FCC: full connect convolution

[yzbx]

fc_net

https://g.gravizo.com/svg?
digraph G {
    subgraph cluster_0 {
        style=filled;
        color=lightgrey;
        node [shape=box,style=filled,color=white];
        data -> backbone
        backbone -> fc_conv_blocks [weight=8];
        fc_conv_blocks -> output;
        label = "fc_net";

    }

    subgraph cluster_1 {
    style=filled;
        color=lightgrey;
        node [shape=box,style=filled,color=white];
        fc_conv_1->fc_conv_2 [label="mul(in_channel_2,k1)"]
        fc_conv_2->fc_conv_k [label="mul(out_channel_2,k2),mul(in_channel_k,k2)"];
        label="fc_conv_blocks"
    }

    subgraph cluster_2 {
        style=filled;
        color=lightgrey;
        node [shape=box,style=filled,color=white];
        label="fc_conv"

        { "conv_1.1" "conv_1.2" "conv_1.k1" } -> { "conv_2.1" "conv_2.2" "conv_2.k2"}

        "conv_1.1.1"->"conv_1.1" [label="mid_channel m1"]
        "conv_1.2.1"->"conv_1.2.2"->"conv_1.2" [label="m2"]
    }
})

[yzbx]

old experiments

for depth in 3 6 9
do 
    python test/train.py --net_name=fc_net --dataset_name=cifar100 \
--note=conv_number_wise_${depth} --fcc_block_number=${depth} \
--fcc_width_wise_ways=conv_number_wise 
done

for depth in 3 6 9
do 
    python test/train.py --net_name=fc_net --dataset_name=cifar100 \
--note=channel_wise_${depth} --fcc_block_number=${depth} \
--fcc_width_wise_ways=channel_wise 
done

sota

for depth in 3 6 9
do 
    python test/train.py --net_name=fc_net --dataset_name=cifar100 \
--note=sota_cnw_${depth} --fcc_block_number=${depth} \
--fcc_width_wise_ways=conv_number_wise 
done

for depth in 3 6 9
do 
    python test/train.py --net_name=fc_net --dataset_name=cifar100 \
--note=sota_cw_${depth} --fcc_block_number=${depth} \
--fcc_width_wise_ways=channel_wise 
done