search

farquasar1 / ConvLSTM

Convolutional LSTM for video segmentation with Keras
108 stars 34 forks source link

issue about time distributed maxpooling #3

Open yl3639 opened 6 years ago

yl3639 commented 6 years ago

Hi @farquasar1 , thanks for the code sharing.

I tried to implement this network to do some 3d image segmentation, my input shape is (80,40,40,3), which 80 is frame size, 40 and 40 are rows and col, 3 is channels. After I try to run the training, I encounter this problem with the first time distributed(maxpooling) layer.

ValueError: Negative dimension size caused by subtracting 2 from 1 for 'time_distributed_2/MaxPool' (op: 'MaxPool') with input shapes: [?,20,1,24].

Thanks for your time! :D

Charlie2048 commented 6 years ago

Hi, @yl3639 I think your frame should be coverted to one channels first which is written in the "READEME.md"

And I have some problem when I run the code in my computer. It seems works well in your computer. Do you have received the error " No module named data_processing"? If so, how to deal with it.

Thanks for your time! :D

yl3639 commented 6 years ago

Hi @Charlie2048 , I didn't use last_fcn_train.py to train the network instead I wrote a simple generator and fit function to train. My problem is solved by setting the image_format to 'channel last' and can run the training but the training process, the loss, seems unreasonable and I do the prediction, it wrongly predict all pixels as foreground. not sure where is wrong.