Function main

cnn_class/cnn_theano.py:63–185

()

def main():
    # step 1: load the data, transform as needed
    train, test = get_data()

    # Need to scale! don't leave as 0..255
    # Y is a N x 1 matrix with values 1..10 (MATLAB indexes by 1)
    # So flatten it and make it 0..9
    # Also need indicator matrix for cost calculation
    Xtrain = rearrange(train['X'])
    Ytrain = train['y'].flatten() - 1
    del train
    Xtrain, Ytrain = shuffle(Xtrain, Ytrain)

    Xtest  = rearrange(test['X'])
    Ytest  = test['y'].flatten() - 1
    del test


    max_iter = 6
    print_period = 10

    lr = np.float32(1e-3)
    mu = np.float32(0.9)

    N = Xtrain.shape[0]
    batch_sz = 500
    n_batches = N // batch_sz

    M = 500
    K = 10
    poolsz = (2, 2)

    # after conv will be of dimension 32 - 5 + 1 = 28
    # after downsample 28 / 2 = 14
    W1_shape = (20, 3, 5, 5) # (num_feature_maps, num_color_channels, filter_width, filter_height)
    W1_init = init_filter(W1_shape, poolsz)
    b1_init = np.zeros(W1_shape[0], dtype=np.float32) # one bias per output feature map

    # after conv will be of dimension 14 - 5 + 1 = 10
    # after downsample 10 / 2 = 5
    W2_shape = (50, 20, 5, 5) # (num_feature_maps, old_num_feature_maps, filter_width, filter_height)
    W2_init = init_filter(W2_shape, poolsz)
    b2_init = np.zeros(W2_shape[0], dtype=np.float32)

    # vanilla ANN weights
    W3_init = np.random.randn(W2_shape[0]*5*5, M) / np.sqrt(W2_shape[0]*5*5 + M)
    b3_init = np.zeros(M, dtype=np.float32)
    W4_init = np.random.randn(M, K) / np.sqrt(M + K)
    b4_init = np.zeros(K, dtype=np.float32)


    # step 2: define theano variables and expressions
    X = T.tensor4('X', dtype='float32')
    Y = T.ivector('T')
    W1 = theano.shared(W1_init, 'W1')
    b1 = theano.shared(b1_init, 'b1')
    W2 = theano.shared(W2_init, 'W2')
    b2 = theano.shared(b2_init, 'b2')