| date | section | title |
|---|---|---|
2021年 九月 1日 星期三 |
3 |
galileo.framework.pytorch.python.trainer.Trainer |
galileo.framework.pytorch.python.trainer.Trainer - pytorch trainer
\
继承自 galileo.framework.python.base_trainer.BaseTrainer .
def __init__ (self, model, inputs=None, module=None, multiprocessing_distributed=False, num_procs_per_shard=None, rank=0, world_size=1, dist_url=None, dist_backend=None, seed=None, zk_server=DefaultValues.ZK_SERVER, zk_path=DefaultValues.ZK_PATH)\
def get_optimizer (self)
return an optimizer
def train (self, **kwargs)\
def evaluate (self, **kwargs)\
def predict (self, **kwargs)\
def run (self, **kwargs)
run train, evaluate, predict
def run_worker (self, local_rank)\
def create_client (self)\
def get_dataset (self, mode)
get an dataset
def do_train (self)\
def do_evaluate (self)\
def do_predict (self)\
module\
run_config\
model\
pytorch trainer
internal attrs:
use_cuda global_rank local_rank is_master
注意
API: galileo.pytoch.Trainer
def galileo.framework.pytorch.python.trainer.Trainer.__init__ ( self, model, inputs = None, module = None, multiprocessing_distributed = False, num_procs_per_shard = None, rank = 0, world_size = 1, dist_url = None, dist_backend = None, seed = None, zk_server = DefaultValues.ZK_SERVER, zk_path = DefaultValues.ZK_PATH)
参数
model instance of torch.nn.Module
inputs Inputs for model
module Module for trainer, use default
multiprocessing_distributed multi-processing distributed training
num_procs_per_shard use default
rank read from env RANK, use default
world_size read from env WORLD_SIZE, use default
dist_url use default
dist_backend use default
seed seed for initializing training
zk_server zookeeper server address
zk_path zookeeper registration node name
参数
model_dir model dir
inputs_fn inputs function, requried when self.inputs is None
log_steps Number of steps to print log
log_max_times_per_epoch log max times per epoch, default is 100
start_epoch start of epoch
num_epochs number epochs
optimizer adam, sgd, momentum
learning_rate learning rate
momentum momentum for optimizer
save_checkpoint_epochs The frequency to save checkpoint per epoch
gpu_status bool show gpu status
save_predict_fn callback for save results of predict save_predict_fn(ids, embeddings, dir, rank)
save_best_model bool, save the best model
spacial params for pytorch
参数
weight_decay weight_decay for optimizer
resume file to checkpoint
hooks hooks for log metrics and so on
params for inputs_fn
参数
batch_size Mini-batch size
max_id max vertex id
batch_num Number of mini-batch, default is [max_id] / [batch_size]
重载 galileo.framework.python.base_trainer.BaseTrainer .
get an dataset
参数
mode train/evaluate/predict
重载 galileo.framework.python.base_trainer.BaseTrainer .
参数
model_dir model dir
inputs_fn inputs function, requried when self.inputs is None
log_steps Number of steps to print log
log_max_times_per_epoch log max times per epoch, default is 100
start_epoch start of epoch
num_epochs number epochs
optimizer adam, sgd, momentum
learning_rate learning rate
momentum momentum for optimizer
save_checkpoint_epochs The frequency to save checkpoint per epoch
gpu_status bool show gpu status
save_predict_fn callback for save results of predict save_predict_fn(ids, embeddings, dir, rank)
save_best_model bool, save the best model
spacial params for pytorch
参数
weight_decay weight_decay for optimizer
resume file to checkpoint
hooks hooks for log metrics and so on
params for inputs_fn
参数
batch_size Mini-batch size
max_id max vertex id
batch_num Number of mini-batch, default is [max_id] / [batch_size]
重载 galileo.framework.python.base_trainer.BaseTrainer .
run train, evaluate, predict
参数
mode str, train, evaluate, predict
model_dir model dir
inputs_fn inputs function, requried when self.inputs is None
log_steps Number of steps to print log
log_max_times_per_epoch log max times per epoch, default is 100
start_epoch start of epoch
num_epochs number epochs
optimizer adam, sgd, momentum
learning_rate learning rate
momentum momentum for optimizer
save_checkpoint_epochs The frequency to save checkpoint per epoch
gpu_status bool show gpu status
save_predict_fn callback for save results of predict save_predict_fn(ids, embeddings, dir, rank)
save_best_model bool, save the best model
spacial params for pytorch
参数
weight_decay weight_decay for optimizer
resume file to checkpoint
hooks hooks for log metrics and so on
params for inputs_fn
参数
batch_size Mini-batch size
max_id max vertex id
batch_num Number of mini-batch, default is [max_id] / [batch_size]
参数
model_dir model dir
inputs_fn inputs function, requried when self.inputs is None
log_steps Number of steps to print log
log_max_times_per_epoch log max times per epoch, default is 100
start_epoch start of epoch
num_epochs number epochs
optimizer adam, sgd, momentum
learning_rate learning rate
momentum momentum for optimizer
save_checkpoint_epochs The frequency to save checkpoint per epoch
gpu_status bool show gpu status
save_predict_fn callback for save results of predict save_predict_fn(ids, embeddings, dir, rank)
save_best_model bool, save the best model
spacial params for pytorch
参数
weight_decay weight_decay for optimizer
resume file to checkpoint
hooks hooks for log metrics and so on
params for inputs_fn
参数
batch_size Mini-batch size
max_id max vertex id
batch_num Number of mini-batch, default is [max_id] / [batch_size]
重载 galileo.framework.python.base_trainer.BaseTrainer .
由 Doyxgen 通过分析 Galileo 的 源代码自动生成.