Martin Han

Cool. Thanks for the PR! Light and not requiring keras or tf is something that I am trying to achieve. Maybe you can take a look at this code I...

Hi, you need to attach the error message and upload your model. Permute is currently not supported by the converter since I don't have model that require permute. I will...

See it, I will take a try.

I think you are not using the latest one. You may need to run pip upgrade.  You may also need to modified you code to remove concat axis -1,...

Cool, let me have a look. It should not be very hard to support hs

Hi. I am not able to find any code that related to _hard_swish activation on keras/tf. Since hard swish need parementers to descripe its shape, and there is none of...

See it... yes, I think you are right. I check the original formula, and I am not able to link the hard swish formula to ncnn's [implementation](https://github.com/Tencent/ncnn/blob/master/src/layer/hardswish.cpp/). Let me ask...

It should support in d7d158e, but to be awared, also mentioned in here: https://github.com/Tencent/ncnn/issues/2826#issue-853650801. Keras and pytorch have different defination for the hardsigmoid. Keras using +/- 2.5, but pytorch is...

> A new problem . > How to support user-defined functions, similar to the load function in keras has custom_objects. > > ` model = k.models.load_model(model_path, compile=False, custom_objects={ 'my_fun': my_fun...

I find a stupid bug in Mul operator, fixed it in e3f90c4. Now I am able to run the model. ```bash Convolution conv2d_1 7.63ms | [160, 160, 3 *1] ->...