This repository is a custom modification of code originally released here:
https://github.com/karpathy/char-rnn
The modification provides improved support for generating Magic the Gathering cards, using the encoding techniques here:
https://github.com/billzorn/mtgencode
For the most part, this code behaves the same as the original char-rnn, but there are slight differences in the behavior of the batcher and the sampling process to take advantage of cool properties of the format. If you want to train neural networks to generate arbitrary text, you're probably better off looking at the original repository; if you only care about MTG, then you're in the right place.
For a detailed tutorial of how to set this code up and get it to work with the tools in mtgencode, refer to the mtgencode readme.
A custom sampling script developed by Talcos is included in sample_hs_v3.lua. It works exactly like the original sampling script, but it accepts additional options that allow you to prime the network with information as it's generating cards, a process we often call whispering. For example, you can specify that all card names should start with 'Garruk', and then whenever the network starts outputting something that appears to be a cardname, it will instead output 'Garruk' and then keep generating the card as if that's what it would have done normally.
This file is due to be updated at some point, and possible renamed or merged into the standard sampling script.
The primary difference between this code and char-rnn is that here we already know a great deal of information about the format. This allows us to do two clever things with the batcher.
First, we can avoid splitting all of the data up and then training on the same, arbitrarily segmented batches for multiple epochs. Instead, the batcher first interprets the data as whole cards, and partitions cards between the splits instead of raw data chunks. To achieve the illusion of epochs, we replicate the training set the desired number of times, but write it out as a coherent sequence of randomly ordered cards before splitting. This way less information is lost when cards are chopped up between different batches, as the chopping will occur in different places for eacy copy. That's the hope, anyway.
The batcher can also dynamically randomize the symbols in mana costs of cards, and the order of the fields in a card if the field's identity is specified by label rather than by order.
NOTE: THE FOLLOWING BATCHER FEATURES ARE STILL BEING IMPLEMENTED AND AREN'T AVAILABLE YET
The other clever feature is the ability to train on dynamically sized batches of exactly one card at a time. This does exactly what you'd expect; unfortunately it's very slow.
A modified LSTM model is available, with modifications that add a learnable bias as suggested by this paper. To use it, specify the option "-model lstmb" when training.
This code implements multi-layer Recurrent Neural Network (RNN, LSTM, and GRU) for training/sampling from character-level language models. In other words the model takes one text file as input and trains a Recurrent Neural Network that learns to predict the next character in a sequence. The RNN can then be used to genrate text character by character that will look like the original training data. The context of this code base is described in detail in my blog post.
If you are new to Torch/Lua/Neural Nets, it might be helpful to know that this code is really just a slightly more fancy version of this 100-line gist that I wrote in Python/numpy. The code in this repo additionally: allows for multiple layers, uses an LSTM instead of a vanilla RNN, has more supporting code for model checkpointing, and is of course much more efficient since it uses mini-batches and can run on a GPU.
This code is written in Lua and requires Torch. If you're on Ubuntu, installing Torch in your home directory may look something like:
$ curl -s https://raw.githubusercontent.com/torch/ezinstall/master/install-deps | bash
$ git clone https://github.com/torch/distro.git ~/torch --recursive
$ cd ~/torch;
$ ./install.sh # and enter "yes" at the end to modify your bashrc
$ source ~/.bashrcSee the Torch installation documentation for more details. After Torch is installed we need to get a few more packages using LuaRocks (which already came with the Torch install). In particular:
$ luarocks install nngraph
$ luarocks install optim
$ luarocks install nnIf you'd like to train on an NVIDIA GPU using CUDA (this can be to about 15x faster), you'll of course need the GPU, and you will have to install the CUDA Toolkit. Then get the cutorch and cunn packages:
$ luarocks install cutorch
$ luarocks install cunnIf you'd like to use OpenCL GPU instead (e.g. ATI cards), you will instead need to install the cltorch and clnn packages, and then use the option -opencl 1 during training (cltorch issues):
$ luarocks install cltorch
$ luarocks install clnnAll input data is stored inside the data/ directory. You'll notice that there is an example dataset included in the repo (in folder data/mtgencode-std) which is the standard encoded corpus of existing Magic the Gathering cards as produced by mtgencode.
Your own data: If you'd like to use your own data then create a single file input.txt and place it into a folder in the data/ directory. For example, data/some_folder/input.txt. The first time you run the training script it will do some preprocessing and write two more convenience cache files into data/some_folder.
Dataset sizes: Note that if your data is too small (1MB is already considered very small) the RNN won't learn very effectively. Remember that it has to learn everything completely from scratch. Conversely if your data is large (more than about 2MB), feel confident to increase rnn_size and train a bigger model (see details of training below). It will work significantly better. For example with 6MB you can easily go up to rnn_size 300 or even more. The biggest that fits on my GPU and that I've trained with this code is rnn_size 700 with num_layers 3 (2 is default).
Start training the model using train.lua. As a sanity check, to run on the included example dataset simply try:
$ th train.lua -gpuid -1
Notice that here we are setting the flag gpuid to -1, which tells the code to train using CPU, otherwise it defaults to GPU 0. There are many other flags for various options. Consult $ th train.lua -help for comprehensive settings. Here's another example that trains a bigger network and also shows how you can run on your own custom dataset (this already assumes that data/some_folder/input.txt exists):
$ th train.lua -data_dir data/some_folder -rnn_size 512 -num_layers 2 -dropout 0.5
Checkpoints. While the model is training it will periodically write checkpoint files to the cv folder. The frequency with which these checkpoints are written is controlled with number of iterations, as specified with the eval_val_every option (e.g. if this is 1 then a checkpoint is written every iteration). The filename of these checkpoints contains a very important number: the loss. For example, a checkpoint with filename lm_lstm_epoch0.95_2.0681.t7 indicates that at this point the model was on epoch 0.95 (i.e. it has almost done one full pass over the training data), and the loss on validation data was 2.0681. This number is very important because the lower it is, the better the checkpoint works. Once you start to generate data (discussed below), you will want to use the model checkpoint that reports the lowest validation loss. Notice that this might not necessarily be the last checkpoint at the end of training (due to possible overfitting).
Another important quantities to be aware of are batch_size (call it B), seq_length (call it S), and the train_frac and val_frac settings. The batch size specifies how many streams of data are processed in parallel at one time. The sequence length specifies the length of each stream, which is also the limit at which the gradients can propagate backwards in time. For example, if seq_length is 20, then the gradient signal will never backpropagate more than 20 time steps, and the model might not find dependencies longer than this length in number of characters. Thus, if you have a very difficult dataset where there are a lot of long-term dependencies you will want to increase this setting. Now, if at runtime your input text file has N characters, these first all get split into chunks of size BxS. These chunks then get allocated across three splits: train/val/test according to the frac settings. By default train_frac is 0.95 and val_frac is 0.05, which means that 95% of our data chunks will be trained on and 5% of the chunks will be used to estimate the validation loss (and hence the generalization). If your data is small, it's possible that with the default settings you'll only have very few chunks in total (for example 100). This is bad: In these cases you may want to decrease batch size or sequence length.
Note that you can also initialize parameters from a previously saved checkpoint using init_from.
Given a checkpoint file (such as those written to cv) we can generate new text. For example:
$ th sample.lua cv/some_checkpoint.t7 -gpuid -1
Make sure that if your checkpoint was trained with GPU it is also sampled from with GPU, or vice versa. Otherwise the code will (currently) complain. As with the train script, see $ th sample.lua -help for full options. One important one is (for example) -length 10000 which would generate 10,000 characters (default = 2000).
Temperature. An important parameter you may want to play with is -temperature, which takes a number in range (0, 1] (0 not included), default = 1. The temperature is dividing the predicted log probabilities before the Softmax, so lower temperature will cause the model to make more likely, but also more boring and conservative predictions. Higher temperatures cause the model to take more chances and increase diversity of results, but at a cost of more mistakes.
Priming. It's also possible to prime the model with some starting text using -primetext. This starts out the RNN with some hardcoded characters to warm it up with some context before it starts generating text. E.g. a fun primetext might be -primetext "the meaning of life is ".
Training with GPU but sampling on CPU. Right now the solution is to use the convert_gpu_cpu_checkpoint.lua script to convert your GPU checkpoint to a CPU checkpoint. In near future you will not have to do this explicitly. E.g.:
$ th convert_gpu_cpu_checkpoint.lua cv/lm_lstm_epoch30.00_1.3950.t7
will create a new file cv/lm_lstm_epoch30.00_1.3950.t7_cpu.t7 that you can use with the sample script and with -gpuid -1 for CPU mode.
Happy sampling!
If you're somewhat new to Machine Learning or Neural Networks it can take a bit of expertise to get good models. The most important quantity to keep track of is the difference between your training loss (printed during training) and the validation loss (printed once in a while when the RNN is run on the validation data (by default every 1000 iterations)). In particular:
- If your training loss is much lower than validation loss then this means the network might be overfitting. Solutions to this are to decrease your network size, or to increase dropout. For example you could try dropout of 0.5 and so on.
- If your training/validation loss are about equal then your model is underfitting. Increase the size of your model (either number of layers or the raw number of neurons per layer)
The two most important parameters that control the model are rnn_size and num_layers. I would advise that you always use num_layers of either 2/3. The rnn_size can be adjusted based on how much data you have. The two important quantities to keep track of here are:
- The number of parameters in your model. This is printed when you start training.
- The size of your dataset. 1MB file is approximately 1 million characters.
These two should be about the same order of magnitude. It's a little tricky to tell. Here are some examples:
- I have a 100MB dataset and I'm using the default parameter settings (which currently print 150K parameters). My data size is significantly larger (100 mil >> 0.15 mil), so I expect to heavily underfit. I am thinking I can comfortably afford to make
rnn_sizelarger. - I have a 10MB dataset and running a 10 million parameter model. I'm slightly nervous and I'm carefully monitoring my validation loss. If it's larger than my training loss then I may want to try to increase dropout a bit and see if that heps the validation loss.
The winning strategy to obtaining very good models (if you have the compute time) is to always err on making the network larger (as large as you're willing to wait for it to compute) and then try different dropout values (between 0,1). Whatever model has the best validation performance (the loss, written in the checkpoint filename, low is good) is the one you should use in the end.
It is very common in deep learning to run many different models with many different hyperparameter settings, and in the end take whatever checkpoint gave the best validation performance.
By the way, the size of your training and validation splits are also parameters. Make sure you have a decent amount of data in your validation set or otherwise the validation performance will be noisy and not very informative.
This code was originally based on Oxford University Machine Learning class practical 6, which is in turn based on learning to execute code from Wojciech Zaremba. Chunks of it were also developed in collaboration with my labmate Justin Johnson.
To learn more about RNN language models I recommend looking at:
- My recent talk on char-rnn
- Generating Sequences With Recurrent Neural Networks by Alex Graves
- Generating Text with Recurrent Neural Networks by Ilya Sutskever
- Tomas Mikolov's Thesis
MIT