Add wandb, make batch and image size configurable, fix some bugs, improve supervised loss function, make use of some RL stuff just implemented

model/exp_synph_real.py

@@ -10,6 +10,7 @@ import matplotlib.pyplot as plt
 import cv2
 import torchvision.transforms as transforms
 
+import wandb
 
 import co
 import torchext
@@ -18,32 +19,39 @@ from data import dataset
 
 
 class Worker(torchext.Worker):
-    def __init__(self, args, num_workers=18, train_batch_size=4, test_batch_size=4, save_frequency=1, **kwargs):
+    def __init__(self, args, num_workers=18, train_batch_size=6, test_batch_size=6, save_frequency=1, **kwargs):
+        if 'batch_size' in dir(args):
+            train_batch_size = args.batch_size
+            test_batch_size = args.batch_size
         super().__init__(args.output_dir, args.exp_name, epochs=args.epochs, num_workers=num_workers,
                          train_batch_size=train_batch_size, test_batch_size=test_batch_size,
                          save_frequency=save_frequency, **kwargs)
-
         self.ms = args.ms
         self.pattern_path = args.pattern_path
         self.lcn_radius = args.lcn_radius
         self.dp_weight = args.dp_weight
         self.data_type = args.data_type
 
-        self.imsizes = [(488, 648)]
-        for iter in range(3):
-            self.imsizes.append((int(self.imsizes[-1][0] / 2), int(self.imsizes[-1][1] / 2)))
-
         with open('config.json') as fp:
             config = json.load(fp)
             data_root = Path(config['DATA_ROOT'])
+            self.imsizes = [tuple(map(int, config['IMSIZE'].split(',')))]
+
+        for iter in range(3):
+            self.imsizes.append((int(self.imsizes[-1][0] / 2), int(self.imsizes[-1][1] / 2)))
+
         self.settings_path = data_root / self.data_type / 'settings.pkl'
         sample_paths = sorted((data_root / self.data_type).glob('0*/'))
 
-        self.train_paths = sample_paths[2 ** 10:]
+        # calc split
-        self.test_paths = sample_paths[:2 ** 8]
+        # since we don't have a lot or RL footage, we compute it as we go
+        train_size = len(sample_paths) * 0.8 // 1
+        test_size = 1 - train_size
+        self.train_paths = sample_paths[test_size:]
+        self.test_paths = sample_paths[:train_size]
 
-        # supervise the edge encoder with only 2**8 samples
+        # don't just supervise the edge encoder with only 2**8 samples
-        self.train_edge = len(self.train_paths) - 2 ** 8
+        self.train_edge = len(self.train_paths)
 
         self.lcn_in = networks.LCN(self.lcn_radius, 0.05)
         self.disparity_loss = networks.DisparityLoss()
@@ -51,6 +59,25 @@ class Worker(torchext.Worker):
         self.sup_disp_loss = torch.nn.MSELoss()
         self.edge_loss = torch.nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([0.1]).to(self.train_device))
 
+        # FIXME L2 Regularization, try it!!
+        # l2_lambda = 0.001
+        # l2_norm = sum(p.pow(2.0).sum()
+                      # for p in net.parameters())
+        # self.sup_disp_loss = torch.nn.MSELoss() + l2_lambda * l2_norm
+        # FIXME try using log of this loss, otherwise it's very large compared to others
+        # self.sup_disp_loss = torch.nn.MSELoss()
+        class RMSLELoss(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.mse = torch.nn.MSELoss()
+
+            def forward(self, pred, actual):
+                # FIXME rename this if log is better than sqrt
+                return torch.log(self.mse(pred, actual))
+                # return torch.sqrt(self.mse(torch.log(pred + 1), torch.log(actual + 1)))
+
+        self.sup_disp_loss = RMSLELoss()
+
         # evaluate in the region where opencv Block Matching has valid values
         self.eval_mask = np.zeros(self.imsizes[0])
         self.eval_mask[13:self.imsizes[0][0] - 13, 140:self.imsizes[0][1] - 13] = 1
@@ -59,14 +86,13 @@ class Worker(torchext.Worker):
         self.eval_w = self.imsizes[0][1] - 13 - 140
 
     def get_train_set(self):
-        train_set = dataset.TrackSynDataset(self.settings_path, self.train_paths, train=True, data_aug=True,
+        train_set = dataset.RealWorldDataset(self.settings_path, self.train_paths, train=True, data_aug=True,
                                             track_length=1)
-
         return train_set
 
     def get_test_sets(self):
         test_sets = torchext.TestSets()
-        test_set = dataset.TrackSynDataset(self.settings_path, self.test_paths, train=False, data_aug=True,
+        test_set = dataset.RealWorldDataset(self.settings_path, self.test_paths, train=False, data_aug=True,
                                            track_length=1)
         test_sets.append('simple', test_set, test_frequency=1)
 
@@ -102,12 +128,12 @@ class Worker(torchext.Worker):
                 self.data[key_std] = im_std.to(device).detach()
 
     def net_forward(self, net, train):
-        # FIXME hier schnibbeln?
         out = net(self.data['im0'])
         return out
 
     def loss_forward(self, out, train):
         out, edge = out
+        losses = {}
         if not (isinstance(out, tuple) or isinstance(out, list)):
             out = [out]
         if not (isinstance(edge, tuple) or isinstance(edge, list)):
@@ -118,23 +144,30 @@ class Worker(torchext.Worker):
         # apply photometric loss
         for s, l, o in zip(itertools.count(), self.losses, out):
             val, pattern_proj = l(o[0], self.data[f'im{s}'][:, 0:1, ...], self.data[f'std{s}'])
+
             if s == 0:
                 self.pattern_proj = pattern_proj.detach()
             vals.append(val)
-
+            losses['photometric'] = val
-        # apply disparity loss
+        # 1-edge as ground truth edge if inverted
-        # 1-edge as ground truth edge if inversed
         if isinstance(edge, tuple):
             edge0 = 1 - torch.sigmoid(edge[0][0])
         else:
             edge0 = 1 - torch.sigmoid(edge[0])
         val = 0
         if isinstance(out[0], tuple):
-            # NOTE use supervised disparity loss
+            sup_loss = self.sup_disp_loss(out[0][1], self.data['disp0'])
-            val += self.sup_disp_loss(out[0][1], self.data['disp0'])
+            val += sup_loss
-            val += self.disparity_loss(out[0][0], edge0)
+
+            disp_loss = self.disparity_loss(out[0][0], edge0)
+            val += disp_loss
+
+            losses['GT Supervised disparity loss'] = sup_loss * self.dp_weight
+            losses['OG disparity loss'] = disp_loss * self.dp_weight
         else:
-            val += self.disparity_loss(out[0], edge0)
+            disp_loss = self.disparity_loss(out[0], edge0)
+            val += disp_loss
+            losses['OG disparity loss'] = disp_loss * self.dp_weight
         if self.dp_weight > 0:
             vals.append(val * self.dp_weight)
 
@@ -159,15 +192,20 @@ class Worker(torchext.Worker):
                     self.edge = e.detach()
                 self.edge = torch.sigmoid(self.edge)
                 self.edge_gt = grad.detach()
+            losses['edge'] = val
             vals.append(val)
 
+        wandb.log(losses)
         return vals
 
     def numpy_in_out(self, output):
         output, edge = output
         if not (isinstance(output, tuple) or isinstance(output, list)):
             output = [output]
+        if isinstance(output[0], tuple):
             es = output[0][0].detach().to('cpu').numpy()
+        else:
+            es = output[0].detach().to('cpu').numpy()
         gt = self.data['disp0'].to('cpu').numpy().astype(np.float32)
         im = self.data['im0'][:, 0:1, ...].detach().to('cpu').numpy()
 
@@ -250,6 +288,7 @@ class Worker(torchext.Worker):
 
         plt.tight_layout()
         plt.savefig(str(out_path))
+        wandb.log({f'results_{"_".join(out_path.stem.split("_")[:-1])}': plt})
         plt.close(fig)
 
     def callback_train_post_backward(self, net, errs, output, epoch, batch_idx, masks=[]):
@@ -293,9 +332,4 @@ class Worker(torchext.Worker):
 
 
 if __name__ == '__main__':
-    # FIXME Nicolas fixe idee
-    #   SGBM nutzen, um GT zu finden
-    #   bei dispnet (oder w/e) letzte paar layer 'dublizieren' (zweiten head bauen) und so mehrere Loss funktionen gleichzeitig trainieren
-    #   L1 + L2 und dann im selben Backwardspass optimieren
-    #   für das ganze forward pass anpassen
     pass