search

jessemelpolio / Faster_RCNN_for_DOTA

Code used for training Faster R-CNN on DOTA
https://arxiv.org/abs/1711.10398
Apache License 2.0
336 stars 102 forks source link

Trying to reproduce pre-trained model of author + problem in training #17

Open swz30 opened 6 years ago

swz30 commented 6 years ago

Hello,

By using the provided code, I am trying to reproduce the model of Faster-RCNN for DOTA. So that I can train it on my data. After training of 60 epochs, the mAP on validation set of my trained model should be similar to the provided pre-trained model, but they are not (mine is much worse). I tried to preprocess the data using the get_best_begin_point function from DOTA Devkit as the authors mentioned in the disclaimer but it did not help either. Does anyone know what could be the problem, what I might be doing wrong?

During training, between epoch 23 to epoch 43, the RCNNLogLoss decreases from 0.18 to 0.17, is it normal? (The authors are decreasing learning rate as the training progresses.)

MaxBazik commented 6 years ago

I also can't seem to reproduce the pre-trained model, even for the axis aligned version.

I am using author's config-file, data, and scripts but the network I get is not of comparable fidelity.

Does anyone know what might be causing this or what hyperparameters (LR etc.) were used to get the author's network?

dingjiansw101 commented 6 years ago

@swz30 @MaxBazik @swz30 How many gpus you used? The learning rate need to be set according to the number of gpus. Our default setting is for 4 gpus. See this https://github.com/msracver/Deformable-ConvNets/issues/146.