Keras的卷积神经网络输入形状

Question

I have 32760 audio spectrums calculated of dimensions = 72 (# frames) x 40 (# frequency bands) that I am trying to feed into a "wide" convolutional neural network (first layer is an ensemble of 4 different conv layers). These spectrums have no depth, therefore they can be represented as a 72 x 40 2D numpy array of floats, and thus the X input to the classifier is an array 32760 elements long, each element being one of these 72 x 40 x 1 spectrums. The Y input is an array of labels, one-hot encoded, with 32760 elements.

When trying to fit the CNN using

model.fit(mono_X, mono_Y, epochs=10, batch_size=None, verbose=2)

I get the following error:

ValueError when checking input: expected input_47 to have 4 dimensions, but got array with shape (32760, 1)

Below is the architecture of my CNN:

spectra = Input(shape=(72, 40, 1)) 

# conv1a
c1a = Conv2D(48, (3,5), activation='relu', padding = 'same')(spectra)
c1a = BatchNormalization()(c1a)
c1a = MaxPooling2D(pool_size=(5, 5), strides = 1)(c1a)
# conv1b
c1b = Conv2D(32, (3,9), activation='relu', padding = 'same')(spectra)
c1b = BatchNormalization()(c1b)
c1b = MaxPooling2D(pool_size=(5, 5), strides = 1)(c1b)
# conv1c
c1c = Conv2D(16, (3,15), activation='relu', padding = 'same')(spectra)
c1c = BatchNormalization()(c1c)
c1c = MaxPooling2D(pool_size=(5, 5), strides = 1)(c1c)
# conv1d
c1d = Conv2D(16, (3,21), activation='relu', padding = 'same')(spectra)
c1d = BatchNormalization()(c1d)
c1d = MaxPooling2D(pool_size=(5, 5), strides = 1)(c1d)

# stack the layers
merged = keras.layers.concatenate([c1a, c1b, c1c, c1d], axis=3)

# conv2
c2 = Conv2D(224, (5,5), activation='relu')(merged)
c2 = BatchNormalization()(c2)
c2 = MaxPooling2D(pool_size=(5, 5), strides = 1)(c2)

# output softmax
out = Dense(15, activation='softmax')(c2)

# create Model
model = Model(spectra, out)

# apply optimization and loss function
adam = Adam(lr=0.002, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)
model.compile(optimizer=adam,
            loss='categorical_crossentropy',
            metrics=['accuracy'])

However, if I try to change the input shape to 32760x1 I receive the following error:

ValueError: Input 0 is incompatible with layer conv2d_203: expected ndim=4, found ndim=3

What am I doing wrong here? Is there a better way to represent my input data? I have already tried using a pandas DataFrame where each row represents one of the spectra and a myriad of other combinations. Using Python 3.6.5 with Keras 2.1.3 with TensorFlow 1.1.0 on the backend.

This is my first CNN, I have only implemented ANNs previously using Keras so there may be a very obvious mistake I'm making. Any help appreciated!

---

Update! Taking the advice from @enumaris, using data_format=channels_last as param on the input layer and adding a Flatten() layer between the last Conv2D and the softmax output layer fixed the latter value error. Now I've come to realize that my training data mono_X is of the wrong shape. The expected input shape if I'm not mistaken should be (#samples, H, W, #channels). mono_X is of shape (32760,) while mono_X[0] is of shape (72, 40). Using numpy's reshape doesn't seem to be able to unpack these nested arrays. How can I properly prepare the input tensor?

Answer 1

The input shape is (72, 40, 1) but you say that elements of mono_X have a shape (72, 40). It needs to be reshaped, possibly when preparing the training data like mono_X = mono_X.reshape(-1, 72, 40, 1) . This assumes that mono_X is a numpy array of shape (#samples, 72, 40), but for some reason it sounds like you have a numpy array of numpy arrays.

You can also reshape the Keras layer like this:

spectra = Input(shape=(72, 40))
spectra = Reshape((72, 40, 1))(spectra)

Personally, I would reshape before training and not in the model to avoid any additional overhead in training.