train
Train a PiNN model given a parameter file.
Usage
pinn train params [options]
Options
| Option [shorthand] | Default | Description |
|---|---|---|
--model-dir [-d] |
None |
model directory (default to 'model' if not set in params) |
--train-ds [-t] |
'train.yml' |
training set (batched PiNN dataset) |
--eval-ds [-e] |
'eval.yml' |
evaluation set (batched PiNN dataset) |
--batch [-b] |
None |
batch size (assume batched dataset by default) |
--(no-)cache |
True |
cache dataset to memory |
--(no-)preprocess |
True |
preprocess the data |
--scatch-dir |
None |
if set, cache the data there instead of RAM |
--train-steps |
1e6 |
max training steps |
--eval-steps |
None |
evaluation steps (defaults to the whole eval set) |
--shuffle-buffer |
100 |
size of shuffle buffer |
--log-every |
1000 |
log every x steps |
--ckpt-every |
10000 |
save checkpoint every x steps |
--max-ckpts |
1 |
max number of checkpoints to save |
--(no-)init |
False |
initialize the params training set |