updated cnn.py

cnn/cnn.py

@@ -453,6 +453,16 @@ def dropout(y, dy, idx, tx):
     return y, dy
 
 def init_params(cnn,opts):
+    """
+    Setup CNN based on network parameters.
+    
+    Args:
+        cnn:    object with cnn parameters
+        opts:   options object
+
+    Returns:
+        cnn object with initial setup
+    """
     numFiltros1 = opts.imageCanal  #number of channels - number of first layer input
     dimInputX = opts.imageDimX     #original image X dimension
     dimInputY = opts.imageDimY     #original image Y dimension
@@ -539,6 +549,18 @@ def init_params(cnn,opts):
     return cnn
 
 def train_cnn(cnn,images,labels, pdir = ''):
+    """
+    Train convolutional neural network.
+    
+    Args:
+        cnn:        cnn object with initial setup and parameters
+        images:     array with images to process
+        labels:     vector with image labels
+        pdir:       directory to save plot if pdir is not empty
+
+    Returns:
+        create 'Total Cost vs Epoch' plot and return cnn object with weights
+    """
     it = 0   # number of iterations
     plotData = pd.DataFrame(columns=['Epoch', 'Iteration', 'Total Cost', 'Cost'])
 
@@ -915,6 +937,7 @@ def loadMNISTLabels(filename):
 # Demo with MNIST dataset
 #=====================================
 if __name__ == '__main__':
+
     opts = obj()
     opts.alpha = 1e-1     # learning rate
     opts.batchsize = 50   # training set size = 150
@@ -982,7 +1005,7 @@ def loadMNISTLabels(filename):
     # cnn.layers[3].dimPool = 2         # filter size
     # cnn.layers[3].criterio = 'max'    # max/mean
 
-    # full connected layer
+    # full connected layer: last layer setup is optional
     # cnn.layers[2] = obj()              #
     # cnn.layers[2].type = 'f'           # f = full connected
     # cnn.layers[2].fativ = 'sig'        # sig/relu: activation funciton