search

hlamba28 / UNET-TGS

Applying UNET Model on TGS Salt Identification Challenge hosted on Kaggle
https://towardsdatascience.com/understanding-semantic-segmentation-with-unet-6be4f42d4b47
MIT License
142 stars 88 forks source link

Wrong Architecture & Inference Pipeline #1

Open lateralus69 opened 5 years ago

lateralus69 commented 5 years ago

In conv2d block function the x coming out from activation function after first layer should go in the second convolution layer instead of original input tensor. The re-written code for second layer in the function should be 

# second layer
x = Conv2D(filters = n_filters, kernel_size = (kernel_size, kernel_size),\
          kernel_initializer = 'he_normal', padding = 'same')(x) #this was the change
if batchnorm:
    x = BatchNormalization()(x)
x = Activation('relu')(x)

return x`

Also in the inference pipeline you are using using contours from ground truth of validation set and plotting them on the predicitions of validation set which makes so sense.

That was a noob way to show good predictions.

hlamba28 commented 5 years ago

Hi @lateralus69 ,

Thanks for pointing out the mistakes. I will try to correct them as I get sometime.

By the way these are just practice projects which I have done for self learning and I completely agree that there can be some mistakes in them. I do not have any intention to show great predictions.

Hope you are doing well. Thanks

hlamba28 commented 5 years ago

Also, I agree for your first point, however there is no problem with the inference pipeline. The contour (boundary) has to be from the ground truth to show how much the predictions overlap with the real boundary. Clearly the validation set predictions are not 100% correct.