signatrix / regnet Goto Github PK

Hi,
Thanks for a GREAT repo!
I think there might be a bug in the creation of RegnetX here:

Why would you multiply the group_width by the bottleneck_ratio?

I will demonstrate through an example:

group_width = 16
block_width = 32
bottleneck_ratio = 2

With these set of parameters I would assume a bottleneck block will be created with 1/2 the channels in the bottleneck and 1 group convolution (i.e. standard convolution)
However: l.25 changes the groups to ls_group_width = ls_group_width.astype(np.int) * bottleneck_ratio => group_width = 32 making this block impossible and having the model FAIL!

Is this intentional or a bug?

Thank you very much

Thanks for your work. Is the model structure as same as original?

First, thanks for the awesome work of re-implementing RegNet.

I am having difficulty of reproducing the results for RegNetY-0.4GF. The configurations are taken from the original repo:

group_width = 8
initial_width = 48
slope = 27.89
quantized_param = 2.09
network_depth = 16

I only get 72.92 top-1 accuracy, but the original paper reported 74.2. Any thoughts on that?

I followed the README to create data folder.
data
├── train
│ ├── n0
│ └── n1
│ └── n2
│ └── n3
│ └── n4
│── val
│ ├── n0
│ └── n1
│ └── n2
│ └── n3
│ └── n4

I modified the NUM_CLASSES = 5 at src/config.py.
And all config are default (RegnetY 200MF).