def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device='cuda',
loss_function='ce',
add_noise=False,
noise_type='Gaussian',
noise_std=0.0,
loss_function_param=None,
**kwargs):
super(StandardClassifierWithNoise, self).__init__(**kwargs)
self.args = None # this will be modified by the decorator
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.loss_function = loss_function
self.add_noise = add_noise
self.noise_type = noise_type
self.noise_std = noise_std
self.loss_function_param = loss_function_param
# initialize the network
self.repr_net = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, device)
self.repr_shape = self.repr_net.output_shape
self.classifier, output_shape = nn_utils.parse_network_from_config(
args=self.architecture_args['classifier'],
input_shape=self.repr_shape)
self.num_classes = output_shape[-1]
self.classifier = self.classifier.to(device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device='cuda',
grad_weight_decay=0.0,
grad_l1_penalty=0.0,
lamb=1.0,
small_qtop=False,
sample_from_q=False,
**kwargs):
super(PredictGradOutputFixedFormWithConfusion, self).__init__(**kwargs)
self.args = None # this will be modified by the decorator
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.grad_weight_decay = grad_weight_decay
self.grad_l1_penalty = grad_l1_penalty
self.lamb = lamb
self.small_qtop = small_qtop
self.sample_from_q = sample_from_q
self.grad_replacement_class = nn_utils.get_grad_replacement_class(
sample=self.sample_from_q,
standard_dev=np.sqrt(1.0 / 2.0 / (self.lamb + 1e-12)))
# initialize the network
self.classifier, output_shape = nn_utils.parse_network_from_config(
args=self.architecture_args['classifier'],
input_shape=self.input_shape)
self.classifier = self.classifier.to(device)
self.num_classes = output_shape[-1]
if self.pretrained_arg is not None:
self.q_base = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, device)
q_base_shape = self.q_base.output_shape
else:
self.q_base, q_base_shape = nn_utils.parse_network_from_config(
args=self.architecture_args['q-base'],
input_shape=self.input_shape)
self.q_base = self.q_base.to(device)
if small_qtop:
self.q_top = torch.nn.Linear(q_base_shape[-1],
self.num_classes).to(device)
else:
self.q_top = torch.nn.Sequential(
torch.nn.Linear(q_base_shape[-1], 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes)).to(device)
# the confusion matrix trainable logits, (true, observed)
Q_init = torch.zeros(size=(self.num_classes, self.num_classes),
device=device,
dtype=torch.float)
Q_init += 1.0 * torch.eye(self.num_classes,
device=device,
dtype=torch.float) # TODO: tune the constant
self.Q_logits = torch.nn.Parameter(Q_init, requires_grad=True)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device='cuda',
loss_function='ce',
add_noise=False,
noise_type='Gaussian',
noise_std=0.0,
loss_function_param=None,
load_from=None,
**kwargs):
super(StandardClassifier, self).__init__(**kwargs)
self.args = {
'input_shape': input_shape,
'architecture_args': architecture_args,
'pretrained_arg': pretrained_arg,
'device': device,
'loss_function': loss_function,
'add_noise': add_noise,
'noise_type': noise_type,
'noise_std': noise_std,
'loss_function_param': loss_function_param,
'load_from': load_from,
'class': 'StandardClassifier'
}
assert len(input_shape) == 3
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.device = device
self.loss_function = loss_function
self.add_noise = add_noise
self.noise_type = noise_type
self.noise_std = noise_std
self.loss_function_param = loss_function_param
self.load_from = load_from
# initialize the network
self.repr_net = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, self.device)
self.repr_shape = self.repr_net.output_shape
self.classifier, output_shape = nn_utils.parse_feed_forward(
args=self.architecture_args['classifier'],
input_shape=self.repr_shape)
self.num_classes = output_shape[-1]
self.classifier = self.classifier.to(self.device)
self.grad_noise_class = nn_utils.get_grad_noise_class(
standard_dev=noise_std, q_dist=noise_type)
if self.load_from is not None:
print("Loading the classifier model from {}".format(load_from))
stored_net = utils.load(load_from, device='cpu')
stored_net_params = dict(stored_net.classifier.named_parameters())
for key, param in self.classifier.named_parameters():
param.data = stored_net_params[key].data.to(self.device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device="cuda",
grad_weight_decay=0.0,
grad_l1_penalty=0.0,
lamb=1.0,
**kwargs):
super(PredictGradOutputGeneralFormUseLabel, self).__init__(**kwargs)
self.args = {
"input_shape": input_shape,
"architecture_args": architecture_args,
"device": device,
"pretrained_arg": pretrained_arg,
"grad_weight_decay": grad_weight_decay,
"grad_l1_penalty": grad_l1_penalty,
"lamb": lamb,
"class": "PredictGradOutputGeneralFormUseLabel",
}
assert len(input_shape) == 3
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.device = device
self.grad_weight_decay = grad_weight_decay
self.grad_l1_penalty = grad_l1_penalty
self.lamb = lamb
# initialize the network
self.classifier, _ = nn_utils.parse_feed_forward(
args=self.architecture_args["classifier"],
input_shape=self.input_shape)
self.classifier = self.classifier.to(self.device)
self.num_classes = self.architecture_args["classifier"][-1]["dim"]
if self.pretrained_arg is not None:
self.q_base = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, self.device)
q_base_shape = self.q_base.output_shape
else:
self.q_base, q_base_shape = nn_utils.parse_feed_forward(
args=self.architecture_args["q-base"],
input_shape=self.input_shape)
self.q_base = self.q_base.to(self.device)
# NOTE: we want to use classifier parameters too
# TODO: find a good parametrization
self.q_top = torch.nn.Sequential(
torch.nn.Linear(q_base_shape[-1] + 2 * self.num_classes, 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes),
).to(self.device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device='cuda',
grad_weight_decay=0.0,
grad_l1_penalty=0.0,
lamb=1.0,
**kwargs):
super(PredictGradOutputGeneralFormUseLabel, self).__init__(**kwargs)
self.args = {
'input_shape': input_shape,
'architecture_args': architecture_args,
'device': device,
'pretrained_arg': pretrained_arg,
'grad_weight_decay': grad_weight_decay,
'grad_l1_penalty': grad_l1_penalty,
'lamb': lamb,
'class': 'PredictGradOutputGeneralFormUseLabel'
}
assert len(input_shape) == 3
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.device = device
self.grad_weight_decay = grad_weight_decay
self.grad_l1_penalty = grad_l1_penalty
self.lamb = lamb
# initialize the network
self.classifier, _ = nn_utils.parse_feed_forward(
args=self.architecture_args['classifier'],
input_shape=self.input_shape)
self.classifier = self.classifier.to(self.device)
self.num_classes = self.architecture_args['classifier'][-1]['dim']
if self.pretrained_arg is not None:
self.q_base = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, self.device)
q_base_shape = self.q_base.output_shape
else:
self.q_base, q_base_shape = nn_utils.parse_feed_forward(
args=self.architecture_args['q-base'],
input_shape=self.input_shape)
self.q_base = self.q_base.to(self.device)
# NOTE: we want to use classifier parameters too
# TODO: find a good parametrization
self.q_top = torch.nn.Sequential(
torch.nn.Linear(q_base_shape[-1] + 2 * self.num_classes, 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes)).to(self.device)
def __init__(
self,
input_shape,
architecture_args,
pretrained_arg=None,
device="cuda",
loss_function="ce",
add_noise=False,
noise_type="Gaussian",
noise_std=0.0,
loss_function_param=None,
**kwargs
):
super(StandardClassifierWithNoise, self).__init__(**kwargs)
self.args = {
"input_shape": input_shape,
"architecture_args": architecture_args,
"pretrained_arg": pretrained_arg,
"device": device,
"loss_function": loss_function,
"add_noise": add_noise,
"noise_type": noise_type,
"noise_std": noise_std,
"loss_function_param": loss_function_param,
"class": "StandardClassifierWithNoise",
}
assert len(input_shape) == 3
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.device = device
self.loss_function = loss_function
self.add_noise = add_noise
self.noise_type = noise_type
self.noise_std = noise_std
self.loss_function_param = loss_function_param
# initialize the network
self.repr_net = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, self.device
)
self.repr_shape = self.repr_net.output_shape
self.classifier, output_shape = nn_utils.parse_feed_forward(
args=self.architecture_args["classifier"], input_shape=self.repr_shape
)
self.num_classes = output_shape[-1]
self.classifier = self.classifier.to(self.device)
def __init__(self, input_shape, architecture_args, pretrained_arg=None, device='cuda',
grad_weight_decay=0.0, grad_l1_penalty=0.0, lamb=1.0, **kwargs):
super(PenalizeLastLayerFixedForm, self).__init__(**kwargs)
self.args = {
'input_shape': input_shape,
'architecture_args': architecture_args,
'pretrained_arg': pretrained_arg,
'device': device,
'grad_weight_decay': grad_weight_decay,
'grad_l1_penalty': grad_l1_penalty,
'lamb': lamb,
'class': 'PenalizeLastLayerFixedForm'
}
assert len(input_shape) == 3
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.device = device
self.grad_weight_decay = grad_weight_decay
self.grad_l1_penalty = grad_l1_penalty
self.lamb = lamb
# initialize the network
self.num_classes = self.architecture_args['classifier'][-1]['dim']
self.last_layer_dim = self.architecture_args['classifier'][-2]['dim']
self.classifier_base, _ = nn_utils.parse_feed_forward(args=self.architecture_args['classifier'][:-1],
input_shape=self.input_shape)
self.classifier_base = self.classifier_base.to(self.device)
self.classifier_last_layer = torch.nn.Linear(self.last_layer_dim,
self.num_classes,
bias=False).to(self.device)
if self.pretrained_arg is not None:
q_base = pretrained_models.get_pretrained_model(self.pretrained_arg, self.input_shape, self.device)
# create the trainable part of the q_network
q_top = torch.nn.Sequential(
torch.nn.Linear(q_base.output_shape[-1], 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes)).to(self.device)
self.q_network = torch.nn.Sequential(q_base, q_top)
else:
self.q_network, _ = nn_utils.parse_feed_forward(args=self.architecture_args['q-network'],
input_shape=self.input_shape)
self.q_network = self.q_network.to(self.device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device='cuda',
grad_weight_decay=0.0,
grad_l1_penalty=0.0,
lamb=1.0,
**kwargs):
super(PenalizeLastLayerFixedForm, self).__init__(**kwargs)
self.args = None # this will be modified by the decorator
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.grad_weight_decay = grad_weight_decay
self.grad_l1_penalty = grad_l1_penalty
self.lamb = lamb
# initialize the network
self.num_classes = self.architecture_args['classifier'][-1]['dim']
self.last_layer_dim = self.architecture_args['classifier'][-2]['dim']
self.classifier_base, _ = nn_utils.parse_network_from_config(
args=self.architecture_args['classifier'][:-1],
input_shape=self.input_shape)
self.classifier_base = self.classifier_base.to(device)
self.classifier_last_layer = torch.nn.Linear(self.last_layer_dim,
self.num_classes,
bias=False).to(device)
if self.pretrained_arg is not None:
q_base = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, device)
# create the trainable part of the q_network
q_top = torch.nn.Sequential(
torch.nn.Linear(q_base.output_shape[-1], 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes)).to(device)
self.q_network = torch.nn.Sequential(q_base, q_top)
else:
self.q_network, _ = nn_utils.parse_network_from_config(
args=self.architecture_args['q-network'],
input_shape=self.input_shape)
self.q_network = self.q_network.to(device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device='cuda',
loss_function='ce',
add_noise=False,
noise_type='Gaussian',
noise_std=0.0,
loss_function_param=None,
**kwargs):
super(StandardClassifierWithNoise, self).__init__(**kwargs)
self.args = {
'input_shape': input_shape,
'architecture_args': architecture_args,
'pretrained_arg': pretrained_arg,
'device': device,
'loss_function': loss_function,
'add_noise': add_noise,
'noise_type': noise_type,
'noise_std': noise_std,
'loss_function_param': loss_function_param,
'class': 'StandardClassifierWithNoise'
}
assert len(input_shape) == 3
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.device = device
self.loss_function = loss_function
self.add_noise = add_noise
self.noise_type = noise_type
self.noise_std = noise_std
self.loss_function_param = loss_function_param
# initialize the network
self.repr_net = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, self.device)
self.repr_shape = self.repr_net.output_shape
self.classifier, output_shape = nn_utils.parse_feed_forward(
args=self.architecture_args['classifier'],
input_shape=self.repr_shape)
self.num_classes = output_shape[-1]
self.classifier = self.classifier.to(self.device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device='cuda',
loss_function='ce',
add_noise=False,
noise_type='Gaussian',
noise_std=0.0,
loss_function_param=None,
load_from=None,
**kwargs):
super(StandardClassifier, self).__init__(**kwargs)
self.args = None # this will be modified by the decorator
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.loss_function = loss_function
self.add_noise = add_noise
self.noise_type = noise_type
self.noise_std = noise_std
self.loss_function_param = loss_function_param
self.load_from = load_from
# initialize the network
self.repr_net = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, device)
self.repr_shape = self.repr_net.output_shape
self.classifier, output_shape = nn_utils.parse_network_from_config(
args=self.architecture_args['classifier'],
input_shape=self.repr_shape)
self.num_classes = output_shape[-1]
self.classifier = self.classifier.to(device)
self.grad_noise_class = nn_utils.get_grad_noise_class(
standard_dev=noise_std, q_dist=noise_type)
if self.load_from is not None:
print("Loading the classifier model from {}".format(load_from))
import methods
stored_net = utils.load(load_from, methods=methods, device='cpu')
stored_net_params = dict(stored_net.classifier.named_parameters())
for key, param in self.classifier.named_parameters():
param.data = stored_net_params[key].data.to(device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device='cuda',
grad_weight_decay=0.0,
grad_l1_penalty=0.0,
lamb=1.0,
**kwargs):
super(PredictGradOutputGeneralFormUseLabel, self).__init__(**kwargs)
self.args = None # this will be modified by the decorator
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.grad_weight_decay = grad_weight_decay
self.grad_l1_penalty = grad_l1_penalty
self.lamb = lamb
# initialize the network
self.classifier, _ = nn_utils.parse_network_from_config(
args=self.architecture_args['classifier'],
input_shape=self.input_shape)
self.classifier = self.classifier.to(device)
self.num_classes = self.architecture_args['classifier'][-1]['dim']
if self.pretrained_arg is not None:
self.q_base = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, device)
q_base_shape = self.q_base.output_shape
else:
self.q_base, q_base_shape = nn_utils.parse_network_from_config(
args=self.architecture_args['q-base'],
input_shape=self.input_shape)
self.q_base = self.q_base.to(device)
# NOTE: we want to use classifier parameters too
# TODO: find a good parametrization
self.q_top = torch.nn.Sequential(
torch.nn.Linear(q_base_shape[-1] + 2 * self.num_classes, 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes)).to(device)
def __init__(
self,
input_shape,
architecture_args,
pretrained_arg=None,
device="cuda",
loss_function="ce",
add_noise=False,
noise_type="Gaussian",
noise_std=0.0,
loss_function_param=None,
load_from=None,
**kwargs
):
super(StandardClassifier, self).__init__(**kwargs)
self.args = {
"input_shape": input_shape,
"architecture_args": architecture_args,
"pretrained_arg": pretrained_arg,
"device": device,
"loss_function": loss_function,
"add_noise": add_noise,
"noise_type": noise_type,
"noise_std": noise_std,
"loss_function_param": loss_function_param,
"load_from": load_from,
"class": "StandardClassifier",
}
assert len(input_shape) == 3
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.device = device
self.loss_function = loss_function
self.add_noise = add_noise
self.noise_type = noise_type
self.noise_std = noise_std
self.loss_function_param = loss_function_param
self.load_from = load_from
# initialize the network
self.repr_net = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, self.device
)
self.repr_shape = self.repr_net.output_shape
self.classifier, output_shape = nn_utils.parse_feed_forward(
args=self.architecture_args["classifier"], input_shape=self.repr_shape
)
self.num_classes = output_shape[-1]
self.classifier = self.classifier.to(self.device)
self.grad_noise_class = nn_utils.get_grad_noise_class(
standard_dev=noise_std, q_dist=noise_type
)
if self.load_from is not None:
print("Loading the classifier model from {}".format(load_from))
stored_net = utils.load(load_from, device="cpu")
stored_net_params = dict(stored_net.classifier.named_parameters())
for key, param in self.classifier.named_parameters():
param.data = stored_net_params[key].data.to(self.device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device="cuda",
grad_weight_decay=0.0,
grad_l1_penalty=0.0,
lamb=1.0,
**kwargs):
super(PenalizeLastLayerFixedForm, self).__init__(**kwargs)
self.args = {
"input_shape": input_shape,
"architecture_args": architecture_args,
"pretrained_arg": pretrained_arg,
"device": device,
"grad_weight_decay": grad_weight_decay,
"grad_l1_penalty": grad_l1_penalty,
"lamb": lamb,
"class": "PenalizeLastLayerFixedForm",
}
assert len(input_shape) == 3
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.device = device
self.grad_weight_decay = grad_weight_decay
self.grad_l1_penalty = grad_l1_penalty
self.lamb = lamb
# initialize the network
self.num_classes = self.architecture_args["classifier"][-1]["dim"]
self.last_layer_dim = self.architecture_args["classifier"][-2]["dim"]
self.classifier_base, _ = nn_utils.parse_feed_forward(
args=self.architecture_args["classifier"][:-1],
input_shape=self.input_shape)
self.classifier_base = self.classifier_base.to(self.device)
self.classifier_last_layer = torch.nn.Linear(self.last_layer_dim,
self.num_classes,
bias=False).to(
self.device)
if self.pretrained_arg is not None:
q_base = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, self.device)
# create the trainable part of the q_network
q_top = torch.nn.Sequential(
torch.nn.Linear(q_base.output_shape[-1], 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes),
).to(self.device)
self.q_network = torch.nn.Sequential(q_base, q_top)
else:
self.q_network, _ = nn_utils.parse_feed_forward(
args=self.architecture_args["q-network"],
input_shape=self.input_shape)
self.q_network = self.q_network.to(self.device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device='cuda',
grad_weight_decay=0.0,
grad_l1_penalty=0.0,
lamb=1.0,
sample_from_q=False,
q_dist='Gaussian',
loss_function='ce',
detach=True,
load_from=None,
warm_up=0,
**kwargs):
super(PredictGradOutput, self).__init__(**kwargs)
self.args = None # this will be modified by the decorator
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.grad_weight_decay = grad_weight_decay
self.grad_l1_penalty = grad_l1_penalty
self.lamb = lamb
self.sample_from_q = sample_from_q
self.q_dist = q_dist
self.detach = detach
self.loss_function = loss_function
self.load_from = load_from
self.warm_up = warm_up
# lamb is the coefficient in front of the H(p,q) term. It controls the variance of predicted gradients.
if self.q_dist == 'Gaussian':
self.grad_replacement_class = nn_utils.get_grad_replacement_class(
sample=self.sample_from_q,
standard_dev=np.sqrt(1.0 / 2.0 / (self.lamb + 1e-12)),
q_dist=self.q_dist)
elif self.q_dist == 'Laplace':
self.grad_replacement_class = nn_utils.get_grad_replacement_class(
sample=self.sample_from_q,
standard_dev=np.sqrt(2.0) / (self.lamb + 1e-6),
q_dist=self.q_dist)
elif self.q_dist in ['dot', 'ce']:
assert not self.sample_from_q
self.grad_replacement_class = nn_utils.get_grad_replacement_class(
sample=False)
else:
raise NotImplementedError()
# initialize the network
self.classifier, output_shape = nn_utils.parse_network_from_config(
args=self.architecture_args['classifier'],
input_shape=self.input_shape)
self.classifier = self.classifier.to(device)
self.num_classes = output_shape[-1]
if self.pretrained_arg is not None:
q_base = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, device)
# create the trainable part of the q_network
q_top = torch.nn.Sequential(
torch.nn.Linear(q_base.output_shape[-1], 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes)).to(device)
self.q_network = torch.nn.Sequential(q_base, q_top)
self.q_network = self.q_network.to(device)
else:
self.q_network, _ = nn_utils.parse_network_from_config(
args=self.architecture_args['q-network'],
input_shape=self.input_shape)
self.q_network = self.q_network.to(device)
if self.load_from is not None:
print("Loading the gradient predictor model from {}".format(
load_from))
import methods
stored_net = utils.load(load_from,
methods=methods,
device='cpu')
stored_net_params = dict(
stored_net.classifier.named_parameters())
for key, param in self.q_network.named_parameters():
param.data = stored_net_params[key].data.to(device)
self.q_loss = None
if self.loss_function == 'none': # predicted gradient has general form
self.q_loss = torch.nn.Sequential(
torch.nn.Linear(2 * self.num_classes, 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes)).to(device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device="cuda",
grad_weight_decay=0.0,
grad_l1_penalty=0.0,
lamb=1.0,
sample_from_q=False,
q_dist="Gaussian",
loss_function="ce",
detach=True,
load_from=None,
warm_up=0,
**kwargs):
super(PredictGradOutput, self).__init__(**kwargs)
self.args = {
"input_shape": input_shape,
"architecture_args": architecture_args,
"pretrained_arg": pretrained_arg,
"device": device,
"grad_weight_decay": grad_weight_decay,
"grad_l1_penalty": grad_l1_penalty,
"lamb": lamb,
"sample_from_q": sample_from_q,
"q_dist": q_dist,
"loss_function": loss_function,
"detach": detach,
"load_from": load_from,
"warm_up": warm_up,
"class": "PredictGradOutput",
}
assert len(input_shape) == 3
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.device = device
self.grad_weight_decay = grad_weight_decay
self.grad_l1_penalty = grad_l1_penalty
self.lamb = lamb
self.sample_from_q = sample_from_q
self.q_dist = q_dist
self.detach = detach
self.loss_function = loss_function
self.load_from = load_from
self.warm_up = warm_up
# lamb is the coefficient in front of the H(p,q) term. It controls the variance of predicted gradients.
if self.q_dist == "Gaussian":
self.grad_replacement_class = nn_utils.get_grad_replacement_class(
sample=self.sample_from_q,
standard_dev=np.sqrt(1.0 / 2.0 / (self.lamb + 1e-12)),
q_dist=self.q_dist,
)
elif self.q_dist == "Laplace":
self.grad_replacement_class = nn_utils.get_grad_replacement_class(
sample=self.sample_from_q,
standard_dev=np.sqrt(2.0) / (self.lamb + 1e-6),
q_dist=self.q_dist,
)
elif self.q_dist == "dot":
assert not self.sample_from_q
self.grad_replacement_class = nn_utils.get_grad_replacement_class(
sample=False)
else:
raise NotImplementedError()
# initialize the network
self.classifier, output_shape = nn_utils.parse_feed_forward(
args=self.architecture_args["classifier"],
input_shape=self.input_shape)
self.classifier = self.classifier.to(self.device)
self.num_classes = output_shape[-1]
if self.pretrained_arg is not None:
q_base = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, self.device)
# create the trainable part of the q_network
q_top = torch.nn.Sequential(
torch.nn.Linear(q_base.output_shape[-1], 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes),
).to(self.device)
self.q_network = torch.nn.Sequential(q_base, q_top)
else:
self.q_network, _ = nn_utils.parse_feed_forward(
args=self.architecture_args["q-network"],
input_shape=self.input_shape)
self.q_network = self.q_network.to(self.device)
if self.load_from is not None:
print("Loading the gradient predictor model from {}".format(
load_from))
stored_net = utils.load(load_from, device="cpu")
stored_net_params = dict(
stored_net.classifier.named_parameters())
for key, param in self.q_network.named_parameters():
param.data = stored_net_params[key].data.to(self.device)
self.q_loss = None
if self.loss_function == "none": # predicted gradient has general form
self.q_loss = torch.nn.Sequential(
torch.nn.Linear(2 * self.num_classes, 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes),
).to(self.device)
def __init__(self,
input_shape,
architecture_args,
pretrained_arg=None,
device="cuda",
grad_weight_decay=0.0,
grad_l1_penalty=0.0,
lamb=1.0,
small_qtop=False,
sample_from_q=False,
**kwargs):
super(PredictGradOutputFixedFormWithConfusion, self).__init__(**kwargs)
self.args = {
"input_shape": input_shape,
"architecture_args": architecture_args,
"pretrained_arg": pretrained_arg,
"device": device,
"grad_weight_decay": grad_weight_decay,
"grad_l1_penalty": grad_l1_penalty,
"lamb": lamb,
"small_qtop": small_qtop,
"sample_from_q": sample_from_q,
"class": "PredictGradOutputFixedFormWithConfusion",
}
assert len(input_shape) == 3
self.input_shape = [None] + list(input_shape)
self.architecture_args = architecture_args
self.pretrained_arg = pretrained_arg
self.device = device
self.grad_weight_decay = grad_weight_decay
self.grad_l1_penalty = grad_l1_penalty
self.lamb = lamb
self.small_qtop = small_qtop
self.sample_from_q = sample_from_q
self.grad_replacement_class = nn_utils.get_grad_replacement_class(
sample=self.sample_from_q,
standard_dev=np.sqrt(1.0 / 2.0 / (self.lamb + 1e-12)),
)
# initialize the network
self.classifier, output_shape = nn_utils.parse_feed_forward(
args=self.architecture_args["classifier"],
input_shape=self.input_shape)
self.classifier = self.classifier.to(self.device)
self.num_classes = output_shape[-1]
if self.pretrained_arg is not None:
self.q_base = pretrained_models.get_pretrained_model(
self.pretrained_arg, self.input_shape, self.device)
q_base_shape = self.q_base.output_shape
else:
self.q_base, q_base_shape = nn_utils.parse_feed_forward(
args=self.architecture_args["q-base"],
input_shape=self.input_shape)
self.q_base = self.q_base.to(self.device)
if small_qtop:
self.q_top = torch.nn.Linear(q_base_shape[-1],
self.num_classes).to(self.device)
else:
self.q_top = torch.nn.Sequential(
torch.nn.Linear(q_base_shape[-1], 128),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(128, self.num_classes),
).to(self.device)
# the confusion matrix trainable logits, (true, observed)
Q_init = torch.zeros(
size=(self.num_classes, self.num_classes),
device=self.device,
dtype=torch.float,
)
Q_init += 1.0 * torch.eye(self.num_classes,
device=self.device,
dtype=torch.float) # TODO: tune the constant
self.Q_logits = torch.nn.Parameter(Q_init, requires_grad=True)