从Keras / Tensor流量模型获得可用的系数权重

Question

I've created a classification model using Keras (with tensorFlow backend). At this point, I'm looking to use model coefficients to create an web API service for my new model. However, when I attempt to get these weights, these are not as simple as multiply these coefficients * features to get expected results.

My training input matrix is a grayscale image of 128 x 128 = 16,384.

 model.get_weights()
 [array([[[[-0.03603082,  0.1334779 ,  0.10018548, -0.14283592, -0.0429921 ,
       -0.03080634,  0.06451669, -0.09407537,  0.04878693, -0.14003059,
        0.04794825,  0.06447313, -0.01520954, -0.10879657,  0.13521752,
       -0.03230923, -0.01395164,  0.04935856,  0.06434418, -0.02601192,
        0.03416487,  0.08788931,  0.0723172 , -0.12923865,  0.04022292,
       -0.1328591 , -0.05803869, -0.01380468, -0.10409287,  0.14212781,
       -0.08511351, -0.03992498]],

     [[-0.11669005,  0.088085  , -0.......

  len(model.get_weights())
  #10

  len(model.get_weights()[0])
  #3

  len(model.get_weights()[0][0])
  #3

  len(model.get_weights()[0][0][0])
  #1

  len(model.get_weights()[0][0][0][0])
  #32

  len(model.get_weights()[0][0][0][0][0])
  #TypeError: object of type 'numpy.float32' has no len()

If I take all those level 10 * 3 * 3 * 1 * 32 = 2880.

So this is telling me the numpy array has 5 dimensions with only 2880 values? Maybe I'm missing something here or perhaps a much easier way. Any thoughts on how this could be applied to a new test image via an API?

Answer 1

The first layer has 288 (3*3*1*32) parameters, but following layers usually have much more parameters especially if your model is CNN.

To print numbers of parameters for all layers (which include bias terms):

weights = model.get_weights()
for i in range(len(weights)):
    print(weights[i].shape)

Alternatively you can check the number of parameters by model.summary()

To apply the model to new images, just run model.predict(image)