def getDevsetAccuracy(self, model=None):
accSum = 0
for xy in self.dev:
xy = to_var_gpu(xy)
accSum += self.batchAccuracy(*xy, model=model)
acc = accSum / len(self.dev)
return acc
def train(self, numEpochs=100):
""" The main training loop called (also for subclasses)"""
for epoch in range(self.epoch, numEpochs):
self.lr_scheduler.step(epoch)
self.epoch = epoch
for i in range(self.numBatchesPerEpoch):
trainData = to_var_gpu(next(self.train_iter))
self.step(*trainData)
self.logStuff(i, epoch, numEpochs, trainData)
def constSWA(self, numEpochs=100, lr=1e-4):
""" runs Stochastic Weight Averaging for numEpochs epochs using const lr """
self.SWAupdates = 0
self.SWA = copy.deepcopy(self.CNN)
## Set the new constant learning rate
new_lr_lambda = lambda epoch: lr / self.hypers['base_lr']
self.lr_scheduler = optim.lr_scheduler.LambdaLR(
self.optimizer, new_lr_lambda)
for epoch in range(self.epoch, numEpochs + self.epoch):
self.lr_scheduler.step(epoch)
self.epoch = epoch
for i in range(self.numBatchesPerEpoch):
trainData = to_var_gpu(next(self.train_iter))
self.step(*trainData)
self.swaLogStuff(i, epoch)
self.logStuff(i, epoch, numEpochs + self.epoch, trainData)
self.updateSWA()
def step(self, x_unlab, *_):
x_unlab = to_var_gpu(x_unlab)
for _ in range(self.hypers['n_critic']):
self.d_optimizer.zero_grad()
z = self.getNoise(self.hypers["ul_BS"]) #*.1
x_fake = self.G(z).detach()
wass_loss = self.D(x_fake).mean() - self.D(x_unlab).mean()
d_loss = wass_loss + self.grad_penalty(x_unlab, x_fake)
d_loss.backward()
self.d_optimizer.step()
self.g_optimizer.zero_grad()
z = self.getNoise(self.hypers["ul_BS"]) #*.1
x_fake = self.G(z)
g_loss = -self.D(x_fake).mean()
g_loss.backward()
self.g_optimizer.step()
return d_loss, g_loss
def updateBatchNorm(self, model):
model.train()
for _ in range(self.numBatchesPerEpoch):
tensors = next(self.train_iter)
trainData = to_var_gpu(tensors, volatile=True)
out = model(self.getLabeledXYonly(trainData)[0])