count number of parameters in CNN numerically in each and every layer

[Vinayakumar R]

nb_filter = 64

filter_length = 3

pool_length=2

dense (128)

dropout(0.5)

Attaching the image

Could you please tell how to compute the numerically. Is there any formula or method to obtain the number of parameters of each and every layer

[Klemen Grm]

params_count = 0
for layer in model.layers:
    weights_list = layer.get_weights()
    for array in weights_list:
        params_count += np.prod(array.shape)

[Vinayakumar R]

No, i need a manual calculation. Already the attached fig shows the total parameters and each layer total parameters

[Klemen Grm]

For 2d convolutional layers: (no. of channels in input tensor) * (no. of channels in output tensor) * (kernel height * kernel width) + (no. of channels in output tensor)
For dense layers: (no. of input features) * (no. of output features + 1)

[Vinayakumar R]

According to your code the displayed parameters are 

192

256

164096

164224

164864

164869

But is there any equation to manually calculate the number of parameters in each layer

[Vinayakumar R]

I am using convolution1d

[Klemen Grm]

1d convolutions are equivalent to 2d convolutions with one of the kernel dimensions set to 1.

On Tuesday, 24 January 2017 14:04:42 UTC+1, Vinayakumar R wrote:

I am using convolution1d

[Vinayakumar R]

For your code my architecture is giving the below parameters. Could you please help how I am getting these many parameters numerically in each layer. I am new to this

192

256

164096

164224

164864

164869

with detailed description as 

____________________________________________________________________________________________________

Layer (type)                     Output Shape          Param #     Connected to

====================================================================================================

convolution1d_1 (Convolution1D)  (None, 41, 64)        256         convolution1d_input_1[0][0]

____________________________________________________________________________________________________

maxpooling1d_1 (MaxPooling1D)    (None, 20, 64)        0           convolution1d_1[0][0]

____________________________________________________________________________________________________

flatten_1 (Flatten)              (None, 1280)          0           maxpooling1d_1[0][0]

____________________________________________________________________________________________________

dense_1 (Dense)                  (None, 128)           163968      flatten_1[0][0]

____________________________________________________________________________________________________

dropout_1 (Dropout)              (None, 128)           0           dense_1[0][0]

____________________________________________________________________________________________________

dense_2 (Dense)                  (None, 5)             645         dropout_1[0][0]

====================================================================================================

Total params: 164,869

Trainable params: 164,869

Non-trainable params: 0

[Klemen Grm]

What's the shape of the input data to your network? Also, if possible post the code that defines your model.

[Vinayakumar R]

cnn = Sequential()

cnn.add(Convolution1D(64, 3, border_mode="same",activation="relu",input_shape=(41, 1)))

cnn.add(MaxPooling1D(pool_length=(2)))

cnn.add(Flatten())

cnn.add(Dense(128, activation="relu"))

cnn.add(Dropout(0.5))

cnn.add(Dense(5, activation="softmax"))

print(cnn.summary())

print(cnn.count_params())

[Klemen Grm]

For the layer convolution2d_1:
input_channels = 1, output_channels = 64, filter_length = 3,
num_params = (input_channels * output_channels * filter_length) + (output_channels) = 256

pooling and flatten layers don't have any free parameters.

for the layer dense_1, the previous layer's output shape is 1344, and its output shape is 128, therefore, it has 1344 * 128 + 128 = 172160 parameters

similarly, the layer dense_2 has 128 * 5 + 5 = 654 parameters.

[Vinayakumar R]

Thank you. This solves the problem

[Vinayakumar R]

The output from your code, I got 

192

256

164096

164224

164864

164869

From your calculated values 192 and 256 are right. But others are wrong