search

ijkguo / mx-rcnn

Parallel Faster R-CNN implementation with MXNet.
Other
671 stars 290 forks source link

How to use the branch for training using gluon cv?? #106

Closed Ram-Godavarthi closed 6 years ago

Ram-Godavarthi commented 6 years ago

@ijkguo Hi, Can You provide me some steps for working on gluon cv in dev branch? I am getting very bad results in master branch. Training script (train.py) is wrong i guess in Master branch.. Test.py working fine if i use it with the other trained model which was trained using previous commit in apache mxnet repo.. Now that repo has also changed... Could you please help me out in this..

Thank You

ijkguo commented 6 years ago

It is not wrong. Training log: https://raw.githubusercontent.com/dmlc/web-data/master/gluoncv/logs/detection/faster_rcnn_resnet50_v2a_voc_train.log

ijkguo commented 6 years ago

I finally understand what you mean.

The master branch is consistent with apache/mxnet. The rcnn example in apache/mxnet was the same with release v5.1 and now the same with release v6. v5.1 is fast but super complex. v6 is simple but slower because it is all numpy. I can reproduce the released models with both v5.1 and v6.

In dev branch, files starting with sym uses symbolic interface while files starting with gluon or nd uses imperative/gluon interface. To use nd or gluon, gluoncv is required as additional dependency.

From your description, you have a model trained with v5.1 on your own dataset. If you do not wish to check how https://github.com/ijkguo/mx-rcnn/issues/105 happened, you can continue using v5.1.

ijkguo commented 6 years ago

Moved to https://github.com/ijkguo/mx-rcnn/issues/105