def __call__(self, trainer=None):
"""Executes the evaluator extension.
Unlike usual extensions, this extension can be executed without passing
a trainer object. This extension reports the performance on validation
dataset using the :func:`~pytorch_trainer.report` function. Thus, users can use
this extension independently from any trainer by manually configuring
a :class:`~pytorch_trainer.Reporter` object.
Args:
trainer (~pytorch_trainer.training.Trainer): Trainer object that invokes
this extension. It can be omitted in case of calling this
extension manually.
Returns:
dict: Result dictionary that contains mean statistics of values
reported by the evaluation function.
"""
# set up a reporter
reporter = reporter_module.Reporter()
if self.name is not None:
prefix = self.name + '/'
else:
prefix = ''
for name, target in six.iteritems(self._targets):
reporter.add_observer(prefix + name, target)
reporter.add_observers(prefix + name, target.named_children())
with reporter:
with torch.no_grad():
result = self.evaluate()
reporter_module.report(result)
return result
def generative_lossfun(self, p_fake):
t_1 = torch.ones(p_fake.shape,
dtype=p_fake.dtype,
device=p_fake.device)
loss = F.mse_loss(p_fake, t_1)
reporter.report({'loss_gen': loss})
return loss
def forward(self, x, t):
y = self.predictor(x)
loss = F.nll_loss(y, t)
reporter.report({'loss': loss}, self)
acc = accuracy(y, t)
reporter.report({'accuracy': acc}, self)
return loss
def discriminative_lossfun(self, p_real, p_fake):
t_1 = torch.ones(p_real.shape,
dtype=p_real.dtype,
device=p_real.device)
t_0 = torch.zeros(p_fake.shape,
dtype=p_fake.dtype,
device=p_fake.device)
loss = (F.mse_loss(p_real, t_1) \
+ F.mse_loss(p_fake, t_0)) * 0.5
reporter.report({'loss_dis': loss})
return loss
def forward(self, x):
h = F.relu(self.l1(x))
h = self.dropout(h)
h = F.relu(self.l2(h))
h = self.dropout(h)
out = self.l3(h)
sigma = self.l3_sigma(h)
reporter.report({'sigma': sigma.mean()}, self)
return out, sigma
def conditional_lossfun(self, y_fake, y_true):
model = self.generator
if hasattr(model, 'lossfun'):
lossfun = model.lossfun
else:
lossfun = self.loss_func
loss = lossfun(y_fake, y_true)
reporter.report({'loss_cond': loss})
return loss
def __call__(self, trainer):
observation = trainer.observation
if not (self._numerator_key in observation and
self._denominator_key in observation):
return
self._numerator += observation[self._numerator_key]
self._denominator += observation[self._denominator_key]
if self._trigger(trainer):
result = float(self._numerator) / self._denominator
self._numerator = 0
self._denominator = 0
reporter.report({self._result_key: result})
def forward(self, x):
h = super().forward(x)
out = self['conv_out'](h)
out = crop(out, x.shape)
if not self._sigma:
return out
sigma = self['conv_sigma'](h)
sigma = crop(sigma, x.shape)
reporter.report({'sigma': torch.mean(sigma)}, self)
return out, sigma
def report(self, trainer):
# set up a reporter
reporter = reporter_module.Reporter()
if self.name is not None:
prefix = self.name + '/'
else:
prefix = ''
for name, target in six.iteritems(self._targets):
reporter.add_observer(prefix + name, target)
reporter.add_observers(prefix + name + '/',
target.named_children())
with reporter:
result = self.evaluate(trainer)
reporter_module.report(result)
return result
def forward(self, *args, **kwargs):
"""Computes the loss value for input and label pair.
It also computes accuracy and stores it to the attribute.
Args:
args (list of ~chainer.Variable): Input minibatch.
kwargs (dict of ~chainer.Variable): Input minibatch.
When ``label_key`` is ``int``, the corresponding element in ``args``
is treated as ground truth labels. And when it is ``str``, the
element in ``kwargs`` is used.
The all elements of ``args`` and ``kwargs`` except the ground truth
labels are features.
It feeds features to the predictor and compare the result
with ground truth labels.
.. note::
We set ``None`` to the attributes ``y``, ``loss`` and ``accuracy``
each time before running the predictor, to avoid unnecessary memory
consumption. Note that the variables set on those attributes hold
the whole computation graph when they are computed. The graph
stores interim values on memory required for back-propagation.
We need to clear the attributes to free those values.
Returns:
~chainer.Variable: Loss value.
"""
self._reset()
n_args = len(args) + len(kwargs)
x = get_values(args, kwargs, self.x_keys)
t = get_values(args, kwargs, self.t_keys) if n_args > 1 else None
# predict, and then apply final activation
y = self.predictor(x)
if self.activation is not None:
y = self.activation(y)
# preserve
self.x = x
self.y = y
self.t = t
# if only input `x` is exist, return the predictions
if t is None:
return y
# if ground-truth label `t` is exist, evaluate the loss and accuracy.
# return the loss during training, otherwise return the predictions.
if self.lossfun is not None:
self.loss = self.lossfun(y, t)
reporter.report({'loss': self.loss}, self)
if self.accfun is not None:
self.accuracy = self.accfun(y, t)
reporter.report({'accuracy': self.accuracy}, self)
if self.training:
if self.loss is None:
raise ValueError('loss is None..')
return self.loss
else:
return self.y
def generative_lossfun(self, p_fake):
size = p_fake.numel() / p_fake.shape[1]
loss = torch.sum(F.softplus(-p_fake)) / size
reporter.report({'loss_gen': loss})
return loss
def discriminative_lossfun(self, p_real, p_fake):
size = p_real.numel() / p_real.shape[1]
loss = (torch.sum(F.softplus(-p_real)) / size \
+ torch.sum(F.softplus(p_fake)) / size) * 0.5 # NOTE: equivalent to binary cross entropy
reporter.report({'loss_dis': loss})
return loss
