Up-Down-Captioner
Up-Down-Captioner copied to clipboard
Not enough memory (RAM) error
Following the author's instruction, if anyone use ./experiments/caption_lstm/train.sh
to train the network without enough RAM, it won't result in meaningful error message -- just produce some empty caffemodel.h5
files and says it cannot load the weights as something got wrong. If you take a look of RAM usage using software such as htop
you will find the memory is exploding.
To solve it, you need to refer to the second points of this segment: replace module: "rcnn_layers"
with module: "efficient_rcnn_layers"
.
I think it's because the Process
managed the memory usage and caught the errors so that it did not produce any useful error message.