Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/harvardnlp/TextFlow
https://github.com/harvardnlp/TextFlow
Last synced: 4 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/harvardnlp/TextFlow
- Owner: harvardnlp
- License: mit
- Created: 2019-02-13T16:54:24.000Z (over 5 years ago)
- Default Branch: master
- Last Pushed: 2019-10-31T18:52:34.000Z (over 4 years ago)
- Last Synced: 2024-01-16T10:46:23.750Z (5 months ago)
- Language: Python
- Size: 1.87 MB
- Stars: 114
- Watchers: 11
- Forks: 15
- Open Issues: 4
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Lists
- awesome-stars - harvardnlp/TextFlow - (Python)
README
# Latent Normalizing Flows for Discrete Sequences
This code provides an implementation for and reproduces results from [Latent Normalizing Flows for Discrete Sequences](https://arxiv.org/abs/1901.10548)
Zachary Ziegler, Alexander Rush
ICML 2019
## Dependencies
The code was tested with `python 3.6`, `pytorch 0.4.1`, `torchtext 0.2.3`, and `CUDA 9.2`.
## Character-level language modeling:
PTB data with [Mikolov preprocessing](http://www.fit.vutbr.cz/~imikolov/rnnlm/char.pdf) is checked in.
Baseline and proposed models can be trained with:
```
python main.py --dataset ptb --run_name charptb_baselinelstm --model_type baseline --dropout_p 0.1 --optim sgd --lr 20
python main.py --dataset ptb --run_name charptb_discreteflow_af-af
python main.py --dataset ptb --run_name charptb_discreteflow_af-scf --hiddenflow_scf_layers
python main.py --dataset ptb --run_name charptb_discreteflow_iaf-scf --dropout_p 0 --hiddenflow_flow_layers 3 --hiddenflow_scf_layers --prior_type IAF
```
Evaluation on the test set is run with e.g.
```
python main.py --dataset ptb --run_name charptb_baselinelstm --model_type baseline --dropout_p 0.1 --optim sgd --lr 20 --load_dir output/charptb_baselinelstm/saves/ --evaluate
```
## Polyphonic music:
Commands to train the baseline and proposed models are
Nottingham:
```
python main.py --indep_bernoulli --dataset nottingham --run_name nottingham_baselinelstm --model_type baseline --dropout_p 0.1 --optim sgd --lr 20 --patience 4
python main.py --indep_bernoulli --dataset nottingham --run_name nottingham_discreteflow_af-af --patience 2 --ELBO_samples 1 --B_train 5 --B_val 1 --initial_kl_zero 20 --kl_rampup_time 15
python main.py --indep_bernoulli --dataset nottingham --run_name nottingham_discreteflow_af-scf --patience 2 --ELBO_samples 1 --B_train 5 --B_val 1 --initial_kl_zero 20 --kl_rampup_time 15 --hiddenflow_scf_layers
python main.py --indep_bernoulli --dataset nottingham --run_name nottingham_discreteflow_iaf-scf --patience 2 --ELBO_samples 1 --B_train 5 --B_val 1 --initial_kl_zero 20 --kl_rampup_time 15 --hiddenflow_scf_layers --prior_type IAF
```
Piano_midi
```
python main.py --indep_bernoulli --dataset piano_midi --run_name piano_midi_baselinelstm --model_type baseline --dropout_p 0.1 --optim sgd --lr 20 --patience 4
python main.py --indep_bernoulli --dataset piano_midi --run_name piano_midi_discreteflow_af-af --patience 2 --ELBO_samples 1 --B_train 1 --B_val 1 --initial_kl_zero 20 --kl_rampup_time 15 --nll_samples 4 --grad_accum 8
python main.py --indep_bernoulli --dataset piano_midi --run_name piano_midi_discreteflow_af-scf --patience 2 --ELBO_samples 1 --B_train 1 --B_val 1 --initial_kl_zero 20 --kl_rampup_time 15 --nll_samples 4 --grad_accum 8 --hiddenflow_scf_layers
python main.py --indep_bernoulli --dataset piano_midi --run_name piano_midi_discreteflow_iaf-scf --patience 2 --ELBO_samples 1 --B_train 1 --B_val 1 --initial_kl_zero 20 --kl_rampup_time 15 --nll_samples 4 --grad_accum 8 --hiddenflow_scf_layers --prior_type IAF
```
Musedata
```
python main.py --indep_bernoulli --dataset muse_data --run_name muse_data_baselinelstm --model_type baseline --dropout_p 0.1 --optim sgd --lr 20 --patience 4
python main.py --indep_bernoulli --dataset muse_data --run_name muse_data_discreteflow_af-af --patience 2 --ELBO_samples 1 --B_train 1 --B_val 1 --initial_kl_zero 20 --kl_rampup_time 15 --nll_samples 4 --grad_accum 8
python main.py --indep_bernoulli --dataset muse_data --run_name muse_data_discreteflow_af-scf --patience 2 --ELBO_samples 1 --B_train 1 --B_val 1 --initial_kl_zero 20 --kl_rampup_time 15 --nll_samples 4 --grad_accum 8 --hiddenflow_scf_layers
python main.py --indep_bernoulli --dataset muse_data --run_name muse_data_discreteflow_iaf-scf --patience 2 --ELBO_samples 1 --B_train 1 --B_val 1 --initial_kl_zero 20 --kl_rampup_time 15 --nll_samples 4 --grad_accum 8 --hiddenflow_scf_layers --prior_type IAF
```
JSB_chorales
```
python main.py --indep_bernoulli --dataset jsb_chorales --run_name jsb_chorales_baselinelstm --model_type baseline --dropout_p 0.1 --optim sgd --lr 20 --patience 4
python main.py --indep_bernoulli --dataset jsb_chorales --run_name jsb_chorales_discreteflow_af-af --patience 2 --ELBO_samples 1 --B_train 5 --B_val 64 --initial_kl_zero 20 --kl_rampup_time 15
python main.py --indep_bernoulli --dataset jsb_chorales --run_name jsb_chorales_discreteflow_af-scf --patience 2 --ELBO_samples 1 --B_train 5 --B_val 64 --initial_kl_zero 20 --kl_rampup_time 15 --hiddenflow_scf_layers
python main.py --indep_bernoulli --dataset jsb_chorales --run_name jsb_chorales_discreteflow_iaf-scf --patience 2 --ELBO_samples 1 --B_train 5 --B_val 64 --initial_kl_zero 20 --kl_rampup_time 15 --hiddenflow_scf_layers --prior_type IAF
```