tropcomplique / mtcnn-pytorch Goto Github PK

I found that it is a bit slow when I run this code. Then I check the usage of gpu and I found this code was run on the cpu. So I want to change the code to run on the gpu. But I found the type of weight of this model is torch.FloatTensor. Is that mean I cannot use the gpu to run this code directly? Is that any solution can help me to run this code on gpu?

Here is the error:
RuntimeError: Expected object of type torch.FloatTensor but found type torch.cuda.FloatTensor for argument #2 'weight'

Hello, thank you so much for your great works.

1. Why do we need use this line img = (img - 127.5)*0.0078125 in preprocess image?
2. I have try to convert it in opencv, but seems like the number of bounding box in the final is same, but in the process is much reduced. Do you think this is because of BGR to RGB conversion?, since opencv is BGR so I convert it first to RGB.
3. Why you not using pnet. eval() and rnet.eval() ?

Python 3.5.2 (default, Nov 23 2017, 16:37:01) 
[GCC 5.4.0 20160609] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from src import detect_faces
>>> from PIL import Image
>>> image = Image.open('images/example.png')
>>> bounding_boxes, landmarks = detect_faces(images)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
NameError: name 'images' is not defined
>>> bounding_boxes, landmarks = detect_faces(image)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/ubuntu/mtcnn/src/detector.py", line 25, in detect_faces
    pnet = PNet()
  File "/home/ubuntu/mtcnn/src/get_nets.py", line 56, in __init__
    for n, p in self.named_parameters():
  File "/usr/local/lib/python3.5/dist-packages/torch/nn/modules/module.py", line 235, in __getattr__
    return object.__getattribute__(self, name)
AttributeError: 'PNet' object has no attribute 'named_parameters'

I can see at get_nets.py line #56 the offending code, and I can see that PyTorch supports that function, so it seems a bit bizarre to me.
I have PyTorch 0.1.10 on a Jetson TX2.

I tested the mtcnn with a image with no face, the program crashed.

It seems that code in detect_faces did not handle the case that there is no bounding box detected in the image. Is that so ?

first thanks for your sharing code, it is good for me to use pytorch get the model test, I want to know can this model be trained directly using pytorch, instead converting the caffe model to npy.

Hi TropComplique,

First, very appreciate for your great work, learned much from your source code.

I have one question, why we need to switch dim2 a dim3 when convert Caffe weights to PyTorch like below:

extract_weights_from_caffe_models.py
all_weights[name + '.weight'] = net.params[p][0].data.transpose((0, 1, 3, 2))

Does the shape of Caffe weight is [N, C, W, H]? So we need to change it to [N, C, H, W], right?

Hope for your reply.

I'm interested in OpenVino lately. And I found the question you asked at here. I would like to try your OpenVino version notebook and learn how to play with a openvino mtcnn model. Unfortunately, the link of notebook have been invalid. So I wish that you can give me some help. May I get your openvino version notebook along with the models that you successfully transferred?
Thx a lot :)

Hi! I just upload this project to PyPI repository for easy installation and made some minor changes. If this is ok for you.

Here is my repos with changes:
https://github.com/khrlimam/mtcnn-pytorch

A few change I made:

1. Delete all the files that unnecessary for the package it self.
2. Rename the package directory.
3. Upload the weights to Github release.
4. Load the weights online.

Hi, from the paper I infer that this method was trained for different datasets. Could you tell me which dataset your model was trained on?