Pytorch->Onnx->Tensorrt 8.5.2.2 conversion

aaronrmm commented 9 months ago

The issue

I am trying to get the model to run with tensorrt==8.5.2.2 because that's the highest tensorrt version I can install on Nvidia Jetson Orin. Conversion to onnx works fine. Conversion to tensorrt results in

[02/01/2024-20:28:37] [E] [TRT] ModelImporter.cpp:731: ERROR: builtin_op_importers.cpp:5427 In function importFallbackPluginImporter:
[8] Assertion failed: creator && "Plugin not found, are the plugin name, version, and namespace correct?"

To Reproduce cuda==11.8 tensorrt==8.5.2.2 torch==2.1.0a0+41361538.nv23.6 onnx==1.15.0

The conversion code:

class Model(nn.Module):
    def __init__(
        self,
    ) -> None:
        super().__init__()
        rtdetr.model.training = False
        self.model = rtdetr.model.eval()

    def forward(self, images, orig_target_sizes):
        outputs = self.model(images, orig_target_sizes)
        return outputs

model = Model().eval()

# as per https://github.com/lyuwenyu/RT-DETR/blob/3330eca679a7d7cce16bbb10509099174a2f40bf/rtdetr_pytorch/tools/export_onnx.py#L48C21-L48C21
dynamic_axes = {
    "images": {
        0: "N",
    },
    "orig_target_sizes": {0: "N"},
}

        image_tensor = torch.rand(4, 3, 640, 640).cuda()
        orig_target_sizes = torch.tensor([[640, 640]], dtype=torch.int32).cuda()

        torch.onnx.export(
            model,
            (image_tensor, orig_target_sizes),
            f"temp/rtdetr.onnx",
            input_names=["images", "orig_target_sizes"],
            output_names=["labels", "boxes", "scores", "features"],
            dynamic_axes=dynamic_axes,
            opset_version=17,
            verbose=False,
        )
        command = (
            f"/usr/src/tensorrt/bin/trtexec "
            f"--onnx={conversion_config.temp_dir/'rtdetr.onnx'}  "
            f"--workspace=16096 "
            f"--minShapes=images:1x3x640x640,orig_target_sizes:1x2 "
            f"--optShapes=images:4x3x640x640,orig_target_sizes:4x2 "
            f"--maxShapes=images:4x3x640x640,orig_target_sizes:4x2 "
            f"--saveEngine=output/model.trt"
        )
        result = subprocess.run(
            command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE
        )

Full error /usr/src/tensorrt/bin/trtexec --onnx=/temp/onnx/rtdetr.onnx --workspace=16096 --minShapes=images:1x3x640x640,orig_target_sizes:1x2 --optShapes=images:4x3x640x640,orig_target_sizes:4x2 --maxShapes=images:4x3x640x640,orig_target_sizes:4x2 --saveEngine=/tests/output/model.trt--avgRuns=10--fp16 [02/01/2024-20:28:33] [I] === Model Options === [02/01/2024-20:28:33] [I] Format: ONNX [02/01/2024-20:28:33] [I] Model: /temp/onnx/rtdetr.onnx [02/01/2024-20:28:33] [I] Output: [02/01/2024-20:28:33] [I] === Build Options === [02/01/2024-20:28:33] [I] Max batch: explicit batch [02/01/2024-20:28:33] [I] Memory Pools: workspace: 16096 MiB, dlaSRAM: default, dlaLocalDRAM: default, dlaGlobalDRAM: default [02/01/2024-20:28:33] [I] minTiming: 1 [02/01/2024-20:28:33] [I] avgTiming: 8 [02/01/2024-20:28:33] [I] Precision: FP32 [02/01/2024-20:28:33] [I] LayerPrecisions: [02/01/2024-20:28:33] [I] Calibration: [02/01/2024-20:28:33] [I] Refit: Disabled [02/01/2024-20:28:33] [I] Sparsity: Disabled [02/01/2024-20:28:33] [I] Safe mode: Disabled [02/01/2024-20:28:33] [I] DirectIO mode: Disabled [02/01/2024-20:28:33] [I] Restricted mode: Disabled [02/01/2024-20:28:33] [I] Build only: Disabled [02/01/2024-20:28:33] [I] Save engine: /tests/output/model.trt--avgRuns=10--fp16 [02/01/2024-20:28:33] [I] Load engine: [02/01/2024-20:28:33] [I] Profiling verbosity: 0 [02/01/2024-20:28:33] [I] Tactic sources: Using default tactic sources [02/01/2024-20:28:33] [I] timingCacheMode: local [02/01/2024-20:28:33] [I] timingCacheFile: [02/01/2024-20:28:33] [I] Heuristic: Disabled [02/01/2024-20:28:33] [I] Preview Features: Use default preview flags. [02/01/2024-20:28:33] [I] Input(s)s format: fp32:CHW [02/01/2024-20:28:33] [I] Output(s)s format: fp32:CHW [02/01/2024-20:28:33] [I] Input build shape: images=1x3x640x640+4x3x640x640+4x3x640x640 [02/01/2024-20:28:33] [I] Input build shape: orig_target_sizes=1x2+4x2+4x2 [02/01/2024-20:28:33] [I] Input calibration shapes: model [02/01/2024-20:28:33] [I] === System Options === [02/01/2024-20:28:33] [I] Device: 0 [02/01/2024-20:28:33] [I] DLACore: [02/01/2024-20:28:33] [I] Plugins: [02/01/2024-20:28:33] [I] === Inference Options === [02/01/2024-20:28:33] [I] Batch: Explicit [02/01/2024-20:28:33] [I] Input inference shape: orig_target_sizes=4x2 [02/01/2024-20:28:33] [I] Input inference shape: images=4x3x640x640 [02/01/2024-20:28:33] [I] Iterations: 10 [02/01/2024-20:28:33] [I] Duration: 3s (+ 200ms warm up) [02/01/2024-20:28:33] [I] Sleep time: 0ms [02/01/2024-20:28:33] [I] Idle time: 0ms [02/01/2024-20:28:33] [I] Streams: 1 [02/01/2024-20:28:33] [I] ExposeDMA: Disabled [02/01/2024-20:28:33] [I] Data transfers: Enabled [02/01/2024-20:28:33] [I] Spin-wait: Disabled [02/01/2024-20:28:33] [I] Multithreading: Disabled [02/01/2024-20:28:33] [I] CUDA Graph: Disabled [02/01/2024-20:28:33] [I] Separate profiling: Disabled [02/01/2024-20:28:33] [I] Time Deserialize: Disabled [02/01/2024-20:28:33] [I] Time Refit: Disabled [02/01/2024-20:28:33] [I] NVTX verbosity: 0 [02/01/2024-20:28:33] [I] Persistent Cache Ratio: 0 [02/01/2024-20:28:33] [I] Inputs: [02/01/2024-20:28:33] [I] === Reporting Options === [02/01/2024-20:28:33] [I] Verbose: Disabled [02/01/2024-20:28:33] [I] Averages: 10 inferences [02/01/2024-20:28:33] [I] Percentiles: 90,95,99 [02/01/2024-20:28:33] [I] Dump refittable layers:Disabled [02/01/2024-20:28:33] [I] Dump output: Disabled [02/01/2024-20:28:33] [I] Profile: Disabled [02/01/2024-20:28:33] [I] Export timing to JSON file: [02/01/2024-20:28:33] [I] Export output to JSON file: [02/01/2024-20:28:33] [I] Export profile to JSON file: [02/01/2024-20:28:33] [I] [02/01/2024-20:28:33] [I] === Device Information === [02/01/2024-20:28:33] [I] Selected Device: Orin [02/01/2024-20:28:33] [I] Compute Capability: 8.7 [02/01/2024-20:28:33] [I] SMs: 8 [02/01/2024-20:28:33] [I] Compute Clock Rate: 1.3 GHz [02/01/2024-20:28:33] [I] Device Global Memory: 62800 MiB [02/01/2024-20:28:33] [I] Shared Memory per SM: 164 KiB [02/01/2024-20:28:33] [I] Memory Bus Width: 256 bits (ECC disabled) [02/01/2024-20:28:33] [I] Memory Clock Rate: 0.612 GHz [02/01/2024-20:28:33] [I] [02/01/2024-20:28:33] [I] TensorRT version: 8.5.2 [02/01/2024-20:28:34] [I] [TRT] [MemUsageChange] Init CUDA: CPU +220, GPU +0, now: CPU 249, GPU 11291 (MiB) [02/01/2024-20:28:37] [I] [TRT] [MemUsageChange] Init builder kernel library: CPU +302, GPU +431, now: CPU 574, GPU 11739 (MiB) [02/01/2024-20:28:37] [I] Start parsing network model [02/01/2024-20:28:37] [I] [TRT] ---------------------------------------------------------------- [02/01/2024-20:28:37] [I] [TRT] Input filename: /temp/onnx/rtdetr.onnx [02/01/2024-20:28:37] [I] [TRT] ONNX IR version: 0.0.8 [02/01/2024-20:28:37] [I] [TRT] Opset version: 17 [02/01/2024-20:28:37] [I] [TRT] Producer name: pytorch [02/01/2024-20:28:37] [I] [TRT] Producer version: 2.1.0 [02/01/2024-20:28:37] [I] [TRT] Domain: [02/01/2024-20:28:37] [I] [TRT] Model version: 0 [02/01/2024-20:28:37] [I] [TRT] Doc string: [02/01/2024-20:28:37] [I] [TRT] ---------------------------------------------------------------- [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_1: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_2: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_2: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_1: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_2: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_1: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_2: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_1: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_2: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_1: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_2: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_1: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_2: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] /model/model/encoder/encoder.0/layers.0/self_attn/MatMul_1: broadcasting input1 to make tensors conform, dims(input0)=[400,-1,256][NONE] dims(input1)=[1,256,256][NONE]. [02/01/2024-20:28:37] [I] [TRT] No importer registered for op: LayerNormalization. Attempting to import as plugin. [02/01/2024-20:28:37] [I] [TRT] Searching for plugin: LayerNormalization, plugin_version: 1, plugin_namespace: [02/01/2024-20:28:37] [I] Finish parsing network model &&&& FAILED TensorRT.trtexec [TensorRT v8502] # /usr/src/tensorrt/bin/trtexec --onnx=/temp/onnx/rtdetr.onnx --workspace=16096 --minShapes=images:1x3x640x640,orig_target_sizes:1x2 --optShapes=images:4x3x640x640,orig_target_sizes:4x2 --maxShapes=images:4x3x640x640,orig_target_sizes:4x2 --saveEngine=/tests/output/model.trt--avgRuns=10--fp16 Error:/libs/conversion/rtdetr_pytorch_tensorrt.py - [02/01/2024-20:28:33] [W] --workspace flag has been deprecated by --memPoolSize flag. [02/01/2024-20:28:37] [W] [TRT] onnx2trt_utils.cpp:375: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32. [02/01/2024-20:28:37] [E] [TRT] ModelImporter.cpp:726: While parsing node number 171 [LayerNormalization -> "/model/model/encoder/encoder.0/layers.0/norm1/LayerNormalization_output_0"]: [02/01/2024-20:28:37] [E] [TRT] ModelImporter.cpp:727: --- Begin node --- [02/01/2024-20:28:37] [E] [TRT] ModelImporter.cpp:728: input: "/model/model/encoder/encoder.0/layers.0/Add_1_output_0" input: "model.model.encoder.encoder.0.layers.0.norm1.weight" input: "model.model.encoder.encoder.0.layers.0.norm1.bias" output: "/model/model/encoder/encoder.0/layers.0/norm1/LayerNormalization_output_0" name: "/model/model/encoder/encoder.0/layers.0/norm1/LayerNormalization" op_type: "LayerNormalization" attribute { name: "axis" i: -1 type: INT } attribute { name: "epsilon" f: 1e-05 type: FLOAT } [02/01/2024-20:28:37] [E] [TRT] ModelImporter.cpp:729: --- End node --- [02/01/2024-20:28:37] [E] [TRT] ModelImporter.cpp:731: ERROR: builtin_op_importers.cpp:5427 In function importFallbackPluginImporter: [8] Assertion failed: creator && "Plugin not found, are the plugin name, version, and namespace correct?" [02/01/2024-20:28:37] [E] Failed to parse onnx file [02/01/2024-20:28:37] [E] Parsing model failed [02/01/2024-20:28:37] [E] Failed to create engine from model or file. [02/01/2024-20:28:37] [E] Engine set up failed

lyuwenyu commented 9 months ago

dynamic_axes=dynamic_axes, opset_version=17,

I have never encountered Plugin not found problem for LayerNormalization before.

I think you can try static input firstly, and try to set opset_version=16

aaronrmm commented 9 months ago

Thanks so much! That did it: I set the opset to 16 and left out all the additional arguments, and that worked.

My working configuration for anyone else trying this:

FROM nvcr.io/nvidia/l4t-pytorch:r35.2.1-pth2.0-py3 # base docker image

torch==2.1.0a0+41361538.nv23.6 torchvision==0.15.2 onnx==1.14.0 tensorrt==8.5.2.2 numpy==1.23.1 # had to downgrade numpy for this version of tensorrt

        command = (
            f"/usr/src/tensorrt/bin/trtexec "
            f"--onnx='rtdetr.onnx' "
            f"--saveEngine='rtdetr.trt' "
            f"--best "
        )

IamShubhamGupto commented 8 months ago

@aaronrmm hello! I am facing similar issues, how did you generate the onnx file?

aaronrmm commented 8 months ago

@IamShubhamGupto Try this and let me know if you have any issues. I simplified my code for loading the pytorch model (first two lines) in order to paste it here, but haven't tested.

# modified from https://github.com/lyuwenyu/RT-DETR/blob/3330eca679a7d7cce16bbb10509099174a2f40bf/rtdetr_pytorch/tools/export_onnx.py#L48C21-L48C21

rtdetr_model = torch.load(model_weights_path, map_location=torch.device(self.device))

class Model(nn.Module):
    def __init__(
        self,
    ) -> None:
        super().__init__()
        rtdetr_model.training = False
        self.model = rtdetr_model.eval()

    def forward(self, images, orig_target_sizes):
        outputs = self.model(images, orig_target_sizes)
        return outputs

model = Model().eval()

image_tensor = torch.rand(1, 3, 640, 640).cuda()
orig_target_sizes = torch.tensor([[640, 640]], dtype=torch.int32).cuda()

# modified to use static axes and opset 16
torch.onnx.export(
    model,
    (image_tensor, orig_target_sizes),
    "rtdetr.onnx",
    input_names=["images", "orig_target_sizes"],
    output_names=["labels", "boxes", "scores"],
    opset_version=16,
    verbose=False,
)

Yuxinyi-Qiyu commented 7 months ago

Thanks so much! That did it: I set the opset to 16 and left out all the additional arguments, and that worked.

My working configuration for anyone else trying this:

FROM nvcr.io/nvidia/l4t-pytorch:r35.2.1-pth2.0-py3 # base docker image

torch==2.1.0a0+41361538.nv23.6 torchvision==0.15.2 onnx==1.14.0 tensorrt==8.5.2.2 numpy==1.23.1 # had to downgrade numpy for this version of tensorrt
        command = (
            f"/usr/src/tensorrt/bin/trtexec "
            f"--onnx='rtdetr.onnx' "
            f"--saveEngine='rtdetr.trt' "
            f"--best "
        )

hey, can you show the full code? i meet the same problem: In node 356 (importFallbackPluginImporter): UNSUPPORTED_NODE: Assertion failed: creator && "Plugin not found, are the plugin name, version, and namespace correct?" , i'd like to try to run your code to see whether it works☺️

lyuwenyu / RT-DETR

Pytorch->Onnx->Tensorrt 8.5.2.2 conversion #200