Lightweight openpose fixes (#242)

* Fixed onnx optimize bugs with half precision and different number of stages

* Fixed height check based on [original repo #202](Daniil-Osokin/lightweight-human-pose-estimation.pytorch#202)

* Update changelog for #242

CHANGELOG.md

@@ -15,10 +15,12 @@ Released on XX, XXth, 2022.
     - `panoptic_segmentation/efficient_ps`: updated dataset preparation scripts to create correct validation ground truth ([#221](https://github.com/opendr-eu/opendr/pull/221)).
     - `panoptic_segmentation/efficient_ps`: added specific configuration files for the provided pretrained models ([#221](https://github.com/opendr-eu/opendr/pull/221)).
     - `c_api/face_recognition`: pass key by const reference in `json_get_key_string()` ([#221](https://github.com/opendr-eu/opendr/pull/221)).
+    - `pose_estimation/lightweight_open_pose`: fixed height check on transformations.py according to original tool repo ([#242](https://github.com/opendr-eu/opendr/pull/242)).
+    - `pose_estimation/lightweight_open_pose`: fixed two bugs where ONNX optimization failed on specific learner parameterization ([#242](https://github.com/opendr-eu/opendr/pull/242)).
   - Dependency Updates:
     - `heart anomaly detection`: upgraded scikit-learn runtime dependency from 0.21.3 to 0.22 ([#198](https://github.com/opendr-eu/opendr/pull/198)).
     - Relaxed all dependencies to allow future versions of non-critical tools to be used ([#201](https://github.com/opendr-eu/opendr/pull/201)).
 
 
 ## Version 1.0
-Released on December 31st, 2021.
+Released on December 31st, 2021.

src/opendr/perception/pose_estimation/lightweight_open_pose/algorithm/datasets/transformations.py

@@ -165,7 +165,7 @@ def __call__(self, sample):
         if image_y_start < 0:
             crop_y_start -= image_y_start
             image_y_start = 0
-        if image_y_start >= w:
+        if image_y_start >= h:
             should_crop = False
 
         if image_x_finish > w:

src/opendr/perception/pose_estimation/lightweight_open_pose/lightweight_open_pose_learner.py

@@ -860,9 +860,14 @@ def __convert_to_onnx(self, output_name, do_constant_folding=False, verbose=Fals
             inp = torch.randn(1, 3, self.base_height, width).cuda()
         else:
             inp = torch.randn(1, 3, self.base_height, width)
+        if self.half:
+            inp = inp.half()
         input_names = ['data']
-        output_names = ['stage_0_output_1_heatmaps', 'stage_0_output_0_pafs',
-                        'stage_1_output_1_heatmaps', 'stage_1_output_0_pafs']
+        if self.num_refinement_stages == 2:
+            output_names = ['stage_0_output_1_heatmaps', 'stage_0_output_0_pafs',
+                            'stage_1_output_1_heatmaps', 'stage_1_output_0_pafs']
+        else:
+            output_names = ['stage_0_output_1_heatmaps', 'stage_0_output_0_pafs']
 
         torch.onnx.export(self.model, inp, output_name, verbose=verbose, enable_onnx_checker=True,
                           do_constant_folding=do_constant_folding, input_names=input_names, output_names=output_names,