def train(args):
data_folder = args.get("data_folder")
save_folder = args.get("save_folder")
image_size = args.get("image_size")
max_epoch = args.get("max_epoch")
save_epoch = args.get("save_epoch") or max_epoch // 10
# Scale lr and steps_per_epoch accordingly.
# Make sure the total number of gradient evaluations is consistent.
n_gpu = args.get("n_gpu") or 1
batch_size = args.get("batch_size") or BATCH
equi_batch_size = max(n_gpu, 1) * batch_size
lr = args.get("lr") or LR
lr *= equi_batch_size
steps_per_epoch = args.get("steps_per_epoch") or 1000
steps_per_epoch /= equi_batch_size
image_steps = args.get("image_steps") or steps_per_epoch // 10
scalar_steps = args.get("scalar_steps")
if scalar_steps > 0:
scalar_steps = max(scalar_steps // equi_batch_size, 1)
else:
scalar_steps = 0 # merge scalar summary every epoch
# lr starts decreasing at half of max epoch
start_dec_epoch = max_epoch // 2
# stops when lr is 0.01 of its initial value
end_epoch = max_epoch - int((max_epoch - start_dec_epoch) * 0.01)
# adjust noise input range according to the input act
zmin, zmax = (0, 1) if args.get("act_input") == "identity" else (-1, 1)
if save_folder == None:
logger.auto_set_dir()
else:
logger.set_logger_dir(save_folder)
df = get_data(data_folder,
image_size,
zmin=zmin,
zmax=zmax,
batch=batch_size)
df = PrintData(df)
data = QueueInput(df)
SynTexTrainer(data, Style2PO(args), n_gpu).train_with_defaults(
callbacks=[
PeriodicTrigger(ModelSaver(), every_k_epochs=save_epoch),
PeriodicTrigger(ModelSaver(),
every_k_epochs=end_epoch), # save model at last
ScheduledHyperParamSetter('learning_rate', [(start_dec_epoch, lr),
(max_epoch, 0)],
interp="linear"),
PeriodicTrigger(VisualizeTestSet(data_folder, image_size),
every_k_epochs=max(1, max_epoch // 100)),
#MergeAllSummaries(period=scalar_steps), # scalar only, slowdown in training, use TCMalloc
MergeAllSummaries(period=image_steps, key="image_summaries"),
MergeAllSummaries(key="acti_summaries"),
],
max_epoch=end_epoch,
steps_per_epoch=steps_per_epoch,
session_init=None)
def fit(self, data):
"""Fit the model to the given data.
Args:
data(pandas.DataFrame): dataset to fit the model.
Returns:
None
"""
self.preprocessor = Preprocessor(
continuous_columns=self.continuous_columns)
data = self.preprocessor.fit_transform(data)
self.metadata = self.preprocessor.metadata
dataflow = TGANDataFlow(data, self.metadata)
batch_data = BatchData(dataflow, self.batch_size)
input_queue = QueueInput(batch_data)
self.model = self.get_model(training=True)
from tensorpack.callbacks import CometMLMonitor
trainer = SeparateGANTrainer(
model=self.model,
input_queue=input_queue,
g_period=6,
)
self.restore_path = os.path.join(self.model_dir, 'checkpoint')
if os.path.isfile(self.restore_path) and self.restore_session:
session_init = SaverRestore(self.restore_path)
with open(os.path.join(self.log_dir, 'stats.json')) as f:
starting_epoch = json.load(f)[-1]['epoch_num'] + 1
else:
session_init = None
starting_epoch = 1
action = 'k' if self.restore_session else 'd'
# logger.set_logger_dir(self.log_dir, action=action)
callbacks = []
monitors = []
if self.save_checkpoints:
callbacks.append(ModelSaver(checkpoint_dir=self.model_dir))
callbacks.append(MergeAllSummaries(period=10))
if self.experiment is not None:
monitors.append(CometMLMonitor(experiment=self.experiment))
trainer.train_with_defaults(callbacks=callbacks,
monitors=monitors,
steps_per_epoch=self.steps_per_epoch,
max_epoch=self.max_epoch,
session_init=session_init,
starting_epoch=starting_epoch)
self.prepare_sampling()
def fit(self, data):
"""Fit the model to the given data.
Args:
data(pandas.DataFrame): dataset to fit the model.
Returns:
None
"""
self.preprocessor = Preprocessor(
continuous_columns=self.continuous_columns)
data = self.preprocessor.fit_transform(data)
self.metadata = self.preprocessor.metadata
dataflow = TGANDataFlow(data, self.metadata)
batch_data = BatchData(dataflow, self.batch_size)
input_queue = QueueInput(batch_data)
self.model = self.get_model(training=True)
if self.trainer == 'GANTrainer':
trainer = GANTrainer(model=self.model, input_queue=input_queue)
elif self.trainer == 'SeparateGANTrainer':
trainer = SeparateGANTrainer(model=self.model,
input_queue=input_queue)
else:
raise ValueError(
'Incorrect trainer name. Use GANTrainer or SeparateGANTrainer')
# trainer = SeparateGANTrainer(model=self.model, input_queue=input_queue)
self.restore_path = os.path.join(self.model_dir, 'checkpoint')
if os.path.isfile(self.restore_path) and self.restore_session:
session_init = SaverRestore(self.restore_path)
with open(os.path.join(self.log_dir, 'stats.json')) as f:
starting_epoch = json.load(f)[-1]['epoch_num'] + 1
else:
session_init = None
starting_epoch = 1
action = 'k' if self.restore_session else None
logger.set_logger_dir(self.log_dir, action=action)
callbacks = []
if self.save_checkpoints:
callbacks.append(ModelSaver(checkpoint_dir=self.model_dir))
trainer.train_with_defaults(callbacks=callbacks,
steps_per_epoch=self.steps_per_epoch,
max_epoch=self.max_epoch,
session_init=session_init,
starting_epoch=starting_epoch)
self.prepare_sampling()
def get_config(
files_list,
input_names=["state_1", "state_2"],
output_names=["Qvalue_1", "Qvalue_2"],
agents=2,
):
"""This is only used during training."""
expreplay = ExpReplay(
predictor_io_names=(input_names, output_names),
player=get_player(task="train", files_list=files_list, agents=agents),
state_shape=IMAGE_SIZE,
batch_size=BATCH_SIZE,
memory_size=MEMORY_SIZE,
init_memory_size=INIT_MEMORY_SIZE,
init_exploration=1.0,
update_frequency=UPDATE_FREQ,
history_len=FRAME_HISTORY,
agents=agents,
)
return TrainConfig(
# dataflow=expreplay,
data=QueueInput(expreplay),
model=Model(agents=agents),
callbacks=[
ModelSaver(),
PeriodicTrigger(
RunOp(DQNModel.update_target_param, verbose=True),
# update target network every 10k steps
every_k_steps=10000 // UPDATE_FREQ,
),
expreplay,
ScheduledHyperParamSetter("learning_rate", [(60, 4e-4),
(100, 2e-4)]),
ScheduledHyperParamSetter(
ObjAttrParam(expreplay, "exploration"),
# 1->0.1 in the first million steps
[(0, 1), (10, 0.1), (320, 0.01)],
interp="linear",
),
PeriodicTrigger(
Evaluator(
nr_eval=EVAL_EPISODE,
input_names=input_names,
output_names=output_names,
files_list=files_list,
get_player_fn=get_player,
agents=agents,
),
every_k_epochs=EPOCHS_PER_EVAL,
),
HumanHyperParamSetter("learning_rate"),
],
steps_per_epoch=STEPS_PER_EPOCH,
max_epoch=1000,
)
def get_config():
"""This is only used during training."""
expreplay = ExpReplay(predictor_io_names=(['state'], ['Qvalue']),
player=get_player(directory=data_dir,
task='train',
files_list=train_data_fpaths),
state_shape=OBSERVATION_DIMS,
batch_size=BATCH_SIZE,
memory_size=MEMORY_SIZE,
init_memory_size=INIT_MEMORY_SIZE,
init_exploration=1.0,
update_frequency=UPDATE_FREQ,
frame_history_len=FRAME_HISTORY)
return TrainConfig(
# dataflow=expreplay,
data=QueueInput(expreplay),
model=Model(),
callbacks=[ # TODO: periodically save videos
ModelSaver(checkpoint_dir="model_checkpoints",
keep_checkpoint_every_n_hours=0.25,
max_to_keep=1000), # TODO: og was just ModelSaver()
PeriodicTrigger(
RunOp(DQNModel.update_target_param, verbose=True),
# update target network every 10k/freq steps
every_k_steps=10000 // UPDATE_FREQ),
# expreplay,
ScheduledHyperParamSetter('learning_rate', [(60, 4e-4),
(100, 2e-4)]),
ScheduledHyperParamSetter(
ObjAttrParam(expreplay, 'exploration'),
# 1->0.1 in the first 10M steps
[(0, 1), (100, 0.1), (120, 0.01)],
interp='linear'),
PeriodicTrigger( # runs exprelay._trigger()
expreplay, every_k_steps=5000),
PeriodicTrigger(
# eval_model_multithread(pred, EVAL_EPISODE, get_player)
Evaluator(nr_eval=EVAL_EPISODE,
input_names=['state'],
output_names=['Qvalue'],
directory=data_dir,
files_list=test_data_fpaths,
get_player_fn=get_player),
every_k_steps=10000 // UPDATE_FREQ),
HumanHyperParamSetter('learning_rate'),
],
steps_per_epoch=STEPS_PER_EPOCH,
max_epoch=NUM_EPOCHS,
)
def train_net(net,
session_init,
batch_size,
num_epochs,
train_dataflow,
val_dataflow):
num_towers = max(get_num_gpu(), 1)
batch_per_tower = batch_size // num_towers
logger.info("Running on {} towers. Batch size per tower: {}".format(num_towers, batch_per_tower))
num_training_samples = 1281167
step_size = num_training_samples // batch_size
max_iter = (num_epochs - 1) * step_size
callbacks = [
ModelSaver(),
ScheduledHyperParamSetter(
'learning_rate',
[(0, 0.5), (max_iter, 0)],
interp='linear',
step_based=True),
EstimatedTimeLeft()]
infs = [ClassificationError('wrong-top1', 'val-error-top1'),
ClassificationError('wrong-top5', 'val-error-top5')]
if num_towers == 1:
# single-GPU inference with queue prefetch
callbacks.append(InferenceRunner(
input=QueueInput(val_dataflow),
infs=infs))
else:
# multi-GPU inference (with mandatory queue prefetch)
callbacks.append(DataParallelInferenceRunner(
input=val_dataflow,
infs=infs,
gpus=list(range(num_towers))))
config = TrainConfig(
dataflow=train_dataflow,
model=net,
callbacks=callbacks,
session_init=session_init,
steps_per_epoch=step_size,
max_epoch=num_epochs)
launch_train_with_config(
config=config,
trainer=SyncMultiGPUTrainerParameterServer(num_towers))
def get_config():
"""This is only used during training."""
expreplay = ExpReplay(predictor_io_names=(['state'], ['Qvalue']),
player=get_player(directory=data_dir,
task='train',
files_list=train_list),
state_shape=IMAGE_SIZE,
batch_size=BATCH_SIZE,
memory_size=MEMORY_SIZE,
init_memory_size=INIT_MEMORY_SIZE,
init_exploration=1.0,
update_frequency=UPDATE_FREQ,
history_len=FRAME_HISTORY)
return TrainConfig(
# dataflow=expreplay,
data=QueueInput(expreplay),
model=Model(),
callbacks=[
ModelSaver(),
PeriodicTrigger(
RunOp(DQNModel.update_target_param, verbose=True),
# update target network every 10k steps
every_k_steps=10000 // UPDATE_FREQ),
expreplay,
ScheduledHyperParamSetter('learning_rate', [(60, 4e-4),
(100, 2e-4)]),
ScheduledHyperParamSetter(
ObjAttrParam(expreplay, 'exploration'),
# 1->0.1 in the first million steps
[(0, 1), (10, 0.1), (320, 0.01)],
interp='linear'),
PeriodicTrigger(Evaluator(nr_eval=EVAL_EPISODE,
input_names=['state'],
output_names=['Qvalue'],
directory=data_dir,
files_list=test_list,
get_player_fn=get_player),
every_k_epochs=EPOCHS_PER_EVAL),
HumanHyperParamSetter('learning_rate'),
],
steps_per_epoch=STEPS_PER_EPOCH,
max_epoch=1000,
)
def get_config(model, conf):
nr_tower = max(get_nr_gpu(), 1)
batch = conf.batch
if conf.fake:
logger.info("For benchmark, batch size is fixed to 64 per tower.")
dataset_train = FakeData([[64, 224, 224, 3], [64]],
1000,
random=False,
dtype='uint8')
callbacks = []
else:
logger.info("Running on {} towers. Batch size per tower: {}".format(
nr_tower, batch))
dataset_train = get_data(conf.data_dir, 'train', batch)
dataset_val = get_data(conf.data_dir, 'val', batch)
callbacks = [
ModelSaver(),
ScheduledHyperParamSetter('learning_rate', [(45, 1e-2), (60, 1e-3),
(65, 1e-4), (70, 1e-5),
(75, 1e-6)]),
HumanHyperParamSetter('learning_rate'),
]
infs = [
ClassificationError('wrong-top1', 'val-error-top1'),
ClassificationError('wrong-top5', 'val-error-top5')
]
if nr_tower == 1:
# single-GPU inference with queue prefetch
callbacks.append(InferenceRunner(QueueInput(dataset_val), infs))
else:
# multi-GPU inference (with mandatory queue prefetch)
callbacks.append(
DataParallelInferenceRunner(dataset_val, infs,
list(range(nr_tower))))
return TrainConfig(model=model,
dataflow=dataset_train,
callbacks=callbacks,
steps_per_epoch=5000,
max_epoch=80,
nr_tower=nr_tower)
def get_config(files_list, data_type, trainable_variables):
"""This is only used during training."""
expreplay = ExpReplay(
predictor_io_names=(['state'], ['Qvalue']),
player=get_player(task='train',
files_list=files_list,
data_type=data_type),
state_shape=IMAGE_SIZE,
batch_size=BATCH_SIZE,
memory_size=MEMORY_SIZE,
init_memory_size=INIT_MEMORY_SIZE,
init_exploration=0.8, #0.0
###############################################################################
# HITL UPDATE
update_frequency=INIT_UPDATE_FREQ,
###############################################################################
history_len=FRAME_HISTORY,
arg_type=data_type)
return TrainConfig(
# dataflow=expreplay,
data=QueueInput(expreplay),
model=Model(IMAGE_SIZE, FRAME_HISTORY, METHOD, NUM_ACTIONS, GAMMA,
trainable_variables),
callbacks=[
ModelSaver(),
PeriodicTrigger(
RunOp(DQNModel.update_target_param, verbose=True),
# update target network every 10k steps
every_k_steps=10000 // UPDATE_FREQ),
expreplay,
ScheduledHyperParamSetter('learning_rate', [(60, 4e-4),
(100, 2e-4)]),
ScheduledHyperParamSetter(
ObjAttrParam(expreplay, 'exploration'),
# 1->0.1 in the first million steps
[(0, 0.8), (1000000, 0.1), (32000000, 0.01)],
interp='linear',
step_based=True),
###############################################################################
# HITL UPDATE
# Here the number of steps taken in the environment is increased from 0, during
# the pretraining phase, to 4 to allow the agent to take 4 steps in the env
# between each TD update.
ScheduledHyperParamSetter(ObjAttrParam(expreplay,
'update_frequency'),
[(0, INIT_UPDATE_FREQ),
(NUM_PRETRAIN, UPDATE_FREQ)],
interp=None,
step_based=True),
###############################################################################
PeriodicTrigger(Evaluator(nr_eval=EVAL_EPISODE,
input_names=['state'],
output_names=['Qvalue'],
files_list=files_list,
data_type=data_type,
get_player_fn=get_player),
every_k_steps=STEPS_PER_EVAL),
HumanHyperParamSetter('learning_rate'),
],
steps_per_epoch=STEPS_PER_EPOCH,
max_epoch=MAX_EPOCHS,
)
start_dec_epoch = max_epoch // 2
# stops when lr is 0.01 of its initial value
end_epoch = max_epoch - int((max_epoch - start_dec_epoch) * 0.01)
# adjust noise input range according to the input act
zmin, zmax = (0, 1) if args.get("act") == "identity" else (-1, 1)
if save_folder == None:
logger.auto_set_dir()
else:
logger.set_logger_dir(save_folder)
df = get_data(data_folder, image_size, zmin=zmin, zmax=zmax)
df = PrintData(df)
data = QueueInput(df)
SynTexTrainer(data, AdaptiveSynTex(args), n_gpu).train_with_defaults(
callbacks=[
PeriodicTrigger(ModelSaver(), every_k_epochs=save_epoch),
PeriodicTrigger(ModelSaver(),
every_k_epochs=end_epoch), # save model at last
ScheduledHyperParamSetter('learning_rate', [(start_dec_epoch, lr),
(max_epoch, 0)],
interp="linear"),
#PeriodicTrigger(VisualizeTestSet(data_folder, image_size), every_k_epochs=10),
MergeAllSummaries(period=scalar_steps), # scalar only
MergeAllSummaries(period=image_steps, key="image_summaries"),
],
max_epoch=end_epoch,
steps_per_epoch=steps_per_epoch,
session_init=None)
vlen, nviews = Ppy.shape[-1], Ppy.shape[0]
os.environ['CUDA_VISIBLE_DEVICES'] = get_visible_device_list(3)
global_step = get_global_step_var()
# set logger directory for checkpoints, etc
logger.set_logger_dir(args.logdir, action='k')
steps_per_epoch = cfg.EPOCH_STEPS
model = Model(vlen, nviews)
# config.gpu_options.allow_growth = True
traincfg = TrainConfig(
model=model,
data=QueueInput(ProjDataFlow(Ppy)),
callbacks=[
PeriodicTrigger(ModelSaver(), every_k_epochs=5),
PeriodicTrigger(VolumeSaver(model), every_k_epochs=5),
# prevent learning in the first epoch
# MemInitHyperParamSetter('learning_rate_mask',(0,1)),
# controls learning rate as a function of epoch
HyperParamSetterWithFunc('learning_rate', learning_rate_fun),
# GraphProfiler()
# PeakMemoryTracker()
# GPUUtilizationTracker(),
],
steps_per_epoch=steps_per_epoch,
max_epoch=200000,
# first time load model from checkpoint and reset GRU state
session_init=ChainInit([TryResumeTraining()]), #,ResetInit(model)])
# session_config=tf.ConfigProto(log_device_placement=True) #config_gpus(1)
)
def train(checkpoint_dir,
model_name,
dataset,
num_epochs,
quant_type,
batch_size_per_gpu,
lr=None,
post_quantize_only=False):
train_data, test_data, (img_shape,
label_shape) = datasets.DATASETS[dataset]()
num_gpus = max(gpu.get_num_gpu(), 1)
effective_batch_size = batch_size_per_gpu * num_gpus
train_data = BatchData(train_data, batch_size_per_gpu)
test_data = BatchData(test_data, batch_size_per_gpu, remainder=True)
steps_per_epoch = len(train_data) // num_gpus
if lr:
if isinstance(lr, str):
lr = ast.literal_eval(lr)
if isinstance(lr, float):
lr_schedule = [(0, lr)]
else:
lr_schedule = lr
else:
lr_schedule = [(0, 0.005), (8, 0.1), (25, 0.005), (30, 0)]
if num_epochs is None:
num_epochs = lr_schedule[-1][0]
if post_quantize_only:
start_quantising_at_epoch = 0
else:
start_quantising_at_epoch = lr_schedule[-2][0] if len(
lr_schedule) > 1 else max(0, num_epochs - 5)
logger.info(f"Training with LR schedule: {str(lr_schedule)}")
logger.info(f"Quantising at epoch {start_quantising_at_epoch}")
# train_data = FakeData([(batch_size_per_gpu,) + img_shape, (batch_size_per_gpu, ) + label_shape])
model_func, input_spec, output_spec = get_model_func(
"train",
model_name,
quant_type,
img_shape,
num_classes=label_shape[0],
quant_delay=steps_per_epoch * start_quantising_at_epoch)
target_spec = [
tf.TensorSpec(t.shape, t.dtype, name=t.name.split("/")[-1] + "_target")
for t in output_spec
]
model = KerasModel(get_model=model_func,
input_signature=input_spec,
target_signature=target_spec,
input=train_data,
trainer=SyncMultiGPUTrainerParameterServer(
num_gpus, ps_device='gpu'))
lr = tf.get_variable('learning_rate',
initializer=lr_schedule[0][1],
trainable=False)
tf.summary.scalar('learning_rate-summary', lr)
model.compile(optimizer=tf.train.MomentumOptimizer(learning_rate=lr,
momentum=0.9),
loss="categorical_crossentropy",
metrics=["categorical_accuracy"])
model.fit(steps_per_epoch=steps_per_epoch,
max_epoch=num_epochs,
callbacks=[
ModelSaver(max_to_keep=1, checkpoint_dir=checkpoint_dir),
DataParallelInferenceRunner(
test_data, ScalarStats(model._stats_to_inference),
num_gpus),
ScheduledHyperParamSetter('learning_rate',
lr_schedule,
interp="linear"),
StatMonitorParamSetter('learning_rate',
'validation_categorical_accuracy',
lambda x: x / 2,
threshold=0.001,
last_k=10,
reverse=True)
],
session_init=SaverRestore(checkpoint_dir + "/checkpoint")
if post_quantize_only else None)
if save_dir is None:
logger.auto_set_dir()
else:
logger.set_logger_dir(save_dir)
dataset_train = get_data('train')
dataset_test = get_data('test')
config = TrainConfig(
model=CifarResNet(n=NUM_UNITS,
mult_decay=mult_decay,
lr_init=lr_base * 0.1),
dataflow=dataset_train,
callbacks=[
ModelSaver(),
InferenceRunner(
dataset_test,
[ScalarStats('cost'),
ClassificationError('wrong_vector')]),
ScheduledHyperParamSetter('learning_rate',
[(1, lr_base), (82, lr_base * 0.1),
(123, lr_base * 0.01),
(164, lr_base * 0.002)])
],
max_epoch=200,
session_init=SmartInit(args.load),
)
num_gpu = max(get_num_gpu(), 1)
launch_train_with_config(config,
SyncMultiGPUTrainerParameterServer(num_gpu))
steps_per_epoch = args.get("steps_per_epoch", 1000) / max(n_gpu, 1)
# lr starts decreasing at half of max epoch
start_dec_epoch = max_epoch // 2
# stops when lr is 0.01 of its initial value
end_epoch = max_epoch - int((max_epoch - start_dec_epoch) * 0.01)
if save_folder == None:
logger.auto_set_dir()
else:
logger.set_logger_dir(save_folder)
df = get_data(data_folder, image_size)
df = PrintData(df)
data = QueueInput(df)
SynTexTrainer(data, ProgressiveSynTex(args), n_gpu).train_with_defaults(
callbacks=[
PeriodicTrigger(ModelSaver(), every_k_epochs=save_epoch),
PeriodicTrigger(ModelSaver(), every_k_epochs=end_epoch), # save model at last
ScheduledHyperParamSetter(
'learning_rate',
[(start_dec_epoch, lr), (max_epoch, 0)], interp="linear"),
#PeriodicTrigger(VisualizeTestSet(data_folder, image_size), every_k_epochs=10),
MergeAllSummaries(period=10), # scalar only
MergeAllSummaries(period=image_steps, key="image_summaries"),
],
max_epoch= end_epoch,
steps_per_epoch=steps_per_epoch,
session_init=None
)
