Members:
- Mihir Rana
- Kenil Tanna
- Diogo Mesquita
- Yassine Kadiri
For ease of setup, we have created a requirements.yaml file which will create a conda environment with the name nlp_project and install all dependencies and requirements into that environment. To do this:
- Install Anaconda and run:
conda env create -f requirements.yaml
- Optionally, if you want to run it on a GPU, install CUDA and cuDNN
Again, for simplicity, we have created a module with the name seq2seq-nlp which can be installed directly into pip by running the following command from the main project directory:
pip install -e .
Note: Please uncomment these lines for getting results on the test set.
usage: main.py [-h] [--project-dir PROJECT_DIR] [--data-dir DATA_DIR]
[--source-dataset SOURCE_DATASET]
[--target-dataset TARGET_DATASET]
[--checkpoints-dir CHECKPOINTS_DIR]
[--load-ckpt LOAD_CHECKPOINT] [--load-enc-ckpt LOAD_ENC_CKPT]
[--load-dec-ckpt LOAD_DEC_CKPT] [--source-vocab SOURCE_VOCAB]
[--target-vocab TARGET_VOCAB] [--max-len-source MAX_LEN_SOURCE]
[--max-len-target MAX_LEN_TARGET]
[--unk-threshold UNK_THRESHOLD] [--batch-size BATCH_SIZE]
[--epochs EPOCHS] [--device DEVICE] [--device-id DEVICE_ID]
[--ngpu NGPU] [--parallel] [--lr LR]
[--encoder-type ENCODER_TYPE] [--decoder-type DECODER_TYPE]
[--num-directions NUM_DIRECTIONS]
[--encoder-num-layers ENCODER_NUM_LAYERS]
[--decoder-num-layers DECODER_NUM_LAYERS]
[--encoder-emb-size ENCODER_EMB_SIZE]
[--decoder-emb-size DECODER_EMB_SIZE]
[--encoder-hid-size ENCODER_HID_SIZE]
[--encoder-dropout ENCODER_DROPOUT]
[--decoder-dropout DECODER_DROPOUT] [--clip-param CLIP_PARAM]
[--beam-size BEAM_SIZE] [--force]
[--teacher-forcing-prob TEACHER_FORCING_PROB] [--use-attn]
optional arguments:
-h, --help show this help message and exit
--project-dir PROJECT_DIR path to project directory
--source-dataset SOURCE_DATASET name of source dataset file in data directory, default='vi'
--target-dataset TARGET_DATASET name of target dataset file in data directory, default='en'
--data-dir DATA_DIR path to data directory, default="data"
--checkpoints-dir CHECKPOINTS_DIR path to checkpoints directory
--load-ckpt LOAD_CHECKPOINT name of checkpoint file to load
--load-enc-ckpt LOAD_ENC_CKPT name of encoder file to load
--load-dec-ckpt LOAD_DEC_CKPT name of decoder file to load
--source-vocab SOURCE_VOCAB source dataset vocabulary size, default=50000
--target-vocab TARGET_VOCAB target dataset vocabulary size, default=50000
--max-len-source MAX_LEN_SOURCE max sentence length of source, default=100
--max-len-target MAX_LEN_TARGET max sentence length of target, default=100
--unk-threshold UNK_THRESHOLD count threshold below which words are to be treated as UNK
--batch-size BATCH_SIZE batch size, default=32
--epochs EPOCHS number of epochs, default=10
--device DEVICE cuda | cpu, default=cuda
device to train on
--device-id DEVICE_ID device id of gpu, default=0
--ngpu NGPU number of GPUs to use (0,1,...,ngpu-1)
--parallel use all GPUs available
--lr LR learning rate, default=1e-4
--encoder-type ENCODER_TYPE type of encoder model (gru | lstm), default="gru"
--decoder-type DECODER_TYPE type of decoder model (gru | lstm), default="gru"
--num-directions NUM_DIRECTIONS number of directions in encoder, default=2
--encoder-num-layers ENCODER_NUM_LAYERS number of layers in encoder, default=2
--decoder-num-layers DECODER_NUM_LAYERS number of layers in decoder, default=2
--encoder-emb-size ENCODER_EMB_SIZE embedding size of encoder, default=256
--decoder-emb-size DECODER_EMB_SIZE embedding size of decoder, default=256
--encoder-hid-size ENCODER_HID_SIZE hidden size of encoder, default=256
--encoder-dropout ENCODER_DROPOUT dropout rate in encoder, default=0.
--decoder-dropout DECODER_DROPOUT dropout rate in decoder FC layer, default=0.
--clip-param CLIP_PARAM clip parameter value for exploding gradients
--beam-size BEAM_SIZE Beam size to use during beam search
--teacher-forcing-prob TEACHER_FORCING_PROB clip parameter value for exploding gradients
--use-attn Use attention in the encoder decoder architecture
--force overwrites all existing dumped data sets