Avoid renaming unwanted part of the path in dataset loading scripts

ml3d/datasets/kitti.py

@@ -266,9 +266,11 @@ def __len__(self):
 
     def get_data(self, idx):
         pc_path = self.path_list[idx]
-        label_path = pc_path.replace('velodyne',
-                                     'label_2').replace('.bin', '.txt')
-        calib_path = label_path.replace('label_2', 'calib')
+
+        # Replace the last instance of "velodyne" in the path by label_2, and the '.bin' by '.txt'
+        label_path = ("label_2".join(pc_path.rsplit("velodyne", 1))).replace(
+            '.bin', '.txt')
+        calib_path = "calib".join(label_path.rsplit("label_2", 1))
 
         pc = self.dataset.read_lidar(pc_path)
         calib = self.dataset.read_calib(calib_path)

ml3d/datasets/matterport_objects.py

@@ -262,7 +262,8 @@ def __len__(self):
 
     def get_data(self, idx):
         pc_path = self.path_list[idx]
-        label_path = pc_path.replace('pc', 'boxes').replace('.bin', '.txt')
+        label_path = ("boxes".join(pc_path.rsplit("pc",
+                                                  1))).replace('.bin', '.txt')
 
         pc = self.dataset.read_lidar(pc_path)
         label = self.dataset.read_label(label_path)

ml3d/datasets/waymo.py

@@ -250,9 +250,9 @@ def __len__(self):
 
     def get_data(self, idx):
         pc_path = self.path_list[idx]
-        label_path = pc_path.replace('velodyne',
-                                     'label_all').replace('.bin', '.txt')
-        calib_path = label_path.replace('label_all', 'calib')
+        label_path = ("label_all".join(pc_path.rsplit("velodyne", 1))).replace(
+            '.bin', '.txt')
+        calib_path = "calib".join(label_path.rsplit("label_all", 1))
 
         pc = self.dataset.read_lidar(pc_path)
         calib = self.dataset.read_calib(calib_path)

ml3d/torch/models/kpconv.py

@@ -557,7 +557,7 @@ def inference_preprocess(self):
 
         return inputs
 
-    def update_probs(self, inputs, results, test_probs, test_labels):
+    def update_probs(self, inputs, results, test_probs):
         self.test_smooth = 0.95
         stk_probs = torch.nn.functional.softmax(results, dim=-1)
         stk_probs = stk_probs.cpu().data.numpy()
@@ -577,16 +577,14 @@ def update_probs(self, inputs, results, test_probs, test_labels):
         for b_i, length in enumerate(lengths):
             # Get prediction
             probs = stk_probs[i0:i0 + length]
-            labels = np.argmax(probs, 1)
 
             proj_inds = r_inds_list[b_i]
             proj_mask = r_mask_list[b_i]
             test_probs[proj_mask] = self.test_smooth * test_probs[proj_mask] + (
                 1 - self.test_smooth) * probs
-            test_labels[proj_mask] = labels
             i0 += length
 
-        return test_probs, test_labels
+        return test_probs
 
     def inference_end(self, inputs, results):
         m_softmax = torch.nn.Softmax(dim=-1)

ml3d/torch/models/point_transformer.py

@@ -304,14 +304,13 @@ def transform(self, data, attr):
 
         return data
 
-    def update_probs(self, inputs, results, test_probs, test_labels):
+    def update_probs(self, inputs, results, test_probs):
         result = results.reshape(-1, self.cfg.num_classes)
         probs = torch.nn.functional.softmax(result, dim=-1).cpu().data.numpy()
-        labels = np.argmax(probs, 1)
 
         self.trans_point_sampler(patchwise=False)
 
-        return probs, labels
+        return probs
 
     def inference_begin(self):
         data = self.preprocess(data, {'split': 'test'})

ml3d/torch/models/pvcnn.py

@@ -250,14 +250,13 @@ def transform(self, data, attr):
 
         return data
 
-    def update_probs(self, inputs, results, test_probs, test_labels):
+    def update_probs(self, inputs, results, test_probs):
         result = results.reshape(-1, self.cfg.num_classes)
         probs = torch.nn.functional.softmax(result, dim=-1).cpu().data.numpy()
-        labels = np.argmax(probs, 1)
 
         self.trans_point_sampler(patchwise=False)
 
-        return probs, labels
+        return probs
 
     def inference_begin(self, data):
         data = self.preprocess(data, {'split': 'test'})

ml3d/torch/models/randlanet.py

@@ -438,17 +438,16 @@ def inference_end(self, inputs, results):
         else:
             return False
 
-    def update_probs(self, inputs, results, test_probs, test_labels):
+    def update_probs(self, inputs, results, test_probs):
         """Update test probabilities with probs from current tested patch.
 
         Args:
             inputs: input to the model.
             results: output of the model.
             test_probs: probabilities for whole pointcloud
-            test_labels: ground truth for whole pointcloud.
 
         Returns:
-            updated probabilities and labels
+            updated probabilities
 
         """
         self.test_smooth = 0.95
@@ -458,14 +457,12 @@ def update_probs(self, inputs, results, test_probs, test_labels):
             result = torch.reshape(results[b], (-1, self.cfg.num_classes))
             probs = torch.nn.functional.softmax(result, dim=-1)
             probs = probs.cpu().data.numpy()
-            labels = np.argmax(probs, 1)
             inds = inputs['data']['point_inds'][b]
 
             test_probs[inds] = self.test_smooth * test_probs[inds] + (
                 1 - self.test_smooth) * probs
-            test_labels[inds] = labels
 
-        return test_probs, test_labels
+        return test_probs
 
 
 MODEL._register_module(RandLANet, 'torch')

ml3d/torch/pipelines/semantic_segmentation.py

@@ -153,7 +153,6 @@ def run_inference(self, data):
         model.trans_point_sampler = infer_sampler.get_point_sampler()
         self.curr_cloud_id = -1
         self.test_probs = []
-        self.test_labels = []
         self.ori_test_probs = []
         self.ori_test_labels = []
 
@@ -219,7 +218,6 @@ def run_test(self):
         model.trans_point_sampler = test_sampler.get_point_sampler()
         self.curr_cloud_id = -1
         self.test_probs = []
-        self.test_labels = []
         self.ori_test_probs = []
         self.ori_test_labels = []
 
@@ -277,8 +275,6 @@ def update_tests(self, sampler, inputs, results):
             self.test_probs.append(
                 np.zeros(shape=[num_points, self.model.cfg.num_classes],
                          dtype=np.float16))
-            self.test_labels.append(np.zeros(shape=[num_points],
-                                             dtype=np.int16))
             self.complete_infer = False
 
         this_possiblility = sampler.possibilities[sampler.cloud_id]
@@ -287,10 +283,11 @@ def update_tests(self, sampler, inputs, results):
             self.pbar_update)
         self.pbar_update = this_possiblility[
             this_possiblility > end_threshold].shape[0]
-        self.test_probs[self.curr_cloud_id], self.test_labels[
-            self.curr_cloud_id] = self.model.update_probs(
-                inputs, results, self.test_probs[self.curr_cloud_id],
-                self.test_labels[self.curr_cloud_id])
+        self.test_probs[self.curr_cloud_id] = self.model.update_probs(
+            inputs,
+            results,
+            self.test_probs[self.curr_cloud_id],
+        )
 
         if (split in ['test'] and
                 this_possiblility[this_possiblility > end_threshold].shape[0]
@@ -303,10 +300,12 @@ def update_tests(self, sampler, inputs, results):
             if proj_inds is None:
                 proj_inds = np.arange(
                     self.test_probs[self.curr_cloud_id].shape[0])
+            test_labels = np.argmax(
+                self.test_probs[self.curr_cloud_id][proj_inds], 1)
+
             self.ori_test_probs.append(
                 self.test_probs[self.curr_cloud_id][proj_inds])
-            self.ori_test_labels.append(
-                self.test_labels[self.curr_cloud_id][proj_inds])
+            self.ori_test_labels.append(test_labels)
             self.complete_infer = True
 
     def run_train(self):

requirements-tensorflow.txt

@@ -1 +1,2 @@
-tensorflow~=2.8.4
+tensorflow~=2.8.4 ; sys_platform != 'darwin' or platform_machine != 'arm64'
+tensorflow-macos==2.8.0 ; sys_platform == 'darwin' and platform_machine == 'arm64'