I was doing some classification with keras, when met this error:
InvalidArgumentError: Dimensions must be equal, but are 256 and 8 for 'dense_185/MatMul' (op: 'MatMul') with input shapes: [?,256], [8,300].
It surprised me because the dimension of the input to the dense is 1.
This is a sequential model with a few custom layers. I have no idea why 8 appears in the error of dense layer.
class Residual(Layer):
def __init__(self,input_shape,**kwargs):
super(Residual, self).__init__(**kwargs)
self.input_shapes = input_shape
def call(self, x):
print(np.shape(x)) # (?, 128, 8)
first_layer = Conv1D(256, 4, activation='relu', input_shape = self.input_shapes)(x)
print(np.shape(first_layer)) (?, 125, 256)
x = Conv1D(256, 4, activation='relu')(first_layer)
print(np.shape(x)) (?, 122, 256)
x = Conv1D(256, 4, activation='relu')(x)
print(np.shape(x)) (?, 119, 256)
x = ZeroPadding1D(padding=3)(x)
residual = Add()([x, first_layer])
x = Activation("relu")(residual)
return x
class Pooling(Layer):
def __init__(self,**kwargs):
super(Pooling, self).__init__(**kwargs)
def call(self, x):
first_layer = GlobalMaxPooling1D(data_format='channels_last')(x)
second_layer = GlobalAveragePooling1D(data_format='channels_last')(x)
pooling = Add()([first_layer, second_layer])
print(np.shape(pooling)) (?, 256)
return pooling
model = Sequential()
model.add(Residual(input_shape=(128,8)))
model.add(Pooling())
model.add(Dense(300, activation='relu'))
model.add(Dense(150, activation='relu'))
model.add(Dense(10, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer=keras.optimizers.Adadelta(), metrics=['accuracy'])
model.fit(np.array(dataset_data), dataset_target, epochs=1000, validation_split=0.1, verbose=1, batch_size=8)
Dimensions:
(1000, 128, 8) - input (1000 audio, 8 features, 128 seq_length)
(1000, 10) - target (1000 audio, 10 classes)
I think there are two edits required:
InputLayer as entrance for the datacompute_output_shape method at least for Pooling layer (link). If this method is not defined, Dense layer can't figure out what's input shape for it, I guess, and then fails.Also there's minor editing - since model have InputLayer, you need no more input_shape kwarg in Residual layer.
class Residual(Layer):
def __init__(self, **kwargs): # remove input shape
super(Residual, self).__init__(**kwargs)
def call(self, x):
print(np.shape(x))
first_layer = Conv1D(256, 4, activation='relu')(x)
print(np.shape(first_layer))
x = Conv1D(256, 4, activation='relu')(first_layer)
print(np.shape(x))
x = Conv1D(256, 4, activation='relu')(x)
print(np.shape(x))
x = ZeroPadding1D(padding=3)(x)
residual = Add()([x, first_layer])
x = Activation("relu")(residual)
return x
class Pooling(Layer):
def __init__(self, **kwargs):
super(Pooling, self).__init__(**kwargs)
def call(self, x):
# !!! I build model without data_format argument - my version of keras
# doesn't support it !!!
first_layer = GlobalMaxPooling1D(data_format='channels_last')(x)
second_layer = GlobalAveragePooling1D(data_format='channels_last')(x)
pooling = Add()([first_layer, second_layer])
print(np.shape(pooling))
self.output_dim = int(np.shape(pooling)[-1]) # save output shape
return pooling
def compute_output_shape(self, input_shape):
# compute output shape here
return (input_shape[0], self.output_dim)
Initialize model:
model = Sequential()
model.add(InputLayer((128,8)))
model.add(Residual())
model.add(Pooling())
model.add(Dense(300, activation='relu'))
model.add(Dense(150, activation='relu'))
model.add(Dense(10, activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
Out:
(?, 128, 8)
(?, 125, 256)
(?, 122, 256)
(?, 119, 256)
(?, 256)
Summary of the model (don't know why Residual and Pooling don't show params the have. I guess some additional method required for this classes to count internal params):
model.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
residual_10 (Residual) (None, 128, 8) 0
_________________________________________________________________
pooling_8 (Pooling) (None, 256) 0
_________________________________________________________________
dense_15 (Dense) (None, 300) 77100
_________________________________________________________________
dense_16 (Dense) (None, 150) 45150
_________________________________________________________________
dense_17 (Dense) (None, 10) 1510
=================================================================
Total params: 123,760
Trainable params: 123,760
Non-trainable params: 0
_________________________________________________________________
Create fake data and check training process:
dataset_data = np.random.randn(1000, 128, 8)
dataset_target = np.zeros((1000, 10))
dataset_target[:, 0] = 1
model.fit(np.array(dataset_data), dataset_target, epochs=1000,
validation_split=0.1, verbose=1, batch_size=8)
Train on 900 samples, validate on 100 samples
Epoch 1/1000
900/900 [==============================] - 2s 2ms/step - loss: 0.0235 - acc: 0.9911 - val_loss: 9.4426e-05 - val_acc: 1.0000
Epoch 2/1000
900/900 [==============================] - 1s 1ms/step - loss: 4.2552e-05 - acc: 1.0000 - val_loss: 1.7458e-05 - val_acc: 1.0000
Epoch 3/1000
900/900 [==============================] - 1s 1ms/step - loss: 1.1342e-05 - acc: 1.0000 - val_loss: 7.3141e-06 - val_acc: 1.0000
... and so on
Looks like it works.