yangfly/sfd.gluoncv

Reproduce SFD face detector using gluon-cv

Description

This work is to reproduce S³FD, a real-time Single Shot Scale-invariant Face Detector.

Environment

• mxnet 1.3.0
• gluon-cv 0.4.0
• MobulaOP required to build custom c++ operator

Preparation

# install MobulaOP following: https://github.com/wkcn/MobulaOP
$$ git clone https://github.com/yangfly/sfd.gluoncv.git
$$ cd sfd.gluoncv

Training your own model

1. download wider face dataset into widerface/downloads/

   $$ tree widerface/downloads
   widerface/downloads
   ├── eval_tools.zip
   ├── Submission_example.zip
   ├── wider_face_split.zip
   ├── WIDER_test.zip
   ├── WIDER_train.zip
   └── WIDER_val.zip
   

2. Parpare data: unzip data, annotations and eval_tools

   $$ python tool/parpare.py
   $$ tree widerface -L 1
   widerface
   ├── downloads
   ├── eval_tools
   ├── wider_face_split
   ├── WIDER_train
   └── WIDER_val
   

3. Prepare custom val dataset for quick validation (crop and resize to 640)

   $$ python tool/build_custom_val.py
   $$ tree widerface -L 1
   widerface
   ├── downloads
   ├── eval_tools
   ├── WIDER_custom
   ├── wider_face_split
   ├── WIDER_train
   └── WIDER_val
   

4. train vgg16 based sfd with 4 gpus

   $$ python sfd/train.py
   

   more supported base models in sfd/nn/sfd.py
5. demo

   $$ python sfd/demo.py --model models/vgg16/sfd_best.params
   

6. eval on WIDER_val

   $$ python sfd/eval.py --model models/vgg16/sfd_best.params
   

sfd.gluoncv's Issues

请问，在运行train.py 跑到validate 报anchor的参数错误，该怎么解决？ 谢谢

names, maps = validate(net, val_data, ctx, eval_metric)

mxnet.base.MXNetError: Error in operator sfd0_sfdanchorgenerator0_slice_like0: [18:11:18] src/operator/tensor/./matrix_op-inl.h:1292: Check failed: ishape[axis] >= from_shape[axis] (160 vs. 170) Slice axis 2 with size 170exceeds limit of input with size 160

请问您训练的结果提供下载嘛？

RT

AttributeError: module 'mobula.func' has no attribute 'compensate'

when i train the model，I get this error:

File "D:\Users\admin\Anaconda3\Lib\multiprocessing\process.py", line 297, in _bootstrap
self.run()
File "D:\Users\admin\Anaconda3\Lib\multiprocessing\process.py", line 99, in run
self._target(*self._args, **self._kwargs)
File "D:\Users\admin\Anaconda3\Lib\site-packages\mxnet\gluon\data\dataloader.py", line 184, in worker_loop
batch = batchify_fn([dataset[i] for i in samples])
File "D:\Users\admin\Anaconda3\Lib\site-packages\mxnet\gluon\data\dataloader.py", line 184, in
batch = batchify_fn([dataset[i] for i in samples])
File "D:\Users\admin\Anaconda3\Lib\site-packages\mxnet\gluon\data\dataset.py", line 128, in getitem
return self._fn(*item)
File "E:\gc\sfd.gluoncv\sfd\data\transform.py", line 91, in call
self._anchors, None, gt_bboxes, gt_ids)
File "D:\Users\admin\Anaconda3\Lib\site-packages\mxnet\gluon\block.py", line 542, in call
out = self.forward(*args)
File "E:\gc\sfd.gluoncv\sfd\nn\anchor.py", line 140, in forward
matches = self._matcher(ious)
File "D:\Users\admin\Anaconda3\Lib\site-packages\mxnet\gluon\block.py", line 542, in call
out = self.forward(*args)
File "E:\gc\sfd.gluoncv\sfd\nn\matcher.py", line 30, in forward
mobula.func.compensate(x.shape[0], gtids, ious, match,
AttributeError: module 'mobula.func' has no attribute 'compensate'