quantdata.dysanos.grid.torch_train#

class quantdata.dysanos.grid.torch_train.TrainConfig(*, num_epochs=5000, batch_size=None, val_batch_size=None, pcnt_validation=0.0, lr=0.001, final_lr=1e-05, weight_decay=0.0001, compile=False, clip_grad_norm=None, clip_grad=None, epoch_interval=200, time_interval=120)[source]#

Bases: object

num_epochs: int = 5000#

total epochs

batch_size: int | None = None#

batch size, or None for all samples

val_batch_size: int | None = None#

batch size for validation, or None for all samples

pcnt_validation: float = 0.0#

fraction of samples to use for validation. Will be automatically adjusted such that there are always at least 2 training samples.

lr: float = 0.001#

initial learning rate

final_lr: float = 1e-05#

final learning rate

weight_decay: float = 0.0001#

weight decay for regularization

compile: bool = False#

compile torch

clip_grad_norm: float | None = None#

max norm for gradient clipping

clip_grad: float | None = None#

max for gradient clipping

epoch_interval: int = 200#

after how many epochs to update visualization and store checkpoint

time_interval: float = 120#

after how many seconds to update visualization and store checkpoint

lr_for_epoch(epoch)[source]#
set_lr_for_epoch(epoch, optimizer)[source]#
get_trn_vld_samples(num_total_samples)[source]#
class quantdata.dysanos.grid.torch_train.MetricTracker(trn_epochs=<factory>, trn_losses=<factory>, lr=<factory>, val_epochs=<factory>, val_losses=<factory>, vld_epochs=<factory>, vld_losses=<factory>)[source]#

Bases: object

quantdata.dysanos.grid.torch_train.train(market, *, mdl_config, trn_config=TrainConfig(num_epochs=5000, batch_size=None, val_batch_size=None, pcnt_validation=0.0, lr=0.001, final_lr=1e-05, weight_decay=0.0001, compile=False, clip_grad_norm=None, clip_grad=None, epoch_interval=200, time_interval=120), viz_config=VizConfig(num_viz_dates=4, seed=1231231, use_nms=False), restart=False, just_load=False, env=<quantdata.torch.BaseEnvironment object>, verbose=<cdxcore.verbose.Context object>)[source]#

Train model

quantdata.dysanos.grid.torch_train.load_model_market(ffn, env=<quantdata.torch.BaseEnvironment object>, *, pool_config=<cdxcore.jcpool.JCPoolConfig object>, verbose=<cdxcore.verbose.Context object>)[source]#

Load model and market from a file ffn.