def get_learner(nb_classes, all=False, args=ARGS):
model = _get_model(nb_classes, args=args)
optimizer = torch.optim.Adam(model.parameters(), lr=0.0001)
from src.metrics import pytorch_metrics
criterion = pytorch_metrics.RMSELoss()
metrics = [pytorch_metrics.RMSELoss(round=True), ]
from torch.optim.lr_scheduler import ReduceLROnPlateau
if all:
from src.pytorch.schedulers import StepLR
scheduler = StepLR(optimizer, step_size=args.all_train_lr_step, gamma=0.1, min_lr=1e-5)
from src.pytorch.early_stoppers import EarlyStopping
early_stopper = None
else:
scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=2, verbose=True, threshold=0.001,
min_lr=1e-5)
from src.pytorch.early_stoppers import EarlyStopping
early_stopper = EarlyStopping(patience=4, verbose=True, delta=0.001, save_model_path=None, wait=1)
from src.pytorch.wrappers import PyTorchNN_vA as PyTorchNN
pytorchmodel = PyTorchNN(model, optimizer, criterion, metrics, scheduler, early_stopper, device=DEVICE,
virtual_batch_size=args.virtual_batch_size)
return pytorchmodel
def get_learner(nb_classes, args=ARGS):
model = _get_model(nb_classes, args=args)
optimizer = torch.optim.Adam(model.parameters(), lr=0.0001)
criterion = nn.CrossEntropyLoss()
from src.metrics import pytorch_metrics
metrics = [
pytorch_metrics.CategoricalAccuracy(),
pytorch_metrics.TopKAccuracy(top_k=2)
]
from src.pytorch.schedulers import StepLR
scheduler = StepLR(optimizer, step_size=8, gamma=0.1, min_lr=1e-5)
from src.pytorch.early_stoppers import EarlyStopping
early_stopper = EarlyStopping(patience=6,
verbose=True,
delta=0.0001,
save_model_path=None,
wait=12)
from src.pytorch.wrappers import PyTorchNN_vA as PyTorchNN
pytorchmodel = PyTorchNN(model,
optimizer,
criterion,
metrics,
scheduler,
early_stopper,
device=DEVICE,
virtual_batch_size=args.virtual_batch_size)
return pytorchmodel
def get_learner(nb_classes, args=ARGS):
model = _get_model(nb_classes, args=args)
optimizer = torch.optim.Adam(model.parameters(), lr=0.0001)
if args.outputs_as_onehotencoder:
criterion = nn.BCEWithLogitsLoss()
from src.metrics import pytorch_metrics
metrics = [
pytorch_metrics.CrossEntropyLoss(one_hot_encoding=True),
]
else:
criterion = nn.CrossEntropyLoss()
from src.metrics import pytorch_metrics
metrics = [
pytorch_metrics.CategoricalAccuracy(),
pytorch_metrics.TopKAccuracy(top_k=2)
]
from torch.optim.lr_scheduler import ReduceLROnPlateau
scheduler = ReduceLROnPlateau(optimizer,
mode='min',
factor=0.1,
patience=3,
verbose=True,
threshold=0.001,
min_lr=1e-5)
from src.pytorch.early_stoppers import EarlyStopping
early_stopper = EarlyStopping(patience=6,
verbose=True,
delta=0.0001,
save_model_path=None)
from src.pytorch.wrappers import PyTorchNN_vA as PyTorchNN
pytorchmodel = PyTorchNN(model,
optimizer,
criterion,
metrics,
scheduler,
early_stopper,
device=DEVICE,
virtual_batch_size=args.virtual_batch_size,
mixup=args.mixup,
mixup_alpha=args.mixup_alpha,
mixup_method=args.mixup_method)
return pytorchmodel
def load_model(filename, nb_classes=None, load_learner=False):
print("-" * 80)
print("LOAD MODEL")
# Loading model
print(f"Loading model from file: {filename}")
checkpoint = torch.load(filename)
# Generate model
model = _get_model(nb_classes=nb_classes)
model.load_state_dict(checkpoint['model_state_dict'])
# Load learner
if load_learner:
pytorchmodel = get_learner(model=model)
else:
from src.pytorch.wrappers import PyTorchNN_vA as PyTorchNN
pytorchmodel = PyTorchNN(model, device=DEVICE)
return pytorchmodel
