AuroraLiang

to keep an eye on the GPU usage: `sudo watch mvidia-smi` you can twist the source code for expending gpu memory usage: try modify the parameters on the line below...

@LMdeLiangMi Try this. 3.95GB * 0.8 would provide 3.16 GB memory for allocation, which meets the minimum requirement of GPU free memory. ``` config = tf.ConfigProto(allow_soft_placement=True) config.gpu_options.allocator_type = 'BFC' config.gpu_options.per_process_gpu_memory_fraction...

@LMdeLiangMi to use batch_generator, first you need to convert your raw input into a specific file via TFRecordWriter. In this implement, the batch number is always to be 1, meaning...

@LMdeLiangMi Check TFFRCNN/lib/roi_data_layer/layer.py. It provides some insights of imdb's mechanism. Hope it could help. [BTW, google research has release its own object detection API : ](https://github.com/tensorflow/models/tree/master/object_detection) Try google's api if...

try clean up the roidb cache file in $TFFRCNN/data/cache. It works for me to solve the 'max_overlaps' problem.