search

toandaominh1997 / EfficientDet.Pytorch

Implementation EfficientDet: Scalable and Efficient Object Detection in PyTorch
MIT License
1.44k stars 306 forks source link

EfficientDet D0 weight SIZE is 40MB??? #100

Open jduki opened 4 years ago

jduki commented 4 years ago

As I read paper, efficient d0 size should be 3.8MB. Why d0 checkpoint weight included on git has 40MB size ??

midasklr commented 4 years ago

As I read paper, efficient d0 size should be 3.8MB. Why d0 checkpoint weight included on git has 40MB size ??

I guess 3.8M, which is 3.9M in the paper, is the number of parameters(see table 2...#Params),not the extact store space in the disk.

nsarang commented 4 years ago

As I read paper, efficient d0 size should be 3.8MB. Why d0 checkpoint weight included on git has 40MB size ??

I guess 3.8M, which is 3.9M in the paper, is the number of parameters(see table 2...#Params),not the extact store space in the disk.

Then, shouldn't it be around 16MB since the weights are in float32?

midasklr commented 4 years ago

As I read paper, efficient d0 size should be 3.8MB. Why d0 checkpoint weight included on git has 40MB size ??

I guess 3.8M, which is 3.9M in the paper, is the number of parameters(see table 2...#Params),not the extact store space in the disk.

Then, shouldn't it be around 16MB since the weights are in float32?

yes,u r right. the weight file is a orderdict include both weights and bias and other param names (orderdict keys)and their values(around 3.9m ).

dvlshah commented 4 years ago

It is 40M because it stores the optimizer state during training in the weight file itself. Store only state dict to disk and you will get the exact weight.

guanshuicheng commented 4 years ago

As I read paper, efficient d0 size should be 3.8MB. Why d0 checkpoint weight included on git has 40MB size ??

3.8M明显是参数的数量