def _build_multigpu_train_op(self, num_gpus):
with self.graph.as_default(), tf.device('/gpu:0'):
tower_grads = []
self.inputs = []
self.labels = []
opt = Optimizer()()
for i in range(num_gpus):
with tf.device('/gpu:%d' % i):
with tf.name_scope('tower_%d' % i) as scope:
# tf.get_variable_scope().reuse_variables()
inputs = self._create_tensor(self.loss_input['inputs'])
labels = self._create_tensor(self.loss_input['labels'])
input = inputs[0]
model_output = self.model(input)
loss = Loss()()
loss = loss(model_output, labels)
# Calculate the gradients for the batch of data on this tower.
varlist = [x for x in tf.trainable_variables() if x.name.startswith('tower_%d' % i)]
grads = opt.compute_gradients(loss, varlist)
tower_grads.append(grads)
if i == 0:
self.actor_var = TFVariables(model_output, self.sess)
self.input = input
self.logits = model_output
self.loss = loss
self.inputs.append(inputs)
self.labels.append(labels)
grads = self._average_gradients(tower_grads)
self.train_op = opt.apply_gradients(grads)
self.sess.run(tf.initialize_all_variables())
class TrainerTf(TrainerBase):
"""Trainer tensorflow class."""
def build(self):
"""Build the trainer by assembling the necessary components."""
super().build()
# Some trainer has different train batch size from valid batch
self.train_metrics = None
self.valid_metrics = self._init_metrics()
self._init_horovod_setting()
def train(self, inputs, labels):
"""Train model."""
feed_dict = {}
with self.graph.as_default():
if self.gpu_nums >= 1:
input_split = [[]] * self.gpu_nums
shape_split = inputs[0].shape[0] // self.gpu_nums
for j in range(self.gpu_nums):
for i in range(len(inputs)):
input_split = inputs[i][j * shape_split:(j + 1) *
shape_split]
feed_dict.update({self.inputs[j][i]: input_split})
for j in range(self.gpu_nums):
for i in range(len(labels)):
input_split = labels[i][j * shape_split:(j + 1) *
shape_split]
feed_dict.update({self.labels[j][i]: input_split})
else:
for i in range(len(inputs)):
feed_dict.update({self.inputs[i]: inputs[i]})
for i in range(len(labels)):
feed_dict.update({self.labels[i]: labels[i]})
_, loss = self.sess.run([self.train_op, self.loss], feed_dict)
return loss
def predict(self, input):
"""Inference model."""
with self.graph.as_default():
feed_dict = {self.input: input}
out = self.sess.run(self.logits, feed_dict)
return out
def save(self, file_name):
"""Save model."""
with self.graph.as_default():
self.actor_var.save_weights(file_name + ".npz")
return file_name + ".npz"
def load(self, model_name, by_name):
"""Load model."""
with self.graph.as_default():
self.actor_var.set_weights_with_npz(model_name)
def set_weights(self, weights):
"""Set weight with memory tensor."""
with self.graph.as_default():
self.actor_var.set_weights(weights)
def get_weights(self):
"""Get the weights."""
with self.graph.as_default():
return self.actor_var.get_weights()
def init_train_op(self):
"""Init Train Op."""
with self.graph.as_default():
self._init_train_op()
def _set_default_funcs(self):
self.model_fn = self._default_model_fn
self.train_input_fn = self._default_train_input_fn
self.valid_input_fn = self._default_valid_input_fn
def _set_condition(self):
self._init_tf_session()
self._init_distributed_setting()
self._init_tf_estimator()
def _train_epoch(self):
self.estimator.train(input_fn=self.train_input_fn,
steps=len(self.train_loader),
hooks=self._init_logging_hook())
def _valid_epoch(self):
self.callbacks.before_valid()
valid_logs = None
eval_metrics = self.estimator.evaluate(input_fn=self.valid_input_fn,
steps=len(self.valid_loader))
self.valid_metrics.update(eval_metrics)
valid_logs = dict()
valid_logs['cur_valid_perfs'] = self.valid_metrics.results
self.callbacks.after_valid(valid_logs)
def _init_distributed_setting(self):
if not self.distributed:
return
if zeus.is_npu_device():
from npu_bridge.estimator import npu_ops
self.npu_init = npu_ops.initialize_system()
self.npu_shutdown = npu_ops.shutdown_system()
self.sess.run(self.npu_init)
import horovod.tensorflow as hvd
if zeus.is_gpu_device():
self._world_size = hvd.size()
self._rank_id = hvd.rank()
self._local_rank_id = hvd.local_rank()
elif zeus.is_npu_device():
from hccl.manage.api import get_local_rank_id
from hccl.manage.api import get_rank_size
from hccl.manage.api import get_rank_id
self._world_size = get_rank_size()
self._rank_id = get_rank_id()
self._local_rank_id = get_local_rank_id()
def _build_multigpu_train_op(self, num_gpus):
with self.graph.as_default(), tf.device('/gpu:0'):
tower_grads = []
self.inputs = []
self.labels = []
opt = Optimizer()()
for i in range(num_gpus):
with tf.device('/gpu:%d' % i):
with tf.name_scope('tower_%d' % i):
# tf.get_variable_scope().reuse_variables()
inputs = self._create_tensor(self.loss_input['inputs'])
labels = self._create_tensor(self.loss_input['labels'])
input = inputs[0]
model_output = self.model(input)
loss = Loss()()
loss = loss(model_output, labels)
# Calculate the gradients for the batch of data on this tower.
varlist = [
x for x in tf.trainable_variables()
if x.name.startswith('tower_%d' % i)
]
grads = opt.compute_gradients(loss, varlist)
tower_grads.append(grads)
if i == 0:
self.actor_var = TFVariables(
model_output, self.sess)
self.input = input
self.logits = model_output
self.loss = loss
self.inputs.append(inputs)
self.labels.append(labels)
grads = self._average_gradients(tower_grads)
self.train_op = opt.apply_gradients(grads)
self.sess.run(tf.initialize_all_variables())
def _average_gradients(self, tower_grads):
avg_grads = []
for grad_and_vars in zip(*tower_grads):
grads = []
for g, _ in grad_and_vars:
expanded_g = tf.expand_dims(g, 0)
grads.append(expanded_g)
all_grad = tf.concat(grads, 0)
avg_grad = tf.reduce_mean(all_grad, 0, keep_dims=False)
for _, v in grad_and_vars:
grad_and_var = (avg_grad, v)
avg_grads.append(grad_and_var)
return avg_grads
def _create_tensor(self, tensor_list):
ret_list = []
for tensor in tensor_list:
tensor_type = tensor['type']
tensor_shape = tensor['shape']
tensor_name = tensor['name']
if type(tensor_shape) is list:
tf_tensor = tf.placeholder(tensor_type,
name=tensor_name,
shape=(None, ) +
tuple(tensor_shape))
else:
tf_tensor = tf.placeholder(tensor_type,
name=tensor_name,
shape=(None, tensor_shape))
ret_list.append(tf_tensor)
return ret_list
def _init_train_op(self):
self.train_count = 0
self.train_time = 0.
import os
if self.gpu_nums >= 1:
if 'CUDA_VISIBLE_DEVICES' in os.environ and os.environ[
'CUDA_VISIBLE_DEVICES'] == '-1':
with tf.name_scope('tower_0'):
self._build_train_op()
else:
self._build_multigpu_train_op(self.gpu_nums)
else:
self._build_train_op()
def _build_train_op(self):
self.inputs = self._create_tensor(self.loss_input['inputs'])
self.labels = self._create_tensor(self.loss_input['labels'])
self.input = self.inputs[0]
logits = self.model(self.input)
self.logits = logits
self.actor_var = TFVariables(logits, self.sess)
loss = Loss()()
self.loss = loss(logits, self.labels)
self.optimizer = Optimizer()(distributed=self.distributed)
grads_and_var = self.optimizer.compute_gradients(self.loss)
grads, var = zip(*grads_and_var)
grads_and_var = list(zip(grads, var))
self.train_op = self.optimizer.apply_gradients(grads_and_var)
self.sess.run(tf.initialize_all_variables())
def _default_train_input_fn(self):
return self.train_loader.input_fn()
def _default_valid_input_fn(self):
return self.valid_loader.input_fn()
def _default_model_fn(self, features, labels, mode):
"""Define model_fn used by TensorFlow Estimator.
:params features: input features
:type features: tensorflow tensors
:params labels: label data
:type labels: tensorflow tensors
:params mode: mode of estimator
:type mode: tf.estimator.ModeKeys
:return: tensorflow EstimatorSpec
:rtype: tf.estimator.EstimatorSpec
"""
logging.info('model function action')
self.model.training = mode == tf.estimator.ModeKeys.TRAIN
logits = self.model(features)
assign_ops = self.model.pretrained()
with tf.control_dependencies(assign_ops):
logits = tf.cast(logits, tf.float32)
if hasattr(self.model, 'add_loss'):
loss_cls = Loss()()
self.model.add_loss(loss_cls)
self.loss = self.model.overall_loss()
else:
self.loss = Loss()()
loss = self.loss(logits, labels)
train_op = None
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.compat.v1.train.get_or_create_global_step()
epoch = tf.cast(global_step, tf.float32) / tf.cast(
len(self.train_loader), tf.float32)
self.optimizer = Optimizer()(distributed=self.distributed)
self.lr_scheduler = LrScheduler()(optimizer=self.optimizer)
self.lr_scheduler.step(epoch)
if self.distributed:
self.optimizer = Optimizer.set_distributed(self.optimizer)
update_ops = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.UPDATE_OPS)
loss_scale = self.config.loss_scale if self.use_amp else 1
minimize_op = self.optimizer.step(loss, loss_scale,
global_step)
train_op = tf.group(minimize_op, update_ops)
eval_metric_ops = None
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = self.valid_metrics(logits, labels)
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)
def _init_tf_estimator(self):
"""Init tensorflow estimator."""
sess_config = self._init_session_config()
if zeus.is_gpu_device():
self._init_gpu_estimator(sess_config)
elif zeus.is_npu_device():
self._init_npu_estimator(sess_config)
def _init_tf_session(self):
sess_config = self._init_session_config()
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf.compat.v1.Session(config=sess_config)
def _init_session_config(self):
sess_config = self._init_gpu_session_config() if zeus.is_gpu_device() else \
self._init_npu_session_config()
return sess_config
def _init_logging_hook(self):
logging_hook = []
if zeus.is_gpu_device() and self.distributed:
import horovod.tensorflow as hvd
logging_hook += [hvd.BroadcastGlobalVariablesHook(0)]
return logging_hook
def _init_gpu_estimator(self, sess_config):
"""Init tensorflow estimator."""
distribution = None
if not self.distributed and General._parallel and General.devices_per_trainer > 1:
distribution = tf.contrib.distribute.MirroredStrategy()
config = tf.estimator.RunConfig(
model_dir=self.get_local_worker_path(),
save_checkpoints_steps=self.config.save_steps,
log_step_count_steps=self.config.report_freq,
session_config=None if distribution else sess_config,
train_distribute=distribution)
self.estimator = tf.estimator.Estimator(model_fn=self.model_fn,
config=config)
def _init_npu_estimator(self, sess_config):
from npu_bridge.estimator.npu.npu_config import NPURunConfig
from npu_bridge.estimator.npu.npu_estimator import NPUEstimator
model_dir = self.get_local_worker_path()
config = NPURunConfig(model_dir=model_dir,
save_checkpoints_steps=self.config.save_steps,
log_step_count_steps=self.config.report_freq,
session_config=sess_config,
enable_data_pre_proc=True,
iterations_per_loop=1)
self.estimator = NPUEstimator(model_fn=self.model_fn, config=config)
def _init_gpu_session_config(self):
sess_config = tf.compat.v1.ConfigProto()
sess_config.gpu_options.allow_growth = True
if self.distributed:
import horovod.tensorflow as hvd
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
return sess_config
def _init_npu_session_config(self):
from tensorflow.core.protobuf.rewriter_config_pb2 import RewriterConfig
sess_config = tf.ConfigProto()
sess_config.graph_options.rewrite_options.remapping = RewriterConfig.OFF
custom_op = sess_config.graph_options.rewrite_options.custom_optimizers.add(
)
custom_op.name = "NpuOptimizer"
if self.use_amp:
custom_op.parameter_map["precision_mode"].s = tf.compat.as_bytes(
"allow_mix_precision")
custom_op.parameter_map["use_off_line"].b = True
return sess_config
class Trainer(DistributedWorker):
"""Trainer class.
:param model: input model, defaults to None
:type model: tf model, optional
:param id: id of the model, defaults to None
:type id: int, optional
:param hps: hyperparameters, defaults to None
:type hps: dict, optional
"""
config = TrainerConfig()
def __init__(self, model=None, id=None, hps=None,
load_ckpt_flag=False, model_desc=None,
lazy_build=True, **kwargs):
super(Trainer, self).__init__()
self.worker_type = WorkerTypes.TRAINER
Trainer.__worker_id__ += 1
if id is not None:
self._worker_id = id
else:
self._worker_id = Trainer.__worker_id__
# Data Memeber list of Trainer
self.is_chief = True
self.use_cuda = self.config.cuda
self.epochs = self.config.epochs
self.do_validation = True
self.auto_save_ckpt = True
self.auto_save_perf = True
self.skip_train = False
self.valid_interval = self.config.valid_interval
self.hps = hps
self.model = model
self.model_desc = model_desc
self.optimizer = None
self.lr_scheduler = None
self.loss = None
self.use_syncbn = self.config.syncbn
self.use_amp = self.config.amp
self.train_metrics = None
self.valid_metrics = None
self.call_metrics_on_train = self.config.call_metrics_on_train
self.train_verbose = self.config.train_verbose
self.valid_verbose = self.config.valid_verbose
self.train_report_steps = self.config.train_report_steps
self.valid_report_steps = self.config.valid_report_steps
self.train_loader = None
self.valid_loader = None
self.train_step = None
self.valid_step = None
self.make_batch = None
self.model_fn = None
self.train_input_fn = None
self.valid_input_fn = None
self.callbacks = None
self.performance = None
self.runtime = None
self.visual_data = {}
self.load_ckpt_flag = load_ckpt_flag
self.distributed = self.config.distributed
# Used by TimmTrainerCallbacks since it builds its trainer in
# the before_train callback
self.lazy_built = self.config.lazy_built
# Indicate whether the necessary components of a trainer
# has been built for running
self._world_size = 1
self._rank_id = 0
self._local_rank_id = 0
self.config.kwargs = kwargs
self.checkpoint_file_name = 'checkpoint.pth'
self.model_pickle_file_name = 'model.pkl'
worker_path = self.get_local_worker_path()
self.model_path = FileOps.join_path(worker_path, self.model_pickle_file_name)
self.checkpoint_file = FileOps.join_path(worker_path, self.checkpoint_file_name)
self.weights_file = FileOps.join_path(worker_path, "model_{}.pth".format(self.worker_id))
self.loss_input = kwargs.get('loss_input', None)
if not lazy_build:
self.init_trainer()
def _set_default_funcs(self):
if zeus.is_torch_backend():
self.make_batch = self._default_make_batch
self.train_step = self._default_train_step
self.valid_step = self._default_valid_step
elif zeus.is_tf_backend():
self.model_fn = self._default_model_fn
self.train_input_fn = self._default_train_input_fn
self.valid_input_fn = self._default_valid_input_fn
def _set_condition(self):
self._init_tf_session()
self._init_distributed_setting()
self._init_cuda_setting()
self._init_tf_estimator()
self._init_ms_context()
def train_process(self):
"""Whole train process of the TrainWorker specified in config.
After training, the model and validation results are saved to local_worker_path and s3_path.
"""
init_log(level=General.logger.level,
log_file="log_worker_{}.txt".format(self.worker_id),
log_path=self.local_log_path)
self._set_default_funcs()
self._set_condition()
self._init_callbacks()
self.callbacks.init_trainer()
if not self.lazy_built:
self.build()
self._train_loop()
def build(self):
"""Build the trainer by assembling the necessary components."""
self._init_hps(self.hps)
logging.debug("Trainer Config: {}".format(self.config))
self.do_validation = self.config.with_valid
self.use_syncbn = self.config.syncbn
if self.use_syncbn and zeus.is_torch_backend():
self.model = apex.parallel.convert_syncbn_model(self.model)
self.train_loader = self._init_dataloader(mode='train')
self.valid_loader = self._init_dataloader(mode='val')
self.batch_num_train = self.train_loader.get_dataset_size() if zeus.is_ms_backend() else len(self.train_loader)
self.batch_num_valid = self.valid_loader.get_dataset_size() if zeus.is_ms_backend() else len(self.valid_loader)
if zeus.is_torch_backend():
self.optimizer = Optimizer()(model=self.model, distributed=self.distributed)
if hasattr(self.model, 'add_loss'):
loss_cls = Loss()()
self.model.add_loss(loss_cls)
self.loss = self.model.overall_loss()
else:
self.loss = Loss()()
self.lr_scheduler = LrScheduler()(self.optimizer)
elif zeus.is_ms_backend():
self.optimizer = Optimizer()(model=self.model)
if hasattr(self.model, 'add_loss'):
loss_cls = Loss()()
self.model.add_loss(loss_cls)
self.loss = self.model.overall_loss()
else:
self.loss = Loss()()
self.metric_name = self.config.metric().type
# Some trainer has different train batch size from valid batch
self.train_metrics = self._init_metrics() if zeus.is_torch_backend() else None
self.valid_metrics = self._init_metrics()
self._init_horovod_setting()
if self.use_amp and zeus.is_torch_backend():
self.model, self.optimizer = amp.initialize(
self.model, self.optimizer, opt_level='O1')
def init_trainer(self):
"""Init Train Op."""
init_log(level=General.logger.level,
log_file="log_worker_{}.txt".format(self.worker_id),
log_path=self.local_log_path)
self._set_default_funcs()
self._set_condition()
self._init_callbacks()
self.callbacks.init_trainer()
self.init_train_op()
def init_train_op(self):
"""Init Train Op."""
if zeus.is_tf_backend():
with self.graph.as_default():
self._init_train_op()
def train(self, inputs, labels):
"""Train model."""
if zeus.is_tf_backend():
feed_dict = {}
with self.graph.as_default():
for i in range(len(inputs)):
feed_dict.update({self.inputs[i]: inputs[i]})
for i in range(len(labels)):
feed_dict.update({self.labels[i]: labels[i]})
_, loss = self.sess.run([self.train_op, self.loss], feed_dict)
return loss
def predict(self, input):
"""Inference model."""
if zeus.is_tf_backend():
with self.graph.as_default():
feed_dict = {self.input: input}
out = self.sess.run(self.logits, feed_dict)
return out
def save(self, file_name):
"""Save model."""
if zeus.is_tf_backend():
with self.graph.as_default():
self.actor_var.save_weights(file_name + ".npz")
return file_name + ".npz"
def load(self, model_name, by_name):
"""Load model."""
if zeus.is_tf_backend():
with self.graph.as_default():
self.actor_var.set_weights_with_npz(model_name)
def set_weights(self, weights):
"""Set weight with memory tensor."""
if zeus.is_tf_backend():
with self.graph.as_default():
self.actor_var.set_weights(weights)
def get_weights(self):
"""Get the weights."""
if zeus.is_tf_backend():
with self.graph.as_default():
return self.actor_var.get_weights()
def _create_tensor(self, tensor_list):
ret_list = []
for tensor in tensor_list:
tensor_type = tensor['type']
tensor_shape = tensor['shape']
tensor_name = tensor['name']
if type(tensor_shape) is list:
tf_tensor = tf.placeholder(tensor_type, name=tensor_name,
shape=(None, ) + tuple(tensor_shape))
else:
tf_tensor = tf.placeholder(tensor_type, name=tensor_name,
shape=(None, tensor_shape))
ret_list.append(tf_tensor)
return ret_list
def _init_train_op(self):
if self.loss_input is not None:
self.inputs = self._create_tensor(self.loss_input['inputs'])
self.labels = self._create_tensor(self.loss_input['labels'])
self.input = self.inputs[0]
logits = self.model(self.input)
self.logits = logits
self.actor_var = TFVariables(logits, self.sess)
loss = Loss()()
self.loss = loss(logits, self.labels)
self.optimizer = Optimizer()(distributed=self.distributed)
grads_and_var = self.optimizer.compute_gradients(self.loss)
grads, var = zip(*grads_and_var)
grads_and_var = list(zip(grads, var))
self.train_op = self.optimizer.apply_gradients(grads_and_var)
self.sess.run(tf.initialize_all_variables())
def _init_cuda_setting(self):
"""Init CUDA setting."""
if not zeus.is_torch_backend():
return
if not self.config.cuda:
self.config.device = -1
return
self.config.device = self.config.cuda if self.config.cuda is not True else 0
self.use_cuda = True
if self.distributed:
torch.cuda.set_device(self._local_rank_id)
torch.cuda.manual_seed(self.config.seed)
def _init_distributed_setting(self):
if not self.distributed:
return
if zeus.is_npu_device():
self.npu_init = npu_ops.initialize_system()
self.npu_shutdown = npu_ops.shutdown_system()
self.sess.run(self.npu_init)
self._world_size = hvd.size() if zeus.is_gpu_device() else get_rank_size()
self._rank_id = hvd.rank() if zeus.is_gpu_device() else get_rank_id()
self._local_rank_id = hvd.local_rank() if zeus.is_gpu_device() else get_local_rank_id()
def _init_horovod_setting(self):
"""Init horovod setting."""
self.is_chief = True
if self.distributed and zeus.is_torch_backend():
hvd.broadcast_parameters(self.model.state_dict(), root_rank=0)
hvd.broadcast_optimizer_state(self.optimizer, root_rank=0)
if hvd.rank() != 0:
self.is_chief = False
else:
self.is_chief = True
def _init_hps(self, hps=None):
"""Load hps from file."""
if hps is not None:
self.hps = hps
elif self.config.hps_file is not None:
desc_file = self.config.hps_file.replace("{local_base_path}", self.local_base_path)
self.hps = Config(desc_file)
elif self.config.hps_folder is not None:
folder = self.config.hps_folder.replace("{local_base_path}", self.local_base_path)
pattern = FileOps.join_path(folder, "desc_*.json")
desc_file = glob.glob(pattern)[0]
self.hps = Config(desc_file)
if self.hps and self.hps.get('trainer'):
self.config.from_json(self.hps.get('trainer'))
self.epochs = self.config.epochs
def _init_metrics(self, metrics=None):
"""Init metrics."""
if metrics is not None:
return metrics
else:
return Metrics()
def _init_dataloader(self, mode, loader=None):
"""Init dataloader."""
if loader is not None:
return loader
if mode == "train" and self.hps is not None and self.hps.get("dataset") is not None:
dataset_cls = ClassFactory.get_cls(ClassType.DATASET)
dataset = dataset_cls(mode=mode, hps=self.hps.get("dataset"))
else:
dataset_cls = ClassFactory.get_cls(ClassType.DATASET)
dataset = dataset_cls(mode=mode)
if self.distributed and mode == "train":
dataset.set_distributed(self._world_size, self._rank_id)
# adapt the dataset to specific backend
dataloader = Adapter(dataset).loader
return dataloader
def _train_loop(self):
"""Do the training with data, callbacks and step functions etc."""
# Allow user to build trainer in before_train() callback, but they
# should set lazy_built in configuration file to True
self.callbacks.before_train()
if self.skip_train:
return
repeat_time = 1 if zeus.is_ms_backend() else self.epochs
for epoch in range(repeat_time):
epoch_logs = {'train_num_batches': self.batch_num_train}
if self.do_validation:
epoch_logs.update({'valid_num_batches': self.batch_num_valid})
self.callbacks.before_epoch(epoch, epoch_logs)
self._train_epoch()
if self.do_validation and self._should_run_validation(epoch):
self._valid_epoch()
self.callbacks.after_epoch(epoch)
self.callbacks.after_train()
if self.distributed:
self._shutdown_distributed()
def _train_epoch(self):
if zeus.is_torch_backend():
self.model.train()
for batch_index, batch in enumerate(self.train_loader):
batch = self.make_batch(batch)
batch_logs = {'train_batch': batch}
self.callbacks.before_train_step(batch_index, batch_logs)
train_batch_output = self.train_step(batch)
batch_logs.update(train_batch_output)
if self.config.is_detection_trainer:
batch_logs.update({'is_detection_trainer': True})
self.callbacks.after_train_step(batch_index, batch_logs)
elif zeus.is_tf_backend():
self.estimator.train(input_fn=self.train_input_fn,
steps=len(self.train_loader),
hooks=self._init_logging_hook())
elif zeus.is_ms_backend():
self.ms_model = MsModel(network=self.model,
loss_fn=self.loss,
optimizer=self.optimizer,
metrics={self.metric_name: self.valid_metrics()})
config_ck = CheckpointConfig(save_checkpoint_steps=self.config.save_steps)
# save the network model and parameters for subsequence fine-tuning
save_path = self.get_local_worker_path(self.step_name, self.worker_id)
ckpoint_cb = ModelCheckpoint(config=config_ck, directory=save_path)
loss_cb = LossMonitor(per_print_times=self.config.report_freq)
eval_cb = EvalCallBack(self.ms_model, self.valid_loader)
self.ms_model.train(epoch=self.epochs,
train_dataset=self.train_loader,
callbacks=[ckpoint_cb, loss_cb, eval_cb],
dataset_sink_mode=self.dataset_sink_mode)
def _valid_epoch(self):
self.callbacks.before_valid()
valid_logs = None
if zeus.is_torch_backend():
self.model.eval()
with torch.no_grad():
for batch_index, batch in enumerate(self.valid_loader):
batch = self.make_batch(batch)
batch_logs = {'valid_batch': batch}
self.callbacks.before_valid_step(batch_index, batch_logs)
valid_batch_output = self.valid_step(batch)
self.callbacks.after_valid_step(batch_index, valid_batch_output)
elif zeus.is_tf_backend():
eval_metrics = self.estimator.evaluate(input_fn=self.valid_input_fn,
steps=len(self.valid_loader))
self.valid_metrics.update(eval_metrics)
valid_logs = dict()
valid_logs['cur_valid_perfs'] = self.valid_metrics.results
elif zeus.is_ms_backend():
eval_metrics = self.ms_model.eval(valid_dataset=self.valid_loader,
dataset_sink_mode=self.dataset_sink_mode)
self.valid_metrics.update(eval_metrics)
valid_logs = dict()
valid_logs['cur_valid_perfs'] = self.valid_metrics.results
self.callbacks.after_valid(valid_logs)
def _default_make_batch(self, batch):
"""Unpack batch to get input and target."""
input, target = batch
if self.use_cuda and not self.config.is_detection_trainer:
input, target = input.cuda(), target.cuda()
return (input, target)
def _default_train_step(self, batch):
input, target = batch
self.optimizer.zero_grad()
output = self.model(input)
loss = self.loss(output, target)
if self.use_amp:
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
self.optimizer.synchronize()
with self.optimizer.skip_synchronize():
self.optimizer.step()
else:
loss.backward()
if self.config.grad_clip:
torch.nn.utils.clip_grad_norm_(
self.model.parameters(), self.config.grad_clip)
self.optimizer.step()
return {'loss': loss.item(),
'train_batch_output': output,
'lr': self.lr_scheduler.get_lr()}
def _default_valid_step(self, batch):
input, target = batch
if self.config.is_detection_trainer:
output = self.model(input, forward_train=False)
else:
output = self.model(input)
return {'valid_batch_output': output}
def _init_minimize_op(self, loss, global_step, var_list=None):
"""Init loss minimize operation, include loss scale method."""
loss_scale = self.config.loss_scale if self.use_amp else 1.
if loss_scale != 1:
scaled_grad_vars = self.optimizer.compute_gradients(loss * loss_scale, var_list=var_list)
unscaled_grad_vars = []
for grad, var in scaled_grad_vars:
unscaled_grad_vars.append((grad, var) if grad is None else (grad / loss_scale, var))
minimize_op = self.optimizer.apply_gradients(unscaled_grad_vars, global_step)
else:
grad_vars = self.optimizer.compute_gradients(loss, var_list=var_list)
minimize_op = self.optimizer.apply_gradients(grad_vars, global_step)
return minimize_op
def _default_train_input_fn(self):
return self.train_loader.input_fn()
def _default_valid_input_fn(self):
return self.valid_loader.input_fn()
def _default_model_fn(self, features, labels, mode):
"""Define model_fn used by TensorFlow Estimator.
:params features: input features
:type features: tensorflow tensors
:params labels: label data
:type labels: tensorflow tensors
:params mode: mode of estimator
:type mode: tf.estimator.ModeKeys
:return: tensorflow EstimatorSpec
:rtype: tf.estimator.EstimatorSpec
"""
logging.info('model function action')
self.model.training = mode == tf.estimator.ModeKeys.TRAIN
logits = self.model(features)
logits = tf.cast(logits, tf.float32)
if hasattr(self.model, 'add_loss'):
loss_cls = Loss()()
self.model.add_loss(loss_cls)
self.loss = self.model.overall_loss()
else:
self.loss = Loss()()
loss = self.loss(logits, labels)
train_op = None
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.compat.v1.train.get_or_create_global_step()
epoch = tf.cast(global_step, tf.float32) / tf.cast(len(self.train_loader), tf.float32)
self.optimizer = Optimizer()(distributed=self.distributed)
self.lr_scheduler = LrScheduler()(optimizer=self.optimizer)
self.lr_scheduler.step(epoch)
update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS)
loss_scale = self.config.loss_scale if self.use_amp else 1
minimize_op = self.optimizer.step(loss, loss_scale, global_step)
train_op = tf.group(minimize_op, update_ops)
eval_metric_ops = None
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = self.valid_metrics(logits, labels)
return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op,
eval_metric_ops=eval_metric_ops)
def _should_run_validation(self, epoch):
# Zero valid_interval means doesn't run _valid_loop of the trainer
# and user may provide _valid_loop in other callbacks
if self.valid_interval == 0:
return False
else:
return epoch % self.valid_interval == 0 or (epoch + 1) == self.epochs
def _init_callbacks(self):
disables = []
customs = self.config.callbacks or []
if customs and not isinstance(customs, list):
customs = [customs]
if not self.config.model_statistics:
disables.append('ModelStatistics')
self.callbacks = CallbackList(customs, disables)
self.callbacks.set_trainer(self)
def _metric_average(self, val, name):
"""Do metric average.
:param val: input value
:param name: metric name
:return:
"""
tensor = torch.tensor(val)
avg_tensor = hvd.allreduce(tensor, name=name)
return avg_tensor.item()
@property
def _first_rank(self):
"""Check if the first rank."""
if self.distributed and hvd.rank() != 0:
return False
else:
return True
def _backup(self):
"""Backup result worker folder."""
if self.need_backup is True and self.backup_base_path is not None:
backup_worker_path = FileOps.join_path(
self.backup_base_path, self.get_worker_subpath())
FileOps.copy_folder(
self.get_local_worker_path(self.step_name, self.worker_id), backup_worker_path)
def _save_visual_data(self, is_train=True, pfms=None, loss=None, lr=None):
# TODO Will move to metric base class later.
for _name, value in pfms.items():
if is_train:
_name = "{}_{}".format("t", _name)
else:
_name = "{}_{}".format("v", _name)
if isinstance(value, list):
for i, _item in enumerate(value):
_name = "{}_{}".format(_name, i)
self.visual_data[_name] = _item.data.item()
elif isinstance(value, dict):
for k, v in value.keys():
_name = "{}_{}".format(_name, k)
self.visual_data[_name] = v
elif value is not None:
self.visual_data[_name] = value.data.item()
if loss is not None:
self.visual_data["loss"] = loss
if lr is not None:
self.visual_data["lr"] = lr
def _init_tf_estimator(self):
"""Init tensorflow estimator."""
if not zeus.is_tf_backend():
return
sess_config = self._init_session_config()
if zeus.is_gpu_device():
self._init_gpu_estimator(sess_config)
elif zeus.is_npu_device():
self._init_npu_estimator(sess_config)
def _init_tf_session(self):
if not zeus.is_tf_backend():
return
sess_config = self._init_session_config()
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf.compat.v1.Session(config=sess_config)
def _init_session_config(self):
sess_config = self._init_gpu_session_config() if zeus.is_gpu_device() else \
self._init_npu_session_config()
return sess_config
def _init_logging_hook(self):
logging_hook = []
if zeus.is_gpu_device() and self.distributed:
logging_hook += [hvd.BroadcastGlobalVariablesHook(0)]
return logging_hook
def _init_gpu_estimator(self, sess_config):
"""Init tensorflow estimator."""
distribution = None
if not self.distributed and General._parallel and General.devices_per_trainer > 1:
distribution = tf.contrib.distribute.MirroredStrategy()
config = tf.estimator.RunConfig(model_dir=self.get_local_worker_path(),
save_checkpoints_steps=self.config.save_steps,
log_step_count_steps=self.config.report_freq,
session_config=None if distribution else sess_config,
train_distribute=distribution)
self.estimator = tf.estimator.Estimator(model_fn=self.model_fn,
config=config)
def _init_npu_estimator(self, sess_config):
model_dir = self.get_local_worker_path()
config = NPURunConfig(model_dir=model_dir,
save_checkpoints_steps=self.config.save_steps,
log_step_count_steps=self.config.report_freq,
session_config=sess_config,
enable_data_pre_proc=True,
iterations_per_loop=1)
self.estimator = NPUEstimator(model_fn=self.model_fn,
config=config)
def _init_gpu_session_config(self):
sess_config = tf.compat.v1.ConfigProto()
sess_config.gpu_options.allow_growth = True
if self.distributed:
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
return sess_config
def _init_npu_session_config(self):
sess_config = tf.ConfigProto()
sess_config.graph_options.rewrite_options.remapping = RewriterConfig.OFF
custom_op = sess_config.graph_options.rewrite_options.custom_optimizers.add()
custom_op.name = "NpuOptimizer"
if self.use_amp:
custom_op.parameter_map["precision_mode"].s = tf.compat.as_bytes("allow_mix_precision")
custom_op.parameter_map["use_off_line"].b = True
# custom_op.parameter_map['hcom_parallel'].b = True
# custom_op.parameter_map["enable_data_pre_proc"].b = True
# custom_op.parameter_map["mix_compile_mode"].b = True # mixed calculation
# custom_op.parameter_map["min_group_size"].b = 1
return sess_config
def _init_ms_context(self):
if not zeus.is_ms_backend():
return
if zeus.is_npu_device():
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
else:
context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
self.dataset_sink_mode = True if zeus.is_npu_device() else False
def _shutdown_distributed(self):
if zeus.is_npu_device() and self.distributed:
self.sess.run(self.npu_shutdown)
self.sess.close()
class TrainerTf(TrainerBase):
"""Trainer tensorflow class."""
def build(self):
"""Build the trainer by assembling the necessary components."""
super().build()
# Some trainer has different train batch size from valid batch
self.train_metrics = None
self.valid_metrics = self._init_metrics()
self._init_horovod_setting()
def train(self, inputs, labels):
"""Train model."""
feed_dict = {}
with self.graph.as_default():
for i in range(len(inputs)):
feed_dict.update({self.inputs[i]: inputs[i]})
for i in range(len(labels)):
feed_dict.update({self.labels[i]: labels[i]})
_, loss = self.sess.run([self.train_op, self.loss], feed_dict)
return loss
def predict(self, input):
"""Inference model."""
with self.graph.as_default():
feed_dict = {self.input: input}
out = self.sess.run(self.logits, feed_dict)
return out
def save(self, file_name):
"""Save model."""
with self.graph.as_default():
self.actor_var.save_weights(file_name + ".npz")
return file_name + ".npz"
def load(self, model_name, by_name):
"""Load model."""
with self.graph.as_default():
self.actor_var.set_weights_with_npz(model_name)
def set_weights(self, weights):
"""Set weight with memory tensor."""
with self.graph.as_default():
self.actor_var.set_weights(weights)
def get_weights(self):
"""Get the weights."""
with self.graph.as_default():
return self.actor_var.get_weights()
def init_train_op(self):
"""Init Train Op."""
with self.graph.as_default():
self._init_train_op()
def _set_default_funcs(self):
self.model_fn = self._default_model_fn
self.train_input_fn = self._default_train_input_fn
self.valid_input_fn = self._default_valid_input_fn
def _set_condition(self):
self._init_tf_session()
self._init_distributed_setting()
self._init_tf_estimator()
def _train_epoch(self):
self.estimator.train(input_fn=self.train_input_fn,
steps=len(self.train_loader),
hooks=self._init_logging_hook())
def _valid_epoch(self):
self.callbacks.before_valid()
valid_logs = None
eval_metrics = self.estimator.evaluate(input_fn=self.valid_input_fn,
steps=len(self.valid_loader))
self.valid_metrics.update(eval_metrics)
valid_logs = dict()
valid_logs['cur_valid_perfs'] = self.valid_metrics.results
self.callbacks.after_valid(valid_logs)
def _init_distributed_setting(self):
if not self.distributed:
return
if zeus.is_npu_device():
self.npu_init = npu_ops.initialize_system()
self.npu_shutdown = npu_ops.shutdown_system()
self.sess.run(self.npu_init)
self._world_size = hvd.size() if zeus.is_gpu_device(
) else get_rank_size()
self._rank_id = hvd.rank() if zeus.is_gpu_device() else get_rank_id()
self._local_rank_id = hvd.local_rank() if zeus.is_gpu_device(
) else get_local_rank_id()
def _create_tensor(self, tensor_list):
ret_list = []
for tensor in tensor_list:
tensor_type = tensor['type']
tensor_shape = tensor['shape']
tensor_name = tensor['name']
if type(tensor_shape) is list:
tf_tensor = tf.placeholder(tensor_type,
name=tensor_name,
shape=(None, ) +
tuple(tensor_shape))
else:
tf_tensor = tf.placeholder(tensor_type,
name=tensor_name,
shape=(None, tensor_shape))
ret_list.append(tf_tensor)
return ret_list
def _init_train_op(self):
self.inputs = self._create_tensor(self.loss_input['inputs'])
self.labels = self._create_tensor(self.loss_input['labels'])
self.input = self.inputs[0]
logits = self.model(self.input)
self.logits = logits
self.actor_var = TFVariables(logits, self.sess)
loss = Loss()()
self.loss = loss(logits, self.labels)
self.optimizer = Optimizer()(distributed=self.distributed)
grads_and_var = self.optimizer.compute_gradients(self.loss)
grads, var = zip(*grads_and_var)
grads_and_var = list(zip(grads, var))
self.train_op = self.optimizer.apply_gradients(grads_and_var)
self.sess.run(tf.initialize_all_variables())
def _default_train_input_fn(self):
return self.train_loader.input_fn()
def _default_valid_input_fn(self):
return self.valid_loader.input_fn()
def _default_model_fn(self, features, labels, mode):
"""Define model_fn used by TensorFlow Estimator.
:params features: input features
:type features: tensorflow tensors
:params labels: label data
:type labels: tensorflow tensors
:params mode: mode of estimator
:type mode: tf.estimator.ModeKeys
:return: tensorflow EstimatorSpec
:rtype: tf.estimator.EstimatorSpec
"""
logging.info('model function action')
self.model.training = mode == tf.estimator.ModeKeys.TRAIN
logits = self.model(features)
logits = tf.cast(logits, tf.float32)
if hasattr(self.model, 'add_loss'):
loss_cls = Loss()()
self.model.add_loss(loss_cls)
self.loss = self.model.overall_loss()
else:
self.loss = Loss()()
loss = self.loss(logits, labels)
train_op = None
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.compat.v1.train.get_or_create_global_step()
epoch = tf.cast(global_step, tf.float32) / tf.cast(
len(self.train_loader), tf.float32)
self.optimizer = Optimizer()(distributed=self.distributed)
self.lr_scheduler = LrScheduler()(optimizer=self.optimizer)
self.lr_scheduler.step(epoch)
update_ops = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.UPDATE_OPS)
loss_scale = self.config.loss_scale if self.use_amp else 1
minimize_op = self.optimizer.step(loss, loss_scale, global_step)
train_op = tf.group(minimize_op, update_ops)
eval_metric_ops = None
if mode == tf.estimator.ModeKeys.EVAL:
eval_metric_ops = self.valid_metrics(logits, labels)
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)
def _init_tf_estimator(self):
"""Init tensorflow estimator."""
sess_config = self._init_session_config()
if zeus.is_gpu_device():
self._init_gpu_estimator(sess_config)
elif zeus.is_npu_device():
self._init_npu_estimator(sess_config)
def _init_tf_session(self):
sess_config = self._init_session_config()
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf.compat.v1.Session(config=sess_config)
def _init_session_config(self):
sess_config = self._init_gpu_session_config() if zeus.is_gpu_device() else \
self._init_npu_session_config()
return sess_config
def _init_logging_hook(self):
logging_hook = []
if zeus.is_gpu_device() and self.distributed:
logging_hook += [hvd.BroadcastGlobalVariablesHook(0)]
return logging_hook
def _init_gpu_estimator(self, sess_config):
"""Init tensorflow estimator."""
distribution = None
if not self.distributed and General._parallel and General.devices_per_trainer > 1:
distribution = tf.contrib.distribute.MirroredStrategy()
config = tf.estimator.RunConfig(
model_dir=self.get_local_worker_path(),
save_checkpoints_steps=self.config.save_steps,
log_step_count_steps=self.config.report_freq,
session_config=None if distribution else sess_config,
train_distribute=distribution)
self.estimator = tf.estimator.Estimator(model_fn=self.model_fn,
config=config)
def _init_npu_estimator(self, sess_config):
model_dir = self.get_local_worker_path()
config = NPURunConfig(model_dir=model_dir,
save_checkpoints_steps=self.config.save_steps,
log_step_count_steps=self.config.report_freq,
session_config=sess_config,
enable_data_pre_proc=True,
iterations_per_loop=1)
self.estimator = NPUEstimator(model_fn=self.model_fn, config=config)
def _init_gpu_session_config(self):
sess_config = tf.compat.v1.ConfigProto()
sess_config.gpu_options.allow_growth = True
if self.distributed:
sess_config.gpu_options.visible_device_list = str(hvd.local_rank())
return sess_config
def _init_npu_session_config(self):
sess_config = tf.ConfigProto()
sess_config.graph_options.rewrite_options.remapping = RewriterConfig.OFF
custom_op = sess_config.graph_options.rewrite_options.custom_optimizers.add(
)
custom_op.name = "NpuOptimizer"
if self.use_amp:
custom_op.parameter_map["precision_mode"].s = tf.compat.as_bytes(
"allow_mix_precision")
custom_op.parameter_map["use_off_line"].b = True
return sess_config