Amazon Sagemaker specific TensorFlow extensions.
Project description
SageMaker specific extensions to TensorFlow, for Python 2.7, 3.4-3.6 and TensorFlow versions 1.7-1.12. This package includes the PipeModeDataset
class, that allows SageMaker Pipe Mode channels to be read using TensorFlow Datasets.
Install
You can build SageMaker TensorFlow into a docker image with the following command:
pip install sagemaker-tensorflow
You can also install sagemaker-tensorflow for a specific version of TensorFlow. The following command will install sagemaker-tensorflow for TensorFlow 1.7:
pip install "sagemaker-tensorflow>=1.7,<1.8"
Build and install from source
The SageMaker TensorFlow build depends on the following:
cmake
tensorflow
curl-dev
To install these run:
pip install cmake tensorflow
On Amazon Linux, curl-dev can be installed with:
yum install curl-dev
On Ubuntu, curl-dev can be installed with:
apt-get install libcurl4-openssl-dev
To build and install this package, run:
pip install .
in this directory.
To build in a SageMaker docker image, you can use the following RUN command in your Dockerfile:
RUN git clone https://github.com/aws/sagemaker-tensorflow-extensions.git && \ cd sagemaker-tensorflow-extensions && \ pip install . && \ cd .. && \ rm -rf sagemaker-tensorflow-extensions
Building for a specific TensorFlow version
Release branching is used to track different versions of TensorFlow. To build for a specific release of TensorFlow, checkout the release branch prior to running a pip install. For example, to build for TensorFlow 1.7, you can run the following command in your Dockerfile:
RUN git clone https://github.com/aws/sagemaker-tensorflow-extensions.git && \ cd sagemaker-tensorflow-extensions && \ git checkout 1.7 && \ pip install . && \ cd .. && \ rm -rf sagemaker-tensorflow-extensions
Requirements
SageMaker TensorFlow extensions builds on Python 2.7, 3.4-3.6 in Linux with a TensorFlow version >= 1.7. Older versions of TensorFlow are not supported. Please make sure to checkout the branch of sagemaker-tensorflow-extensions that matches your TensorFlow version.
SageMaker Pipe Mode
SageMaker Pipe Mode is a mechanism for providing S3 data to a training job via Linux fifos. Training programs can read from the fifo and get high-throughput data transfer from S3, without managing the S3 access in the program itself.
SageMaker Pipe Mode is enabled when a SageMaker training job is created. Multiple S3 datasets can be mapped to individual fifos, configured in the training request. Pipe Mode is covered in more detail in the SageMaker documentation: https://docs.aws.amazon.com/sagemaker/latest/dg/your-algorithms-training-algo.html#your-algorithms-training-algo-running-container-inputdataconfig
Using the PipeModeDataset
The PipeModeDataset
is a TensorFlow Dataset
for reading SageMaker Pipe Mode channels. After installing this package, the PipeModeDataset
can be imported from a moduled named sagemaker_tensorflow
.
To construct a PipeModeDataset
that reads TFRecord encoded records from a “training” channel, do the following:
from sagemaker_tensorflow import PipeModeDataset
ds = PipeModeDataset(channel='training', record_format='TFRecord')
A PipeModeDataset
should be created for a SageMaker Pipe Mode channel. Each channel corresponds to a single S3 dataset, configured when the training job is created. You can create multiple PipeModeDataset
instances over different channels to read from multiple S3 datasets in the same training job.
A PipeModeDataset
can read TFRecord, RecordIO, or text line records, by using the record_format
constructor argument. The record_format
keyword argument can be set to either RecordIO
, TFRecord
, or TextLine
to differentiate between the three encodings. RecordIO
is the default.
A PipeModeDataset
is a regular TensorFlow Dataset
and as such can be used in TensorFlow input processing pipelines, and in TensorFlow Estimator input_fn
definitions. All Dataset
operations are supported on PipeModeDataset
. The following code snippet shows how to create a batching and parsing Dataset
that reads data from a SageMaker Pipe Mode channel:
features = {
'data': tf.FixedLenFeature([], tf.string),
'labels': tf.FixedLenFeature([], tf.int64),
}
def parse(record):
parsed = tf.parse_single_example(record, features)
return ({
'data': tf.decode_raw(parsed['data'], tf.float64)
}, parsed['labels'])
ds = PipeModeDataset(channel='training', record_format='TFRecord')
num_epochs = 20
ds = ds.repeat(num_epochs)
ds = ds.prefetch(10)
ds = ds.map(parse, num_parallel_calls=10)
ds = ds.batch(64)
Using the PipeModeDataset with the SageMaker Python SDK
The sagemaker_tensorflow
module is available for TensorFlow scripts to import when launched on SageMaker via the SageMaker Python SDK. If you are using the SageMaker Python SDK TensorFlow
Estimator to launch TensorFlow training on SageMaker, note that the default channel name is training
when just a single S3 URI is passed to fit
.
Using the PipeModeDataset with SageMaker Augmented Manifest Files
SageMaker Augmented Manifest Files provide a mechanism to associate metdata (such as labels) with binary data (like images) for training. An Augmented Manifest File is a single json-lines file, stored as an object in S3. During training, SageMaker reads the data from an Augmented Manifest File and passes the data to the running training job, through a SageMaker Pipe Mode channel.
To learn more about preparing and using an Augmented Manifest File, please consult the SageMaker documentation on Augmented Manifest Files here.
You can use the PipeModeDataset to read data from a Pipe Mode channel that is backed by an Augmented Manifest, by following these guidelines:
First, use a Dataset batch
operation to combine successive records into a single tuple. Each attribute in an Augmented Manifest File record is queued into the Pipe Mode’s fifo as a separate record. By batching, you can combine these successive per-attribute records into a single per-record tuple. In general, if your Augmented Manifest File contains n attributes, then you should issue a call to batch(n)
on your PipeModeDataset and then use a simple combining function applied with a map
to combine each per-attribute record in the batch into a single tuple. For example, assume your Augmented Manifest File contains 3 attributes, the following code sample will read Augmented Manifest records into a 3-tuple of string Tensors when applied to a PipeModeDataset.
ds = PipeModeDataset("my_channel")
def combine(records):
return (records[0], records[1], records[2])
ds = ds.batch(3) # Batch series of three attributes together.
ds = ds.map(combine) # Convert each batch of three records into a single tuple with three Tensors.
# Perform other operations on the Dataset - e.g. subsequent batching, decoding
...
Second, pass "RecordIO"
as the value for RecordWrapperType
when you launch the SageMaker training job with an Augmented Manifest File. Doing this will cause SageMaker to wrap each per-attribute record in a RecordIO wrapper, enabling the PipeModeDataset to separate these records.
Third, ensure your PipeModeDataset splits records using RecordIO decoding in your training script. You can do this by simply constructing the PipeModeDataset with no record_format
argument, as RecordIO is the default record wrapping type for the PipeModeDataset.
If you follow these steps then the PipeModeDataset will produce tuples of string Tensors that you can then decode or process further (for example, by doing a jpeg decode if your data are images).
Support
We’re here to help. Have a question? Please open a GitHub issue, we’d love to hear from you.
License
SageMaker TensorFlow is licensed under the Apache 2.0 License. It is copyright 2018 Amazon.com, Inc. or its affiliates. All Rights Reserved. The license is available at: http://aws.amazon.com/apache2.0/
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distributions
Built Distributions
Hashes for sagemaker_tensorflow-2.1.0.1.0.0-cp36-cp36m-manylinux1_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 4a889a5327a3149a8a4a9be63284ed55b3dccd330f26ceffd44b65023dd78a53 |
|
MD5 | cd98e4351f889bec650b9a579b7fef40 |
|
BLAKE2b-256 | 3ef24c9796793a06d9fd48497c00f1f5d518d6fb207b3aa6fe0aaeed39bc88cc |
Hashes for sagemaker_tensorflow-2.1.0.1.0.0-cp27-cp27mu-manylinux1_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 1cf755467a1b698c3e7bafabad6f56ca5363ca2e1bae014dafccda3a50e2c375 |
|
MD5 | 7d7213a605c3ada95fd77cad58f5d02f |
|
BLAKE2b-256 | 8be15d3a56ed3b4ade695393af43a6cbf7d781c34d5e587c4c1ca026969f0301 |