fix(YOLOX) fix bugs in utils and evalutor, add user-friendly const value. (#538)

demo/OpenVINO/cpp/yolox_openvino.cpp

@@ -23,6 +23,7 @@ using namespace InferenceEngine;
 
 static const int INPUT_W = 416;
 static const int INPUT_H = 416;
+static const int NUM_CLASSES = 80; // COCO has 80 classes. Modify this value on your own dataset.
 
 cv::Mat static_resize(cv::Mat& img) {
     float r = std::min(INPUT_W / (img.cols*1.0), INPUT_H / (img.rows*1.0));
@@ -97,7 +98,6 @@ static void generate_grids_and_stride(const int target_size, std::vector<int>& s
 
 static void generate_yolox_proposals(std::vector<GridAndStride> grid_strides, const float* feat_ptr, float prob_threshold, std::vector<Object>& objects)
 {
-    const int num_class = 80;  // COCO has 80 classes. Modify this value on your own dataset.
 
     const int num_anchors = grid_strides.size();
 
@@ -107,7 +107,7 @@ static void generate_yolox_proposals(std::vector<GridAndStride> grid_strides, co
         const int grid1 = grid_strides[anchor_idx].grid1;
         const int stride = grid_strides[anchor_idx].stride;
 
-	const int basic_pos = anchor_idx * 85;
+	const int basic_pos = anchor_idx * (NUM_CLASSES + 5);
 
         // yolox/models/yolo_head.py decode logic
         //  outputs[..., :2] = (outputs[..., :2] + grids) * strides
@@ -120,7 +120,7 @@ static void generate_yolox_proposals(std::vector<GridAndStride> grid_strides, co
         float y0 = y_center - h * 0.5f;
 
         float box_objectness = feat_ptr[basic_pos + 4];
-        for (int class_idx = 0; class_idx < num_class; class_idx++)
+        for (int class_idx = 0; class_idx < NUM_CLASSES; class_idx++)
         {
             float box_cls_score = feat_ptr[basic_pos + 5 + class_idx];
             float box_prob = box_objectness * box_cls_score;

demo/TensorRT/cpp/yolox.cpp

@@ -30,6 +30,7 @@ using namespace nvinfer1;
 // stuff we know about the network and the input/output blobs
 static const int INPUT_W = 640;
 static const int INPUT_H = 640;
+static const int NUM_CLASSES = 80;
 const char* INPUT_BLOB_NAME = "input_0";
 const char* OUTPUT_BLOB_NAME = "output_0";
 static Logger gLogger;
@@ -163,7 +164,6 @@ static void nms_sorted_bboxes(const std::vector<Object>& faceobjects, std::vecto
 
 static void generate_yolox_proposals(std::vector<GridAndStride> grid_strides, float* feat_blob, float prob_threshold, std::vector<Object>& objects)
 {
-    const int num_class = 80;
 
     const int num_anchors = grid_strides.size();
 
@@ -173,7 +173,7 @@ static void generate_yolox_proposals(std::vector<GridAndStride> grid_strides, fl
         const int grid1 = grid_strides[anchor_idx].grid1;
         const int stride = grid_strides[anchor_idx].stride;
 
-        const int basic_pos = anchor_idx * (num_class + 5);
+        const int basic_pos = anchor_idx * (NUM_CLASSES + 5);
 
         // yolox/models/yolo_head.py decode logic
         float x_center = (feat_blob[basic_pos+0] + grid0) * stride;
@@ -184,7 +184,7 @@ static void generate_yolox_proposals(std::vector<GridAndStride> grid_strides, fl
         float y0 = y_center - h * 0.5f;
 
         float box_objectness = feat_blob[basic_pos+4];
-        for (int class_idx = 0; class_idx < num_class; class_idx++)
+        for (int class_idx = 0; class_idx < NUM_CLASSES; class_idx++)
         {
             float box_cls_score = feat_blob[basic_pos + 5 + class_idx];
             float box_prob = box_objectness * box_cls_score;

yolox/evaluators/voc_evaluator.py

@@ -79,7 +79,7 @@ class VOCEvaluator:
 
         inference_time = 0
         nms_time = 0
-        n_samples = len(self.dataloader) - 1
+        n_samples = max(len(self.dataloader) - 1, 1)
 
         if trt_file is not None:
             from torch2trt import TRTModule

yolox/utils/ema.py

@@ -12,12 +12,9 @@ __all__ = ["ModelEMA", "is_parallel"]
 
 def is_parallel(model):
     """check if model is in parallel mode."""
-    import apex
-
     parallel_type = (
         nn.parallel.DataParallel,
         nn.parallel.DistributedDataParallel,
-        apex.parallel.distributed.DistributedDataParallel,
     )
     return isinstance(model, parallel_type)