Prioritized Subnet Sampling for Resource-Adaptive Supernet Training (paper)

Tips

Any problem, please contact the first author (Email: [email protected]).

Dependencies

• Python 3.8
• Pytorch 1.7

Supernet Training

1. Prepare your ImageNet dataset
2. Run training script

The training scripts are located in the ./scripts/train directory.

./scripts/train/
├── pss-mbv1-f.sh	# train PSS-MBV1 under FLOPS Constraint
├── pss-mbv1-c.sh	# train PSS-MBV1 under CPU Latency Constraint
├── pss-mbv1-g.sh	# train PSS-MBV1 under GPU Latency Constraint
├── pss-mbv1-m.sh	# train PSS-MBV1 under FLOPS / CPU Latency / GPU Latency Constraint
├── pss-mbv2-f.sh	# train PSS-MBV2 under FLOPs Constraint
├── pss-mbv2-c.sh	# train PSS-MBV2 under CPU Latency Constraint
├── pss-mbv2-g.sh	# train PSS-MBV2 under GPU Latency Constraint
└── pss-mbv2-m.sh	# train PSS-MBV2 under FLOPS / CPU Latency / GPU Latency Constraint

Evaluate Our Result

1. Prepare your ImageNet dataset
2. Download PSS-Net checkpoint from the link in the table below
3. Run evaluation script

The evaluation scripts are located in the ./scripts/test directory.

./scripts/test/
├── pss-mbv1-f.sh	# evaluate PSS-MBV1 under FLOPS Constraint
├── pss-mbv1-c.sh	# evaluate PSS-MBV1 under CPU Latency Constraint
├── pss-mbv1-g.sh	# evaluate PSS-MBV1 under GPU Latency Constraint
├── pss-mbv1-m.sh	# evaluate PSS-MBV1 under FLOPS / CPU Latency / GPU Latency Constraint
├── pss-mbv2-f.sh	# evaluate PSS-MBV2 under FLOPs Constraint
├── pss-mbv2-c.sh	# evaluate PSS-MBV2 under CPU Latency Constraint
├── pss-mbv2-g.sh	# evaluate PSS-MBV2 under GPU Latency Constraint
└── pss-mbv2-m.sh	# evaluate PSS-MBV2 under FLOPS / CPU Latency / GPU Latency Constraint

ImageNet Results

Our Trained Supernets and Typical Subnets Reported in the Paper

SuperNet Link	Constraint Type	Subnet1	Subnet2	Subnet3	Subnet4	Subnet5	Subnet6	Subnet7	Subnet8	Subnet9	Subnet10	...
PSS-MBV1-F	FLOPs	555M 74.2%	496M 73.8%	443M 73.6%	375M 73.2%	343M 72.7%	310M 72.4%	306M 72.3%	273M 72.1%	126M 68.1%	107M 67.9%	...
PSS-MBV1-G	GPU Latency	238us 74.4%	220us 73.92%	201us 73.5%	172us 73.3%	148us 72.6%	128us 71.9%	111us 70.9%	94us 69.8%	76us 68.1%	67us 67.9%	...
PSS-MBV1-C	CPU Latency	32ms 74.2%	30ms 73.9%	28ms 73.6%	26ms 73.5	23ms 73.2%	20ms 72.5%	18ms 71.5%	16ms 70.2%	13ms 68.3%	11ms 68.0%	...
PSS-MBV1-M	FLOPs	562M 74.5%	511M 74.1%	458M 73.6%	419M 73.3%	360M 73.1%	309M 73.5%	273M 72.1%	203M 70.6%	118M 68.1%	107M 68.0%	...
	GPU Latency	236us 74.5%	221us 74.1%	197us 73.4%	166us 73.1%	141us 72.4%	120us 71.5%	100us 70.0%	88us 68.8%	71us 68.1%	69us 68.0%	...
	CPU Latency	30ms 74.5%	28ms 73.8%	26ms 73.4%	23ms 73.1%	21ms 72.5%	18ms 71.5%	16ms 70.8%	14ms 69.6%	13ms 68.6%	11ms 68.0%	...
PSS-MBV2-F	FLOPs	301M 73.4%	294M 73.3%	277M 72.8%	241M 72.4%	203M 72.1%	163M 70.9%	154M 70.7%	134M 70.0%	84M 66.8%	69M 66.4%	...
PSS-MBV2-G	GPU Latency	260us 73.4%	252us 72.8%	234us 72.7%	225us 72.6%	197us 72.5%	179us 72.0%	165us 71.0%	145us 70.4%	126us 68.7%	101us 66.5%	...
PSS-MBV2-C	CPU Latency	42ms 73.3%	30ms 73.2%	27ms 73.0%	25ms 72.5%	23ms 72.2%	21ms 71.0%	19ms 70.3%	18ms 69.5%	17ms 69.0%	16ms 66.7%	...
PSS-MBV2-M	FLOPs	301M 73.4%	295M 73.4%	249M 72.7%	192M 71.9%	163M 70.9%	154M 70.7%	132M 70.0%	104M 68.7%	83M 66.7%	69M 66.4%	...
	GPU Latency	260us 73.4%	226us 72.9%	202us 72.5%	197us 72.4%	174us 71.9%	150us 70.7%	139us 70.0%	121us 67.5%	102us 66.7%	101us 66.4%	...
	CPU Latency	33ms 73.4%	32ms 73.2%	31ms 73.1%	29ms 72.9%	27ms 72.4%	24ms 72.2%	22ms 70.7%	20ms 69.4%	18ms 68.4%	16ms 66.6%	...

We only list the results of some subnets of the supernet, if you want to get the results of all subnets in the supernet, you can check the training logs or download the supernet weights for evaluation by using the links in the first column of the table.

Running the following command to get the above figure:

python ./scripts/plot_main_result.py