char-rnn.pytorch
char-rnn.pytorch copied to clipboard
spro
PyTorch implementation of char-rnn (character-level language model)
- Its possible to add validation data and monitor overfitting by means of early stopping, as well as saving the model training at best checkpoints. - Its possible to launch...
I'm impressed with how you can get reasonable results with this after just a few minutes. However I tried running it overnight with 3 layers and 20'000 iterations and while...
Command ``` python train.py runoja.txt --n_epochs 5000 --model lstm --cuda ``` gives ``` Traceback (most recent call last): File "train_.py", line 95, in loss = train(*random_training_set(args.chunk_len, args.batch_size)) File "train_.py", line...
1. Fix out-of-array bug in training. 2. Allow to generate on a different architecture. 3. Fix lstm.
What's the difference between applying the loss function after each cell of the RNN vs applying on the entire sequence? In your example you feed only 1 character at a...
While training at a seemingly random point, it fails with this error (both lstm and gru): ``` Traceback (most recent call last): File "train.py", line 98, in loss = train(*random_training_set(args.chunk_len,...
The `--n_epochs` argument is in fact an iteration, which updates the weight based on a random sample of data points. There is no conclusive definition, but the community generally define...
This modifies the `read_file` method to extract the character vocabulary from the input file, which is then passed as an argument to the `generate` and `char_tensor` methods. This also modifies...
1. Fix index out of range error. 2. About 8x (for me) training speed-up by feeding whole sample sequence through CuDNN, not char-by-char. 3. Inference: GPU memory requirement not growing...
Right now it seems like generate.py using a lot of cuda memory during inference. For example. I trained a small 2 layer 150 hidden sized GRU network on the Shakespeare...
