def train_model():
"""
Trains a CIANet model on the following params.
"""
epochs = 100
learning_rate = 0.001
momentum = 0.9
expand_by_k = 5
iter_to_change = 30
err_stop_threshold = 0.99
# sizes = {"encoder": [28 * 28, 50, 50, 10],
# sizes = {"encoder": [28 * 28, 20, 20, 15, 15, 10, 10, 10],
# "action": [10, 10]}
sizes = {"classifier": [28 * 28, 312, 128, 10]}
drift_thresholds = {
"classifier": [0.45, 0.4, 10]
} # Drift threshold for split in DENz
drift_deltas = {"classifier": [0, 0, 10]}
trainer = DENTrainer(data_loaders, FF, sizes, learning_rate, momentum,
criterion, penalty, iter_to_change, device,
error_function, number_of_tasks, drift_thresholds,
err_stop_threshold, drift_deltas)
results = trainer.train_all_tasks_sequentially(epochs, with_den=True)
loss, err = trainer.test_model(list(range(number_of_tasks)), False)[0]
print("Net has final shape:" + str(trainer.model.sizes))
print("Done training with total net accuracy:" + str(err))
print("Done training with results from error function:" + str(results))
return trainer.model, results
def train_model():
"""
Trains a CIANet model on the following params.
"""
epochs = 1000
learning_rate = 0.001
momentum = 0.9
iter_to_change = 50
err_stop_threshold = 0.99
# sizes = {"classifier": [3072, 120, 60, 10],
sizes = {"encoder": [32 * 32 * 3, 100, 50, 100], "action": [100, 100]}
# }
trainer = DENTrainer(data_loaders, ActionEncoder, sizes, learning_rate,
momentum, criterion, penalty, iter_to_change, device,
error_function, number_of_tasks, drift_threshold,
err_stop_threshold)
results = trainer.train_all_tasks_sequentially(epochs, with_den=True)
loss, err = trainer.test_model(list(range(number_of_tasks)), False)[0]
print("Net has final shape:" + str(trainer.model.sizes))
print("Done training with total net accuracy:" + str(err))
print("Done training with results from error function:" + str(results))
return trainer.model, results
def train_model():
"""
Trains a CIANet model on the following params.
"""
epochs = 30
learning_rate = 0.05
momentum = 0.9
expand_by_k = 5
err_stop_threshold = 0.99
sizes = {
"classifier": [3072, 120, 60, 10],
# sizes = {"encoder": [28 * 28, 20, 20, 15, 15, 10, 10, 10],
# "action": [10, 10]}
}
trainer = DENTrainer(data_loaders, FFConv, sizes, learning_rate, momentum,
criterion, penalty, expand_by_k, device,
error_function, number_of_tasks, drift_threshold,
err_stop_threshold)
results = trainer.train_all_tasks_sequentially(epochs, with_den=True)
loss, err = trainer.test_model(list(range(number_of_tasks)), False)[0]
print("Net has final shape:" + str(trainer.model.sizes))
print("Done training with total net accuracy:" + str(err))
print("Done training with results from error function:" + str(results))
return trainer.model, results
def _setup(self, config):
trainer_args = config.get("DENTrainerArgs") # Type: list
self.trainer = DENTrainer(*trainer_args)
self.trainer.optimizer = torch.optim.SGD(
self.trainer.model.parameters(),
lr=config.get("lr", 0.01),
momentum=config.get("momentum", 0.9))
self.trainer.expand_by_k = config.get("expand_by_k", 10)
self.trainer.drift_threshold = config.get("drift_threshold", 0.02)
if hasattr(self.trainer.penalty, "l1_coeff"):
self.trainer.penalty.l1_coeff = config.get("l1_coeff", 0)
if hasattr(self.trainer.penalty, "l2_coeff"):
self.trainer.penalty.l2_coeff = config.get("l2_coeff", 0)
def train_model():
"""
Trains a CIANet model on the following params.
"""
epochs = 5
learning_rate = 0.002
momentum = 0
expand_by_k = 10
sizes = {
"encoder": [img_size[0] * img_size[1] * 3, 1000, 522],
"action": [522, 80, 2]
}
trainer = DENTrainer(data_loaders, sizes, learning_rate, momentum,
criterion, penalty, expand_by_k, device,
error_function, number_of_tasks, drift_threshold)
print(trainer.model.sizes)
results = trainer.train_all_tasks_sequentially(epochs, with_den=True)
print("Done training with results from error function:" + str(results))
trainer.save_model("banana_car.pt")
return trainer.model, results
class PytorchTrainable(tune.Trainable):
"""
Train a Pytorch net with Trainable and PopulationBasedTraining scheduler.
Finds hypers for one task.
"""
def _setup(self, config):
trainer_args = config.get("DENTrainerArgs") # Type: list
self.trainer = DENTrainer(*trainer_args)
self.trainer.optimizer = torch.optim.SGD(
self.trainer.model.parameters(),
lr=config.get("lr", 0.01),
momentum=config.get("momentum", 0.9))
self.trainer.expand_by_k = config.get("expand_by_k", 10)
self.trainer.drift_threshold = config.get("drift_threshold", 0.02)
if hasattr(self.trainer.penalty, "l1_coeff"):
self.trainer.penalty.l1_coeff = config.get("l1_coeff", 0)
if hasattr(self.trainer.penalty, "l2_coeff"):
self.trainer.penalty.l2_coeff = config.get("l2_coeff", 0)
def _train(self):
# Do one epoch for all tasks
for i in range(self.trainer.number_of_tasks):
self.trainer.train_tasks([i], 5, (not i == 0))
loss_err_list = self.trainer.test_model(
tasks=list(range(self.trainer.number_of_tasks)))
mean_err = 0
mean_loss = 0
for loss, err in loss_err_list:
mean_err += err
mean_loss += loss
mean_loss /= len(loss_err_list)
mean_err /= len(loss_err_list)
return {"mean_accuracy": mean_err, "mean_loss": mean_loss}
def _save(self, checkpoint_dir):
return self.trainer.save_model("model.pth", checkpoint_dir)
def _restore(self, checkpoint_path):
self.trainer.load_model(checkpoint_path)
def _export_model(self, export_formats, export_dir):
if export_formats == [ExportFormat.MODEL]:
path = os.path.join(export_dir, "exported_actionnet.pt")
torch.save(self.trainer.model.state_dict(), path)
return {export_formats[0]: path}
raise ValueError("unexpected formats: " + str(export_formats))
def reset_config(self, new_config):
for param_group in self.trainer.optimizer.param_groups:
if "lr" in new_config:
param_group["lr"] = new_config["lr"]
if "momentum" in new_config:
param_group["momentum"] = new_config["momentum"]
self.config = new_config
return True