meijieru / atomnas Goto Github PK

Both Top-1 and Top-5 accuracies of three distinct models are being improved uniformely with respect to the baseline accuracies after the introduction of the swish activation and SE modules:
AtomNAS-A:74.6->76.3 (+1.7), 92.1->93.0 (+0.9)
AtomNAS-B:75.5->77.2 (+1.7), 92.6->93.5 (+0.9)
AtomNAS-C:75.9->77.6 (+1.7), 92.7->93.6 (+0.9)
which is an eye-catching uniformity that rarely appears. Is there a reasonable explaination for this?
Thanks very much for your reply!

Hi, thank you for your interesting work.

1. Why the FLOPs of MobileNetV2 your implementation (301M) is different from the original paper (300M) in Table 1?
2. Can you provide the pre-trained model?

I find running script will create directory recursively, i.e., establish result dir many times. I think it attribute to the following code block but do not know why, could u kindly give me a hint?

    shutil.copytree(scripts_dir,
                    dst_dir,
                    ignore=functools.partial(
                        ignore_func,
                        suffixs=['.py', '.pyx'],
                        blacklist_dirs_abs=blacklist_dirs_abs))
    config_path = os.path.relpath(config_path, scripts_dir)
    if not config_path.startswith(config_dir):
        raise RuntimeError('Config files assume to live in `apps`')
    shutil.copytree(os.path.join(scripts_dir, config_dir),
                    os.path.join(path, config_dir),
                    ignore=functools.partial(ignore_func,
                                             suffixs=['.yml', '.yaml']))

I use this command to test AtomNas-a.
FILE=$(realpath pretrained/atomnas-a) CHECKPOINT=ckpt ATOMNAS_VAL=True bash scripts/run.sh apps/eval/eval_shrink.yml

Missing keys and size mismatch occur.

RuntimeError: Error(s) in loading state_dict for AllReduceDistributedDataParallel:
        Missing key(s) in state_dict: "module.features.5.ops.2.0.0.weight", "module.features.5.ops.2.0.1.weight", "module.features.5.ops.2.0.1.bias", "module.features.5.ops.2.0.1.running_mean", "module.features.5.ops.2.0.1.running_var", "module.features.5.ops.2.1.0.weight", "module.features.5.ops.2.1.1.weight", "module.features.5.ops.2.1.1.bias", "module.features.5.ops.2.1.1.running_mean", "module.features.5.ops.2.1.1.running_var", "module.features.5.ops.2.2.weight".
        size mismatch for module.features.4.ops.0.0.0.weight: copying a param with shape torch.Size([9, 24, 1, 1]) from checkpoint, the shape in current model is torch.Size([10, 24, 1, 1]).
        size mismatch for module.features.4.ops.0.0.1.weight: copying a param with shape torch.Size([9]) from checkpoint, the shape in current model is torch.Size([10]).
        size mismatch for module.features.4.ops.0.0.1.bias: copying a param with shape torch.Size([9]) from checkpoint, the shape in current model is torch.Size([10]).
        size mismatch for module.features.4.ops.0.0.1.running_mean: copying a param with shape torch.Size([9]) from checkpoint,
the shape in current model is torch.Size([10]).
...

I can test AtomNas-b and AtomNas-c correctly, so I think the code is right.
Is there something wrong with AtomNas-a ckpt?

In the code, it seems that you use the model stored in ema to do the validation and get the final result.Here i have two questions:
1.Why do you use the ema model to do the validation, but do not use the model_wrapper which is directely preuned?
2.Is it supposed that ema model will get better result than model_wrapper?And why?
Thanks very much for your reply!

you use nccl in the distributed training, my problem is do you use nccl in pytorch or do you install nccl
seperately?And how do you set your environment variable?I am queite confused about it.Thanks very much!I meet the following problem when i use two machine to run the code.

1. INFO NET/Plugin : No plugin found (libnccl-net.so)
   2.NCCL error in: /pytorch/torch/lib/c10d/ProcessGroupNCCL.cpp:400, unhandled cuda error.
   3.NCCL INFO NET/IB : No device found

Thx！！

Hi, thanks for your great work!
Do you plan to release the search code?

@meijieru
Hi, thanks for your great work! The first block of AtomNAS-C in the code only have single kernel size: 3 (first block:[1, 16, 1, 1, [3]]), but the Figure 5 in the paper show there have three kernel size(show as three difference color)，why? looking forward your reply!

File "train.py", line 375, in run_one_epoch
data_iterator.data_queue.qsize(),
AttributeError: '_MultiProcessingDataLoaderIter' object has no attribute 'data_queue'

In torch.utils.data.DataLoader, different point is a data_queue
ver 1.2.0: data_queue
ver 1.3.0: _data_queue