Skip to content

Commit

Permalink
fix tests
Browse files Browse the repository at this point in the history
  • Loading branch information
@wolny
wolny committed Dec 4, 2023
1 parent 5f50cb0 commit 42b6549
Show file tree
Hide file tree
Showing 2 changed files with 54 additions and 61 deletions.
4 changes: 2 additions & 2 deletions pytorch3dunet/unet3d/buildingblocks.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -349,10 +349,10 @@ def __init__(self, in_channels, out_channels, conv_kernel_size=3, scale_factor=2
if upsample is not None and upsample != 'none':
if upsample == 'default':
if basic_module == DoubleConv:
upsample = 'nearest' # use nearest neighbot interpolation for upsampling
upsample = 'nearest' # use nearest neighbor interpolation for upsampling
concat = True # use concat joining
adapt_channels = False # don't adapt channels
elif basic_module == ResNetBlock:
elif basic_module == ResNetBlock or basic_module == ResNetBlockSE:
upsample = 'deconv' # use deconvolution upsampling
concat = False # use summation joining
adapt_channels = True # adapt channels after joining
Expand Down
111 changes: 52 additions & 59 deletions pytorch3dunet/unet3d/utils.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -273,82 +273,75 @@ def _to_numpy(i):


def create_optimizer(optimizer_config, model):
optim_name = optimizer_config['name']

optim_name = optimizer_config.get('name', 'Adam')
# common optimizer settings
learning_rate = optimizer_config['learning_rate']
learning_rate = optimizer_config.get('learning_rate', 1e-3)
weight_decay = optimizer_config.get('weight_decay', 0)

# grab optimizer specific settings and init
# optimizer
if optim_name == 'Adadelta':
rho = optimizer_config.get('rho', 0.9)
optimizer = optim.Adadelta(model.parameters(), lr=learning_rate, rho=rho,
weight_decay=weight_decay)
rho = optimizer_config.get('rho', 0.9)
optimizer = optim.Adadelta(model.parameters(), lr=learning_rate, rho=rho,
weight_decay=weight_decay)
elif optim_name == 'Adagrad':
lr_decay = optimizer_config.get('lr_decay', 0)
optimizer = optim.Adagrad(model.parameters(), lr=learning_rate, lr_decay=lr_decay,
weight_decay=weight_decay)
lr_decay = optimizer_config.get('lr_decay', 0)
optimizer = optim.Adagrad(model.parameters(), lr=learning_rate, lr_decay=lr_decay,
weight_decay=weight_decay)
elif optim_name == 'AdamW':
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
optimizer = optim.AdamW(model.parameters(), lr=learning_rate, betas=betas,
weight_decay=weight_decay)
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
optimizer = optim.AdamW(model.parameters(), lr=learning_rate, betas=betas,
weight_decay=weight_decay)
elif optim_name == 'SparseAdam':
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
optimizer = optim.SparseAdam(model.parameters(), lr=learning_rate, betas=betas,
weight_decay=weight_decay)
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
optimizer = optim.SparseAdam(model.parameters(), lr=learning_rate, betas=betas)
elif optim_name == 'Adamax':
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
optimizer = optim.Adamax(model.parameters(), lr=learning_rate, betas=betas,
weight_decay=weight_decay)
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
optimizer = optim.Adamax(model.parameters(), lr=learning_rate, betas=betas,
weight_decay=weight_decay)
elif optim_name == 'ASGD':
lambd = optimizer_config.get('lambd', 0.0001)
alpha = optimizer_config.get('alpha', 0.75)
t0 = optimizer_config.get('t0', 1e6)
optimizer = optim.Adamax(model.parameters(), lr=learning_rate, lambd=lambd,
alpha=alpha, t0=t0, weight_decay=weight_decay)
lambd = optimizer_config.get('lambd', 0.0001)
alpha = optimizer_config.get('alpha', 0.75)
t0 = optimizer_config.get('t0', 1e6)
optimizer = optim.Adamax(model.parameters(), lr=learning_rate, lambd=lambd,
alpha=alpha, t0=t0, weight_decay=weight_decay)
elif optim_name == 'LBFGS':
max_iter = optimizer_config.get('max_iter', 20)
max_eval = optimizer_config.get('max_eval', None)
tolerance_grad = optimizer.config.get('tolerance_grad', 1e-7)
tolerance_change = optimizer.config.get('tolerance_change', 1e-9)
history_size = optimizer.config.get('history_size', 100)
optimizer = optim.LBFGS(model.parameters(), lr=learning_rate, max_iter=max_iter,
max_eval=max_eval, tolerance_grad=tolerance_grad,
tolerance_change=tolerance_change, history_size=history_size,
weight_decay=weight_decay)
max_iter = optimizer_config.get('max_iter', 20)
max_eval = optimizer_config.get('max_eval', None)
tolerance_grad = optimizer_config.get('tolerance_grad', 1e-7)
tolerance_change = optimizer_config.get('tolerance_change', 1e-9)
history_size = optimizer_config.get('history_size', 100)
optimizer = optim.LBFGS(model.parameters(), lr=learning_rate, max_iter=max_iter,
max_eval=max_eval, tolerance_grad=tolerance_grad,
tolerance_change=tolerance_change, history_size=history_size)
elif optim_name == 'NAdam':
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
momentum_decay = optimizer_config.get('momentum_decay', 4e-3)
decoupled_weight_decay = optimizer_config.get('decoupled_weight_decay', False)
optimizer = optim.NAdam(model.parameters(), lr=learning_rate, betas=betas,
momentum_decay=momentum_decay,
decoupled_weight_decay=decoupled_weight_decay,
weight_decay=weight_decay)
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
momentum_decay = optimizer_config.get('momentum_decay', 4e-3)
optimizer = optim.NAdam(model.parameters(), lr=learning_rate, betas=betas,
momentum_decay=momentum_decay,
weight_decay=weight_decay)
elif optim_name == 'RAdam':
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
optimizer = optim.RAdam(model.parameters(), lr=learning_rate, betas=betas,
weight_decay=weight_decay)
elif optim_name == 'RMSprop':
alpha = optimizer_config.get('alpha', 0.99)
optimizer = optim.RMSprop(model.parameters(), lr=learning_rate, alpha=alpha,
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
optimizer = optim.RAdam(model.parameters(), lr=learning_rate, betas=betas,
weight_decay=weight_decay)
elif optim_name == 'RMSprop':
alpha = optimizer_config.get('alpha', 0.99)
optimizer = optim.RMSprop(model.parameters(), lr=learning_rate, alpha=alpha,
weight_decay=weight_decay)
elif optim_name == 'Rprop':
etas = tuple(optimizer_config.get('etas', (0.5, 1.2)))
step_sizes = tuple(optimizer_config.get('step_sizes', (1e-6, 50)))
optimizer = optim.RMSprop(model.parameters(), lr=learning_rate, etas=etas,
step_sizes=step_sizes, weight_decay=weight_decay)
momentum = optimizer_config.get('momentum', 0)
optimizer = optim.RMSprop(model.parameters(), lr=learning_rate, weight_decay=weight_decay, momentum=momentum)
elif optim_name == 'SGD':
momentum = optimizer_config.get('momentum', 0)
dampening = optimizer_config.get('dampening', 0)
nesterov = optimizer_config.get('nesterov', False)
optimizer = optim.SGD(model.parameters(), lr=learning_rate, momentum=momentum,
dampening=dampening, nesterov=nesterov,
weight_decay=weight_decay)
else: # Adam is default
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
optimizer = optim.Adam(model.parameters(), lr=learning_rate, betas=betas,
weight_decay=weight_decay)
momentum = optimizer_config.get('momentum', 0)
dampening = optimizer_config.get('dampening', 0)
nesterov = optimizer_config.get('nesterov', False)
optimizer = optim.SGD(model.parameters(), lr=learning_rate, momentum=momentum,
dampening=dampening, nesterov=nesterov,
weight_decay=weight_decay)
else: # Adam is default
betas = tuple(optimizer_config.get('betas', (0.9, 0.999)))
optimizer = optim.Adam(model.parameters(), lr=learning_rate, betas=betas,
weight_decay=weight_decay)

return optimizer

Expand Down

0 comments on commit 42b6549

Please sign in to comment.