Initial commit

datasets/S3DIS.py

@@ -1191,17 +1191,32 @@ def S3DISCollate(batch_data):
 #       \*********************/
 
 
-def debug_sampling(dataset, sampler, loader):
+def debug_upsampling(dataset, loader):
     """Shows which labels are sampled according to strategy chosen"""
-    label_sum = np.zeros((dataset.num_classes), dtype=np.int32)
+
+
     for epoch in range(10):
 
-        for batch_i, (points, normals, labels, indices, in_sizes) in enumerate(loader):
+        for batch_i, batch in enumerate(loader):
-            # print(batch_i, tuple(points.shape),  tuple(normals.shape), labels, indices, in_sizes)
 
-            label_sum += np.bincount(labels.numpy(), minlength=dataset.num_classes)
+            pc1 = batch.points[1].numpy()
-            print(label_sum)
+            pc2 = batch.points[2].numpy()
-            #print(sampler.potentials[:6])
+            up1 = batch.upsamples[1].numpy()
+
+            print(pc1.shape, '=>', pc2.shape)
+            print(up1.shape, np.max(up1))
+
+            pc2 = np.vstack((pc2, np.zeros_like(pc2[:1, :])))
+
+            # Get neighbors distance
+            p0 = pc1[10, :]
+            neighbs0 = up1[10, :]
+            neighbs0 = pc2[neighbs0, :] - p0
+            d2 = np.sum(neighbs0 ** 2, axis=1)
+
+            print(neighbs0.shape)
+            print(neighbs0[:5])
+            print(d2[:5])
 
             print('******************')
         print('*******************************************')
@@ -1210,7 +1225,7 @@ def debug_sampling(dataset, sampler, loader):
     print(counts)
 
 
-def debug_timing(dataset, sampler, loader):
+def debug_timing(dataset, loader):
     """Timing of generator function"""
 
     t = [time.time()]
@@ -1252,7 +1267,7 @@ def debug_timing(dataset, sampler, loader):
     print(counts)
 
 
-def debug_show_clouds(dataset, sampler, loader):
+def debug_show_clouds(dataset, loader):
 
 
     for epoch in range(10):
@@ -1307,7 +1322,7 @@ def debug_show_clouds(dataset, sampler, loader):
     print(counts)
 
 
-def debug_batch_and_neighbors_calib(dataset, sampler, loader):
+def debug_batch_and_neighbors_calib(dataset, loader):
     """Timing of generator function"""
 
     t = [time.time()]

datasets/common.py

@@ -469,7 +469,8 @@ class PointCloudDataset(Dataset):
             # Reduce size of neighbors matrices by eliminating furthest point
             conv_i = self.big_neighborhood_filter(conv_i, len(input_points))
             pool_i = self.big_neighborhood_filter(pool_i, len(input_points))
-            up_i = self.big_neighborhood_filter(up_i, len(input_points))
+            if up_i.shape[0] > 0:
+                up_i = self.big_neighborhood_filter(up_i, len(input_points)+1)
 
             # Updating input lists
             input_points += [stacked_points]

models/architectures.py

@@ -53,6 +53,7 @@ class KPCNN(nn.Module):
                 break
 
             # Apply the good block function defining tf ops
+            print(block, r)
             self.block_ops.append(block_decider(block,
                                                 r,
                                                 in_dim,
@@ -60,6 +61,7 @@ class KPCNN(nn.Module):
                                                 layer,
                                                 config))
 
+
             # Index of block in this layer
             block_in_layer += 1
 
@@ -329,6 +331,7 @@ class KPFCNN(nn.Module):
             if block_i in self.encoder_skips:
                 skip_x.append(x)
             x = block_op(x, batch)
+            print(block_op)
 
         for block_i, block_op in enumerate(self.decoder_blocs):
             if block_i in self.decoder_concats:

models/blocks.py

@@ -355,6 +355,8 @@ class KPConv(nn.Module):
         # Convolution sum [n_points, out_fdim]
         return torch.sum(kernel_outputs, dim=0)
 
+    def __repr__(self):
+        return "KPConv(radius: %.2f, in_feat: %i, out_feat: %i)"% (self.radius, self.in_channels, self.out_channels)
 
 # ----------------------------------------------------------------------------------------------------------------------
 #

plot_convergence.py

@@ -1374,7 +1374,9 @@ def ModelNet40_deform(old_result_limit):
 
 def S3DIS_first(old_result_limit):
     """
-    Test first S3DIS
+    Test first S3DIS. First two test have all symetries (even vertical), which is not good). We corecct for
+    the following.
+    Then we try some experiments with different input scalea and the results are not as high as expected. WHY?
     """
 
     # Using the dates of the logs, you can easily gather consecutive ones. All logs should be of the same dataset.
@@ -1395,6 +1397,8 @@ def S3DIS_first(old_result_limit):
                   'Fin=5_R=1.2_r=0.02',
                   'Fin=5_R=1.8_r=0.03',
                   'Fin=5_R=2.5_r=0.04',
+                  'original_normal',
+                  'original_deform',
                   'test']
 
     logs_names = np.array(logs_names[:len(logs)])

train_S3DIS.py

@@ -62,7 +62,7 @@ class S3DISConfig(Config):
     dataset_task = ''
 
     # Number of CPU threads for the input pipeline
-    input_threads = 10
+    input_threads = 20
 
     #########################
     # Architecture definition
@@ -70,14 +70,15 @@ class S3DISConfig(Config):
 
     # Define layers
     architecture = ['simple',
-                    'resnetb_strided',
                     'resnetb',
                     'resnetb_strided',
                     'resnetb',
                     'resnetb_strided',
                     'resnetb',
                     'resnetb_strided',
-                    'resnetb',
+                    'resnetb_deformable',
+                    'resnetb_deformable_strided',
+                    'resnetb_deformable',
                     'nearest_upsample',
                     'unary',
                     'nearest_upsample',
@@ -92,7 +93,7 @@ class S3DISConfig(Config):
     ###################
 
     # Radius of the input sphere
-    in_radius = 2.5
+    in_radius = 2.0
 
     # Number of kernel points
     num_kernel_points = 15
@@ -107,7 +108,7 @@ class S3DISConfig(Config):
     deform_radius = 6.0
 
     # Radius of the area of influence of each kernel point in "number grid cell". (1.0 is the standard value)
-    KP_extent = 1.5
+    KP_extent = 1.2
 
     # Behavior of convolutions in ('constant', 'linear', 'gaussian')
     KP_influence = 'linear'
@@ -119,11 +120,11 @@ class S3DISConfig(Config):
     in_features_dim = 5
 
     # Can the network learn modulations
-    modulated = True
+    modulated = False
 
     # Batch normalization parameters
     use_batch_norm = True
-    batch_norm_momentum = 0.05
+    batch_norm_momentum = 0.02
 
     # Offset loss
     # 'permissive' only constrains offsets inside the deform radius (NOT implemented yet)
@@ -145,7 +146,7 @@ class S3DISConfig(Config):
     grad_clip_norm = 100.0
 
     # Number of batch
-    batch_num = 8
+    batch_num = 10
 
     # Number of steps per epochs
     epoch_steps = 500
@@ -163,7 +164,7 @@ class S3DISConfig(Config):
     augment_scale_min = 0.9
     augment_scale_max = 1.1
     augment_noise = 0.001
-    augment_color = 0.9
+    augment_color = 0.8
 
     # The way we balance segmentation loss TODO: implement and test 'class' and 'batch' modes
     #   > 'none': Each point in the whole batch has the same contribution.
@@ -188,6 +189,11 @@ if __name__ == '__main__':
     # Initialize the environment
     ############################
 
+    # TODO: 9 millions de parametres au lieu de 14 millions... Pourquoi?
+    # TODO: radius des strided 2 fois trop grand
+    # TODO: implement un sampler plus simple
+    # TODO: test batch size a 16
+
     # Set which gpu is going to be used
     GPU_ID = '1'
 
@@ -264,9 +270,9 @@ if __name__ == '__main__':
     training_sampler.calibration(training_loader, verbose=True)
     test_sampler.calibration(test_loader, verbose=True)
 
-    #debug_timing(training_dataset, training_sampler, training_loader)
+    #debug_timing(training_dataset, training_loader)
-    #debug_timing(test_dataset, test_sampler, test_loader)
+    #debug_timing(test_dataset, test_loader)
-    #debug_show_clouds(training_dataset, training_sampler, training_loader)
+    #debug_upsampling(training_dataset, training_loader)
 
     print('\nModel Preparation')
     print('*****************')
@@ -274,6 +280,8 @@ if __name__ == '__main__':
     # Define network model
     t1 = time.time()
     net = KPFCNN(config)
+    print(net)
+    print("Model size %i" % sum(param.numel() for param in net.parameters() if param.requires_grad))
 
     # Define a trainer class
     trainer = ModelTrainer(net, config, chkp_path=chosen_chkp)

utils/config.py

@@ -144,7 +144,7 @@ class Config:
     augment_color = 0.7
 
     # Augment with occlusions (not implemented yet)
-    augment_occlusion = 'planar'
+    augment_occlusion = 'none'
     augment_occlusion_ratio = 0.2
     augment_occlusion_num = 1
 
@@ -233,7 +233,7 @@ class Config:
         # Class variable dictionary
         for line in lines:
             line_info = line.split()
-            if len(line_info) > 1 and line_info[0] != '#':
+            if len(line_info) > 2 and line_info[0] != '#':
 
                 if line_info[2] == 'None':
                     setattr(self, line_info[0], None)

utils/trainer.py

@@ -262,7 +262,6 @@ class ModelTrainer:
                     checkpoint_path = join(checkpoint_directory, 'chkp_{:04d}.tar'.format(self.epoch))
                     torch.save(save_dict, checkpoint_path)
 
-
             # Validation
             net.eval()
             self.validation(net, val_loader, config)