Skip to content

tensorrt

sleap_nn.export.predictors.tensorrt

TensorRT predictor for exported models.

Classes:

Name Description
TensorRTEngine

Low-level wrapper for native TensorRT engine files (.trt).
TensorRTPredictor

TensorRT inference for exported models.

TensorRTEngine

Low-level wrapper for native TensorRT engine files (.trt).

Provides a callable interface similar to PyTorch models, returning torch.Tensor outputs directly.

Parameters:

Name Type Description Default
engine_path str | Path

Path to the TensorRT engine file (.trt).

 required
Example

engine = TensorRTEngine("model.trt") input_tensor = torch.randn(1, 1, 512, 512, device="cuda") outputs = engine(input_tensor) # Dict[str, torch.Tensor]

Methods:

Name Description
__call__

Run inference.
__init__

Initialize TensorRT engine wrapper.
Source code in sleap_nn/export/predictors/tensorrt.py 
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
class TensorRTEngine:
    """Low-level wrapper for native TensorRT engine files (.trt).

    Provides a callable interface similar to PyTorch models, returning
    torch.Tensor outputs directly.

    Args:
        engine_path: Path to the TensorRT engine file (.trt).

    Example:
        >>> engine = TensorRTEngine("model.trt")
        >>> input_tensor = torch.randn(1, 1, 512, 512, device="cuda")
        >>> outputs = engine(input_tensor)  # Dict[str, torch.Tensor]
    """

    def __init__(self, engine_path: str | Path) -> None:
        """Initialize TensorRT engine wrapper.

        Args:
            engine_path: Path to the serialized TensorRT engine file.
        """
        import tensorrt as trt

        self.engine_path = Path(engine_path)
        self.logger = trt.Logger(trt.Logger.WARNING)

        # Load engine
        with open(engine_path, "rb") as f:
            self.engine = trt.Runtime(self.logger).deserialize_cuda_engine(f.read())

        if self.engine is None:
            raise RuntimeError(f"Failed to load TensorRT engine: {engine_path}")

        # Create execution context
        self.context = self.engine.create_execution_context()

        # Get input/output info
        self.input_names: List[str] = []
        self.output_names: List[str] = []
        self.input_shapes: Dict[str, tuple] = {}
        self.output_shapes: Dict[str, tuple] = {}

        for i in range(self.engine.num_io_tensors):
            name = self.engine.get_tensor_name(i)
            shape = tuple(self.engine.get_tensor_shape(name))
            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
                self.input_names.append(name)
                self.input_shapes[name] = shape
            else:
                self.output_names.append(name)
                self.output_shapes[name] = shape

    def __call__(self, *args, **kwargs) -> Dict[str, torch.Tensor]:
        """Run inference.

        Args:
            *args: Input tensors (positional) or
            **kwargs: Input tensors by name

        Returns:
            Dict of output tensors on the same device as input.
        """
        import tensorrt as trt

        # Handle inputs
        if args:
            inputs = {self.input_names[i]: arg for i, arg in enumerate(args)}
        else:
            inputs = kwargs

        # Set input shapes and allocate outputs
        outputs: Dict[str, torch.Tensor] = {}
        bindings: Dict[str, int] = {}

        for name in self.input_names:
            tensor = inputs[name]
            if not isinstance(tensor, torch.Tensor):
                raise ValueError(f"Input {name} must be a torch.Tensor")
            # Set actual shape for dynamic dimensions
            self.context.set_input_shape(name, tuple(tensor.shape))
            bindings[name] = tensor.contiguous().data_ptr()

        # Allocate output tensors
        device = next(iter(inputs.values())).device
        for name in self.output_names:
            shape = self.context.get_tensor_shape(name)
            dtype = self._trt_dtype_to_torch(self.engine.get_tensor_dtype(name))
            outputs[name] = torch.empty(tuple(shape), dtype=dtype, device=device)
            bindings[name] = outputs[name].data_ptr()

        # Set tensor addresses
        for name, ptr in bindings.items():
            self.context.set_tensor_address(name, ptr)

        # Run inference
        stream = torch.cuda.current_stream().cuda_stream
        self.context.execute_async_v3(stream)
        torch.cuda.current_stream().synchronize()

        return outputs

    @staticmethod
    def _trt_dtype_to_torch(trt_dtype):
        """Convert TensorRT dtype to PyTorch dtype."""
        import tensorrt as trt

        mapping = {
            trt.DataType.FLOAT: torch.float32,
            trt.DataType.HALF: torch.float16,
            trt.DataType.INT32: torch.int32,
            trt.DataType.INT8: torch.int8,
            trt.DataType.BOOL: torch.bool,
        }
        return mapping.get(trt_dtype, torch.float32)

__call__(*args, **kwargs)

Run inference.

Parameters:

Name Type Description Default
*args

Input tensors (positional) or

 ()
**kwargs

Input tensors by name

 {}

Returns:

Type Description
Dict[str, Tensor]

Dict of output tensors on the same device as input.
Source code in sleap_nn/export/predictors/tensorrt.py 
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
def __call__(self, *args, **kwargs) -> Dict[str, torch.Tensor]:
    """Run inference.

    Args:
        *args: Input tensors (positional) or
        **kwargs: Input tensors by name

    Returns:
        Dict of output tensors on the same device as input.
    """
    import tensorrt as trt

    # Handle inputs
    if args:
        inputs = {self.input_names[i]: arg for i, arg in enumerate(args)}
    else:
        inputs = kwargs

    # Set input shapes and allocate outputs
    outputs: Dict[str, torch.Tensor] = {}
    bindings: Dict[str, int] = {}

    for name in self.input_names:
        tensor = inputs[name]
        if not isinstance(tensor, torch.Tensor):
            raise ValueError(f"Input {name} must be a torch.Tensor")
        # Set actual shape for dynamic dimensions
        self.context.set_input_shape(name, tuple(tensor.shape))
        bindings[name] = tensor.contiguous().data_ptr()

    # Allocate output tensors
    device = next(iter(inputs.values())).device
    for name in self.output_names:
        shape = self.context.get_tensor_shape(name)
        dtype = self._trt_dtype_to_torch(self.engine.get_tensor_dtype(name))
        outputs[name] = torch.empty(tuple(shape), dtype=dtype, device=device)
        bindings[name] = outputs[name].data_ptr()

    # Set tensor addresses
    for name, ptr in bindings.items():
        self.context.set_tensor_address(name, ptr)

    # Run inference
    stream = torch.cuda.current_stream().cuda_stream
    self.context.execute_async_v3(stream)
    torch.cuda.current_stream().synchronize()

    return outputs

__init__(engine_path)

Initialize TensorRT engine wrapper.

Parameters:

Name Type Description Default
engine_path str | Path

Path to the serialized TensorRT engine file.

 required
Source code in sleap_nn/export/predictors/tensorrt.py 
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
def __init__(self, engine_path: str | Path) -> None:
    """Initialize TensorRT engine wrapper.

    Args:
        engine_path: Path to the serialized TensorRT engine file.
    """
    import tensorrt as trt

    self.engine_path = Path(engine_path)
    self.logger = trt.Logger(trt.Logger.WARNING)

    # Load engine
    with open(engine_path, "rb") as f:
        self.engine = trt.Runtime(self.logger).deserialize_cuda_engine(f.read())

    if self.engine is None:
        raise RuntimeError(f"Failed to load TensorRT engine: {engine_path}")

    # Create execution context
    self.context = self.engine.create_execution_context()

    # Get input/output info
    self.input_names: List[str] = []
    self.output_names: List[str] = []
    self.input_shapes: Dict[str, tuple] = {}
    self.output_shapes: Dict[str, tuple] = {}

    for i in range(self.engine.num_io_tensors):
        name = self.engine.get_tensor_name(i)
        shape = tuple(self.engine.get_tensor_shape(name))
        if self.engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
            self.input_names.append(name)
            self.input_shapes[name] = shape
        else:
            self.output_names.append(name)
            self.output_shapes[name] = shape

TensorRTPredictor

Bases: ExportPredictor

TensorRT inference for exported models.

This predictor loads a native TensorRT engine file (.trt) and provides inference capabilities using CUDA for high-throughput predictions.

Parameters:

Name Type Description Default
engine_path str | Path

Path to the TensorRT engine file (.trt).

 required
device str

Device to run inference on (only "cuda" supported for TRT).

 'cuda'
Example

predictor = TensorRTPredictor("model.trt") outputs = predictor.predict(images) # uint8 [B, C, H, W]

Methods:

Name Description
__init__

Initialize TensorRT predictor with a serialized engine.
benchmark

Benchmark inference performance.
predict

Run TensorRT inference.
Source code in sleap_nn/export/predictors/tensorrt.py 
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
class TensorRTPredictor(ExportPredictor):
    """TensorRT inference for exported models.

    This predictor loads a native TensorRT engine file (.trt) and provides
    inference capabilities using CUDA for high-throughput predictions.

    Args:
        engine_path: Path to the TensorRT engine file (.trt).
        device: Device to run inference on (only "cuda" supported for TRT).

    Example:
        >>> predictor = TensorRTPredictor("model.trt")
        >>> outputs = predictor.predict(images)  # uint8 [B, C, H, W]
    """

    def __init__(
        self,
        engine_path: str | Path,
        device: str = "cuda",
    ) -> None:
        """Initialize TensorRT predictor with a serialized engine.

        Args:
            engine_path: Path to the TensorRT engine file.
            device: Device for inference ("cuda" or "auto"). TensorRT requires CUDA.
        """
        import tensorrt as trt

        if device not in ("cuda", "auto"):
            raise ValueError(f"TensorRT only supports CUDA devices, got: {device}")

        self.engine_path = Path(engine_path)
        if not self.engine_path.exists():
            raise FileNotFoundError(f"TensorRT engine not found: {engine_path}")

        self.logger = trt.Logger(trt.Logger.WARNING)

        # Load engine
        with open(self.engine_path, "rb") as f:
            self.engine = trt.Runtime(self.logger).deserialize_cuda_engine(f.read())

        if self.engine is None:
            raise RuntimeError(f"Failed to load TensorRT engine: {engine_path}")

        # Create execution context
        self.context = self.engine.create_execution_context()

        # Get input/output info
        self.input_names: List[str] = []
        self.output_names: List[str] = []
        self.input_shapes: Dict[str, tuple] = {}
        self.output_shapes: Dict[str, tuple] = {}

        for i in range(self.engine.num_io_tensors):
            name = self.engine.get_tensor_name(i)
            shape = tuple(self.engine.get_tensor_shape(name))
            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
                self.input_names.append(name)
                self.input_shapes[name] = shape
            else:
                self.output_names.append(name)
                self.output_shapes[name] = shape

        self.device = torch.device("cuda")

    def predict(self, image: np.ndarray) -> Dict[str, np.ndarray]:
        """Run TensorRT inference.

        Args:
            image: Input image(s) as numpy array [B, C, H, W] with uint8 dtype.

        Returns:
            Dict mapping output names to numpy arrays.
        """
        import tensorrt as trt

        # Convert input to torch tensor on GPU
        input_tensor = torch.from_numpy(image).to(self.device)

        # Check if engine expects uint8 or float32 and convert if needed
        input_name = self.input_names[0]
        expected_dtype = self.engine.get_tensor_dtype(input_name)
        if expected_dtype == trt.DataType.UINT8:
            # Engine expects uint8 - keep as uint8
            if input_tensor.dtype != torch.uint8:
                input_tensor = input_tensor.to(torch.uint8)
        else:
            # Engine expects float - convert uint8 to float32
            if input_tensor.dtype == torch.uint8:
                input_tensor = input_tensor.to(torch.float32)

        # Ensure contiguous memory
        input_tensor = input_tensor.contiguous()

        # Set input shape for dynamic dimensions
        input_name = self.input_names[0]
        self.context.set_input_shape(input_name, tuple(input_tensor.shape))

        # Allocate output tensors
        outputs: Dict[str, torch.Tensor] = {}
        bindings: Dict[str, int] = {}

        # Set input binding
        bindings[input_name] = input_tensor.data_ptr()

        # Allocate outputs
        for name in self.output_names:
            shape = self.context.get_tensor_shape(name)
            dtype = self._trt_dtype_to_torch(self.engine.get_tensor_dtype(name))
            outputs[name] = torch.empty(tuple(shape), dtype=dtype, device=self.device)
            bindings[name] = outputs[name].data_ptr()

        # Set tensor addresses
        for name, ptr in bindings.items():
            self.context.set_tensor_address(name, ptr)

        # Run inference
        stream = torch.cuda.current_stream().cuda_stream
        success = self.context.execute_async_v3(stream)
        torch.cuda.current_stream().synchronize()

        if not success:
            raise RuntimeError("TensorRT inference failed")

        # Convert outputs to numpy
        return {name: tensor.cpu().numpy() for name, tensor in outputs.items()}

    def benchmark(
        self, image: np.ndarray, n_warmup: int = 50, n_runs: int = 200
    ) -> Dict[str, float]:
        """Benchmark inference performance.

        Args:
            image: Input image(s) as numpy array [B, C, H, W].
            n_warmup: Number of warmup runs (not timed).
            n_runs: Number of timed runs.

        Returns:
            Dict with timing statistics:
                - mean_ms: Mean inference time in milliseconds
                - std_ms: Standard deviation of inference time
                - min_ms: Minimum inference time
                - max_ms: Maximum inference time
                - fps: Frames per second (based on mean time and batch size)
        """
        batch_size = image.shape[0]

        # Warmup
        for _ in range(n_warmup):
            _ = self.predict(image)

        # Timed runs
        times = []
        for _ in range(n_runs):
            start = time.perf_counter()
            _ = self.predict(image)
            times.append((time.perf_counter() - start) * 1000)

        times_arr = np.array(times)
        mean_ms = float(np.mean(times_arr))

        return {
            "mean_ms": mean_ms,
            "std_ms": float(np.std(times_arr)),
            "min_ms": float(np.min(times_arr)),
            "max_ms": float(np.max(times_arr)),
            "fps": (batch_size * 1000) / mean_ms if mean_ms > 0 else 0.0,
        }

    @staticmethod
    def _trt_dtype_to_torch(trt_dtype):
        """Convert TensorRT dtype to PyTorch dtype."""
        import tensorrt as trt

        mapping = {
            trt.DataType.FLOAT: torch.float32,
            trt.DataType.HALF: torch.float16,
            trt.DataType.INT32: torch.int32,
            trt.DataType.INT8: torch.int8,
            trt.DataType.BOOL: torch.bool,
        }
        return mapping.get(trt_dtype, torch.float32)

__init__(engine_path, device='cuda')

Initialize TensorRT predictor with a serialized engine.

Parameters:

Name Type Description Default
engine_path str | Path

Path to the TensorRT engine file.

 required
device str

Device for inference ("cuda" or "auto"). TensorRT requires CUDA.

 'cuda'
Source code in sleap_nn/export/predictors/tensorrt.py 
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
def __init__(
    self,
    engine_path: str | Path,
    device: str = "cuda",
) -> None:
    """Initialize TensorRT predictor with a serialized engine.

    Args:
        engine_path: Path to the TensorRT engine file.
        device: Device for inference ("cuda" or "auto"). TensorRT requires CUDA.
    """
    import tensorrt as trt

    if device not in ("cuda", "auto"):
        raise ValueError(f"TensorRT only supports CUDA devices, got: {device}")

    self.engine_path = Path(engine_path)
    if not self.engine_path.exists():
        raise FileNotFoundError(f"TensorRT engine not found: {engine_path}")

    self.logger = trt.Logger(trt.Logger.WARNING)

    # Load engine
    with open(self.engine_path, "rb") as f:
        self.engine = trt.Runtime(self.logger).deserialize_cuda_engine(f.read())

    if self.engine is None:
        raise RuntimeError(f"Failed to load TensorRT engine: {engine_path}")

    # Create execution context
    self.context = self.engine.create_execution_context()

    # Get input/output info
    self.input_names: List[str] = []
    self.output_names: List[str] = []
    self.input_shapes: Dict[str, tuple] = {}
    self.output_shapes: Dict[str, tuple] = {}

    for i in range(self.engine.num_io_tensors):
        name = self.engine.get_tensor_name(i)
        shape = tuple(self.engine.get_tensor_shape(name))
        if self.engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
            self.input_names.append(name)
            self.input_shapes[name] = shape
        else:
            self.output_names.append(name)
            self.output_shapes[name] = shape

    self.device = torch.device("cuda")

benchmark(image, n_warmup=50, n_runs=200)

Benchmark inference performance.

Parameters:

Name Type Description Default
image ndarray

Input image(s) as numpy array [B, C, H, W].

 required
n_warmup int

Number of warmup runs (not timed).

 50
n_runs int

Number of timed runs.

 200

Returns:

Type Description
Dict[str, float]

Dict with timing statistics: - mean_ms: Mean inference time in milliseconds - std_ms: Standard deviation of inference time - min_ms: Minimum inference time - max_ms: Maximum inference time - fps: Frames per second (based on mean time and batch size)
Source code in sleap_nn/export/predictors/tensorrt.py 
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
def benchmark(
    self, image: np.ndarray, n_warmup: int = 50, n_runs: int = 200
) -> Dict[str, float]:
    """Benchmark inference performance.

    Args:
        image: Input image(s) as numpy array [B, C, H, W].
        n_warmup: Number of warmup runs (not timed).
        n_runs: Number of timed runs.

    Returns:
        Dict with timing statistics:
            - mean_ms: Mean inference time in milliseconds
            - std_ms: Standard deviation of inference time
            - min_ms: Minimum inference time
            - max_ms: Maximum inference time
            - fps: Frames per second (based on mean time and batch size)
    """
    batch_size = image.shape[0]

    # Warmup
    for _ in range(n_warmup):
        _ = self.predict(image)

    # Timed runs
    times = []
    for _ in range(n_runs):
        start = time.perf_counter()
        _ = self.predict(image)
        times.append((time.perf_counter() - start) * 1000)

    times_arr = np.array(times)
    mean_ms = float(np.mean(times_arr))

    return {
        "mean_ms": mean_ms,
        "std_ms": float(np.std(times_arr)),
        "min_ms": float(np.min(times_arr)),
        "max_ms": float(np.max(times_arr)),
        "fps": (batch_size * 1000) / mean_ms if mean_ms > 0 else 0.0,
    }

predict(image)

Run TensorRT inference.

Parameters:

Name Type Description Default
image ndarray

Input image(s) as numpy array [B, C, H, W] with uint8 dtype.

 required

Returns:

Type Description
Dict[str, ndarray]

Dict mapping output names to numpy arrays.
Source code in sleap_nn/export/predictors/tensorrt.py 
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
def predict(self, image: np.ndarray) -> Dict[str, np.ndarray]:
    """Run TensorRT inference.

    Args:
        image: Input image(s) as numpy array [B, C, H, W] with uint8 dtype.

    Returns:
        Dict mapping output names to numpy arrays.
    """
    import tensorrt as trt

    # Convert input to torch tensor on GPU
    input_tensor = torch.from_numpy(image).to(self.device)

    # Check if engine expects uint8 or float32 and convert if needed
    input_name = self.input_names[0]
    expected_dtype = self.engine.get_tensor_dtype(input_name)
    if expected_dtype == trt.DataType.UINT8:
        # Engine expects uint8 - keep as uint8
        if input_tensor.dtype != torch.uint8:
            input_tensor = input_tensor.to(torch.uint8)
    else:
        # Engine expects float - convert uint8 to float32
        if input_tensor.dtype == torch.uint8:
            input_tensor = input_tensor.to(torch.float32)

    # Ensure contiguous memory
    input_tensor = input_tensor.contiguous()

    # Set input shape for dynamic dimensions
    input_name = self.input_names[0]
    self.context.set_input_shape(input_name, tuple(input_tensor.shape))

    # Allocate output tensors
    outputs: Dict[str, torch.Tensor] = {}
    bindings: Dict[str, int] = {}

    # Set input binding
    bindings[input_name] = input_tensor.data_ptr()

    # Allocate outputs
    for name in self.output_names:
        shape = self.context.get_tensor_shape(name)
        dtype = self._trt_dtype_to_torch(self.engine.get_tensor_dtype(name))
        outputs[name] = torch.empty(tuple(shape), dtype=dtype, device=self.device)
        bindings[name] = outputs[name].data_ptr()

    # Set tensor addresses
    for name, ptr in bindings.items():
        self.context.set_tensor_address(name, ptr)

    # Run inference
    stream = torch.cuda.current_stream().cuda_stream
    success = self.context.execute_async_v3(stream)
    torch.cuda.current_stream().synchronize()

    if not success:
        raise RuntimeError("TensorRT inference failed")

    # Convert outputs to numpy
    return {name: tensor.cpu().numpy() for name, tensor in outputs.items()}