).to(device)
if times:
# OBSOLETE: DO NOT WORK ANYMORE, WILL CAUSE CRASHES TODO
# overrides default mode !
# tests different values of workers and batch sizes to check which is the fastest
num_workers = [16, 32, 64, 128]
batch_sizes = [16, 32]
perfs = []
for nw, bs in product(num_workers, batch_sizes):
tl, vl, _ = create_loaders(
data_path,
train_size,
bs,
max_subj,
ch_type,
data_type,
num_workers=nw,
debug=debug,
chunkload=chunkload,
permute_labels=permute_labels,
)
tpb, et = train(net, tl, vl, "", lr=learning_rate, timing=True)
perfs.append((nw, bs, tpb, et))
for x in sorted(perfs, key=lambda x: x[-1]):
logging.info(f"\n{x[0]} {x[1]} {nt(x[2])} {nt(x[3])}")
else:
# We create loaders and datasets (see dataloaders.py)
trainloader, validloader, testloader = create_loaders(
logging.info(net)
file_exists = False
if os.path.exists(save_path + f"Guided_Bprop_{net_name}.mat"):
logging.warning("GuidedBackprop for this architecture already exists")
file_exists = True
if mode == "overwrite" or not file_exists:
# We create loaders and datasets (see dataloaders.py)
_, validloader, _ = create_loaders(
data_path,
train_size,
batch_size,
max_subj,
ch_type,
data_type,
seed=seed,
num_workers=num_workers,
chunkload=chunkload,
include=(0, 1, 0),
ages=ages,
)
model_filepath = save_path + f"{net_name}.pt"
logging.info(net.name)
_, net_state, _ = load_checkpoint(model_filepath)
net.load_state_dict(net_state)
gbp = GuidedBackprop(net)
gbps = []
).to(device)
lin_size = input_size[0] * input_size[1] * input_size[2]
if torchsum:
logging.info(summary(net, (1, lin_size)))
else:
logging.info(net)
# We create loaders and datasets (see dataloaders.py)
trainloader, validloader, testloader = create_loaders(
data_path,
train_size,
batch_size,
max_subj,
ch_type,
data_type,
seed=seed,
num_workers=num_workers,
chunkload=chunkload,
debug=debug,
printmem=printmem,
)
if mode == "overwrite":
save = True
load = False
elif mode in ("continue", "evaluate"):
save = True
load = True
else:
save = False
# num_workers=num_workers,
# chunkload=chunkload,
# debug=debug,
# printmem=printmem,
# include=(1, 1, 0),
# ages=ages,
# dattype="task",
# )
rest_train, rest_valid, rest_test = create_loaders(
data_path,
train_size,
batch_size,
max_subj,
ch_type,
data_type,
seed=seed,
num_workers=num_workers,
chunkload=chunkload,
debug=debug,
printmem=printmem,
include=(1, 1, 0),
ages=ages,
)
# Testing wether ERM works as intended:
train_loaders = [rest_train, rest_train]
valid_loaders = [rest_valid, rest_valid]
if torchsum:
logging.info(summary(net, input_size))
else:
logging.info(net)
nchan,
linear,
dropout,
dropout_option,
).to(device)
# We create loaders and datasets (see dataloaders.py)
task_train, task_valid, _ = create_loaders(
data_path,
train_size,
batch_size,
max_subj,
ch_type,
data_type,
seed=seed,
num_workers=num_workers,
chunkload=chunkload,
debug=debug,
printmem=printmem,
include=(1, 1, 0),
ages=ages,
dattype="task",
samples=samples,
infinite=True,
)
rest_train, rest_valid, _ = create_loaders(
data_path,
train_size,
batch_size,
max_subj,
ch_type,
data_type,
dropout=dropout,
).to(device)
if times:
# overrides default mode !
# tests different values of workers and batch sizes to check which is the fastest
num_workers = [16, 32, 64, 128]
batch_sizes = [16, 32]
perfs = []
for nw, bs in product(num_workers, batch_sizes):
tl, vl, _ = create_loaders(
data_path,
train_size,
bs,
max_subj,
ch_type,
data_type,
num_workers=nw,
debug=debug,
chunkload=chunkload,
)
tpb, et = train(net, tl, vl, "", lr=learning_rate, timing=True)
perfs.append((nw, bs, tpb, et))
for x in sorted(perfs, key=lambda x: x[-1]):
print("\n", (x[0], x[1], nt(x[2]), nt(x[3])))
else:
print(net)
# We create loaders and datasets (see dataloaders.py)