers.Dense(10, activation='softmax')
])
model.compile(optimizer='adam', # learning rate will be set by LearningRateScheduler
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
with strategy.scope(): # the new way of handling distribution strategies in Tensorflow 1.14+
model = make_model()
# print model layers
model.summary()
# set up learning rate decay
lr_decay = tf.keras.callbacks.LearningRateScheduler(lambda epoch: learning_rate * math.pow(0.5, 1+epoch) + learning_rate/200, verbose=True)
# set up Tensorboard logs
timestamp = time.strftime("%Y-%m-%d-%H-%M-%S")
log_dir=os.path.join(BUCKET, 'mnist-logs', timestamp)
tb_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir, update_freq=50*global_batch_size)
print("Tensorboard loggs written to: ", log_dir)
In [14]:
EPOCHS = 10
steps_per_epoch = 60000//global_batch_size # 60,000 items in this dataset
print("Step (batches) per epoch: ", steps_per_epoch)
# Counting steps and batches on TPU: the tpu.keras_to_tpu_model API regards the batch size of the input dataset
# as the per-core batch size. The effective batch size is 8x more because Cloud TPUs have 8 cores. It increments
# the step by +8 everytime a global batch (8 per-core batches) is processed. Therefore batch size and steps_per_epoch
# settings can stay as they are for TPU training. The training will just go faster.
# Warning: this might change in the final version of the Keras/TPU API.
history = model.fit(training_dataset, steps_per_epoch=steps_per_epoch, epochs=EPOCHS,
validation_data=validation_dataset, validation_steps=1, callbacks=[lr_decay, tb_callback])
In [15]:
# recognize digits from local fonts
probabilities = model.predict(font_digits, steps=1)
predicted_labels = np.argmax(probabilities, axis=1)
display_digits(font_digits, predicted_labels, font_labels, "predictions from local fonts (bad predictions in red)", N)
# recognize validation digits
probabilities = model.predict(validation_digits, steps=1)
predicted_labels = np.argmax(probabilities, axis=1)
display_top_unrecognized(validation_digits, predicted_labels, validation_labels, N, 7)
In [16]:
class ServingInput(tf.keras.layers.Layer):
# the important detail in this boilerplate code is "trainable=False"
def __init__(self, name, dtype, batch_input_shape=None):
super(ServingInput, self).__init__(trainable=False, name=name, dtype=dtype, batch_input_shape=batch_input_shape)
def get_config(self):
return {'batch_input_shape': self._batch_input_shape, 'dtype': self.dtype, 'name': self.name }
def call(self, inputs):
# When the deployed model is called through its REST API,
# the JSON payload is parsed automatically, transformed into
# a tensor and passed to this input layer. You can perform
# additional transformations, such as decoding JPEGs for example,
# before sending the data to your model. However, you can only
# use tf.xxxx operations.
return inputs
# little wrinkle: must copy the model from TPU to CPU manually. This is a temporary workaround.
restored_model = make_model()
restored_model.set_weights(model.get_weights()) # this copied the weights from TPU, does nothing on GPU
# add the serving input layer
serving_model = tf.keras.Sequential()
serving_model.add(ServingInput('serving', tf.float32, (None, 28*28)))
serving_model.add(restored_model)
export_path = os.path.join(BUCKET, 'mnist-export', timestamp)
tf.saved_model.save(serving_model, export_path)
print("Model exported to: ", export_path)
Push your trained model to production on AI Platform for a serverless, autoscaled, REST API experience.
You will need a GCS bucket and a GCP project for this. Models deployed on AI Platform autoscale to zero if not used. There will be no ML Engine charges after you are done testing. Google Cloud Storage incurs charges. Empty the bucket after deployment if you want to avoid these. Once the model is deployed, the bucket is not useful anymore.
In [17]:
# Enable model deployment here
DEPLOY = False # #@param {type:"boolean"}
# Create the model only once, after that, create new versions of the same model
CREATE_MODEL = True #@param {type:"boolean"}
# Models are deployed in your cloud project
PROJECT = "" #@param {type:"string"}
MODEL_NAME = "mnist" #@param {type:"string"}
MODEL_VERSION = "v0" #@param {type:"string"}
if DEPLOY:
assert PROJECT, 'For this part, you need a GCP project. Head to http://console.cloud.google.com/ and create one.'
assert re.search(r'gs://.+', export_path), 'For this part, the model must have been exported to a GCS bucket.'
This uses the command-line interface. You can do the same thing through the ML Engine UI at https://console.cloud.google.com/mlengine/models
In [18]:
# Create the model
if DEPLOY and CREATE_MODEL:
!gcloud ai-platform models create {MODEL_NAME} --project={PROJECT} --regions=us-central1
In [19]:
# Create a version of this model (you can add --async at the end of the line to make this call non blocking)
# Additional config flags are available: https://cloud.google.com/ml-engine/reference/rest/v1/projects.models.versions
# You can also deploy a model that is stored locally by providing a --staging-bucket=... parameter
if DEPLOY:
!echo "Deployment takes a couple of minutes. You can watch your deployment here: https://console.cloud.google.com/mlengine/models/{MODEL_NAME}"
!gcloud ai-platform versions create {MODEL_VERSION} --model={MODEL_NAME} --origin="{export_path}" --project={PROJECT} --runtime-version=1.13 --python-version=3.5
In [27]:
# prepare digits to send to online prediction endpoint
digits = np.concatenate((font_digits, validation_digits[:100-N]))
labels = np.concatenate((font_labels, validation_labels[:100-N]))
with open("digits.json", "w") as f:
for digit in digits:
# the format for ML Engine online predictions is: one JSON object per line
data = json.dumps({"serving_input": digit.tolist()}) # "serving_input" because the ServingInput layer was named "serving". Keras appends "_input"
f.write(data+'\n')
In [28]:
if DEPLOY: # Request online predictions from deployed model (REST API) using the "gcloud ai-platform" command line.
predictions = !gcloud ai-platform predict --model={MODEL_NAME} --json-instances digits.json --project={PROJECT} --version {MODEL_VERSION}
print(predictions)
probabilities = np.stack([json.loads(p) for p in predictions[1:]]) # first line is the name of the input layer: drop it, parse the rest
predictions = np.argmax(probabilities, axis=1)
display_top_unrecognized(digits, predictions, labels, N, 100//N)
author: Martin Gorner
twitter: @martin_gorner
Copyright 2019 Google LLC
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
This is not an official Google product but sample code provided for an educational purpose
In [ ]: