ValueError: "concat" mode can only merge layers with matching output shapes except for the concat axis. Layer shapes: [(None, 2, 14, 512), (None, 2, 14, 256)] about kaggle_dstl_submission HOT 7 CLOSED

@leader80 there are two ways to organize image data: NCHW or channels_first and NHWC or channels_last. First N stands for N images in batch, C stands for channels and H, W is height and width of image. As you can see in the code, this network defaults to NCHW input (for example, look here https://github.com/ternaus/kaggle_dstl_submission/blob/master/src/unet_buildings.py#L63) and Theano backend used during training does the same too. It means that all convolutions change C axis of tensor (second place, index 1 starting from zero) and pooling layers change only H and W, thus concat_axis everywhere is 1, as we concat tensors of the same height and width and only different number of channels and batchnorm axis also 1 (along channels).

New versions of Keras use Tensorflow backend and Tensorflow uses NHWC by default because it is faster on CPU (sic!) (see discussion here). So this change messes up all the axes in the network. There are to ways dealing with it:

• first option, add old ordering option to Keras config file (lookup at google how to do it in current version), -
• second option, change input to be NHWC format and replace all concat and batch norm axes to -1.

That's it, hope this helps

from kaggle_dstl_submission.

@dmpetrov Thanks a lot.
I resolved this problem by set "unet_buildings.py" code :
up6 = merge([UpSampling2D(size=(2, 2))(conv5), conv4], mode='concat', concat_axis=1)
change "concat_axis=1" to "concat_axis= -1" to match [2, 14, 512] and [2, 14, 256].

from kaggle_dstl_submission.

plus: I use LINUX system.

from kaggle_dstl_submission.

@WenwenQi, you can try to set image_dim_ordering to Theano th in ~/.keras/keras.json.

from kaggle_dstl_submission.

I meet the same question as you. and I change "concat_axis=1" to "concat_axis= -1" from up6 to up9, but I got an error as following:

File "/home/chen-h/PycharmProjects/3th kaggle_dstl_submission-master/kaggle_dstl_submission-master/src/unet_trees.py", line 250, in
model = get_unet0()
File "/home/chen-h/PycharmProjects/3th kaggle_dstl_submission-master/kaggle_dstl_submission-master/src/unet_trees.py", line 133, in get_unet0
conv9 = BatchNormalization(mode=0, axis=1)(crop9)
File "/home/chen-h/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 575, in call
self.build(input_shapes[0])
File "/home/chen-h/anaconda3/lib/python3.6/site-packages/keras/layers/normalization.py", line 103, in build
constraint=self.gamma_constraint)
File "/home/chen-h/anaconda3/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 87, in wrapper
return func(*args, **kwargs)
File "/home/chen-h/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 396, in add_weight
weight = K.variable(initializer(shape),
File "/home/chen-h/anaconda3/lib/python3.6/site-packages/keras/initializers.py", line 41, in call
return K.constant(1, shape=shape, dtype=dtype)
File "/home/chen-h/anaconda3/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 358, in constant
return tf.constant(value, dtype=dtype, shape=shape, name=name)
File "/home/chen-h/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/constant_op.py", line 99, in constant
tensor_util.make_tensor_proto(value, dtype=dtype, shape=shape, verify_shape=verify_shape))
File "/home/chen-h/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/tensor_util.py", line 421, in make_tensor_proto
(shape_size, nparray.size))
ValueError: Too many elements provided. Needed at most -16, but received 1

It seems to caused by the following code:

conv9 = Convolution2D(32, 3, 3, border_mode='same', init='he_uniform')(conv9)
crop9 = Cropping2D(cropping=((16, 16), (16, 16)))(conv9)
conv9 = BatchNormalization(mode=0, axis=1)(crop9)

Can you give me some advice?
thanks a lot.

from kaggle_dstl_submission.

@cug-chenhui you have to carefully check where do you place your channels and correct channel_axis everywhere starting from your input, then check every layer to the output. Or use the same channel as was originally intended :)

from kaggle_dstl_submission.

@cepera-ang do you mind being more specific? sorry I'm also facing this issue

from kaggle_dstl_submission.