Dynamic batch size support for TensorRT (#8526)

export.py

@@ -216,8 +216,9 @@ def export_coreml(model, im, file, int8, half, prefix=colorstr('CoreML:')):
         return None, None
 
 
-def export_engine(model, im, file, train, half, simplify, workspace=4, verbose=False, prefix=colorstr('TensorRT:')):
+def export_engine(model, im, file, train, half, dynamic, simplify, workspace=4, verbose=False):
     # YOLOv5 TensorRT export https://developer.nvidia.com/tensorrt
+    prefix = colorstr('TensorRT:')
     try:
         assert im.device.type != 'cpu', 'export running on CPU but must be on GPU, i.e. `python export.py --device 0`'
         try:
@@ -230,11 +231,11 @@ def export_engine(model, im, file, train, half, simplify, workspace=4, verbose=F
         if trt.__version__[0] == '7':  # TensorRT 7 handling https://github.com/ultralytics/yolov5/issues/6012
             grid = model.model[-1].anchor_grid
             model.model[-1].anchor_grid = [a[..., :1, :1, :] for a in grid]
-            export_onnx(model, im, file, 12, train, False, simplify)  # opset 12
+            export_onnx(model, im, file, 12, train, dynamic, simplify)  # opset 12
             model.model[-1].anchor_grid = grid
         else:  # TensorRT >= 8
             check_version(trt.__version__, '8.0.0', hard=True)  # require tensorrt>=8.0.0
-            export_onnx(model, im, file, 13, train, False, simplify)  # opset 13
+            export_onnx(model, im, file, 13, train, dynamic, simplify)  # opset 13
         onnx = file.with_suffix('.onnx')
 
         LOGGER.info(f'\n{prefix} starting export with TensorRT {trt.__version__}...')
@@ -263,6 +264,14 @@ def export_engine(model, im, file, train, half, simplify, workspace=4, verbose=F
         for out in outputs:
             LOGGER.info(f'{prefix}\toutput "{out.name}" with shape {out.shape} and dtype {out.dtype}')
 
+        if dynamic:
+            if im.shape[0] <= 1:
+                LOGGER.warning(f"{prefix}WARNING: --dynamic model requires maximum --batch-size argument")
+            profile = builder.create_optimization_profile()
+            for inp in inputs:
+                profile.set_shape(inp.name, (1, *im.shape[1:]), (max(1, im.shape[0] // 2), *im.shape[1:]), im.shape)
+            config.add_optimization_profile(profile)
+
         LOGGER.info(f'{prefix} building FP{16 if builder.platform_has_fast_fp16 and half else 32} engine in {f}')
         if builder.platform_has_fast_fp16 and half:
             config.set_flag(trt.BuilderFlag.FP16)
@@ -460,7 +469,7 @@ def run(
         keras=False,  # use Keras
         optimize=False,  # TorchScript: optimize for mobile
         int8=False,  # CoreML/TF INT8 quantization
-        dynamic=False,  # ONNX/TF: dynamic axes
+        dynamic=False,  # ONNX/TF/TensorRT: dynamic axes
         simplify=False,  # ONNX: simplify model
         opset=12,  # ONNX: opset version
         verbose=False,  # TensorRT: verbose log
@@ -520,7 +529,7 @@ def run(
     if jit:
         f[0] = export_torchscript(model, im, file, optimize)
     if engine:  # TensorRT required before ONNX
-        f[1] = export_engine(model, im, file, train, half, simplify, workspace, verbose)
+        f[1] = export_engine(model, im, file, train, half, dynamic, simplify, workspace, verbose)
     if onnx or xml:  # OpenVINO requires ONNX
         f[2] = export_onnx(model, im, file, opset, train, dynamic, simplify)
     if xml:  # OpenVINO
@@ -579,7 +588,7 @@ def parse_opt():
     parser.add_argument('--keras', action='store_true', help='TF: use Keras')
     parser.add_argument('--optimize', action='store_true', help='TorchScript: optimize for mobile')
     parser.add_argument('--int8', action='store_true', help='CoreML/TF INT8 quantization')
-    parser.add_argument('--dynamic', action='store_true', help='ONNX/TF: dynamic axes')
+    parser.add_argument('--dynamic', action='store_true', help='ONNX/TF/TensorRT: dynamic axes')
     parser.add_argument('--simplify', action='store_true', help='ONNX: simplify model')
     parser.add_argument('--opset', type=int, default=12, help='ONNX: opset version')
     parser.add_argument('--verbose', action='store_true', help='TensorRT: verbose log')

models/common.py

@@ -384,19 +384,24 @@ class DetectMultiBackend(nn.Module):
             logger = trt.Logger(trt.Logger.INFO)
             with open(w, 'rb') as f, trt.Runtime(logger) as runtime:
                 model = runtime.deserialize_cuda_engine(f.read())
+            context = model.create_execution_context()
             bindings = OrderedDict()
             fp16 = False  # default updated below
+            dynamic_input = False
             for index in range(model.num_bindings):
                 name = model.get_binding_name(index)
                 dtype = trt.nptype(model.get_binding_dtype(index))
-                shape = tuple(model.get_binding_shape(index))
+                if model.binding_is_input(index):
+                    if -1 in tuple(model.get_binding_shape(index)):  # dynamic
+                        dynamic_input = True
+                        context.set_binding_shape(index, tuple(model.get_profile_shape(0, index)[2]))
+                    if dtype == np.float16:
+                        fp16 = True
+                shape = tuple(context.get_binding_shape(index))
                 data = torch.from_numpy(np.empty(shape, dtype=np.dtype(dtype))).to(device)
                 bindings[name] = Binding(name, dtype, shape, data, int(data.data_ptr()))
-                if model.binding_is_input(index) and dtype == np.float16:
-                    fp16 = True
             binding_addrs = OrderedDict((n, d.ptr) for n, d in bindings.items())
-            context = model.create_execution_context()
-            batch_size = bindings['images'].shape[0]
+            batch_size = bindings['images'].shape[0]  # if dynamic, this is instead max batch size
         elif coreml:  # CoreML
             LOGGER.info(f'Loading {w} for CoreML inference...')
             import coremltools as ct
@@ -466,7 +471,12 @@ class DetectMultiBackend(nn.Module):
             im = im.cpu().numpy()  # FP32
             y = self.executable_network([im])[self.output_layer]
         elif self.engine:  # TensorRT
-            assert im.shape == self.bindings['images'].shape, (im.shape, self.bindings['images'].shape)
+            if im.shape != self.bindings['images'].shape and self.dynamic_input:
+                self.context.set_binding_shape(self.model.get_binding_index('images'), im.shape)  # reshape if dynamic
+                self.bindings['images'] = self.bindings['images']._replace(shape=im.shape)
+            assert im.shape == self.bindings['images'].shape, (
+                f"image shape {im.shape} exceeds model max shape {self.bindings['images'].shape}" if self.dynamic_input
+                else f"image shape {im.shape} does not match model shape {self.bindings['images'].shape}")
             self.binding_addrs['images'] = int(im.data_ptr())
             self.context.execute_v2(list(self.binding_addrs.values()))
             y = self.bindings['output'].data