CODEWITHSUNDEEP
python-3.xtensorflow
yangxiang_li
yangxiang_li

Reputation: 179

TypeError: 'Conv2dLayer' object is not subscriptable

I am using TensorFlow to reconstruct a CNN model called 'Deep Voxel Flow': https://github.com/liuziwei7/voxel-flow.

The output of the model is a tensor with the shape of [16,256,256,3] (16 is the batch size, 256 is the size of feature map, and 3 is the number of channels). When I tried to extract first 2 channels from the 4D tensor('flow=deconv4[:, :, :, 0:2]'), but I got an error: 'TypeError: 'Conv2dLayer' object is not subscriptable'.

deconv4 = Conv2d(deconv3_bn_relu, 3, [5, 5], act=tf.tanh, padding='SAME', 
W_init=w_init, name='deconv4')

flow = deconv4[:, :, :, 0:2]

The expected result is that the 'flow' is a 4D tensor with the shape of [16,256,256,2].

Upvotes: 2

Views: 2383

Answers (1)

Tezirg
Tezirg

Reputation: 1639

You should replace the use of Conv2D by a convolution layer, that returns a tensor:

deconv4 = tf.layers.conv2d(deconv3_bn_relu, 3, [5, 5], activation=tf.tanh, 
                           padding='SAME', kernel_initializer=w_init,
                           name='deconv4')

Then, you cannot use the [] to access the result of a Tensorflow operation directly. But you can use other operations to reshape/modify/split/truncate..etc the tensor.

In your case I think what you're looking for might be:

flow = tf.slice(deconv4, [0,0,0,0], [16, 256, 256, 2])

Upvotes: 1

Related Questions