Haonan Chen

The model you use (model_id1-501-1448236541.t7) is a GPU model, but you run it with CPU mode (`-gpuid -1`). You can use `convert_checkpoint_gpu_to_cpu.lua` to replace CudaTensor with normal Tensor first.

works pretty well. when do you think it will be merged? @karpathy

did you get exactly the same result compared with cloned version? I just have the same confusion...

oh, this may be the answer: you need to use `self.input` in backward process. ([line370](https://github.com/karpathy/neuraltalk2/blob/master/misc/LanguageModel.lua#L370)). Basically, this variable saves `self.state` in it ([line343](https://github.com/karpathy/neuraltalk2/blob/master/misc/LanguageModel.lua#L343)), and `self.state` is extracted from the output...

via PyPI

Good point. I opened this PR because when I was using `librosa.load`, I got "ValueError: frames must be specified for non-seekable files" from `soundfile` causing `librosa.load` failed. But audioread seems...

Thanks! What I mean by crash is that when I "stop search" after searching has started for some time, sublime is frozen and I have to do a force quit....

That's `335804` files

Can this be implemented as a relative path to the current file, or add a configuration file to choose between searching or not? I guess traversal of a large directory...

In my case the two sub policies learns to aim for different goal points, But the master policy leads the agent to a single goal.