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.
TPUs are located in Google Cloud, for optimal performance, they read data directly from Google Cloud Storage.
In [0]:
import os, re, time, json
import PIL.Image, PIL.ImageFont, PIL.ImageDraw
import numpy as np
import tensorflow as tf
from matplotlib import pyplot as plt
print("Tensorflow version " + tf.__version__)
In [0]:
#@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
batch_train_ds = training_dataset.apply(tf.data.experimental.unbatch()).batch(N)
# eager execution: loop through datasets normally
if tf.executing_eagerly():
for validation_digits, validation_labels in validation_dataset:
validation_digits = validation_digits.numpy()
validation_labels = validation_labels.numpy()
break
for training_digits, training_labels in batch_train_ds:
training_digits = training_digits.numpy()
training_labels = training_labels.numpy()
break
else:
v_images, v_labels = validation_dataset.make_one_shot_iterator().get_next()
t_images, t_labels = batch_train_ds.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)
# utility to display training and validation curves
def display_training_curves(training, validation, title, subplot):
if subplot%10==1: # set up the subplots on the first call
plt.subplots(figsize=(10,10), facecolor='#F0F0F0')
plt.tight_layout()
ax = plt.subplot(subplot)
ax.grid(linewidth=1, color='white')
ax.plot(training)
ax.plot(validation)
ax.set_title('model '+ title)
ax.set_ylabel(title)
ax.set_xlabel('epoch')
ax.legend(['train', 'valid.'])
(you can double-ckick on collapsed cells to view the non-essential code inside)
In [0]:
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
# Authenticates the Colab machine and also the TPU using your
# credentials so that they can access your private GCS buckets.
auth.authenticate_user()
In [0]:
# Detect hardware
try:
tpu = tf.distribute.cluster_resolver.TPUClusterResolver() # TPU detection
except ValueError:
tpu = None
gpus = tf.config.experimental.list_logical_devices("GPU")
# Select appropriate distribution strategy
if tpu:
tf.config.experimental_connect_to_cluster(tpu)
tf.tpu.experimental.initialize_tpu_system(tpu)
strategy = tf.distribute.experimental.TPUStrategy(tpu, steps_per_run=128) # Going back and forth between TPU and host is expensive. Better to run 128 batches on the TPU before reporting back.
print('Running on TPU ', tpu.cluster_spec().as_dict()['worker'])
elif len(gpus) > 1:
strategy = tf.distribute.MirroredStrategy([gpu.name for gpu in gpus])
print('Running on multiple GPUs ', [gpu.name for gpu in gpus])
elif len(gpus) == 1:
strategy = tf.distribute.get_strategy() # default strategy that works on CPU and single GPU
print('Running on single GPU ', gpus[0].name)
else:
strategy = tf.distribute.get_strategy() # default strategy that works on CPU and single GPU
print('Running on CPU')
print("Number of accelerators: ", strategy.num_replicas_in_sync)
In [0]:
BATCH_SIZE = 64 * strategy.num_replicas_in_sync # Gobal batch size.
# The global batch size will be automatically sharded across all
# replicas by the tf.data.Dataset API. A single TPU has 8 cores.
# The best practice is to scale the batch size by the number of
# replicas (cores). The learning rate should be increased as well.
LEARNING_RATE = 0.01
LEARNING_RATE_EXP_DECAY = 0.6 if strategy.num_replicas_in_sync == 1 else 0.7
# Learning rate computed later as LEARNING_RATE * LEARNING_RATE_EXP_DECAY**epoch
# 0.7 decay instead of 0.6 means a slower decay, i.e. a faster learnign rate.
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'
Please read the best practices for building input pipelines with tf.data.Dataset
In [0]:
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
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
dataset = dataset.batch(10000, drop_remainder=True) # 10000 items in eval dataset, all in one batch
dataset = dataset.repeat() # Mandatory for Keras for now
return dataset
# instantiate the datasets
training_dataset = get_training_dataset(training_images_file, training_labels_file, BATCH_SIZE)
validation_dataset = get_validation_dataset(validation_images_file, validation_labels_file)
In [0]:
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)
If you are not sure what cross-entropy, dropout, softmax or batch-normalization mean, head here for a crash-course: Tensorflow and deep learning without a PhD
In [0]:
# This model trains to 99.4% 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), name="image"),
tf.keras.layers.Conv2D(filters=12, 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=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.Conv2D(filters=32, 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(0.4), # 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():
model = make_model()
# print model layers
model.summary()
# set up learning rate decay
lr_decay = tf.keras.callbacks.LearningRateScheduler(
lambda epoch: LEARNING_RATE * LEARNING_RATE_EXP_DECAY**epoch,
verbose=True)
In [0]:
EPOCHS = 10
steps_per_epoch = 60000//BATCH_SIZE # 60,000 items in this dataset
print("Steps per epoch: ", steps_per_epoch)
# Little wrinkle: in the present version of Tensorfow (1.14), switching a TPU
# between training and evaluation is slow (approx. 10 sec). For small models,
# it is recommeneded to run a single eval at the end.
history = model.fit(training_dataset,
steps_per_epoch=steps_per_epoch, epochs=EPOCHS,
callbacks=[lr_decay])
final_stats = model.evaluate(validation_dataset, steps=1)
print("Validation accuracy: ", final_stats[1])
In [0]:
# 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)
Push your trained model to production on AI Platform for a serverless, autoscaled, REST API experience.
You will need a GCS (Google Cloud Storage) bucket and a GCP project for this. Models deployed on AI Platform autoscale to zero if not used. There will be no AI Platform 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 [0]:
PROJECT = "" #@param {type:"string"}
BUCKET = "gs://" #@param {type:"string", default:"jddj"}
NEW_MODEL = True #@param {type:"boolean"}
MODEL_NAME = "mnist" #@param {type:"string"}
MODEL_VERSION = "v1" #@param {type:"string"}
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://.+', BUCKET), 'For this part, you need a GCS bucket. Head to http://console.cloud.google.com/storage and create one.'
In [0]:
# Wrap the model so that we can add a serving function
class ExportModel(tf.keras.Model):
def __init__(self, model):
super().__init__(self)
self.model = model
# The serving function performig data pre- and post-processing.
# Pre-processing: images are received in uint8 format converted
# to float32 before being sent to through the model.
# Post-processing: the Keras model outputs digit probabilities. We want
# the detected digits. An additional tf.argmax is needed.
# @tf.function turns the code in this function into a Tensorflow graph that
# can be exported. This way, the model itself, as well as its pre- and post-
# processing steps are exported in the SavedModel and deployed in a single step.
@tf.function(input_signature=[tf.TensorSpec([None, 28*28], dtype=tf.uint8)])
def my_serve(self, images):
images = tf.cast(images, tf.float32)/255 # pre-processing
probabilities = self.model(images) # prediction from model
classes = tf.argmax(probabilities, axis=-1) # post-processing
return {'digits': classes}
# Must copy the model from TPU to CPU to be able to compose them.
restored_model = make_model()
restored_model.set_weights(model.get_weights()) # this copies the weights from TPU, does nothing on GPU
# create the ExportModel and export it to the Tensorflow standard SavedModel format
serving_model = ExportModel(restored_model)
export_path = os.path.join(BUCKET, 'keras_export', str(time.time()))
tf.keras.backend.set_learning_phase(0) # inference only
tf.saved_model.save(serving_model, export_path, signatures={'serving_default': serving_model.my_serve})
print("Model exported to: ", export_path)
# Note: in Tensorflow 2.0, it will also be possible to
# export to the SavedModel format using model.save():
# serving_model.save(export_path, save_format='tf')
In [0]:
# saved_model_cli: a useful too for troubleshooting SavedModels (the tool is part of the Tensorflow installation)
!saved_model_cli show --dir {export_path}
!saved_model_cli show --dir {export_path} --tag_set serve
!saved_model_cli show --dir {export_path} --tag_set serve --signature_def serving_default
# A note on naming:
# The "serve" tag set (i.e. serving functionality) is the only one exported by tf.saved_model.save
# All the other names are defined by the user in the fllowing lines of code:
# def myserve(self, images):
# ******
# return {'digits': classes}
# ******
# tf.saved_model.save(..., signatures={'serving_default': serving_model.myserve})
# ***************
This uses the command-line interface. You can do the same thing through the AI Platform UI at https://console.cloud.google.com/mlengine/models
In [0]:
# Create the model
if NEW_MODEL:
!gcloud ai-platform models create {MODEL_NAME} --project={PROJECT} --regions=us-central1
In [0]:
# 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
!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=1.14 --python-version=3.5
In [0]:
# prepare digits to send to online prediction endpoint
digits_float32 = np.concatenate((font_digits, validation_digits[:100-N])) # pixel values in [0.0, 1.0] float range
digits_uint8 = np.round(digits_float32*255).astype(np.uint8) # pixel values in [0, 255] int range
labels = np.concatenate((font_labels, validation_labels[:100-N]))
with open("digits.json", "w") as f:
for digit in digits_uint8:
# the format for AI Platform online predictions is: one JSON object per line
data = json.dumps({"images": digit.tolist()}) # "images" because that was the name you gave this parametr in the serving funtion my_serve
f.write(data+'\n')
In [0]:
# Request online predictions from deployed model (REST API) using the "gcloud ml-engine" command line.
predictions = !gcloud ai-platform predict --model={MODEL_NAME} --json-instances digits.json --project={PROJECT} --version {MODEL_VERSION}
print(predictions)
predictions = np.stack([json.loads(p) for p in predictions[1:]]) # first elemet is the name of the output layer: drop it, parse the rest
display_top_unrecognized(digits_float32, 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