The size of ckpt is 7.5G?

[cumtwb]

I trained the network on my own dataset. But the size of saved checkpoint is 7.5G. Would you like to tell how much space your checkpoint occupies?

[heart4lor]

in my case one single ckpt occupies ~1.5G, but I set --light parameter to true because of poor GPU memory.

[HLearning]

in my case one single ckpt occupies ~1.6G, but I set --light parameter to true because of poor GPU memory.

[thewaifuai]

Using the light parameter my checkpoints were about 1.5G each.

[cumtwb]

Thank you! But I still have two problem:

1. Why the checkpoint of UGATIT is so huge? The ckpt of MUNIT is 100~150M, of CycleGAN is only ~40M.
2. The unique difference of setting the --light parameter to true or false is in the MLP function of AdaLIN. If --light==true, global average pooling will be used to input x. The output unit number of next fc layer is 64 4(ch n_res)=256.

light	false	true
x	64 64 256	256
fc	64 64 256 * 256	256 * 64

A float32 parameter occupies 4B, all parameters of one full fc layer occupy 64 64 256 256 4B = 1G. Therefore, the full a2b and b2a generators should occupy ~2G than the light generators. But my full ckpt occupies ~7.5G, is 6G than your light ckpt. Why? @heart4lor @HLearning @thewaifuai

[mchalecki]

Original checkpoint(not light model) size is 7.5G