Coldmooon

我写了一个在 Xcode 里编译这个 repository 的方法，这样就可以进行单步调试了。 对于 Linux 下的 eclipse， 同样的思路也应该适用。 http://coldmooon.github.io/2015/08/14/compile_caffe_cpp/

@vadiklukmanov They are different tasks. The default torchvision dataset requires a single digit prediction, not a sequence of digits.

Open the file `./cmake/MakePythonWheel.cmake`, and you will find the code: ``` message(STATUS "Python module `setuptools` required for correct wheel filename generation. Please install if needed.") set(wheel_filename "unknown;unknown") ``` This is...

Sorry, The code is currently for Caffe and Torch. I haven't implemented the PyTorch version. MPELU can be factorized as the production of ELU and PRELU. So it is convenient...

Now MPELU has supported pytorch.

It has already been clear. It's a shape mismatch problem. Modify your prototxt file if you want to finetune conv4, or rename the layer if you want to learn this...

@seandales This is a different question. Yes, whenever the model is tested, some `comp4_det_test` files are outputted. These files are in your data directory which is defined by the `output_result_dir`...

@SikhaS SSD has considered the class of background. If you have only one class, just set `num_classes` in `./examples/ssd/ssd_xxx.py` to 2.

It seems like a bug or a mistake somewhere... Can you use `pycaffe` to print intermediate output? If loss is always 0, the error might occur earlier..

@guoqiang01486 Try decreasing the learning rate or tuing other hyper-parameters. If this does not work, there may be other bugs...