Overview

DLPy is a high-level Python library for the SAS Deep learning features available in SAS Viya. DLPy is designed to provide an efficient way to apply deep learning methods to image, text, and audio data. DLPy APIs created following the Keras APIs with a touch of PyTorch flavor.

What's Recently Added

• Text, audio, and time series support in addition to image
• New APIs for:
• RNN based tasks: text classification, text generation, and sequence labeling
• Object detection
• Image segmentation
• Time series processing and modeling
• Processing audio files and creating speech recognition models
• Additional pre-defined network architectures such as DenseNet, DarkNet, Inception, Yolo, FasterRCNN, U-Net, MobileNet, and ShuffleNet
• Enhanced data visualization and metadata handling

Prerequisites

• Python version 3 or greater is required
• Install SAS Scripting Wrapper for Analytics Transfer (SWAT) for Python using pip install swat or conda install -c sas-institute swat
• Access to a SAS Viya 3.4 environment with Visual Data Mining and Machine Learning (VDMML) is required
• To use timeseries APIs, access to either SAS Econometrics or SAS Visual Forecasting is required
• A user login to your SAS Viya back-end is required. See your system administrator for details if you do not have a SAS Viya account.
• It is recommended that you install the open source graph visualization software called Graphviz to enable graphic visualizations of the DLPy deep learning models
• Install DLPy using pip install sas-dlpy or conda install -c sas-institute sas-dlpy

SAS Viya and VDMML versions vs. DLPY versions

DLPy versions are aligned with the SAS Viya and VDMML versions. Below is the versions matrix.

DLPy	SAS Viya	VDMML
1.2.x	3.5	8.5
1.1.x	3.4	8.4
1.0.x	3.4	8.3

The table above can be read as follows: DLPy versions between 1.0 (inclusive) to 1.1 (exclusive) are designed to work with the SAS Viya 3.4 and VDMML 8.3.

External Libraries

The following versions of external libraries are supported: - ONNX: versions >= 1.5.0 - Keras: versions >= 2.1.3

Getting Started

To connect to a SAS Viya server, import SWAT and use the swat.CAS class to create a connection:

Note: The default CAS port is 5570.

>>> import swat
>>> sess = swat.CAS('mycloud.example.com', 5570)

Next, import the DLPy package, and then build a simple convolutional neural network (CNN) model.

Import DLPy model functions:

>>> from dlpy import Model, Sequential
>>> from dlpy.layers import *

Use DLPy to create a sequential model and name it Simple_CNN:

>>> model1 = Sequential(sess, model_table = 'Simple_CNN')

Define an input layer to add to model1:

# The input shape contains RGB images (3 channels)
# The model images are 224 px in height and 224 px in width

>>> model1.add(InputLayer(3,224,224))

NOTE: Input layer added.

Add a 2D convolution layer and a pooling layer:

# Add 2-Dimensional Convolution Layer to model1
# that has 8 filters and a kernel size of 7.

>>> model1.add(Conv2d(8,7))

NOTE: Convolutional layer added.

# Add Pooling Layer of size 2

>>> model1.add(Pooling(2))

NOTE: Pooling layer added.

Add an additional pair of 2D convolution and pooling layers:

# Add another 2D convolution Layer that has 8 filters and a kernel size of 7

>>> model1.add(Conv2d(8,7))

NOTE: Convolutional layer added.

# Add a pooling layer of size 2 to # complete the second pair of layers.

>>> model1.add(Pooling(2))

NOTE: Pooling layer added.

Add a fully connected layer:

# Add Fully-Connected Layer with 16 units

>>> model1.add(Dense(16))

NOTE: Fully-connected layer added.

Finally, add the output layer:

# Add an output layer that has 2 nodes and uses
# the Softmax activation function

>>> model1.add(OutputLayer(act='softmax',n=2))

NOTE: Output layer added.
NOTE: Model compiled successfully

Additional Resources

• DLPy examples: https://github.com/sassoftware/python-dlpy/tree/master/examples
• DLPy API documentation sassoftware.github.io/python-dlpy.
• SAS SWAT for Python
• SAS ESPPy
• Watch: DLPy videos:
• DLPy v1.0 examples:
  • Image classification using CNNs
  • Object detection using TinyYOLOv2
  • Import and export deep learning models with ONNX
  • Text classification and text generation using RNNs
  • Time series forecasting using RNNs
• DLPy v1.1 examples:
  • Leverage Functional APIs to Build Complex Models
  • Image Segmentation with U-Net
  • Object Detection with Faster-RCNN
  • Image Classification with ShuffleNet and MobileNet
  • Multi-class Deep learning
• SAS Deep Learning with Python made easy using DLPy

Contributing

Have something cool to share? SAS gladly accepts pull requests on GitHub! See the Contributor Agreement for details.

Licensing

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 LICENSE.txt

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.

for more products do visit our github and contribute.

Issues

cannot load weights of bert model when i use multiple targets (target sequence)

opened on 2023-03-17 09:50:59 by witkophdw

i have project where i have to predict multiple keywords based on description of a movie.

I am able to train the bert model with 1 (binary target) , but I have train the model with multiple (431) binary targets. I got an error when I try to load the weight to cas. parameter:

• Final classification layer number of class predictions n_classes = 2. (but i tried as well with 1 (regression)

• Number of target variables (num_tgt_var = 435)

bert.load_weights('/opt/sas/viya/config/data/cas/default/public/bert-base-uncased.kerasmodel.h5' , num_target_var=num_tgt_var, # Do not freeze base model weights. # Allow layer updates with model tuning. freeze_base_model=False )

error: ERROR: Target sequence for text input is not supported. ERROR: The action stopped due to errors.

---

DLPyError Traceback (most recent call last) Input In [17], in () ----> 1 bert.load_weights('/opt/sas/viya/config/data/cas/default/public/bert-base-uncased.kerasmodel.h5' , 2 num_target_var=num_tgt_var, 3 # Do not freeze base model weights. 4 # Allow layer updates with model tuning. 5 freeze_base_model=False 6 )

File ~/python-dlpy/dlpy/transformers/bert_model.py:967, in BERT_Model.load_weights(self, path, num_target_var, freeze_base_model, use_gpu, last_frozen_layer) 964 data_spec = self.get_data_spec(num_target_var) 966 # attach layer weights --> 967 super(BERT_Model, self).load_weights(path, 968 data_spec=data_spec, 969 use_gpu=use_gpu, 970 embedding_dim=self._config['hidden_size'])
972 # determine which layers to freeze 973 self._freeze_layers = self._rnn_layer

File ~/python-dlpy/dlpy/network.py:1003, in Network.load_weights(self, path, labels, data_spec, label_file_name, label_length, use_gpu, embedding_dim) 1000 self.load_weights_from_caffe(path, labels=labels, data_spec=data_spec, label_file_name=label_file_name, 1001 label_length=label_length) 1002 elif file_name.lower().endswith('kerasmodel.h5'): -> 1003 self.load_weights_from_keras(path, labels=labels, data_spec=data_spec, label_file_name=label_file_name, 1004 label_length=label_length, use_gpu=use_gpu, embedding_dim=embedding_dim) 1005 elif file_name.lower().endswith('onnxmodel.h5'): 1006 self.load_weights_from_keras(path, labels=labels, data_spec=data_spec, label_file_name=label_file_name, 1007 label_length=label_length, use_gpu=use_gpu, embedding_dim=embedding_dim)

File ~/python-dlpy/dlpy/network.py:1077, in Network.load_weights_from_keras(self, path, labels, data_spec, label_file_name, label_length, use_gpu, embedding_dim) 1073 self.load_weights_from_file_with_labels(path=path, format_type='KERAS', data_spec=data_spec, 1074 label_file_name=label_file_name, label_length=label_length, 1075 use_gpu=use_gpu, embedding_dim=embedding_dim) 1076 else: -> 1077 self.load_weights_from_file(path=path, format_type='KERAS', data_spec=data_spec, use_gpu=use_gpu, 1078 embedding_dim=embedding_dim)

File ~/python-dlpy/dlpy/network.py:1187, in Network.load_weights_from_file(self, path, format_type, data_spec, use_gpu, embedding_dim) 1185 for msg in rt.messages: 1186 print(msg) -> 1187 raise DLPyError('Cannot import model weights, there seems to be a problem.') 1189 # create attributes if necessary 1190 if not has_data_spec:

DLPyError: Cannot import model weights, there seems to be a problem.

load audio files server and local path issues

opened on 2022-01-21 15:24:24 by tomstocksasinstitute

modelPath = '/srv/nfs/kubedata/cas-landingzone/sbxtot/sounds'

s.addcaslib(activeonadd=True,datasource={'srctype':'path'},name='mycaslib',path=modelPath,subdirectories=True) s.setsessopt(caslib='mycaslib')

local_data_test = '/srv/nfs/kubedata/cas-landingzone/sbxtot/sounds/test' server_data_test = '/srv/nfs/kubedata/cas-landingzone/sbxtot/sounds/test' local_data_train = '/srv/nfs/kubedata/cas-landingzone/sbxtot/sounds/train' server_data_train = '/srv/nfs/kubedata/cas-landingzone/sbxtot/sounds/train' local_data_validate = '/srv/nfs/kubedata/compute-landingzone/sbxtot/sounds/validate' server_data_validate = 'srv/nfs/kubedata/compute-landingzone/sbxtot/sounds/validate'

audio_table_train = AudioTable.load_audio_files(s, casout={'name':'audio_files_train','replace':'True'},
local_audio_path=local_data_train, server_audio_path=server_playground_data)

I can't understand why this procedure can't process those files and constantly throw that error:

Cannot convert file /srv/nfs/kubedata/cas-landingzone/sbxtot/sounds/train/abnormal/00000110.wav Number of files processed: 0 File conversions are finished. ERROR: A PATH type caslib is required when using filetype DOCUMENT. ERROR: The action stopped due to errors. ERROR: A PATH type caslib is required when using filetype DOCUMENT. ERROR: The action stopped due to errors.

While in the meantime, the following procedure works and create my table but it includes all the subdirectories of sounds (test, train, validate) content in one file

rt2 = s.retrieve('table.loadtable', _messagelevel='error', casout= {'name':'audio_files_train','replace':'True'}, caslib='mycaslib', path='', importOptions=dict(fileType='AUDIO', contents=True, recurse=True))

Can you let me know how to solve this issue ? Thanks, Tom.

Export ONNX

opened on 2020-10-07 03:57:48 by vorodrigues

Hi,

When you try to export a model in ONNX format, the deploy method expects the InputLayer parameters height and n_channels not to be None, even though they are meant to be used only with image data.

It would be nice not to have to specify dummy values for them when training RNNs and DNNs for example.

Best regards, Victor

model.deploy to table and s.table.loadTable don't work well together

opened on 2020-02-19 23:45:10 by smatton

If I deploy a model to table using, model.deploy(path='/path/to/dir', output_format='table') it produces 3 tables

modelname.sashdat modelname_weights.sashdat modelname_weights_attr.sashdat

If I try to load these tables with

s.table.loadTable it will complain that it can't find modelname.ATTRS.sashdat , if I rename the attributes table to what it expects it will work.

modeldetect = s.table.loadTable(caslib='scz', path='detectmobile/Model_M43t70.sashdat', loadattrs=True) modeldetectname = s.table.loadTable(caslib='scz',casout={'name':'mobiledetect'}, path='detectmobile/Model_M43t70_weights.sashdat')

create_object_detection_table error

opened on 2020-02-11 08:27:18 by teoland

Hi,

I've been trying to execute the create_object_detection_table function with no luck so far. I am using only the data_path option as iamges and cas server are in the same machine.

Attached the error I am receiving. I really don't get why it is looking for a object_x file or path.

Any ideas?

Thanks!!

Matteo

Support DepthwiseConv2d, SeparableConv2d, PointwiseConv2d

opened on 2020-02-10 21:18:22 by wenyushi

The issue relates to implementing DepthwiseConv2d, SeparableConv2d, PointwiseConv2d.

Releases

DLPy v1.2.0 2020-03-13 19:08:45

• sas-dlpy-1.2.0-doc.tar.gz

Highlights Include:

• Supporting features released with SAS Viya 3.5
• APIs for Image embeddings
• Survival analysis support
• More example notebooks
• Various bug fixes

The latest DLPy release can be installed using the following pip command:

pip install sas-dlpy

Or, if you use Anaconda:

conda install -c sas-institute sas-dlpy

DLPy v1.1.2 2019-08-13 14:01:06

• sas-dlpy-1.1.2-doc.tar.gz

Highlights Include:

• Updating and expanding the documentation
• Various bug fixes

The latest DLPy release can be installed using the following pip command:

pip install sas-dlpy

Or, if you use Anaconda:

conda install -c sas-institute sas-dlpy

DLPy v1.1.1 2019-08-05 19:17:35

• sas-dlpy-1.1.1-doc.tar.gz

Highlights Include:

• New speech-to-text APIs that enables us doing language model post processing
• APIs for Image captioning
• More example notebooks
• Updated weights (located: http://support.sas.com/documentation/prod-p/vdmml/zip/index.html) supporting new version of ONNX
• Various bug fixes

The latest DLPy release can be installed using the following pip command:

pip install sas-dlpy

Or, if you use Anaconda:

conda install -c sas-institute sas-dlpy

DLPy v1.1.0 2019-07-16 18:21:31

• sas-dlpy-1.1.0-doc.tar.gz

Highlights Include:

• New layers: transpose convolution and group convolution
• New architectures: U-Net, MobileNet, and ShuffleNet
• Supporting FasterRCNN for object detection
• Supporting image segmentation
• Various bug fixes
• Improved documentation
• Improved code coverage

The latest DLPy release can be installed using the following pip command:

pip install sas-dlpy

Or, if you use Anaconda:

conda install -c sas-institute sas-dlpy

DLPy v1.0.2 2019-03-22 17:53:13

• sas-dlpy-1.0.2-doc.tar.gz

Highlights Include:

• Various bug fixes
• A new way of creating models
• Multi-step forecasting

The latest DLPy release can be installed using the following pip command:

pip install sas-dlpy

Or, if you use Anaconda:

conda install -c sas-institute sas-dlpy

DLPy v1.0.1 2018-12-04 19:15:03

• sas-dlpy-1.0.1-doc.tar.gz
• sas-dlpy-1.0.2-doc.tar.gz

Highlights Include:

• Various bug fixes
• ONNX import and export support
• Re-organization of examples

The latest DLPy release can be installed using the following pip command:

pip install sas-dlpy

Or, if you use Anaconda:

conda install -c sas-institute sas-dlpy