I'm trying to use an autoencoder on spatiotemporal data.
My data shape is: batches , filters, timesteps, rows, columns
. I have problem with setting the autoencoder to the right shape.
This is my model:
input_imag = Input(shape=(3, 81, 4, 4))
x = Conv3D(16, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(input_imag)
x = MaxPooling3D((3, 2, 2), data_format='channels_first', padding='same')(x)
x = Conv3D(8, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(x)
x = MaxPooling3D((3, 2, 2), data_format='channels_first', padding='same')(x)
x = Conv3D(4, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(x)
encoded = MaxPooling3D((3, 2, 2), data_format='channels_first', padding='same', name='encoder')(x)
x = Conv3D(4, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(encoded)
x = UpSampling3D((3, 2, 2), data_format='channels_first')(x)
x = Conv3D(8, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(x)
x = UpSampling3D((3, 2, 2), data_format='channels_first')(x)
x = Conv3D(16, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(x)
x = UpSampling3D((3, 2, 2), data_format='channels_first')(x)
decoded = Conv3D(3, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(x)
autoencoder = Model(input_imag, decoded)
autoencoder.compile(optimizer='adam', loss='mse')
autoencoder.summary()
This is the summary:
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 3, 81, 4, 4)] 0
_________________________________________________________________
conv3d (Conv3D) (None, 16, 81, 4, 4) 2176
_________________________________________________________________
max_pooling3d (MaxPooling3D) (None, 16, 27, 2, 2) 0
_________________________________________________________________
conv3d_1 (Conv3D) (None, 8, 27, 2, 2) 5768
_________________________________________________________________
max_pooling3d_1 (MaxPooling3 (None, 8, 9, 1, 1) 0
_________________________________________________________________
conv3d_2 (Conv3D) (None, 4, 9, 1, 1) 1444
_________________________________________________________________
encoder (MaxPooling3D) (None, 4, 3, 1, 1) 0
_________________________________________________________________
conv3d_3 (Conv3D) (None, 4, 3, 1, 1) 724
_________________________________________________________________
up_sampling3d (UpSampling3D) (None, 4, 9, 2, 2) 0
_________________________________________________________________
conv3d_4 (Conv3D) (None, 8, 9, 2, 2) 1448
_________________________________________________________________
up_sampling3d_1 (UpSampling3 (None, 8, 27, 4, 4) 0
_________________________________________________________________
conv3d_5 (Conv3D) (None, 16, 27, 4, 4) 5776
_________________________________________________________________
up_sampling3d_2 (UpSampling3 (None, 16, 81, 8, 8) 0
_________________________________________________________________
conv3d_6 (Conv3D) (None, 3, 81, 8, 8) 2163
=================================================================
Total params: 19,499
Trainable params: 19,499
Non-trainable params: 0
What I should change to have the decoder output shape as [?,3,81,4,4]
not [?,3,81,8,8]
?
CodePudding user response:
It looks like you want the MaxPooling3D and UpSampling3D operations to be symmetrical (at least in terms of output shapes). Let's look at the input shape of the last MaxPooling3D layer:
conv3d_2 (Conv3D) (None, 4, 9, 1, 1) 1444
_________________________________________________________________
encoder (MaxPooling3D) (None, 4, 3, 1, 1) 0
The shape is (None, 4, 9, 1, 1)
. The last two dimensions are already 1, so they can't be divided by 2, as specified in pool_size
. So MaxPooling3D layer, despite having a pool_size=(3, 2, 2)
, effectively does an operation with pool_size=(3, 1, 1)
. At least I think that is what happens under the hood.
I'm a bit surprised there is no error or warning when specifying pool_size greater than input size.
To fix that you can set the first UpSampling3D layer's shape to (3, 1, 1)
x = UpSampling3D((3, 1, 1), data_format='channels_first')(x)
So, the complete solution:
input_imag = Input(shape=(3, 81, 4, 4))
x = Conv3D(16, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(input_imag)
x = MaxPooling3D((3, 2, 2), data_format='channels_first', padding='same')(x)
x = Conv3D(8, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(x)
x = MaxPooling3D((3, 2, 2), data_format='channels_first', padding='same')(x)
x = Conv3D(4, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(x)
encoded = MaxPooling3D((3, 2, 2), data_format='channels_first', padding='same', name='encoder')(x)
x = Conv3D(4, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(encoded)
x = UpSampling3D((3, 1, 1), data_format='channels_first')(x)
x = Conv3D(8, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(x)
x = UpSampling3D((3, 2, 2), data_format='channels_first')(x)
x = Conv3D(16, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(x)
x = UpSampling3D((3, 2, 2), data_format='channels_first')(x)
decoded = Conv3D(3, (5, 3, 3), data_format='channels_first', activation='relu', padding='same')(x)
autoencoder = Model(input_imag, decoded)
autoencoder.compile(optimizer='adam', loss='mse')
autoencoder.summary()
Output:
Model: "model_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_3 (InputLayer) [(None, 3, 81, 4, 4)] 0
conv3d_14 (Conv3D) (None, 16, 81, 4, 4) 2176
max_pooling3d_4 (MaxPooling (None, 16, 27, 2, 2) 0
3D)
conv3d_15 (Conv3D) (None, 8, 27, 2, 2) 5768
max_pooling3d_5 (MaxPooling (None, 8, 9, 1, 1) 0
3D)
conv3d_16 (Conv3D) (None, 4, 9, 1, 1) 1444
encoder (MaxPooling3D) (None, 4, 3, 1, 1) 0
conv3d_17 (Conv3D) (None, 4, 3, 1, 1) 724
up_sampling3d_6 (UpSampling (None, 4, 9, 1, 1) 0
3D)
conv3d_18 (Conv3D) (None, 8, 9, 1, 1) 1448
up_sampling3d_7 (UpSampling (None, 8, 27, 2, 2) 0
3D)
conv3d_19 (Conv3D) (None, 16, 27, 2, 2) 5776
up_sampling3d_8 (UpSampling (None, 16, 81, 4, 4) 0
3D)
conv3d_20 (Conv3D) (None, 3, 81, 4, 4) 2163
=================================================================
Total params: 19,499
Trainable params: 19,499
Non-trainable params: 0