Add flops, num_params, inference speed logging and best.pt logging (#84)

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

ultralytics/yolo/engine/validator.py

@@ -30,6 +30,7 @@ class BaseValidator:
         self.device = None
         self.batch_i = None
         self.training = True
+        self.speed = None
         self.save_dir = save_dir if save_dir is not None else \
                 increment_path(Path(self.args.project) / self.args.name, exist_ok=self.args.exist_ok)
 
@@ -110,12 +111,14 @@ class BaseValidator:
 
         self.print_results()
 
-        # print speeds
-        if not self.training:
+        # calculate speed only once when training
+        if not self.training or trainer.epoch == 0:
             t = tuple(x.t / len(self.dataloader.dataset) * 1E3 for x in dt)  # speeds per image
-            # shape = (self.dataloader.batch_size, 3, imgsz, imgsz)
-            self.logger.info(
-                'Speed: %.1fms pre-process, %.1fms inference, %.1fms loss, %.1fms post-process per image at shape ' % t)
+            self.speed = t
+
+            if not self.training:  # print only at inference
+                self.logger.info(
+                    'Speed: %.1fms pre-process, %.1fms inference, %.1fms loss, %.1fms post-process per image' % t)
 
         if self.training:
             model.float()

ultralytics/yolo/utils/callbacks/clearml.py

@@ -1,3 +1,5 @@
+from ultralytics.yolo.utils.torch_utils import get_flops, get_num_params
+
 try:
     import clearml
     from clearml import Task
@@ -38,8 +40,23 @@ def on_val_end(trainer):
     _log_scalers(val_loss_dict, "val", trainer.epoch)
     _log_scalers(metrics, "metrics", trainer.epoch)
 
+    if trainer.epoch == 0:
+        infer_speed = trainer.validator.speed[1]
+        model_info = {
+            "inference_speed": infer_speed,
+            "flops@640": get_flops(trainer.model),
+            "params": get_num_params(trainer.model)}
+        _log_scalers(model_info, "model")
+
+
+def on_train_end(trainer):
+    task = Task.current_task()
+    if task:
+        task.update_output_model(model_path=str(trainer.best), model_name='Best Model', auto_delete_file=False)
+
 
 callbacks = {
     "before_train": before_train,
     "on_val_end": on_val_end,
-    "on_batch_end": on_batch_end,}
+    "on_batch_end": on_batch_end,
+    "on_train_end": on_train_end}

ultralytics/yolo/utils/torch_utils.py

@@ -125,8 +125,8 @@ def fuse_conv_and_bn(conv, bn):
 
 def model_info(model, verbose=False, imgsz=640):
     # Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320]
-    n_p = sum(x.numel() for x in model.parameters())  # number parameters
-    n_g = sum(x.numel() for x in model.parameters() if x.requires_grad)  # number gradients
+    n_p = get_num_params(model)
+    n_g = get_num_gradients(model)  # number gradients
     if verbose:
         print(f"{'layer':>5} {'name':>40} {'gradient':>9} {'parameters':>12} {'shape':>20} {'mu':>10} {'sigma':>10}")
         for i, (name, p) in enumerate(model.named_parameters()):
@@ -134,18 +134,31 @@ def model_info(model, verbose=False, imgsz=640):
             print('%5g %40s %9s %12g %20s %10.3g %10.3g' %
                   (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))
 
-    try:  # FLOPs
+    flops = get_flops(model, imgsz)
+    fs = f', {flops:.1f} GFLOPs' if flops else ''
+    name = Path(model.yaml_file).stem.replace('yolov5', 'YOLOv5') if hasattr(model, 'yaml_file') else 'Model'
+    LOGGER.info(f"{name} summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}")
+
+
+def get_num_params(model):
+    return sum(x.numel() for x in model.parameters())
+
+
+def get_num_gradients(model):
+    return sum(x.numel() for x in model.parameters() if x.requires_grad)
+
+
+def get_flops(model, imgsz=640):
+    try:
         p = next(model.parameters())
         stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32  # max stride
         im = torch.empty((1, p.shape[1], stride, stride), device=p.device)  # input image in BCHW format
         flops = thop.profile(deepcopy(model), inputs=(im,), verbose=False)[0] / 1E9 * 2  # stride GFLOPs
         imgsz = imgsz if isinstance(imgsz, list) else [imgsz, imgsz]  # expand if int/float
-        fs = f', {flops * imgsz[0] / stride * imgsz[1] / stride:.1f} GFLOPs'  # 640x640 GFLOPs
+        flops = flops * imgsz[0] / stride * imgsz[1] / stride  # 640x640 GFLOPs
+        return flops
     except Exception:
-        fs = ''
-
-    name = Path(model.yaml_file).stem.replace('yolov5', 'YOLOv5') if hasattr(model, 'yaml_file') else 'Model'
-    LOGGER.info(f"{name} summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}")
+        return 0
 
 
 def initialize_weights(model):