From 59ccf61fed4d95b7fe91bb9552f0deb2f2c75b76 Mon Sep 17 00:00:00 2001
From: Jordan Gong <jordan.gong@protonmail.com>
Date: Sat, 23 Jan 2021 16:04:35 +0800
Subject: Add late start support for non-disentangling parts

---
 config.py              |  2 ++
 models/model.py        | 24 +++++++++++++++++-------
 utils/configuration.py |  1 +
 3 files changed, 20 insertions(+), 7 deletions(-)

diff --git a/config.py b/config.py
index 991a3a6..641e8fb 100644
--- a/config.py
+++ b/config.py
@@ -70,6 +70,8 @@ config: Configuration = {
         },
         'optimizer': {
             # Global parameters
+            # Iteration start to optimize non-disentangling parts
+            # 'start_iter': 10,
             # Initial learning rate of Adam Optimizer
             'lr': 1e-4,
             # Coefficients used for computing running averages of
diff --git a/models/model.py b/models/model.py
index 6b799ad..cccb6c4 100644
--- a/models/model.py
+++ b/models/model.py
@@ -141,6 +141,7 @@ class Model:
         # Prepare for model, optimizer and scheduler
         model_hp = self.hp.get('model', {})
         optim_hp: dict = self.hp.get('optimizer', {}).copy()
+        start_iter = optim_hp.pop('start_iter', 0)
         ae_optim_hp = optim_hp.pop('auto_encoder', {})
         pn_optim_hp = optim_hp.pop('part_net', {})
         hpm_optim_hp = optim_hp.pop('hpm', {})
@@ -151,9 +152,6 @@ class Model:
         self.rgb_pn = self.rgb_pn.to(self.device)
         self.optimizer = optim.Adam([
             {'params': self.rgb_pn.ae.parameters(), **ae_optim_hp},
-            {'params': self.rgb_pn.pn.parameters(), **pn_optim_hp},
-            {'params': self.rgb_pn.hpm.parameters(), **hpm_optim_hp},
-            {'params': self.rgb_pn.fc_mat, **fc_optim_hp},
         ], **optim_hp)
         self.scheduler = optim.lr_scheduler.StepLR(self.optimizer, **sched_hp)
         self.writer = SummaryWriter(self._log_name)
@@ -173,8 +171,18 @@ class Model:
         start_time = datetime.now()
         running_loss = torch.zeros(4).to(self.device)
         print(f"{'Iter':^5} {'Loss':^6} {'Xrecon':^8} {'PoseSim':^8}",
-              f"{'CanoCons':^8} {'BATrip':^8} {'LR':^9}")
+              f"{'CanoCons':^8} {'BATrip':^8} LR(s)")
         for (batch_c1, batch_c2) in dataloader:
+            if self.curr_iter == start_iter:
+                self.optimizer.add_param_group(
+                    {'params': self.rgb_pn.pn.parameters(), **pn_optim_hp}
+                )
+                self.optimizer.add_param_group(
+                    {'params': self.rgb_pn.hpm.parameters(), **hpm_optim_hp}
+                )
+                self.optimizer.add_param_group(
+                    {'params': self.rgb_pn.fc_mat, **fc_optim_hp}
+                )
             self.curr_iter += 1
             # Zero the parameter gradients
             self.optimizer.zero_grad()
@@ -186,8 +194,6 @@ class Model:
             loss = losses.sum()
             loss.backward()
             self.optimizer.step()
-            # Step scheduler
-            self.scheduler.step()
 
             # Statistics and checkpoint
             running_loss += losses.detach()
@@ -199,11 +205,15 @@ class Model:
             ], losses)), self.curr_iter)
 
             if self.curr_iter % 100 == 0:
+                lrs = self.scheduler.get_last_lr()
                 print(f'{self.curr_iter:5d} {running_loss.sum() / 100:6.3f}',
                       '{:f} {:f} {:f} {:f}'.format(*running_loss / 100),
-                      f'{self.scheduler.get_last_lr()[0]:.3e}')
+                      ' '.join(('{:.3e}'.format(lr) for lr in lrs)))
                 running_loss.zero_()
 
+            # Step scheduler
+            self.scheduler.step()
+
             if self.curr_iter % 1000 == 0:
                 torch.save({
                     'iter': self.curr_iter,
diff --git a/utils/configuration.py b/utils/configuration.py
index 8b265e8..c4c4b4d 100644
--- a/utils/configuration.py
+++ b/utils/configuration.py
@@ -52,6 +52,7 @@ class SubOptimizerHPConfiguration(TypedDict):
 
 
 class OptimizerHPConfiguration(TypedDict):
+    start_iter: int
     lr: int
     betas: tuple[float, float]
     eps: float
-- 
cgit v1.2.3