Skip to content

Commit 1c46424

Browse files
committed
Added: TensorFlow local mode example
1 parent baf726f commit 1c46424

File tree

3 files changed

+363
-0
lines changed

3 files changed

+363
-0
lines changed
Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,10 @@
1+
2+
{
3+
"default-runtime": "nvidia",
4+
"runtimes": {
5+
"nvidia": {
6+
"path": "/usr/bin/nvidia-container-runtime",
7+
"runtimeArgs": []
8+
}
9+
}
10+
}
Lines changed: 68 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,68 @@
1+
#!/bin/bash
2+
3+
# Do we have GPU support?
4+
nvidia-smi > /dev/null 2>&1
5+
if [ $? -eq 0 ]; then
6+
# check if we have nvidia-docker
7+
NVIDIA_DOCKER=`rpm -qa | grep -c nvidia-docker2`
8+
if [ $NVIDIA_DOCKER -eq 0 ]; then
9+
# Install nvidia-docker2
10+
#sudo pkill -SIGHUP dockerd
11+
sudo yum -y remove docker
12+
sudo yum -y install docker-17.09.1ce-1.111.amzn1
13+
14+
sudo /etc/init.d/docker start
15+
16+
curl -s -L https://nvidia.github.io/nvidia-docker/amzn1/nvidia-docker.repo | sudo tee /etc/yum.repos.d/nvidia-docker.repo
17+
sudo yum install -y nvidia-docker2
18+
sudo cp daemon.json /etc/docker/daemon.json
19+
sudo pkill -SIGHUP dockerd
20+
echo "installed nvidia-docker2"
21+
else
22+
echo "nvidia-docker2 already installed. We are good to go!"
23+
fi
24+
fi
25+
26+
# This is common for both GPU and CPU instances
27+
28+
# check if we have docker-compose
29+
docker-compose version >/dev/null 2>&1
30+
if [ $? -ne 0 ]; then
31+
# install docker compose
32+
pip install docker-compose
33+
fi
34+
35+
# check if we need to configure our docker interface
36+
SAGEMAKER_NETWORK=`docker network ls | grep -c sagemaker-local`
37+
if [ $SAGEMAKER_NETWORK -eq 0 ]; then
38+
docker network create --driver bridge sagemaker-local
39+
fi
40+
41+
# Notebook instance Docker networking fixes
42+
RUNNING_ON_NOTEBOOK_INSTANCE=`sudo iptables -S OUTPUT -t nat | grep -c 169.254.0.2`
43+
44+
# Get the Docker Network CIDR and IP for the sagemaker-local docker interface.
45+
SAGEMAKER_INTERFACE=br-`docker network ls | grep sagemaker-local | cut -d' ' -f1`
46+
DOCKER_NET=`ip route | grep $SAGEMAKER_INTERFACE | cut -d" " -f1`
47+
DOCKER_IP=`ip route | grep $SAGEMAKER_INTERFACE | cut -d" " -f12`
48+
49+
# check if both IPTables and the Route Table are OK.
50+
IPTABLES_PATCHED=`sudo iptables -S PREROUTING -t nat | grep -c 169.254.0.2`
51+
ROUTE_TABLE_PATCHED=`sudo ip route show table agent | grep -c $SAGEMAKER_INTERFACE`
52+
53+
if [ $RUNNING_ON_NOTEBOOK_INSTANCE -gt 0 ]; then
54+
55+
if [ $ROUTE_TABLE_PATCHED -eq 0 ]; then
56+
# fix routing
57+
sudo ip route add $DOCKER_NET via $DOCKER_IP dev $SAGEMAKER_INTERFACE table agent
58+
else
59+
echo "SageMaker instance route table setup is ok. We are good to go."
60+
fi
61+
62+
if [ $IPTABLES_PATCHED -eq 0 ]; then
63+
sudo iptables -t nat -A PREROUTING -i $SAGEMAKER_INTERFACE -d 169.254.169.254/32 -p tcp -m tcp --dport 80 -j DNAT --to-destination 169.254.0.2:9081
64+
echo "iptables for Docker setup done"
65+
else
66+
echo "SageMaker instance routing for Docker is ok. We are good to go!"
67+
fi
68+
fi
Lines changed: 285 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,285 @@
1+
{
2+
"cells": [
3+
{
4+
"cell_type": "markdown",
5+
"metadata": {},
6+
"source": [
7+
"# Tensorflow MNIST local training \n",
8+
"\n",
9+
"## Pre-requisites\n",
10+
"\n",
11+
"This notebook shows how to use the SageMaker Python SDK to run your code in a local container before deploying to SageMaker's managed training or hosting environments. This can speed up iterative testing and debugging while using the same familiar Python SDK interface. Just change your estimator's `train_instance_type` to `local` (or `local_gpu` if you're using an ml.p2 or ml.p3 notebook instance).\n",
12+
"\n",
13+
"In order to use this feature you'll need to install docker-compose (and nvidia-docker if training with a GPU).\n",
14+
"\n",
15+
"**Note, you can only run a single local notebook at one time.**"
16+
]
17+
},
18+
{
19+
"cell_type": "code",
20+
"execution_count": null,
21+
"metadata": {},
22+
"outputs": [],
23+
"source": [
24+
"!/bin/bash ./setup.sh"
25+
]
26+
},
27+
{
28+
"cell_type": "markdown",
29+
"metadata": {},
30+
"source": [
31+
"## Overview\n",
32+
"\n",
33+
"The **SageMaker Python SDK** helps you deploy your models for training and hosting in optimized, productions ready containers in SageMaker. The SageMaker Python SDK is easy to use, modular, extensible and compatible with TensorFlow and MXNet. This tutorial focuses on how to create a convolutional neural network model to train the [MNIST dataset](http://yann.lecun.com/exdb/mnist/) using **TensorFlow in local mode**.\n",
34+
"\n",
35+
"### Set up the environment Set up the environment"
36+
]
37+
},
38+
{
39+
"cell_type": "code",
40+
"execution_count": null,
41+
"metadata": {},
42+
"outputs": [],
43+
"source": [
44+
"import os\n",
45+
"import subprocess\n",
46+
"import sagemaker\n",
47+
"from sagemaker import get_execution_role\n",
48+
"\n",
49+
"sagemaker_session = sagemaker.Session()\n",
50+
"\n",
51+
"instance_type = 'local'\n",
52+
"\n",
53+
"if subprocess.call('nvidia-smi') == 0:\n",
54+
" ## Set type to GPU if one is present\n",
55+
" instance_type = 'local_gpu'\n",
56+
" \n",
57+
"print(\"Instance type = \" + instance_type)\n",
58+
"\n",
59+
"role = get_execution_role()"
60+
]
61+
},
62+
{
63+
"cell_type": "markdown",
64+
"metadata": {},
65+
"source": [
66+
"### Download the MNIST dataset"
67+
]
68+
},
69+
{
70+
"cell_type": "code",
71+
"execution_count": null,
72+
"metadata": {
73+
"scrolled": false
74+
},
75+
"outputs": [],
76+
"source": [
77+
"import utils\n",
78+
"from tensorflow.contrib.learn.python.learn.datasets import mnist\n",
79+
"import tensorflow as tf\n",
80+
"\n",
81+
"data_sets = mnist.read_data_sets('data', dtype=tf.uint8, reshape=False, validation_size=5000)\n",
82+
"\n",
83+
"utils.convert_to(data_sets.train, 'train', 'data')\n",
84+
"utils.convert_to(data_sets.validation, 'validation', 'data')\n",
85+
"utils.convert_to(data_sets.test, 'test', 'data')"
86+
]
87+
},
88+
{
89+
"cell_type": "markdown",
90+
"metadata": {},
91+
"source": [
92+
"### Upload the data\n",
93+
"We use the ```sagemaker.Session.upload_data``` function to upload our datasets to an S3 location. The return value inputs identifies the location -- we will use this later when we start the training job."
94+
]
95+
},
96+
{
97+
"cell_type": "code",
98+
"execution_count": null,
99+
"metadata": {},
100+
"outputs": [],
101+
"source": [
102+
"inputs = sagemaker_session.upload_data(path='data', key_prefix='data/mnist')"
103+
]
104+
},
105+
{
106+
"cell_type": "markdown",
107+
"metadata": {},
108+
"source": [
109+
"# Construct a script for training \n",
110+
"Here is the full code for the network model:"
111+
]
112+
},
113+
{
114+
"cell_type": "code",
115+
"execution_count": null,
116+
"metadata": {
117+
"scrolled": false
118+
},
119+
"outputs": [],
120+
"source": [
121+
"!cat 'mnist.py'"
122+
]
123+
},
124+
{
125+
"cell_type": "markdown",
126+
"metadata": {},
127+
"source": [
128+
"The script here is and adaptation of the [TensorFlow MNIST example](https://github.com/tensorflow/models/tree/master/official/mnist). It provides a ```model_fn(features, labels, mode)```, which is used for training, evaluation and inference. \n",
129+
"\n",
130+
"## A regular ```model_fn```\n",
131+
"\n",
132+
"A regular **```model_fn```** follows the pattern:\n",
133+
"1. [defines a neural network](https://github.com/tensorflow/models/blob/master/official/mnist/mnist.py#L96)\n",
134+
"- [applies the ```features``` in the neural network](https://github.com/tensorflow/models/blob/master/official/mnist/mnist.py#L178)\n",
135+
"- [if the ```mode``` is ```PREDICT```, returns the output from the neural network](https://github.com/tensorflow/models/blob/master/official/mnist/mnist.py#L186)\n",
136+
"- [calculates the loss function comparing the output with the ```labels```](https://github.com/tensorflow/models/blob/master/official/mnist/mnist.py#L188)\n",
137+
"- [creates an optimizer and minimizes the loss function to improve the neural network](https://github.com/tensorflow/models/blob/master/official/mnist/mnist.py#L193)\n",
138+
"- [returns the output, optimizer and loss function](https://github.com/tensorflow/models/blob/master/official/mnist/mnist.py#L205)\n",
139+
"\n",
140+
"## Writing a ```model_fn``` for distributed training\n",
141+
"When distributed training happens, the same neural network will be sent to the multiple training instances. Each instance will predict a batch of the dataset, calculate loss and minimize the optimizer. One entire loop of this process is called **training step**.\n",
142+
"\n",
143+
"### Syncronizing training steps\n",
144+
"A [global step](https://www.tensorflow.org/api_docs/python/tf/train/global_step) is a global variable shared between the instances. It necessary for distributed training, so the optimizer will keep track of the number of **training steps** between runs: \n",
145+
"\n",
146+
"```python\n",
147+
"train_op = optimizer.minimize(loss, tf.train.get_or_create_global_step())\n",
148+
"```\n",
149+
"\n",
150+
"That is the only required change for distributed training!"
151+
]
152+
},
153+
{
154+
"cell_type": "markdown",
155+
"metadata": {},
156+
"source": [
157+
"## Create a training job using the sagemaker.TensorFlow estimator\n",
158+
"\n",
159+
"The `TensorFlow` class allows us to run our training function on SageMaker. We need to configure it with our training script, an IAM role, the number of training instances, and the training instance type. Here is the the only difference from [tensorflow_distributed_mnist.ipynb](./tensorflow_distributed_mnist.ipynb) is that instead of ``train_instance_type='ml.c4.xlarge'``, we set it to ``train_instance_type='local'``. For local training with GPU, we could set this to \"local_gpu\". In this case, `instance_type` was set above based on your whether you're running a GPU instance.\n",
160+
"\n",
161+
"After we've constructed our `TensorFlow` object, we fit it using the data we uploaded to S3. Even though we're in local mode, using S3 as our data source makes sense because it maintains consistency with how SageMaker's distributed, managed training ingests data."
162+
]
163+
},
164+
{
165+
"cell_type": "code",
166+
"execution_count": null,
167+
"metadata": {
168+
"scrolled": false
169+
},
170+
"outputs": [],
171+
"source": [
172+
"from sagemaker.tensorflow import TensorFlow\n",
173+
"\n",
174+
"mnist_estimator = TensorFlow(entry_point='mnist.py',\n",
175+
" role=role,\n",
176+
" training_steps=10, \n",
177+
" evaluation_steps=10,\n",
178+
" train_instance_count=1,\n",
179+
" train_instance_type=instance_type)\n",
180+
"\n",
181+
"mnist_estimator.fit(inputs)"
182+
]
183+
},
184+
{
185+
"cell_type": "markdown",
186+
"metadata": {},
187+
"source": [
188+
"The **```fit```** method will create a training job in two **ml.c4.xlarge** instances. The logs above will show the instances doing training, evaluation, and incrementing the number of **training steps**. \n",
189+
"\n",
190+
"In the end of the training, the training job will generate a saved model for TF serving."
191+
]
192+
},
193+
{
194+
"cell_type": "markdown",
195+
"metadata": {
196+
"collapsed": true
197+
},
198+
"source": [
199+
"# Deploy the trained model to prepare for predictions\n",
200+
"\n",
201+
"The deploy() method creates an endpoint (in this case locally) which serves prediction requests in real-time."
202+
]
203+
},
204+
{
205+
"cell_type": "code",
206+
"execution_count": null,
207+
"metadata": {},
208+
"outputs": [],
209+
"source": [
210+
"mnist_predictor = mnist_estimator.deploy(initial_instance_count=1,\n",
211+
" instance_type=instance_type)"
212+
]
213+
},
214+
{
215+
"cell_type": "markdown",
216+
"metadata": {},
217+
"source": [
218+
"# Invoking the endpoint"
219+
]
220+
},
221+
{
222+
"cell_type": "code",
223+
"execution_count": null,
224+
"metadata": {},
225+
"outputs": [],
226+
"source": [
227+
"import numpy as np\n",
228+
"from tensorflow.examples.tutorials.mnist import input_data\n",
229+
"\n",
230+
"mnist = input_data.read_data_sets(\"/tmp/data/\", one_hot=True)\n",
231+
"\n",
232+
"for i in range(10):\n",
233+
" data = mnist.test.images[i].tolist()\n",
234+
" tensor_proto = tf.make_tensor_proto(values=np.asarray(data), shape=[1, len(data)], dtype=tf.float32)\n",
235+
" predict_response = mnist_predictor.predict(tensor_proto)\n",
236+
" \n",
237+
" print(\"========================================\")\n",
238+
" label = np.argmax(mnist.test.labels[i])\n",
239+
" print(\"label is {}\".format(label))\n",
240+
" prediction = predict_response['outputs']['classes']['int64Val'][0]\n",
241+
" print(\"prediction is {}\".format(prediction))"
242+
]
243+
},
244+
{
245+
"cell_type": "markdown",
246+
"metadata": {},
247+
"source": [
248+
"# Clean-up\n",
249+
"\n",
250+
"Deleting the local endpoint when you're finished is important since you can only run one local endpoint at a time."
251+
]
252+
},
253+
{
254+
"cell_type": "code",
255+
"execution_count": null,
256+
"metadata": {},
257+
"outputs": [],
258+
"source": [
259+
"mnist_estimator.delete_endpoint()"
260+
]
261+
}
262+
],
263+
"metadata": {
264+
"kernelspec": {
265+
"display_name": "conda_tensorflow_p27",
266+
"language": "python",
267+
"name": "conda_tensorflow_p27"
268+
},
269+
"language_info": {
270+
"codemirror_mode": {
271+
"name": "ipython",
272+
"version": 2
273+
},
274+
"file_extension": ".py",
275+
"mimetype": "text/x-python",
276+
"name": "python",
277+
"nbconvert_exporter": "python",
278+
"pygments_lexer": "ipython2",
279+
"version": "2.7.14"
280+
},
281+
"notice": "Copyright 2017 Amazon.com, Inc. or its affiliates. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the \"License\"). You may not use this file except in compliance with the License. A copy of the License is located at http://aws.amazon.com/apache2.0/ or in the \"license\" file accompanying this file. This file 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."
282+
},
283+
"nbformat": 4,
284+
"nbformat_minor": 2
285+
}

0 commit comments

Comments
 (0)