Devin Yang
Devin Yang
Hi @iamhankai thanks for sharing this good work. I success training GhostNet 1.3x to `75.78/92.77 top1/top5`, it's almost your paper mentioned. Details [here](https://github.com/PistonY/ModelZoo.pytorch) But I use the same training setting...
And you not specify `dw_kernel_size` for GhostNet in paper Table 7, are you using `dw_kernel_size=3` by default?
Hi @iamhankai, thanks for reply and open source. But I still confuse about how to get TinyGhostNet-X(B/C/D/E) from TinyGhostNet-A, are using using same param with EfficientNet?
Hi @lucidrains, thanks for reply. I tested them all, I think you're right. When apply `bn+relu` the val accuracy doesn't grow. This is my final [implement](https://gist.github.com/PistonY/ad33ab9e3d5f9a6a38345eb184e68cb4). Now I'm training `LambdaResnet50`,it's...
@lucidrains Unfortunately, I only got 76.1 best top1 on val set(79.2 on train set). I'd better wait author release their code.
I try out `LambdaResnet50` with 64 batch_size about cost 9-10GB gpu memory in FP32 precision,it's much larger than `Resnet50`
M@sriharikarnam Does this has any progress?Is this work still going on?
Same question
Recently I use [distribute train](https://github.com/PistonY/ModelZoo.pytorch/blob/master/scripts/distribute_train_script.py) more often. You need to make sure single gpu has same batch size with me, you should get same result but may take more time...
No need to change I think. This paper should mean batch size on one device, normally batch size in paper just mean on device hold, take care of the difference...