lizhengwei1992/mobile_phone_human_matting

human matting on mobile phone

mobile_phone_human_matting's Issues

conver to ncnn

Can you tell me how to convert ncnn, i used the pytorchconverter tool failed。

RuntimeError: Assertion `cur_target >= 0 && cur_target < n_classes' failed.

@lizhengwei1992 ,
在训练的时候，计算交叉熵这里cross_entropy_loss = criterion(seg, mask_gt[:, 0, :, :].long())报错，查了些资料没搞明白什么原因，其中criterion的参数shape如下
seg.shape torch.Size([1, 2, 256, 256])
mask_gt torch.Size([1, 256, 256])
能帮忙看看如何解决么

split not supported yet when convert to ncnn

hi, your network use torch.split operation, however when I convert trained onnx model to ncnn using the ./onnx2ncnn tool, there is some error: "split not supported yet" when convert to ncnn.
Could you provide the detail about how you convert your network to ncnn? Or does ncnn support the torch.split operation?

Ios coreml model

Hi,
Do you have the converted model that you demoed on your iphone 6?

License

Is this repo for research only? Can we get a use license commited?

fusion loss

hi, thank you for your share.
In the fusion loss, you refer to "paper loss", could you tell me the name of paper?

problem in converting pth to onnx

I have tried pytorch 0.4.0, 0.4.1, 1.0 , 1.2 to export the model to onnx file, but all of them dont work.
Will you share the onnx file in the future? Thanks a lot

Weird speed performance on CPU and worse accuracy than your provided weights.

Thank your for sharing the awesome project.

I tried to train the model on supervisely dataset by following the steps:

1. Resize downloaded images to the long side of 512 by keeping original ratio and generate train list.
2. Modify line 14 in main.py net-> segnet
3. Modify args('dataDir') of train.sh and run train.sh

After training, I found:

1. Inference time on CPU is 3-4 times of your provided weights on (test on both Windows and Mac, PyTorch0.4/0.41/1.0.1). After adding the "torch.set_flush_denormal(True)", speed performs normally, but I don't know what caused this difference.
2. Accuracy is not as good as your provided weights.

I'm wondering if you had ever found these phenomenons.
It would be great if you could share more details and suggestions.

Thanks.

Format of Train List

What must be the format of the train list?
My data has:

|--Train
|---Image
|---mask (segmentation mask)
|---alpha
|---trimap

So what must be the train List format:

• image/mask
• or
• image/trimap
• or
• image/alpha

Detailed explanation about the training data

Hello
How are you?
Thanks for contributing this project.
Could u provide a detailed explanation about the training data?
Thanks

details about how to prepare dataset?

I don't know how to prepare dataset for the training, could you tell me some details about it? Thanks!

convert to onnx

Hello， i converted pth model to onnx model, but i cannot get the same nice result when i used the onnx model. Could you give me some advices? The following code is mine.

      session = onnxruntime.InferenceSession(args.model)
      input_name = [input.name for input in session.get_inputs()][0] 
      output_names = [output.name for output in session.get_outputs()]

      origin_h, origin_w, c = image.shape
      image_resize = cv2.resize(image, (INPUT_SIZE,INPUT_SIZE), interpolation=cv2.INTER_CUBIC)
      image_resize = (image_resize - (104., 112., 121.,)) / 255.0
      image_resize = image_resize.astype(np.float32)
      image_resize = np.transpose(image_resize, [2, 0, 1])
      image_resize = image_resize[np.newaxis,:,:,:]
      seg, alpha = session.run(output_names, {input_name: image_resize})
      alpha_np = alpha[0, 0, :, :]
      fg_alpha = cv2.resize(alpha_np, (origin_w, origin_h), interpolation=cv2.INTER_CUBIC)

Corresponding papers for each component of the network

Hello
How are you?
Thanks for contributing this project.
We want to know about the corresponding papers for each components (encoder, decoder and feathering module) of your network.
Thanks

ckpt_lastest.pth error

It seems to be the best model， when i use it， got error：
myModel.eval()
AttributeError: 'dict' object has no attribute 'eval'

I wonder if this model is better than model_obj.pth

how do you create tensor_4D input for using in ncnn?