VDIGPKU/M2Det

Reproduce of vgg16_m2det320

Closed this issue · 15 comments

I tried to reproduce the result of vgg16_m2det320 twice,
the first time I used the setting in the paper:

lr = [0.002, 0.0002, 0.00002, 0.00002]
COCO = [6, 90, 120, 150]

the second time I used your new setting:

lr = [0.004, 0.002, 0.0004, 0.00004, 0.000004]
COCO = [90, 110, 130, 150, 160]

But both time I got mAP of 22.2, I don't know why.

Show me the more detailed information. batch size, num of GPUs

Show me the more detailed information. batch size, num of GPUs

@qijiezhao The first time, batch_size is 8 and num of GPUS is 7

Show me the more detailed information. batch size, num of GPUs

The second time training, batch_size is 8, GPUs is 4

You may contact with: #5 (comment)

By the way, test-set accuracy is always 0.2~0.4 higher than val-set for MS-COCO.

I found that it is the problem of batch size, if the total batch size is larger than 32, it may cause the problem.

Have you reproduced?

Have you reproduced?

Yes, I got mAP33.1 for vgg16+320