basecls.engine.hooks#

class basecls.engine.hooks.CheckpointHook(save_dir=None, save_every_n_epoch=1)[源代码]#

基类：BaseHook

Hook for managing checkpoints during training.

Effect during after_epoch and after_train procedure.

参数

save_dir (Optional[str]) – checkpoint directory.
save_every_n_epoch (int) – interval for saving checkpoint. Default: 1

after_epoch()[源代码]#: Called after each epoch.

after_train()[源代码]#: Called after training process.

class basecls.engine.hooks.EvalHook(save_dir=None, eval_every_n_epoch=1)[源代码]#

基类：BaseHook

Hook for evaluating during training.

Effect during after_epoch and after_train procedure.

参数

save_dir (Optional[str]) – checkpoint directory.
eval_every_n_epoch (int) – interval for evaluating. Default: 1

after_epoch()[源代码]#: Called after each epoch.

after_train()[源代码]#: Called after training process.

test(cfg, model, ema=None)[源代码]#

class basecls.engine.hooks.LoggerHook(log_every_n_iter=20)[源代码]#

基类：BaseHook

Hook for logging during training.

Effect during before_train, after_train, before_iter and after_iter procedure.

参数: log_every_n_iter (int) – interval for logging. Default: 20

before_train()[源代码]#: Called before training process.

after_train()[源代码]#: Called after training process.

before_iter()[源代码]#: Called before each iteration.

after_iter()[源代码]#: Called after each iteration.

get_loss_str(meter)[源代码]#: Get loss information during trainging process.

get_stat_str(meter)[源代码]#: Get stat information during trainging process.

get_memory_str(meter)[源代码]#: Get memory information during trainging process.

get_train_info_str()[源代码]#: Get training process related information such as learning rate.

get_time_str(left_iters)[源代码]#

Get time related information sucn as data_time, train_time, ETA and so on.

返回类型: str

class basecls.engine.hooks.LRSchedulerHook[源代码]#

基类：BaseHook

Hook for learning rate scheduling during training.

Effect during before_epoch procedure.

before_epoch()[源代码]#: Called before each epoch.

get_lr_factor(cfg, epoch_id)[源代码]#

Calculate learning rate factor.

It supports "step", "linear", "cosine", "exp", and "rel_exp" schedule.

参数

cfg (ConfigDict) – config for training.
epoch_id (int) – current epoch.

返回类型

float

返回

Learning rate factor.

total_lr#: Total learning rate.

class basecls.engine.hooks.PreciseBNHook(precise_every_n_epoch=1)[源代码]#

基类：BaseHook

Hook for precising BN during training.

Effect during after_epoch procedure.

参数: precise_every_n_epoch (int) – interval for precising BN. Default: 1

before_train()[源代码]#: Called before training process.

after_epoch()[源代码]#: Called after each epoch.

class basecls.engine.hooks.ResumeHook(save_dir=None, resume=False)[源代码]#

基类：BaseHook

Hook for resuming training process.

Effect during before_train procedure.

参数

save_dir (Optional[int]) – checkpoint directory.
resume (bool) – enable resume or not. Default: False

before_train()[源代码]#: Called before training process.

class basecls.engine.hooks.TensorboardHook(log_dir, log_every_n_iter=20, scalar_type='latest')[源代码]#

基类：BaseHook

Hook for tensorboard during training.

Effect during before_train, after_train and after_iter procedure.

参数

log_dir (str) – tensorboard directory.
log_every_n_iter (int) – interval for logging. Default: 20
scalar_type (str) – statistic to record, supports "latest", "avg", "global_avg" and "median". Default: "latest"

before_train()[源代码]#: Called before training process.

after_train()[源代码]#: Called after training process.

after_iter()[源代码]#: Called after each iteration.

write(context)[源代码]#

classmethod calc_iter(progress)[源代码]#