MobileOne: An Improved One millisecond Mobile Backbone

Installation

conda create -n PyTorch python=3.8
conda activate PyTorch
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch-lts
pip install opencv-python==4.5.5.64
pip install tqdm

Note

• The default training configuration is for mobile_one-s0
• The test results including accuracy, params and FLOP are obtained by using fused model

Parameters and FLOPS

Number of parameters: 2078504
Time per operator type:
        15.0684 ms.    91.0851%. Conv
        1.20933 ms.     7.3101%. Relu
       0.242441 ms.     1.4655%. FC
      0.0117301 ms.  0.0709057%. AveragePool
     0.00421935 ms.   0.025505%. Reshape
     0.00261659 ms.  0.0158167%. Gather
     0.00200163 ms.  0.0120994%. ExpandDims
     0.00170158 ms.  0.0102857%. Concat
      0.0007769 ms. 0.00469618%. Shape
        16.5432 ms in Total
FLOP per operator type:
       0.548173 GFLOP.    99.6276%. Conv
       0.002049 GFLOP.   0.372395%. FC
              0 GFLOP.          0%. Concat
              0 GFLOP.          0%. Relu
       0.550222 GFLOP in Total
Feature Memory Read per operator type:
        19.7686 MB.    50.6551%. Conv
        15.1532 MB.    38.8285%. Relu
         4.1041 MB.    10.5164%. FC
        1.2e-05 MB. 3.07489e-05%. Concat
        39.0258 MB in Total
Feature Memory Written per operator type:
        15.1532 MB.    49.9934%. Conv
        15.1532 MB.    49.9934%. Relu
          0.004 MB.  0.0131968%. FC
          8e-06 MB. 2.63937e-05%. Concat
        30.3103 MB in Total
Parameter Memory per operator type:
         4.1801 MB.    50.4837%. Conv
            4.1 MB.    49.5163%. FC
              0 MB.          0%. Concat
              0 MB.          0%. Relu
         8.2801 MB in Total

Train

• Configure your IMAGENET dataset path in main.py for training
• Run bash main.sh $ --train for training, $ is number of GPUs

Test

• Configure your IMAGENET path in main.py for testing
• Run python main.py --test for testing

Results

Version	Epochs	Top-1 Acc	Top-5 Acc	Params (M)	FLOP (G)	Download
mobile_one-s0	300	-	-	2.08	0.550	-
mobile_one-s0*	300	71.4	89.9	2.08	0.550	model
mobile_one-s1*	300	75.8	92.8	4.76	1.650	model
mobile_one-s2*	300	77.4	93.2	7.80	2.596	model
mobile_one-s3*	300	77.9	93.9	10.07	3.791	model
mobile_one-s4*	300	79.3	94.4	14.83	5.960	model

• * means that weights are ported from original repo, see reference

Reference

• https://github.com/apple/ml-mobileone