Add timing cache to accelerate consequent .engine export (#13386)

* fix: typos

* feat: enable timing cache for engine export

* Auto-format by https://ultralytics.com/actions

---------

Co-authored-by: UltralyticsAssistant <web@ultralytics.com>

export.py

@@ -593,7 +593,9 @@ def export_coreml(model, im, file, int8, half, nms, mlmodel, prefix=colorstr("Co
 
 
 @try_export
-def export_engine(model, im, file, half, dynamic, simplify, workspace=4, verbose=False, prefix=colorstr("TensorRT:")):
+def export_engine(
+    model, im, file, half, dynamic, simplify, workspace=4, verbose=False, cache="", prefix=colorstr("TensorRT:")
+):
     """
     Export a YOLOv5 model to TensorRT engine format, requiring GPU and TensorRT>=7.0.0.
 
@@ -606,6 +608,7 @@ def export_engine(model, im, file, half, dynamic, simplify, workspace=4, verbose
         simplify (bool): Set to True to simplify the model during export.
         workspace (int): Workspace size in GB (default is 4).
         verbose (bool): Set to True for verbose logging output.
+        cache (str): Path to save the TensorRT timing cache.
         prefix (str): Log message prefix.
 
     Returns:
@@ -660,6 +663,11 @@ def export_engine(model, im, file, half, dynamic, simplify, workspace=4, verbose
         config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, workspace << 30)
     else:  # TensorRT versions 7, 8
         config.max_workspace_size = workspace * 1 << 30
+    if cache:  # enable timing cache
+        Path(cache).parent.mkdir(parents=True, exist_ok=True)
+        buf = Path(cache).read_bytes() if Path(cache).exists() else b""
+        timing_cache = config.create_timing_cache(buf)
+        config.set_timing_cache(timing_cache, ignore_mismatch=True)
     flag = 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
     network = builder.create_network(flag)
     parser = trt.OnnxParser(network, logger)
@@ -688,6 +696,9 @@ def export_engine(model, im, file, half, dynamic, simplify, workspace=4, verbose
     build = builder.build_serialized_network if is_trt10 else builder.build_engine
     with build(network, config) as engine, open(f, "wb") as t:
         t.write(engine if is_trt10 else engine.serialize())
+    if cache:  # save timing cache
+        with open(cache, "wb") as c:
+            c.write(config.get_timing_cache().serialize())
     return f, None
 
 
@@ -1277,6 +1288,7 @@ def run(
     int8=False,  # CoreML/TF INT8 quantization
     per_tensor=False,  # TF per tensor quantization
     dynamic=False,  # ONNX/TF/TensorRT: dynamic axes
+    cache="",  # TensorRT: timing cache path
     simplify=False,  # ONNX: simplify model
     mlmodel=False,  # CoreML: Export in *.mlmodel format
     opset=12,  # ONNX: opset version
@@ -1306,6 +1318,7 @@ def run(
         int8 (bool): Apply INT8 quantization for CoreML or TensorFlow models. Default is False.
         per_tensor (bool): Apply per tensor quantization for TensorFlow models. Default is False.
         dynamic (bool): Enable dynamic axes for ONNX, TensorFlow, or TensorRT exports. Default is False.
+        cache (str): TensorRT timing cache path. Default is an empty string.
         simplify (bool): Simplify the ONNX model during export. Default is False.
         opset (int): ONNX opset version. Default is 12.
         verbose (bool): Enable verbose logging for TensorRT export. Default is False.
@@ -1341,6 +1354,7 @@ def run(
             int8=False,
             per_tensor=False,
             dynamic=False,
+            cache="",
             simplify=False,
             opset=12,
             verbose=False,
@@ -1378,7 +1392,8 @@ def run(
     # Input
     gs = int(max(model.stride))  # grid size (max stride)
     imgsz = [check_img_size(x, gs) for x in imgsz]  # verify img_size are gs-multiples
-    im = torch.zeros(batch_size, 3, *imgsz).to(device)  # image size(1,3,320,192) BCHW iDetection
+    ch = next(model.parameters()).size(1)  # require input image channels
+    im = torch.zeros(batch_size, ch, *imgsz).to(device)  # image size(1,3,320,192) BCHW iDetection
 
     # Update model
     model.eval()
@@ -1402,7 +1417,7 @@ def run(
     if jit:  # TorchScript
         f[0], _ = export_torchscript(model, im, file, optimize)
     if engine:  # TensorRT required before ONNX
-        f[1], _ = export_engine(model, im, file, half, dynamic, simplify, workspace, verbose)
+        f[1], _ = export_engine(model, im, file, half, dynamic, simplify, workspace, verbose, cache)
     if onnx or xml:  # OpenVINO requires ONNX
         f[2], _ = export_onnx(model, im, file, opset, dynamic, simplify)
     if xml:  # OpenVINO
@@ -1497,6 +1512,7 @@ def parse_opt(known=False):
     parser.add_argument("--int8", action="store_true", help="CoreML/TF/OpenVINO INT8 quantization")
     parser.add_argument("--per-tensor", action="store_true", help="TF per-tensor quantization")
     parser.add_argument("--dynamic", action="store_true", help="ONNX/TF/TensorRT: dynamic axes")
+    parser.add_argument("--cache", type=str, default="", help="TensorRT: timing cache file path")
     parser.add_argument("--simplify", action="store_true", help="ONNX: simplify model")
     parser.add_argument("--mlmodel", action="store_true", help="CoreML: Export in *.mlmodel format")
     parser.add_argument("--opset", type=int, default=17, help="ONNX: opset version")

train.py

@@ -717,10 +717,10 @@ def main(opt, callbacks=Callbacks()):
             "perspective": (True, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
             "flipud": (True, 0.0, 1.0),  # image flip up-down (probability)
             "fliplr": (True, 0.0, 1.0),  # image flip left-right (probability)
-            "mosaic": (True, 0.0, 1.0),  # image mixup (probability)
+            "mosaic": (True, 0.0, 1.0),  # image mosaic (probability)
             "mixup": (True, 0.0, 1.0),  # image mixup (probability)
-            "copy_paste": (True, 0.0, 1.0),
+            "copy_paste": (True, 0.0, 1.0),  # segment copy-paste (probability)
-        }  # segment copy-paste (probability)
+        }
 
         # GA configs
         pop_size = 50