Week #21

18 Mar 2019

Results

Gourp Id	Parameter num	initial lr	step_size	gamma	epoch	Loss	Figure
0	13177026	0.001	100	0.8	350	0.0402
1	13177026	0.001	100	0.85	750	0.1314
2	13177026	0.001	100	0.85	950	0.52985
3	13177026	0.001	100	0.85	900	1.23616
3	52633794	0.001	100	0.85	450	0.979685

Try a simple network

After discussing with Prof. Xavier, with his recommendation, I tried to conduct the experiments using the simpliest Conv-BatchNorm-Relu concatenated architecture. The experiments are as belows:

1. Group 0, simple network with 7 layers, with 7,932,630 parameters.
2. Group 1, simple network with 4 layers, with 7,932,630 parameters.
3. Group 2, simple network with 4 layers, with 298,710 parameters.
4. Group 2, simple network with 7 layers, with 7,932,630 parameters.
5. Group 3, simple network with 7 layers, with 7,932,630 parameters.

Since we have the experiments done with QuickNAT using the group 0,1,2,3. So here the same group number is used for comparison.

Paper Construction

We concluded that the paper should be constructed in order to find out the contributions we want to make.