MobileNetV1 from HuggingFace

[deJQK]

I want to train mobilenetv1 with timm and I found it only provide v2 and v3. I am trying to use the model from huggingface named Matthijs/mobilenet_v1_1.0_224. However, after downloading it will report error saying the model name not exist in the factory, at this step. I am new to timm but wonder how to solve this. Thanks!

[deJQK]

The reason is that in the provided huggingface model, the pretrained_cfg has 'architectures' instead of 'architecture' as a key, so this line can be followed by if model_name is None: model_name = pretrained_cfg.get('architectures')[0]. However, the model_name of MobileNetV1ForImageClassification is not defined nor registered so this line will raise an error. I do not know how to use the MobileNetV1ForImageClassification as stated on this page. Specifically, I am not able to do from transformers import MobileNetV1ForImageClassification.

My final solution to use MobileNetV1 is to define it by myself in a local models directory, copying and modifying the module EfficientNet to a new module MobileNet by removing the argument num_features and the conv_head block, and copying and modifying this function for MobileNetV2 by using several blocks of dsa_r*_k*_s*_c*, and modifying the other necessary functions like this, together with modified default_cfgs and __init__.py.

[rwightman]

@deJQK thanks for posting your own solution, heh.... I've actually wanted to provide a factory that could load hugging face pretrained models for eval and train with the timm scripts, as there aren't scripts with the same capability for augmentation, etc included in HF repos.

However, ended up being a bit more work than I expected, having to adapt args, do weight adaptation for head changes, setting dropout, etc. I will tackle it again at some point and talk with some other people on the HF team for help.

The efficientnet.py does fully support mobilenetv1 if you copy the v2 creation fn as you say and switch to ds_ down the stages, change act and features/widths, I think I actually had it in there at one point a long long time ago, but removed it because I never planned to train it and it was just taking up code space :) v2 and v3 and other related (mnasnet, harcorenas, tinynet, etc) are much more capable...

[deJQK]

Thanks @rwightman for your reply. I am trying to train MobileNetV1 because it is faster for experiment (due to much less layers). Also, there are more results from literature for comparison, compared to other architectures.

[rwightman]

Found it, but some of the fns, args might have changed... dsa is for depthwise separable w/ final activation, the ds used in others is similar to InvertedResidual in that it has linear output

def _gen_mobilenet_v1(channel_multiplier, num_classes=1000, **kwargs):
    """ Generate MobileNet-V1 network
    """
    arch_def = [
        ['dsa_r1_k3_s1_c64'],
        ['dsa_r2_k3_s2_c128'],
        ['dsa_r2_k3_s2_c256'],
        ['dsa_r6_k3_s2_c512'],
        ['dsa_r2_k3_s2_c1024'],
    ]
    bn_momentum, bn_eps = _resolve_bn_params(kwargs)
    model = EfficientNet(
        _decode_arch_def(arch_def),
        num_classes=num_classes,
        stem_size=32,
        num_features=1024,
        channel_multiplier=channel_multiplier,
        channel_divisor=8,
        channel_min=None,
        bn_momentum=bn_momentum,
        bn_eps=bn_eps,
        act_fn=F.relu6,
        head_conv='none',
        **kwargs
        )
    return model

[deJQK]

Thanks. Yes, this might be an early version of EfficientNet as the current one does not have argument head_conv. Also, MobileNetV1 used ReLU activation instead of ReLU6. But thanks for the sharing. I have edited my original answer from ds_* to dsa_* and thanks for pointing out this.

[rwightman]

I'm pretty sure it's ReLU6...
https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.py#L428-L479
https://github.com/keras-team/keras/blob/v2.9.0/keras/applications/mobilenet.py#L363-L439

[deJQK]

Thanks @rwightman . You are right, it uses ReLU6. It is interesting that the author did not mention this in the original paper. They said 'MobileNets use both batchnorm and ReLU nonlinearities for both layers.' The first time I can find about ReLU6 is in the paper of MobileNetV2 but the authors reference to MobileNetV1 when they were discussing about this. Also, I checked the earlier version of the tensorflow git and it seems the original version uses ReLU6.