Business Context
A financial institution needs to verify customer signatures by comparing a new signature image with a reference image on file. This requires a Siamese Network, which uses two identical, parallel CNN branches that share the same weights to process the two images and produce comparable feature vectors.
About the Set : deepLearn
Creation and training of deep neural networks.
Discover all actions of deepLearn Data Preparation
This scenario focuses on building the complex model architecture. We will create an empty model table named 'siamese_signature_model'.
Copied!
| 1 | /* Model architecture test. No |
| 2 | data step needed. */ |
Étapes de réalisation
1
Initialize the model and add two separate INPUT layers, one for the 'reference_signature' and one for the 'new_signature'.
Copied!
| 1 | PROC CAS; |
| 2 | DEEPLEARN.addLayer / modelTable={name='siamese_signature_model', replace=true} name='input_ref' layer={type='input', nchannels=1, width=150, height=50}; |
| 3 | DEEPLEARN.addLayer / modelTable={name='siamese_signature_model'} name='input_new' layer={type='input', nchannels=1, width=150, height=50}; |
| 4 | RUN; |
2
Build the first branch of the network. Add a Convolutional layer ('conv1_ref') and a Pooling layer ('pool1_ref') connected to the 'input_ref' layer.
Copied!
| 1 | PROC CAS; |
| 2 | DEEPLEARN.addLayer / modelTable={name='siamese_signature_model'} name='conv1_ref' layer={type='convolution', nFilters=8, width=5, act='relu'} srcLayers={'input_ref'}; |
| 3 | DEEPLEARN.addLayer / modelTable={name='siamese_signature_model'} name='pool1_ref' layer={type='pooling', width=2, pool='max'} srcLayers={'conv1_ref'}; |
| 4 | RUN; |
3
Build the second branch. Add a Convolutional layer ('conv1_new') connected to 'input_new'. Critically, use the 'sharingWeights' parameter to share weights from 'conv1_ref'.
Copied!
| 1 | PROC CAS; |
| 2 | DEEPLEARN.addLayer / |
| 3 | modelTable={name='siamese_signature_model'} |
| 4 | name='conv1_new' |
| 5 | layer={type='convolution', nFilters=8, width=5, act='relu'} |
| 6 | srcLayers={'input_new'} |
| 7 | sharingWeights='conv1_ref'; |
| 8 | RUN; |
4
Add the corresponding Pooling layer for the second branch, also sharing weights from its counterpart in the first branch (though pooling layers have no weights, this tests syntax).
Copied!
| 1 | PROC CAS; |
| 2 | DEEPLEARN.addLayer / |
| 3 | modelTable={name='siamese_signature_model'} |
| 4 | name='pool1_new' |
| 5 | layer={type='pooling', width=2, pool='max'} |
| 6 | srcLayers={'conv1_new'} |
| 7 | sharingWeights='pool1_ref'; |
| 8 | RUN; |
5
Add a CONCAT layer to merge the outputs from the two parallel branches ('pool1_ref' and 'pool1_new') before feeding them to the final decision layers.
Copied!
| 1 | PROC CAS; |
| 2 | DEEPLEARN.addLayer / |
| 3 | modelTable={name='siamese_signature_model'} |
| 4 | name='concat_features' |
| 5 | layer={type='concat'} |
| 6 | srcLayers={'pool1_ref', 'pool1_new'}; |
| 7 | RUN; |
Expected Result
A Siamese network architecture is created in the 'siamese_signature_model' table. The model summary should show two input layers and two parallel branches ('conv1_ref'/'pool1_ref' and 'conv1_new'/'pool1_new') whose outputs are merged by the 'concat_features' layer. The model definition should confirm that 'conv1_new' shares weights with 'conv1_ref'.