text stringlengths 1 93.6k |
|---|
# for visualization
|
self.checkpoint_pcd = [] # to save the staged checkpoints
|
self.checkpoint_flags = [] # plot subtitle
|
if len(args.w_D_loss) == 1:
|
self.w_D_loss = args.w_D_loss * len(args.G_lrs)
|
else:
|
self.w_D_loss = args.w_D_loss
|
def finetune(self, bool_gt=False, save_curve=False, ith=-1):
|
# forward
|
if bool_gt:
|
self.args.G_lrs = [2e-7, 1e-6, 1e-6, 2e-7]
|
self.args.z_lrs = [9e-3, 2e-3, 1e-3, 1e-6]
|
self.iterations = [0, 0, 2, 1]
|
#self.iterations = [24, 8, 2, 4]
|
#self.iterations = [12, 2, 1, 1]
|
self.k_mask_k = [1, 1, 1, 1]
|
else:
|
self.args.G_lrs = [2e-7, 1e-6, 1e-6, 2e-7]
|
self.args.z_lrs = [9e-3, 2e-3, 1e-3, 1e-6]
|
self.iterations = [1, 4, 4, 1]
|
self.k_mask_k = [1, 1, 1, 1]
|
tree = [self.partial]
|
for model_name in ["Encoder", "Decoder", "DI", "DS", "MS"]:
|
self.models[model_name].eval()
|
with torch.no_grad():
|
hidden_z = self.Encoder(tree)
|
f_di = self.DI_Disentangler(hidden_z)
|
#f_di_c = self.Z_Mapper(f_di)
|
f_ms = self.MS_Disentangler(hidden_z)
|
f_ds = self.DS_Disentangler(hidden_z)
|
f_combine_c = torch.cat([f_di, f_ms*0., f_ds], 1)
|
self.z.copy_(f_combine_c)
|
loss_dict = {}
|
curr_step = 0
|
count = 0
|
if save_curve:
|
cd_curve_list = []
|
for stage, iteration in enumerate(self.iterations):
|
for i in range(iteration):
|
curr_step += 1
|
# setup learning rate
|
self.schedulers['Decoder'].update(curr_step, self.args.G_lrs[stage])
|
self.z_scheduler.update(curr_step, self.args.z_lrs[stage])
|
# forward
|
self.z_optim.zero_grad()
|
if self.update_G_stages[stage]:
|
self.Decoder.optim.zero_grad()
|
tree = self.z
|
x = self.Decoder(tree)
|
# masking
|
x_map = self.pre_process(x,stage=stage)
|
### compute losses
|
ftr_loss = self.criterion(self.ftr_net, x_map, self.partial)
|
dist1, dist2 , _, _ = distChamfer(self.partial, x)
|
ucd_loss = dist1.mean()
|
dist1, dist2 , _, _ = distChamfer(x_map, self.partial)
|
cd_loss = dist1.mean() + dist2.mean()
|
if self.gt is not None:
|
dist1, dist2 , _, _ = distChamfer(x, self.gt)
|
gt_cd_loss = dist1.mean() + dist2.mean()
|
# optional early stopping
|
if self.args.early_stopping:
|
if cd_loss.item() < self.args.stop_cd:
|
break
|
# nll corresponds to a negative log-likelihood loss
|
nll = self.z**2 / 2
|
nll = nll.mean()
|
### loss
|
loss = ftr_loss * self.w_D_loss[0] + nll * self.args.w_nll \
|
+ cd_loss * 1
|
# optional to use directed_hausdorff
|
directed_hausdorff_loss = self.directed_hausdorff(self.partial.permute([0,2,1]), x.permute([0,2,1]))
|
if self.args.directed_hausdorff:
|
print("Using Hausdorff")
|
loss += directed_hausdorff_loss*self.args.w_directed_hausdorff_loss
|
# backward
|
loss.backward()
|
self.z_optim.step()
|
if self.update_G_stages[stage]:
|
self.Decoder.optim.step()
|
if save_curve:
|
if self.gt is not None:
|
dist1, dist2 , _, _ = distChamfer(x,self.gt)
|
test_cd = dist1.mean() + dist2.mean()
|
cd_curve_list.append(test_cd.item())
|
self.x = x
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.