Keras模型网络3D输入到2D输出

Question

I have a 3D Input (Samples, Steps, Features). So each sample has a chain of steps that have different features. Now, I want a 2D Output (Samples, Steps) where I have samples and at each step of the sample a 0 or a 1 calculated from the model.

So I think it is a sequential binary classification problem. I have some difficult to define the model, especially the output layer.

Here are the shapes of numpy arrays:

x_train.shape 
# (200, 1657, 669)

x_test.shape
# (41, 1657, 669)

y_train.shape
# (200, 1657)

y_test.shape
# (41, 1657)

I tried this model but the output was not the one I was expecting

n_residues, n_features, n_outputs = trainX.shape[1], trainX.shape[2], trainy.shape[1]
model = Sequential()
model.add(Conv1D(filters=64, kernel_size=3, activation='relu', input_shape=(n_residues,n_features)))
model.add(Conv1D(filters=64, kernel_size=3, activation='relu'))
model.add(Dropout(0.1))
model.add(MaxPooling1D(pool_size=2))
model.add(Flatten())
model.add(Dense(100, activation='relu'))
model.add(Dense(n_outputs, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
# fit network
model.fit(trainX, trainy, epochs=epochs, batch_size=batch_size, verbose=verbose)

m_classes = model.predict_classes(x_test, verbose=0)
print(m_classes)
[  36   36   59   32   16   32   36  804 1047   16   16   36   32   36
   36   36   16   16   16   16   16   16   36   16   36   36   36   16
   59   36   36   36   16   16   16  804   16   16   16   36   36]

The output is a 41 long vector for the 41 samples in the test set with classes 0 -1657 I assume.

My desired output for the test set would be 41 binary vectors that are 1657 long.

Thanks!

Answer 1

When you are dealing with Conv1D, RNN or sequence models the output can be one to one, many to one, (or) many to many. In this case, the model is acting many to one. Generally, In Keras, there is a return_sequence or stateful parameter. If these parameters are False then your model behaves like many to one . (ie.., the output shape is (batch_size, unit_length). In this case, unit length is an output). To make a model many to many then the output should be like (batch_size, time_step, unit_length) . Just initialize the stateful==True will help you to solve this problem.

some of the helpful links to understand about sequence output data
Understanding LSTM
Feeding one prediction after another

check the conv1D initialization official documentation and set the parameter stateful=True . This is the Theoretical Idea behind LSTM or Conv1D. This links will help you to get architecture Idea behind LSTM's.