Torch Starter

This is a simple Torch7 starter package. It can be used as a simplified kickoff point for a Torch project.

I pieced together this package largely from Torch7 resources online. I mostly just copied the code, and stripped a lot of extra functionality out, to make it easier to hack on.

If something is not clear, or could be made more simple, please let me know. The goal is to be simple.

Installation

If you are at CSAIL, you can use my Torch installation:

. /data/vision/torralba/commonsense/torch/install/bin/torch-activate
export LD_LIBRARY_PATH=/data/vision/torralba/commonsense/cudnnv5/cuda/lib64:$LD_LIBRARY_PATH

Otherwise, installation is fairly simple. You need to install:

• Torch7
• cunn for training on GPU
• cudnn for faster training on GPU
• tds for some data structures
• display for graphs

You can install all of these with the commands:

# install torch first
git clone https://github.com/torch/distro.git ~/torch --recursive
cd ~/torch; bash install-deps;
./install.sh

# install libraries
luarocks install cunn
luarocks install cudnn
luarocks install tds
luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec

Learning Resources

• Torch Cheat Sheet
• 60 minute blitz

Model

I trained an AlexNet-esque network on Places365 with this code, which you can download here. This model obtains 50% top-1 accuracy on the validation set. This is slightly worse than the published result because we didn't do averaging over 10 crops.

If you use this model, please cite the Places2 paper (of which I am not affiliated). Note this model is slightly different from the AlexNet in Caffe. Notable differences: no groups in the convolutions, no spatial normalization, batch normalizaiton, trained with Adam instead of SGD, and sampling with replacement. It is unclear to me whether these changes have a significant impact on performance.

Data Setup

By default, we assume you have a text file that lists your dataset. This text does not store your dataset; it just lists filepaths to it, and any meta data. Each line in this text file represents one training example, and its associated category ID. The syntax of the line should be:

<filename><tab><number>

For example:

bedroom/IMG_050283.jpg    5
bedroom/IMG_237761.jpg    5
office/IMG_838222.jpg     10

The <number> should start counting at 1.

After you create this file, open main.lua and change data_list to point to this file. You can specify a data_root too, which will be prepended to each filename.

Training

Define your model in the net variable. By default, it is AlexNet. To learn more about the modules you can use, see nn. You can also adjust your loss with the criterion variable.

Remember to also adjust any options in opt, such as the learning rate and the number of classes. Setting these hyperparameters is a bit of an art, but generally it is recommended to use a learning rate of 0.001 and batch size of at least 64, but 128 or 256 may be better if you have the memory. For a systematic study, see this paper.

Finally, to start training, just do:

$ CUDA_VISIBLE_DEVICES=0 th main.lua

where you replace the number after CUDA_VISIBLE_DEVICES with the GPU you want to run on. You can find which GPU to use with $ nvidia-smi on our GPU cluster. Note: this number is 0-indexed, unlike the rest of Torch!

During training, it will dump snapshots to the checkpoints/ directory every epoch. Each time you start a new experiment, you should change the name (in opt), to avoid overwriting previous experiments. The code will not warn you about this (to keep things simple).

Evaluation

To evaluate your model, you can use the eval.lua script. It mostly follows the same format as main.lua. It reads your validation/testing dataset from a file similar to before, and sequentially runs through it, calculating both the top-1 and top-5 accuracy.

Graphics, Logs

If you want to see graphics and the loss over time, in a different shell on the same machine, run this command:

$ th -ldisplay.start 8000 0.0.0.0

then navigate to http://HOST.csail.mit.edu:8000 in your browser. Every 10th iteration it will push graphs.

On the CSAIL vision cluster, you can run this code out-of-the-box, and it will start to train AlexNet on the Places2 database.