def train(self, dataloader):
for epoch in range(self.args.epochs):
for idx, (x, y) in enumerate(dataloader):
x, y = x.to(self.device), y.to(self.device)
# train D with real
self.D.zero_grad()
minibs = x.size(0)
label = torch.full((minibs, ),
self.real_label,
device=self.device)
y = self.D(x)
loss_real = self.criterion(y, label)
loss_real.backward()
D_real_avg = y.mean().item()
# train D with fake
noise = torch.randn(minibs, self.args.rvs, 1,
1).to(self.device)
label.fill_(self.fake_label)
fake_y = self.G(noise)
# print(fake_y.shape)
y = self.D(fake_y.detach())
# print(y.shape)
loss_fake = self.criterion(y, label)
loss_fake.backward()
D_fake_avg = y.mean().item()
loss_D = loss_real + loss_fake
self.optimizerD.step()
# train G
self.G.zero_grad()
label.fill_(self.real_label)
y = self.D(fake_y)
loss_G = self.criterion(y, label)
loss_G.backward()
G_avg = y.mean().item()
self.optimizerG.step()
# print log
if idx % 100 == 0:
print(
'Epoch [{}/{}], [{}/{}], Loss_D: {:.4f}, Loss_G: {:.4f}, Loss_D(x): '
'{:.4f}, Loss_D(G(z)): {:.4f}/{:.4f}'.format(
epoch, self.args.epochs, idx, len(dataloader),
loss_D.item(), loss_G.item(), D_real_avg, G_avg,
D_fake_avg))
get_grid(imgs=self.G(
self.test_noise).detach().cpu().numpy(),
args=self.args,
epoch=epoch,
it=idx)
self.save()
def get_feasible_centers_brute(pivot, pi_hat, mu_hat, S_hat, Rho,
base_intervals):
pivot_cluster, pivot_dim = pivot
dim = mu_hat.shape[0]
pi_hat_pivot = pi_hat[pivot_cluster]
mu_hat_pivot = mu_hat[pivot_dim]
pi_hat_other = pi_hat[1 - pivot_cluster]
def is_base_feasible(center, b, j):
return (center >= base_intervals[b][j].lo
and center <= base_intervals[b][j].hi)
def other_center(center, prob, mean_hat):
return (mean_hat - prob * center) / (1 - prob)
pivot_interval = get_grid(
get_first_order_constraints(pi_hat_pivot, mu_hat_pivot), 1000)
feasible_solutions = []
for pivot_center in pivot_interval:
candidate_solution = np.zeros((2, dim))
candidate_solution[pivot_cluster, pivot_dim] = pivot_center
candidate_solution[1 - pivot_cluster,
pivot_dim] = other_center(pivot_center,
pi_hat_pivot,
mu_hat_pivot)
feasible = True
for j in range(dim):
if j == pivot_dim: continue
jth_center = ((S_hat[pivot_dim, j] - Rho[pivot_dim, j] -
mu_hat[j] * pivot_center) /
(mu_hat_pivot - pivot_center))
if not is_base_feasible(jth_center, 1 - pivot_cluster, j):
feasible = False
break
else:
candidate_solution[1 - pivot_cluster, j] = jth_center
candidate_solution[pivot_cluster,
j] = other_center(pivot_center,
pi_hat_other, mu_hat[j])
if feasible:
feasible_solutions.append(candidate_solution)
return feasible_solutions
def get_second_order_constraints(S, rho, mu_2dim, other_interval):
other_grid = get_grid(other_interval)
this_grid = (S - rho - mu_2dim[1] * other_grid) / (mu_2dim[0] - other_grid)
return clip(Interval(lo=this_grid.min(), hi=this_grid.max()))
optimizer.step()
if (idx + 1) % 100 == 0:
print(
'Epoch: [{}/{}], Step: [{}/{}], Loss: {:.4f}, Acc: {:.4f}'.
format(epoch + 1, args.epochs, idx + 1, len(train_loader),
loss.item(), 100 * correct / total))
# 保存模型参数
torch.save(
model.state_dict(),
os.path.join(
'./log', '{}_{}_{}.ckpt'.format(args.model, args.dataset,
args.epochs)))
model.eval()
with torch.no_grad():
correct = 0
total = 0
for idx, (x, y) in enumerate(test_loader):
x, y = x.to(device), y.to(device)
y_pred = model(x)
_, y_pred = torch.max(y_pred.data, 1)
total += y.size(0)
correct += (y_pred == y).sum().item()
# print(result)
if idx % 100 == 0:
get_grid(x.cpu().numpy(), args, args.epochs, idx)
print(y_pred.data.cpu().numpy(), y.data.cpu().numpy)
print('Test Acc: {:.4f}%, Model: {}, Epochs: {}'.format(
correct / total * 100, args.model, args.epochs))
download=True,
transform=transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, ))])),
batch_size=args.bs,
shuffle=True)
test_loader = torch.utils.data.DataLoader(datasets.MNIST(
args.root,
train=False,
transform=transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, ))])),
batch_size=args.bs,
shuffle=True)
return train_loader, test_loader
from options import get_args
from util import get_grid
if __name__ == '__main__':
args = get_args()
# print(args.bs)
# print(args.root)
train_loader, test_loader = get_data(args=args)
for idx, (x, y) in enumerate(train_loader):
# print(x)
# print(x.shape)
imgs = get_grid(x, args, 1, idx)
# imgs.show()
def reconstruct(dmap, kmap, mask_r, pl=10, ps=4):
h, w, _ = dmap.shape
grid_h, grid_w = util.get_grid(h, w)
connections = util.get_connections()
limbs = util.get_limbs()
patch = util.get_patch(pl, ps)
humans = [] # store the annotation for each human
q = queue.Queue() # store the layers to be extended from
used = [False] * 14 # store wether the layer has been explored
starters, layer = find_outstander(kmap, mask_r, patch)
# print(layer)
# print(starters)
# print(dmap[8,41,0])
# print(dmap[8,41,1])
# print(dmap[8,41,26])
# print(dmap[8,41,27])
for p in starters:
dic = {}
dic[layer] = p
humans.append(dic)
q.put(layer)
used[layer] = True
debug_cnt = 0
while True:
if q.empty():
# print('empty')
break
layer1 = q.get()
# print('from', layer1)
starters = []
for i,h in enumerate(humans):
if layer1 in h.keys():
starters.append((h[layer1], i))
for piece in connections[layer1]:
debug_cnt += 1
# if debug_cnt == 2:
# return humans
layer2, sign, d_layer = piece
if used[layer2]:
continue
used[layer2] = True
q.put(layer2)
# print('finding', layer2)
# print('sign', sign)
# print('dmap layer', d_layer)
k_slice = kmap[:,:,layer2]
# in limbs, the vector is from START to END
# in connections, sign = -1 means layer2 is START
# forward is from layer1's view to layer2
# backward is from layer2's view to layer1
# d_layer is the layer stores vector from START to END
# so, if sign = 1, layer1 is START, layer2 is END
# and, dx_forward is from layer1 to layer2
# so
dx_forward = dmap[:,:,d_layer*2]
dy_forward = dmap[:,:,d_layer*2+1]
dx_backward = dmap[:,:,d_layer*2+26]
dy_backward = dmap[:,:,d_layer*2+26+1]
# else, if sign = -1, layer1 is END, layer2 is START
# and, dx_forward is from layer1 to layer2
# so
if sign == -1:
dx_forward, dx_backward = dx_backward, dx_forward
dy_forward, dy_backward = dy_backward, dy_forward
mask_log = group_match(starters, dx_forward, dy_forward, dx_backward, dy_backward, grid_h, grid_w, k_slice, mask_r, patch)
trans_mask_log(mask_log, layer2, humans)
# for h in humans:
# print(h)
# print('\n')
# print(humans)
# print('mask_log')
# print(mask_log)
return humans
import numpy as np
import os
def get_data(args):
train_dataset = ImageFolder(args.root, transform=transforms.Compose([
transforms.Resize((32, 32)),
transforms.CenterCrop((32, 32)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]))
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=args.bs, shuffle=True)
test_loader = []
return train_loader, test_loader
from options import get_args
from util import get_grid
if __name__ == '__main__':
args = get_args()
# print(args.bs)
# print(args.root)
train_loader, test_loader = get_data(args=args)
for idx, (x, y) in enumerate(train_loader):
# print(x)
# print(x.shape)
imgs = get_grid(x.numpy(), args, 1, idx)
# imgs.show()
'Epoch [{}/{}], [{}/{}], Loss_D: {:.4f}, Loss_G: {:.4f}, Loss_D(x): '
'{:.4f}, Loss_D(G(z)): {:.4f}/{:.4f}'.format(
epoch, self.args.epochs, idx, len(dataloader),
loss_D.item(), loss_G.item(), D_real_avg, G_avg,
D_fake_avg))
get_grid(imgs=self.G(
self.test_noise).detach().cpu().numpy(),
args=self.args,
epoch=epoch,
it=idx)
self.save()
def test(self, dataloader):
pass
def simple(self):
pass
from options import get_args
from util import get_grid
if __name__ == '__main__':
args = get_args()
model = GAN(args=args)
noise = torch.randn(args.bs, args.rvs, 1, 1).to(model.device)
gg = model.G(noise)
print(gg.shape)
get_grid(imgs=gg.detach().cpu().numpy(), args=args, epoch=1, it=1)
# print(model.D(gg))