ASPP model without extra blocks about deeplabv3-tensorflow

[eveningdong]

Nice spot! Thanks! You are right about 7 blocks!!!
For RELU, I think as an activation function, without specific explanation, it should be added automatically.
I will fix the bug and retrain the model, the previous train mIOU seems to be a result of overfitting.

[John1231983]

You are welcome. How about weight decay of BN ?

[eveningdong]

Yes, I will add it to config.py.

[John1231983]

Hi, I found some changing architecture in your code in ASPP. You changed from concat to tf.add_n in the convolution (1×1) and convolutions (3x3) with difference rate, called (a) part. The ouput of addition will feed to global average pooling (b) part. Then results will concat to the (a). Why did you change from concat to add in (a)? And why is not input of global pooling features of block 4? And I also did not find that Batchnorm did not train in your ASPP. Thanks

[eveningdong]

Hi, John, you are right. For the current version of validation mIOU, I trained with Imaging Pooling over the ASPP modules, this is different, I reverted to the original one shown in the paper since this is a reimplementation.

But I do find that the modified Imaging Pooling has some effect. I will retrain a little bit.

[John1231983]

Very surprised, When you train your model, do you also train the parameter of BN? The paper mentioned that he trained the BN with batch size 16 for 30k. After that, he froze the BN parameter and train with low learning rate. I asked this because you just use batch size 4, but still got very high performance (better than batch size 12 in the paper). I am running your code in my machine and I will lets you know my performance.

[eveningdong]

@John1231983 I am rethinking the relu function in ASPP module. I heard you achieved 73% in validation data. So are you using relu or not?

[bhack]

@NanqingD there are some variation over the Deeplabv3 arch if you are interested. See tensorflow/tensorflow#16452 (comment)

[John1231983]

Yes, it should include Relu as default option.