EfficientNet 测试

需要torch1.10版本

最快的需要25ms gpu

https://github.com/zsef123/EfficientNets-PyTorch

还有这个：

https://github.com/jacke121/efficientnet-pytorch

 

1070上：efficientnet-b0

512  batch_seze 1需要25ms， 4需要56ms，batch_size为6就内存溢出

416  batch_seze 1需要25ms， 4需要56ms，batch_size为10就内存溢出

proxyless_gpu batch_size 20 可以，再大就不行了。

def test():
    x = torch.FloatTensor(4, 3, 512, 512).cuda()

    w, d, _, p = efficientnet_params('efficientnet-b0')
    # note: all models have drop connect rate = 0.2
    blocks_args, global_params = efficientnet(width_coefficient=w, depth_coefficient=d, dropout_rate=p)

    model=EfficientNet(blocks_args, global_params)
    model.cuda()
    model.eval()
    for i in range(2000):
        t1 = time.time()
        out3= model(x)
        # print(out3)
        cnt = time.time() - t1
        print(cnt, out3.size())

if __name__ == '__main__':
    test()

https://github.com/lukemelas/EfficientNet-PyTorch/blob/master/efficientnet_pytorch/model.py