File size: 3,263 Bytes
59630ba | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 | # inherites from base_experiment.yaml
# most of the options have docs at https://lightning.ai/docs/pytorch/stable/common/trainer.html
defaults:
- base_exp
tasks: [training] # tasks to run sequantially, change when your project has multiple stages and you want to run only a subset of them.
num_nodes: 1 # number of gpu servers used in large scale distributed training
training:
precision: 16-mixed # set float precision, 16-mixed is faster while 32 is more stable
compile: False # whether to compile the model with torch.compile
lr: 0.001 # learning rate
batch_size: 16 # training batch size; effective batch size is this number * gpu * nodes iff using distributed training
max_epochs: 1000 # set to -1 to train forever
max_steps: -1 # set to -1 to train forever, will override max_epochs
max_time: null # set to something like "00:12:00:00" to enable
data:
num_workers: 16 # number of CPU threads for data preprocessing.
shuffle: True # whether training data will be shuffled
optim:
accumulate_grad_batches: 1 # accumulate gradients for n batches before backprop
gradient_clip_val: 0 # clip gradients with norm above this value, set to 0 to disable
checkpointing:
# these are arguments to pytorch lightning's callback, `ModelCheckpoint` class
every_n_train_steps: 5000 # save a checkpoint every n train steps
every_n_epochs: null # mutually exclusive with ``every_n_train_steps`` and ``train_time_interval``
train_time_interval: null # in format of "00:12:00:00", mutually exclusive with ``every_n_train_steps`` and ``every_n_epochs``.
enable_version_counter: False # If this is ``False``, later checkpoint will be overwrite previous ones.
save_top_k: -1
validation:
precision: 16-mixed
compile: False # whether to compile the model with torch.compile
batch_size: 16 # validation batch size per GPU; effective batch size is this number * gpu * nodes iff using distributed training
val_every_n_step: 2000 # controls how frequent do we run validation, can be float (fraction of epoches) or int (steps) or null (if val_every_n_epoch is set)
val_every_n_epoch: null # if you want to do validation every n epoches, requires val_every_n_step to be null.
num_sanity_val_steps: 1 # number of batches to run validation before training, if null, set to int(debug)
limit_batch: null # if null, run through validation set. Otherwise limit the number of batches to use for validation.
data:
num_workers: 16 # number of CPU threads for data preprocessing, for validation.
shuffle: False # whether validation data will be shuffled
inference_mode: True
test:
precision: 16-mixed
compile: False # whether to compile the model with torch.compile
batch_size: 16 # test batch size per GPU; effective batch size is this number * gpu * nodes iff using distributed training
limit_batch: null # if null, run through test set. Otherwise limit the number of batches to use for test.
data:
num_workers: 16 # number of CPU threads for data preprocessing, for test.
shuffle: False # whether test data will be shuffled
inference_mode: True
find_unused_parameters: False
reload_dataloaders_every_n_epochs: 0
ema:
enable: False
decay: 0.9999
validate_original_weights: False |