Cask (shader run failed)

[DaraOrange]

Hello! I'm trying to convert model to int8. trtexec converts in successfully, but while converting with C++ API and my own calibrator I get the following error. 1: [softMaxV2Runner.cpp::execute::226] Error Code 1: Cask (shader run failed) Polygraphy convert with implemented dataloader.py also gives this error. Could you suppose what could be wrong?

[zerollzeng]

This looks like dup of https://github.com/NVIDIA/TensorRT/issues/3339 which I've already filed internal to track.

[zerollzeng]

Would be great if you can also provide a reproduce too.

[DaraOrange]

I can't provide the code, but I localized the reason of crash - NmsOp from MMDet. However, I don't know how to rewrite it yet. Is that bug fixed?

[DaraOrange]

Basically, is it possible to use nms in tensorrt int8? Could you advice me some working option?

[zerollzeng]

NmsOp from MMDet

is it a standard onnx operator or a customer op?

[DaraOrange]

It is an operator from MMDet library. Now I rewrote it in Pytorch and trying to convert to onnx:

def nms(boxes: array_like_type,
        scores: array_like_type,
        iou_threshold: float,
        offset: int = 0,
        score_threshold: float = 0,
        max_num: int = -1) -> Tuple[array_like_type, array_like_type]:
    assert isinstance(boxes, Tensor)
    assert isinstance(scores, Tensor)
    assert boxes.size(1) == 4
    assert boxes.size(0) == scores.size(0)
    assert offset in (0, 1)

    if score_threshold > 0:
        boxes = boxes[scores > score_threshold]
        scores = scores[scores > score_threshold]

    N = len(boxes)
    max_l = torch.max(boxes[:, 0].unsqueeze(-1).repeat(1, N).flatten(), boxes[:, 0].repeat(N))
    min_r = torch.min(boxes[:, 2].unsqueeze(-1).repeat(1, N).flatten(), boxes[:, 2].repeat(N))
    max_u = torch.max(boxes[:, 1].unsqueeze(-1).repeat(1, N).flatten(), boxes[:, 1].repeat(N))
    min_d = torch.min(boxes[:, 3].unsqueeze(-1).repeat(1, N).flatten(), boxes[:, 3].repeat(N))

    diff_l = (min_r - max_l).reshape((N, N))
    diff_r = (min_d - max_u).reshape((N, N))
    areas = (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1])
    sum_areas = (areas.unsqueeze(-1).repeat(1, N).flatten() +
                 areas.repeat(N)).reshape((N, N))
    iou = (diff_l * diff_r) / (sum_areas - diff_l * diff_r)

    mask = (diff_l > 0).float() * (diff_r > 0).float() * (iou > iou_threshold).float()
    inds = torch.nonzero(mask)
    bad_inds = inds[inds[:,1] < inds[:,0]][:,0].flatten()#.unique()
    mask_inds = torch.ones(N)

    mask_inds[bad_inds] = 0
    inds = torch.nonzero(mask_inds).flatten()

    dets = torch.cat((boxes[inds], scores[inds].reshape(-1, 1)), dim=1)
    return dets, inds

I got the same error in dets = torch.cat((boxes[inds], scores[inds].reshape(-1, 1)), dim=1) or even just boxes[inds]

[DaraOrange]

I prepared a toy example of a mistake in this function.

1. Code of onnx model creation:

import torch
import numpy as np

def nms(boxes,
        scores,
        iou_threshold: float,
        offset: int = 0,
        score_threshold: float = 0,
        max_num: int = -1):
    assert boxes.size(1) == 4
    assert boxes.size(0) == scores.size(0)
    assert offset in (0, 1)

    if score_threshold > 0:
        boxes = boxes[scores > score_threshold]
        scores = scores[scores > score_threshold]

    N = len(boxes)
    max_l = torch.max(boxes[:, 0].unsqueeze(-1).repeat(1, N).flatten(), boxes[:, 0].repeat(N))
    min_r = torch.min(boxes[:, 2].unsqueeze(-1).repeat(1, N).flatten(), boxes[:, 2].repeat(N))
    max_u = torch.max(boxes[:, 1].unsqueeze(-1).repeat(1, N).flatten(), boxes[:, 1].repeat(N))
    min_d = torch.min(boxes[:, 3].unsqueeze(-1).repeat(1, N).flatten(), boxes[:, 3].repeat(N))

    diff_l = (min_r - max_l).reshape((N, N))
    diff_r = (min_d - max_u).reshape((N, N))
    areas = (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1])
    sum_areas = (areas.unsqueeze(-1).repeat(1, N).flatten() +
                 areas.repeat(N)).reshape((N, N))
    iou = (diff_l * diff_r) / (sum_areas - diff_l * diff_r)

    mask = (diff_l > 0).float() * (diff_r > 0).float() * (iou > iou_threshold).float()
    inds = torch.nonzero(mask)
    bad_inds = inds[inds[:,1] < inds[:,0]][:,0].flatten()#.unique()
    mask_inds = torch.ones(N)

    mask_inds[bad_inds] = 0
    inds = torch.nonzero(mask_inds).flatten()

    res = torch.zeros(100, 4)
    boxes_cnt = torch.LongTensor([inds.shape[0], 100]).min()
    res[:boxes_cnt] = boxes[inds][:boxes_cnt]
    return res

class DummyModel(torch.nn.Module):
    def __init__(self):
        super(DummyModel, self).__init__()
        self.w_b = torch.nn.Linear(16, 4)
        self.w_s = torch.nn.Linear(16, 1)

    def forward(self, x):
        x_b = self.w_b(x)
        x_s = self.w_s(x)
        x_s = x_s.sigmoid()
        x_b[:,2] += x_b[:,0]
        x_b[:,3] += x_b[:,1]
        return nms(x_b[0], x_s[0], 0.7, 0, 0)

dummy_boxes = torch.randn((1, 4300, 16)).cuda()
dummy_model = DummyModel().cuda()
torch.onnx.export(dummy_model, dummy_boxes, "model.onnx", verbose=True, 
                  input_names=["x_in"], 
                  output_names=["x_out"], 
                  dynamic_axes=None, opset_version=11)

2. And then I use polygraphy tool with a following dummy dataloader:

import numpy as np

def load_data():
    for i in range(100):
        img = np.random.randn(1, 4300, 16).astype(np.float32)
        yield {"x_in": img}

Polygraphy command:

polygraphy convert model.onnx --int8  --data-loader-script ./dummy_dataloader.py     --calibration-cache calib_dummy.cache -o dummy.engine --pool-limit workspace:2G --verbose

I got the following error (it differs, but proves that there is some bug in tensorrt while such simple nms code conversion)

[E] 1: [executionContext.cpp::commonEmitDebugTensor::1821] Error Code 1: Cuda Runtime (invalid argument)
[E] 3: [engine.cpp::~Engine::289] Error Code 3: API Usage Error (Parameter check failed at: runtime/api/engine.cpp::~Engine::289, condition: mExecutionContextCounter.use_count() == 1. Destroying an engine object before destroying the IExecutionContext objects it created leads to undefined behavior.
    )
[E] 2: [calibrator.cpp::calibrateEngine::1181] Error Code 2: Internal Error (Assertion context->executeV2(&bindings[0]) failed. )
[!] Invalid Engine. Please ensure the engine was built correctly

[Data-Iab]

@DaraOrange for nms, you can replace the operation with this

https://github.com/meituan/YOLOv6/blob/d512ce7c4f103e8887960198505518bed404abdc/yolov6/models/end2end.py#L235

[Steven-stars]

I got the same error and still can't solve it, please help me!

[Steven-stars]

Have you solved this problem now?

[DaraOrange]

I just removed nms layer from my model. I still don't know about good decision :(

[Egorundel]

@DaraOrange Hello, can you help me, please?

https://github.com/NVIDIA/TensorRT/issues/4053