Can only get a max pass score of 57

I have trained the model and even the parameters are the same values as the expected output for both FCN8 function and conv_block. But yet, the maximum value I can get is 57 and I fail. May i know what I can do to get it to more than 60?

Joel,

When you are not getting the desired accuracy or value, try looking into your model again where you could adjust parameters like input shape, layers in the model algorithm, what other parameters for model fit could be changed, batch_size.

In such assignment query post, any mentor would want to know how many layer of conv block you used and what padding or unit you used, what other parameters like input shape you used, loss detail, optimizer detail and at last if any epoch setting was set by you. Share the image of the model training image once you completed the training, even that can help a bit.

You didn’t give any detail about your fcn8 function codes you used. I am not asking you to mention the codes but roughly give an idea about your model, so mentors know where things can be changed.

Regards
DP

I have called the conv_block function at first with 32 filters, kernel size 3,3, pool size 2,2 and strides 2,2.
then 64 filters, 128 filters, 256 and 256 filters again.
Then the decoder starts with the prewritten code given by the author.
after that it has conv2d transpose with kernel size 4,4 and strides 2,2 and kept bias as false.
cropping of 1,1.
loaded pool 4 prediction f4 and did convolution with size 1,1.
added o,o2 since i had to add the results of upsampling and pool 4 prediction.
then i used conv2dtransp with the same size and stides 2,2 as before.
cropping 1,1.

then loaded f3, conv2d with 1,1 size and dataformat IMAGE_ORDERING.
added o,o2.
in the end i kept conv2d tranpose with kernel size 8,8 and strides 8,8.
compiling with loss categorical crossentropy optimizer SGD.

I have solved the issue and passed the assignment, I changed from sgd to adam optimizer with an lr of 0.001.
Thank you for your time ma’am.

2 Likes

That’s great joel, you debug your issue again and just needed hints. Welcome programmer.

Regards
DP

1 Like