tsn-tensorflow
tsn-tensorflow copied to clipboard
shuangshuangguo
BNInception converted from caffe/pytorch
Hi,
Have you converted the BNinception model from TSN to TensorFlow format ?
Sorry, I have not. I am training the tensorflow implementation. Now I got 82% accuracy for RGB stream with 25 segments and one random crop. I am debugging still...
I am also trying to reproduce TSN using TensorFlow. However, the performance is not satisfactory (~7% vs ~17% in the Something-Something V1 validation dataset).
I tried to keep all configurations the same as tsn-pytorch except that I am using InceptionV2 instead of BNinception. I guess this might be the problem.
As far as I am known, BNInception is InceptionV2 actually. But I found slim InceptionV2 has a different implementation compared to pytorch BNInception.
slim InceptionV2 implements the first layer as separable convolution.
You are right. In principle, InceptionV2 should work as well as BNInception since BNInception is just a reproduction of InceptionV2 by xiong.
The differences are:
-
InceptionV2 inputs = 2.0 * (RGB_images / 255 - 0.5) BNinception inputs = BGR_images - [104, 117, 128]
-
InceptionV2 does not have biases in convolution and gamma in BN.
-
InceptionV2 weight decay 4e-5 BNInception weight decay 5e-4
-
InceptionV2 first layer: separable convolution
-
InceptionV2 Mixed_4d branch 3 Convolution has 96 channels BNInception has 128 channels
You are right. In principle, InceptionV2 should work as well as BNInception since BNInception is just a reproduction of InceptionV2 by xiong.
The differences are:
- InceptionV2 inputs = 2.0 * (RGB_images / 255 - 0.5) BNinception inputs = BGR_images - [104, 117, 128]
- InceptionV2 does not have biases in convolution and gamma in BN.
- InceptionV2 weight decay 4e-5 BNInception weight decay 5e-4
- InceptionV2 first layer: separable convolution
- InceptionV2 Mixed_4d branch 3 Convolution has 96 channels BNInception has 128 channels
are you succeed？？？i also want to do that
