megvii-model/singlepathoneshot

singlepathoneshot's Issues

One called function in EA search has no effect and should be removed

Hi, thanks for your excellent work while I have one confusion below

In src/Search/search.py, line230, below called function has no effect in every search loop

self.get_random(self.population_num)

After the mutation and crossover, the total counts of the self.candidates is 50. When calling the above function, the while loop will not be executed because the bool expression in line 139 is False.

Besides, line 131, variable info pointed to self.vis_dict[cand] is never used which should be removed too maybe😉

Regards,

Supernet accuracy

I train the supernet with the default setting, and got an accuracy of

TRAIN Iter 150000: lr = 0.000003, loss = 2.805815, Top-1 err = 0.417285, Top-5 err = 0.202197, data_time = 0.050527, train_time = 0.565107

Is this accuracy right?

Channel search

Thanks for your excellent work!

I seem to only see the block search, no channel search found.

Evolution search results

Hi. Sorry to bother again.

I tried to run your evolution search code, with checkpoints-150000 provided. The results look like this:

which is nowhere near the results shown in paper (over 68 in general).

Trying to dig out why, I found that recalculating BN on 20k images brings little improvement (less than one percent) to validation performance, compared to inheriting BN directly from the supernet. I'm not sure whether there is something wrong here. Have you guys observed a similar phenomenon?

Or, my wild guess is, the supernet you have shared is not a well-trained supernet?

Why set `affine=False` in all `nn.BatchNorm2d`

I fail to figure out a reasonable explanation. Could anyone help me out?

reproduce retrain result

Using the architecture results from the README search((2, 1, 0, 1, 2, 0, 2, 0, 2, 0, 2, 3, 0, 0, 0, 0, 3, 2, 3, 3)), and using the Retrain code from the repository (with no additional modifications), I ended up with [11 02:26:43] te s t iter 301250: Loss = 2.017752, top-1 ERR = 0.262860, Top-5 ERR = 0.084000, val_time = 71.723475. This is a little different from the TOP-1 error=25.6 reported by the repo. Have you re-run the evaluation code of the repo to get similar results?

link 进不去

代码运行时间

请问默认 Supernet/train.py 在 imagenet上要运行多久？我用两块V100，20个小时只跑了 8880 Ite。预计完整跑完需要两周？查看了一下，发现GPU的利用率很低，不到10%。是哪里设置的有问题吗？不知你们如何解决的？原版代码，没有改动。

Ckpt

1

Use BGR tensor in data preprocessing

Hi. Thanks for this great repo.

I'm interested in why you did convert to BGR tensor intentionally and keep the values between 0 - 255 instead of normalization (div 255, minus mean, div std) as usual.

Related code:

SinglePathOneShot/src/Supernet/train.py

Line 37 in 36eed6c

class ToBGRTensor(object):

What's the minimum flops-limit in the search?

Thx！！！

Evaluate pretrained model

Hello, I follow your order to get the arch of the model and load the weight in Onedrive/Evaluate, but I get the 0.368 accuracy. Did I make mistakes?

resource in onedrive missing

hi, i want to download the flops-dict and surpernet resource, but they are missing in the Link.

SuperNet Training 代码优化

您好，在supernet training中，看到关于梯度收集这块做了优化，如下：

        for p in model.parameters():
            if p.grad is not None and p.grad.sum() == 0:
                p.grad = None

请问这个原理是什么？

Why not shuffle validation data during search

Hi. Thanks for this great repo.

SinglePathOneShot/src/Search/imagenet_dataset.py

Line 94 in 36eed6c

batch_size=batch_size, shuffle=False,

This seems to have produced very unstable validation results.

What is the purpose of setting flops bins when training SuperNet?

SinglePathOneShot/src/Supernet/train.py

Line 213 in 36eed6c

bins = [[i, i+flops_step] for i in range(flops_l, flops_r, flops_step)]

It seems that author want the candidates' flops distribute uniformly among [290, 360].

However, I random sampled 12500 candidates and calculate their flops. The nature distribution of the ShuffleNet SuperNet is like [ 132. 828. 2366. 3941. 3482. 1406. 312.], each number indicates the random sampled candidates whose flops are among [[290, 300], [310, 320], ..., [350, 360]].

If you don't put any constrain on candidates' flops, the occurrence probability of each choice block in each layer is very close to a uniform distribution(because of numpy.random.randint) which is consistent with the paper.

However, once the flops constrain is added, the occurrence probability of choice block will not strictly follow uniform distribution.

So I am curious about the motivation？Or it is just a experimental experience？

how can I constraint FLOPs?

thank you for providing the code of your paper!
i'm starting to study autoML these day and i've found your great work through this paper

and i've seen many expression 'FLOPs' in your paper and 'mobile setting(<600M FLOPs)' in others.
and i encountered this question.

Q. constraint FLOPs manually in the code is the normal(ordinary) way to experiment in 'mobile setting' (or something)?

thank you for answering in advance despite your busy schedule :)

Missing arguments in `validate`.

SinglePathOneShot/src/Supernet/train.py

Lines 263 to 273 in 36eed6c

	model.eval()
	max_val_iters = 250
	t1 = time.time()
	with torch.no_grad():
	for _ in range(1, max_val_iters + 1):
	data, target = val_dataprovider.next()
	target = target.type(torch.LongTensor)
	data, target = data.to(device), target.to(device)
	output = model(data)
	loss = loss_function(output, target)

As above, there is a missing argument in model(data, arch).

Can not open the $Link

Could you upload your $Link resources to BaiduYun?