| This post was kindly contributed by SAS Users - go there to comment and to read the full post. |
Editor’s Note: This article was translated and edited by SAS USA and was originally written by Makoto Unemi. The original text is here.
SAS previously provided SAS Scripting Wrapper for Analytics Transfer (SWAT), a package for using SAS Viya functions from various general-purpose programming languages such as Python.
In addition to SWAT, SAS launched Deep Learning Python (DLPy), a higher-level API package for Python, making it possible to use SAS Viya functions more efficiently from Python. In this article I outline more about what DLPy is and how it’s implementation.
About DLPy
DLPy is a high-level package for the Python API created for deep learning and image action set after Viya3.3. DLPy provides an API similar to Keras to improve the efficiency of deep learning and image processing coding. With just a little rewriting of the existing Keras code, it is possible to execute the processing on SAS Viya.
For example, below is an example of a Convolutional Neural Network (CNN) layer definition; you can see that it is very similar to Keras.
The layers supported by DLPy are: InputLayer, Conv2d, Pooling, Dense, Recurrent, BN, Res, Proj, and OutputLayer. The following is an example of learning.
DLPy functions
Introducing DLPy’s functions (partial excerpts), taking as an example the learning of multiple dolphins and giraffe images using CNN and applying test images to the model.
Implementation of major deep learning networks
DLPy offers the following pre-built deep learning models: VGG11/13/16/19, ResNet34/50/101/152, wide_resnet, and dense_net.
The following models also offer pre-trained weights using ImageNet data (these weights can be used for unique tasks by transfer learning): VGG16, VGG19, ResNet50, ResNet101, and ResNet152. The following is an example of transferring ResNet50 pre-trained weights.
CNN judgment basis information
Using the heat_map_analysis() method, you can output a colorful heat map and check where you focused on the image.
In addition, the get_feature_maps() method is used to get the feature map of each layer of CNN, and feature_maps.display() method is used to specify and display the obtained feature map layer and check can also do.
The following is the output result of layer 1 feature map.
The following is the output result of layer 18 feature map.
Deep learning & image processing related task support function
resize() method: Resize image data
as_patches() method: Image data expansion (generates a patch from the original image)
two_way_split() method: Data split (learning, testing)
plot_network() method: draws the structure of the defined deep learning layer (network) as a graphical diagram
plot_training_history() method: Iterative learning history display
predict() method: Display prediction (scoring) results
plot_predict_res() method: Display classification results
And of course, you can use DLPy to get data from a SAS Viya in-memory session, pass it to your local client, and convert it to common data formats like numpy arrays and Pandas DataFrames. The converted data can be smoothly supplied to models of other open source packages such as scikit-learn.
Regarding image classification using DLPy, videos are also available in the Deep Learning with Python (DLPy) Demo Series section of the DLPy product page.
SAS Viya: Package for Python API for deep learning and image processing: DLPy was published on SAS Users.
| This post was kindly contributed by SAS Users - go there to comment and to read the full post. |