Training fails when trying to build conv2d_7 layer (AssertionError: assert height % 2 == 0)

[chuckyee]

Training fails when trying to build conv2d_7 layer (AssertionError: assert height % 2 == 0)

[AloshkaD]

@DarylWM I've had the same issue. I've tried to change padding type, size of images, depth of features, layers, dialation, but nothing worked. If you remove the assertion you endup with an error like this 👍 ValueError: Negative dimension size caused by subtracting 3 from 1 for 'conv2d_49/convolution' (op: 'Conv2D') with input shapes: [?,1,1,1024], [3,3,1024,1024].

Have you managed to solve this issue?

[DarylWM]

Hi Ali. I haven't resolved the issue and had to move on. I've used this: https://github.com/nicolov/segmentation_keras

…

On 24 Oct. 2017 7:13 am, "Ali (Alex) Darwish" ***@***.***> wrote: @DarylWM <https://github.com/darylwm> I've had the same issue. I've tried to change padding type, size of images, depth of features, layers, dialation, but nothing worked. If you remove the assertion you endup with an error like this 👍 ValueError: Negative dimension size caused by subtracting 3 from 1 for 'conv2d_49/convolution' (op: 'Conv2D') with input shapes: [?,1,1,1024], [3,3,1024,1024]. Have you managed to solve this issue? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#4 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA-B7z1AB16hbDdQr8S55fGm4lw57hhIks5svPNXgaJpZM4PzOfD> .

[AloshkaD]

Thanks Daryl,
Also you may want to look at this repo. It has several networks that use dilated conv.

[chuckyee]

Set the `depth' option to 3. The original u-net had 4 downsampling blocks, but we only use 3 in this work since the dimensions of the images don't allow for 4 divisions by 2. Alternatively, resize the images so they are a power of 2^4 = 16 in both width and height.

I probably should have set the default to 3 in defaults.config.

Sry for delay -- been traveling.

[AloshkaD]

Thanks @chuckyee , I'll test it.

[DarylWM]

@AloshkaD How did you go?

[AloshkaD]

@DarylWM If I recall correctly there was another issue after changing the dim. In all cases I quit pursuing this and instead I've built a new network for a similar task. It has achieved 98% recall and 91% accuracy with and IoU of 50%.

[chuckyee]

@AloshkaD Nice. What do you define accuracy? I'm more familiar with precision and recall. If the IOU is 50% but the recall is 98%, then that means the model must be including lots of false positive areas.