search

nikhilroxtomar / Deep-Residual-Unet

ResUNet, a semantic segmentation model inspired by the deep residual learning and UNet. An architecture that take advantages from both(Residual and UNet) models.
https://idiotdeveloper.com/what-is-resunet/
221 stars 56 forks source link

ResUNet improvement #1

Closed foobar167 closed 5 years ago

foobar167 commented 5 years ago

Thank you for the implementation and explanation of ResUNet! That's really interesting.

It seems to me that you should have a different "bridge" for ResUNet. According to the article, the "bridge" is the same as the "residual block" of ResUNet. The "bridge" and "residual block" are the same, except that dashed arrow from the entrance of "residual block" to the "Addition" block. But I think this is an error in the article :-) There must be a dashed arrow otherwise we don't need "Addition" block in the "bridge". So I've changed your code to this:

def ResUNet():
f = [16, 32, 64, 128, 256]
inputs = keras.layers.Input((image_size, image_size, 3))

## Encoder
e0 = inputs
e1 = stem(e0, f[0])
e2 = residual_block(e1, f[1], strides=2)
e3 = residual_block(e2, f[2], strides=2)
e4 = residual_block(e3, f[3], strides=2)

## Bridge
b0 = residual_block(e4, f[4], strides=2)

## Decoder
u1 = upsample_concat_block(b0, e4)
d1 = residual_block(u1, f[4])

u2 = upsample_concat_block(d1, e3)
d2 = residual_block(u2, f[3])

u3 = upsample_concat_block(d2, e2)
d3 = residual_block(u3, f[2])

u4 = upsample_concat_block(d3, e1)
d4 = residual_block(u4, f[1])

outputs = keras.layers.Conv2D(1, (1, 1), padding="same", activation="sigmoid")(d4)
model = keras.models.Model(inputs, outputs)
return model

As you can see I just put "e5" to the "bridge" and that's all :-) Also I've set 120 epochs and obtained dice coefficient equal to 0.9825 237/237 [==============================] - 11s 48ms/step - loss: 0.0175 - dice_coef: 0.9825 - val_loss: 0.1533 - val_dice_coef: 0.8467

The result is on 8.6% higher than the 1st place on Kaggle, team name "b.e.s. & phalanx". It means that YOU can be a Kaggle winner next time. Keep doing don't stop :-)

arghadeep25 commented 5 years ago

The training loss seems to be much lower than the validation loss or vice versa. i.e. I guess your model is overfitting. Which dataset you used?

foobar167 commented 5 years ago

Yes, arghadeep25, I think you write, there are too many epochs (120 epochs). So it seems overfitting.

The dataset is from Kaggle TGS Salt Identification Challenge: https://www.kaggle.com/c/tgs-salt-identification-challenge Original Jupyter Notebook: https://github.com/nikhilroxtomar/Deep-Residual-Unet/blob/master/Deep%20Residual%20UNet.ipynb You can open and run it via Google Colab: https://github.com/foobar167/articles/blob/master/Machine_Learning/code_examples/deep_residual_unet_segmentation.ipynb

arghadeep25 commented 5 years ago

The data seems to be quite sparse than the networks. I guess this can be solved by using earlyStopping, dropout and augmenting the data.