Business Context
A retail company wants to build a Convolutional Neural Network (CNN) to automatically classify images of its products into categories like 'shirts', 'shoes', and 'accessories'. This test scenario simulates the foundational step of constructing the model architecture layer by layer.
About the Set : deepLearn
Creation and training of deep neural networks.
Discover all actions of deepLearn Data Preparation
This scenario does not require pre-existing data in a table, as the 'addLayer' action operates on a model definition table. We will create an empty model table named 'product_cnn_model' to start.
Copied!
| 1 | /* The action builds the model architecture in a CAS table. No |
| 2 | data step is needed for this specific test. */ |
Étapes de réalisation
1
Initialize the model by adding the first layer, an INPUT layer, to a new model table named 'product_cnn_model'. This layer is configured for 128x128 pixel color images.
Copied!
| 1 | PROC CAS; |
| 2 | DEEPLEARN.addLayer / |
| 3 | modelTable={name='product_cnn_model', replace=true} |
| 4 | layer={type='input', nchannels=3, width=128, height=128, scale=1.0/255.0} |
| 5 | name='data'; |
| 6 | RUN; |
2
Add the first Convolutional layer, named 'conv1', which takes the 'data' layer as input. It uses 16 filters and a ReLU activation function.
Copied!
| 1 | PROC CAS; |
| 2 | DEEPLEARN.addLayer / |
| 3 | modelTable={name='product_cnn_model'} |
| 4 | name='conv1' |
| 5 | layer={type='convolution', nFilters=16, width=3, height=3, stride=1, act='relu'} |
| 6 | srcLayers={'data'}; |
| 7 | RUN; |
3
Add a Max Pooling layer, named 'pool1', to down-sample the feature maps from the 'conv1' layer.
Copied!
| 1 | PROC CAS; |
| 2 | DEEPLEARN.addLayer / |
| 3 | modelTable={name='product_cnn_model'} |
| 4 | name='pool1' |
| 5 | layer={type='pooling', width=2, height=2, stride=2, pool='max'} |
| 6 | srcLayers={'conv1'}; |
| 7 | RUN; |
4
Add a Fully Connected layer, named 'fc1', with 128 neurons.
Copied!
| 1 | PROC CAS; |
| 2 | DEEPLEARN.addLayer / |
| 3 | modelTable={name='product_cnn_model'} |
| 4 | name='fc1' |
| 5 | layer={type='fullconnect', n=128, act='relu'} |
| 6 | srcLayers={'pool1'}; |
| 7 | RUN; |
5
Add the final OUTPUT layer for classification, with 3 neurons (for 3 product categories) and a Softmax activation.
Copied!
| 1 | PROC CAS; |
| 2 | DEEPLEARN.addLayer / |
| 3 | modelTable={name='product_cnn_model'} |
| 4 | name='output' |
| 5 | layer={type='output', n=3, act='softmax'} |
| 6 | srcLayers={'fc1'}; |
| 7 | RUN; |
Expected Result
A complete, sequential CNN model architecture is successfully created in the 'product_cnn_model' CAS table. The final model summary should show five layers: 'data' (Input), 'conv1' (Convolution), 'pool1' (Pooling), 'fc1' (FullConnect), and 'output' (Output), connected in that order.