This notebook demonstrates how to implement different image models on MNIST using the tf.keras API.
First things first. Configure the parameters below to match your own Google Cloud project details.
In [1]:
from datetime import datetime
import os
PROJECT = "your-project-id-here" # REPLACE WITH YOUR PROJECT ID
BUCKET = "your-bucket-id-here" # REPLACE WITH YOUR BUCKET NAME
REGION = "us-central1" # REPLACE WITH YOUR BUCKET REGION e.g. us-central1
MODEL_TYPE = "cnn" # "linear", "dnn", "dnn_dropout", or "dnn"
# Do not change these
os.environ["PROJECT"] = PROJECT
os.environ["BUCKET"] = BUCKET
os.environ["REGION"] = REGION
os.environ["MODEL_TYPE"] = MODEL_TYPE
os.environ["TFVERSION"] = "2.1" # Tensorflow version
os.environ["IMAGE_URI"] = os.path.join("gcr.io", PROJECT, "mnist_models")
In the previous notebook, mnist_linear.ipynb, we ran our code directly from the notebook. In order to run it on the AI Platform, it needs to be packaged as a python module.
The boilerplate structure for this module has already been set up in the folder mnist_models. The module lives in the sub-folder, trainer, and is designated as a python package with the empty __init__.py (mnist_models/trainer/__init__.py) file. It still needs the model and a trainer to run it, so let's make them.
Let's start with the trainer file first. This file parses command line arguments to feed into the model.
In [2]:
%%writefile mnist_models/trainer/task.py
import argparse
import json
import os
import sys
from . import model
def _parse_arguments(argv):
"""Parses command-line arguments."""
parser = argparse.ArgumentParser()
parser.add_argument(
'--model_type',
help='Which model type to use',
type=str, default='linear')
parser.add_argument(
'--epochs',
help='The number of epochs to train',
type=int, default=10)
parser.add_argument(
'--steps_per_epoch',
help='The number of steps per epoch to train',
type=int, default=100)
parser.add_argument(
'--job-dir',
help='Directory where to save the given model',
type=str, default='mnist_models/')
return parser.parse_known_args(argv)
def main():
"""Parses command line arguments and kicks off model training."""
args = _parse_arguments(sys.argv[1:])[0]
# Configure path for hyperparameter tuning.
trial_id = json.loads(
os.environ.get('TF_CONFIG', '{}')).get('task', {}).get('trial', '')
output_path = args.job_dir if not trial_id else args.job_dir + '/'
model_layers = model.get_layers(args.model_type)
image_model = model.build_model(model_layers, args.job_dir)
model_history = model.train_and_evaluate(
image_model, args.epochs, args.steps_per_epoch, args.job_dir)
if __name__ == '__main__':
main()
Next, let's group non-model functions into a util file to keep the model file simple. We'll copy over the scale and load_dataset functions from the previous lab.
In [3]:
%%writefile mnist_models/trainer/util.py
import tensorflow as tf
def scale(image, label):
"""Scales images from a 0-255 int range to a 0-1 float range"""
image = tf.cast(image, tf.float32)
image /= 255
image = tf.expand_dims(image, -1)
return image, label
def load_dataset(
data, training=True, buffer_size=5000, batch_size=100, nclasses=10):
"""Loads MNIST dataset into a tf.data.Dataset"""
(x_train, y_train), (x_test, y_test) = data
x = x_train if training else x_test
y = y_train if training else y_test
# One-hot encode the classes
y = tf.keras.utils.to_categorical(y, nclasses)
dataset = tf.data.Dataset.from_tensor_slices((x, y))
dataset = dataset.map(scale).batch(batch_size)
if training:
dataset = dataset.shuffle(buffer_size).repeat()
return dataset
Finally, let's code the models! The tf.keras API accepts an array of layers into a model object, so we can create a dictionary of layers based on the different model types we want to use. The below file has two functions: get_layers and create_and_train_model. We will build the structure of our model in get_layers. Last but not least, we'll copy over the training code from the previous lab into train_and_evaluate.
TODO 1: Define the Keras layers for a DNN model
TODO 2: Define the Keras layers for a dropout model
TODO 3: Define the Keras layers for a CNN model
Hint: These models progressively build on each other. Look at the imported tensorflow.keras.layers modules and the default values for the variables defined in get_layers for guidance.
In [4]:
%%writefile mnist_models/trainer/model.py
import os
import shutil
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
from tensorflow.keras import Sequential
from tensorflow.keras.callbacks import TensorBoard
from tensorflow.keras.layers import (
Conv2D, Dense, Dropout, Flatten, MaxPooling2D, Softmax)
from . import util
# Image Variables
WIDTH = 28
HEIGHT = 28
def get_layers(
model_type,
nclasses=10,
hidden_layer_1_neurons=400,
hidden_layer_2_neurons=100,
dropout_rate=0.25,
num_filters_1=64,
kernel_size_1=3,
pooling_size_1=2,
num_filters_2=32,
kernel_size_2=3,
pooling_size_2=2):
"""Constructs layers for a keras model based on a dict of model types."""
model_layers = {
'linear': [
Flatten(),
Dense(nclasses),
Softmax()
],
'dnn': [
Flatten(),
Dense(hidden_layer_1_neurons, activation='relu'),
Dense(hidden_layer_2_neurons, activation='relu'),
Dense(nclasses),
Softmax()
],
'dnn_dropout': [
Flatten(),
Dense(hidden_layer_1_neurons, activation='relu'),
Dense(hidden_layer_2_neurons, activation='relu'),
Dropout(dropout_rate),
Dense(nclasses),
Softmax()
],
'cnn': [
Conv2D(num_filters_1, kernel_size=kernel_size_1,
activation='relu', input_shape=(WIDTH, HEIGHT, 1)),
MaxPooling2D(pooling_size_1),
Conv2D(num_filters_2, kernel_size=kernel_size_2,
activation='relu'),
MaxPooling2D(pooling_size_2),
Flatten(),
Dense(hidden_layer_1_neurons, activation='relu'),
Dense(hidden_layer_2_neurons, activation='relu'),
Dropout(dropout_rate),
Dense(nclasses),
Softmax()
]
}
return model_layers[model_type]
def build_model(layers, output_dir):
"""Compiles keras model for image classification."""
model = Sequential(layers)
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def train_and_evaluate(model, num_epochs, steps_per_epoch, output_dir):
"""Compiles keras model and loads data into it for training."""
mnist = tf.keras.datasets.mnist.load_data()
train_data = util.load_dataset(mnist)
validation_data = util.load_dataset(mnist, training=False)
callbacks = []
if output_dir:
tensorboard_callback = TensorBoard(log_dir=output_dir)
callbacks = [tensorboard_callback]
history = model.fit(
train_data,
validation_data=validation_data,
epochs=num_epochs,
steps_per_epoch=steps_per_epoch,
verbose=2,
callbacks=callbacks)
if output_dir:
export_path = os.path.join(output_dir, 'keras_export')
model.save(export_path, save_format='tf')
return history
With everything set up, let's run locally to test the code. Some of the previous tests have been copied over into a testing script mnist_models/trainer/test.py to make sure the model still passes our previous checks. On line 13, you can specify which model types you would like to check. line 14 and line 15 has the number of epochs and steps per epoch respectively.
Moment of truth! Run the code below to check your models against the unit tests. If you see "OK" at the end when it's finished running, congrats! You've passed the tests!
In [11]:
!python3 -m mnist_models.trainer.test
Now that we know that our models are working as expected, let's run it on the Google Cloud AI Platform. We can run it as a python module locally first using the command line.
The below cell transfers some of our variables to the command line as well as create a job directory including a timestamp.
In [12]:
current_time = datetime.now().strftime("%y%m%d_%H%M%S")
model_type = 'cnn'
os.environ["MODEL_TYPE"] = model_type
os.environ["JOB_DIR"] = "mnist_models/models/{}_{}/".format(
model_type, current_time)
The cell below runs the local version of the code. The epochs and steps_per_epoch flag can be changed to run for longer or shorther, as defined in our mnist_models/trainer/task.py file.
In [13]:
%%bash
python3 -m mnist_models.trainer.task \
--job-dir=$JOB_DIR \
--epochs=5 \
--steps_per_epoch=50 \
--model_type=$MODEL_TYPE
Since we're using an unreleased version of TensorFlow on AI Platform, we can instead use a Deep Learning Container in order to take advantage of libraries and applications not normally packaged with AI Platform. Below is a simple Dockerlife which copies our code to be used in a TF2 environment.
In [14]:
%%writefile mnist_models/Dockerfile
FROM gcr.io/deeplearning-platform-release/tf2-cpu
COPY mnist_models/trainer /mnist_models/trainer
ENTRYPOINT ["python3", "-m", "mnist_models.trainer.task"]
The below command builds the image and ships it off to Google Cloud so it can be used for AI Platform. When built, it will show up here with the name mnist_models. (Click here to enable Cloud Build)
In [15]:
!docker build -f mnist_models/Dockerfile -t $IMAGE_URI ./
In [16]:
!docker push $IMAGE_URI
Finally, we can kickoff the AI Platform training job. We can pass in our docker image using the master-image-uri flag.
In [17]:
current_time = datetime.now().strftime("%y%m%d_%H%M%S")
model_type = 'cnn'
os.environ["MODEL_TYPE"] = model_type
os.environ["JOB_DIR"] = "gs://{}/mnist_{}_{}/".format(
BUCKET, model_type, current_time)
os.environ["JOB_NAME"] = "mnist_{}_{}".format(
model_type, current_time)
In [18]:
%%bash
echo $JOB_DIR $REGION $JOB_NAME
gcloud ai-platform jobs submit training $JOB_NAME \
--staging-bucket=gs://$BUCKET \
--region=$REGION \
--master-image-uri=$IMAGE_URI \
--scale-tier=BASIC_GPU \
--job-dir=$JOB_DIR \
-- \
--model_type=$MODEL_TYPE
Can't wait to see the results? Run the code below and copy the output into the Google Cloud Shell to follow.
Once you have a model you're proud of, let's deploy it! All we need to do is give AI Platform the location of the model. Below uses the keras export path of the previous job, but ${JOB_DIR}keras_export/ can always be changed to a different path.
Uncomment the delete commands below if you are getting an "already exists error" and want to deploy a new model.
In [21]:
%%bash
MODEL_NAME="mnist"
MODEL_VERSION=${MODEL_TYPE}
MODEL_LOCATION=${JOB_DIR}keras_export/
echo "Deleting and deploying $MODEL_NAME $MODEL_VERSION from $MODEL_LOCATION ... this will take a few minutes"
#yes | gcloud ai-platform versions delete ${MODEL_VERSION} --model ${MODEL_NAME}
#yes | gcloud ai-platform models delete ${MODEL_NAME}
gcloud ai-platform models create ${MODEL_NAME} --regions $REGION
gcloud ai-platform versions create ${MODEL_VERSION} \
--model ${MODEL_NAME} \
--origin ${MODEL_LOCATION} \
--framework tensorflow \
--runtime-version=2.1
To predict with the model, let's take one of the example images.
TODO 4: Write a .json file with image data to send to an AI Platform deployed model
In [22]:
import json, codecs
import tensorflow as tf
import matplotlib.pyplot as plt
from mnist_models.trainer import util
HEIGHT = 28
WIDTH = 28
IMGNO = 12
mnist = tf.keras.datasets.mnist.load_data()
(x_train, y_train), (x_test, y_test) = mnist
test_image = x_test[IMGNO]
jsondata = test_image.reshape(HEIGHT, WIDTH, 1).tolist()
json.dump(jsondata, codecs.open("test.json", "w", encoding = "utf-8"))
plt.imshow(test_image.reshape(HEIGHT, WIDTH));
Finally, we can send it to the prediction service. The output will have a 1 in the index of the corresponding digit it is predicting. Congrats! You've completed the lab!
In [23]:
%%bash
gcloud ai-platform predict \
--model=mnist \
--version=${MODEL_TYPE} \
--json-instances=./test.json
Copyright 2020 Google Inc. 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.