def fmix(data, targets, alpha, decay_power, shape, max_soft=0.0):
lam, mask = sample_mask(alpha, decay_power, shape, max_soft)
indices = torch.randperm(data.size(0))
shuffled_data = data[indices]
# shuffled_targets = targets[indices]
x1 = torch.from_numpy(mask)*data
x2 = torch.from_numpy(1-mask)*shuffled_data
# targets=(targets, shuffled_targets, lam)
return (x1+x2), targets
def __call__(self, x):
shape = [int(s) for s in x.shape][1:-1]
lam, mask = sample_mask(self.alpha, self.decay_power, shape, self.max_soft, self.reformulate)
index = np.random.permutation(int(x.shape[0]))
index = tf.constant(index)
mask = np.expand_dims(mask, -1)
x1 = x * mask
x2 = tf.gather(x, index) * (1 - mask)
self.index = index
self.lam = lam
return x1 + x2
def __call__(self, x):
# Sample mask and generate random permutation
lam, mask = sample_mask(self.alpha, self.decay_power, self.size,
self.max_soft, self.reformulate)
index = torch.randperm(x.size(0)).to(x.device)
mask = torch.from_numpy(mask).float().to(x.device)
# Mix the images
x1 = mask * x
x2 = (1 - mask) * x[index]
self.index = index
self.lam = lam
return x1 + x2
def fmix(data, target, alpha, decay_power, shape, max_soft=0.0, reformulate=False):
lam, mask = sample_mask(alpha, decay_power, shape, max_soft, reformulate)
indices = torch.randperm(data.size(0))
shuffled_data = data[indices]
shuffled_target = target[indices]
x1 = torch.from_numpy(mask).to(data.device) * data
x2 = torch.from_numpy(1 - mask).to(data.device) * shuffled_data
targets = {
'target': target,
'shuffled_target': shuffled_target,
'lam': lam
}
return x1 + x2, targets
def __call__(self, x):
size = []
for i, s in enumerate(self.size):
if s != -1:
size.append(s)
else:
size.append(x.shape[i + 1])
lam, mask = sample_mask(self.alpha, self.decay_power, size,
self.max_soft, self.reformulate)
index = torch.randperm(x.size(0)).to(x.device)
mask = torch.from_numpy(mask).float().to(x.device)
# Mix the images
x1 = mask * x
x2 = (1 - mask) * x[index]
self.index = index
self.lam = lam
return x1 + x2
def fmix(data,
targets,
alpha,
decay_power,
shape,
device,
max_soft=0.0,
reformulate=False):
data = data.detach().cpu()
lam, mask = sample_mask(alpha, decay_power, shape, max_soft, reformulate)
indices = torch.randperm(data.size(0))
shuffled_data = data[indices]
shuffled_targets = targets[indices]
x1 = torch.from_numpy(mask) * data
x2 = torch.from_numpy(1 - mask) * shuffled_data
# x1 = x1.to(device)
# x2 = x2.to(device)
targets = (targets, shuffled_targets, lam)
return (x1 + x2), targets
def __call__(self, x):
# Sample mask and generate random permutation
lam, mask = sample_mask(self.alpha, self.decay_power, self.size,
self.max_soft, self.reformulate)
index = torch.randperm(x.size(0)).to(x.device)
mask = torch.from_numpy(mask).float().to(x.device)
# Mix the images
x1 = mask * x
x2 = (1 - mask) * x[index]
self.index = index
self.lam = lam
if lam >= 0.5:
x_cut = x1
else:
x_cut = x2
masky = mask.cpu()
#print(masky.numpy())
#np.save("mask.txt",masky.numpy())
return x1 + x2, x1, x2