Paddle3D
Paddle3D copied to clipboard
PaddlePaddle
SMOKE模型修改相机参数Ks后，转静态模型出错
由于paddle模型转ONNX不支持inverse操作，因此将smoker_coder.py 186行的 Ks_inv = Ks.inverse()[obj_id] 修改成了Ks_inv=paddle.to_tensor([[[0.00138593 , 0. , -0.84480590],[0. , 0.00138593 , -0.23956336],[0. ,0. ,1.]]], dtype='float32')[obj_id]，如图一所示。
之后使用导出工具导出模型时发生了错误，错误如图2。请问应该怎么解决。
@Truebazinga 刚刚看了下，之前改错代码了，确实会有这个问题，这个是因为Ks这个输入Tensor由于没有被用到，在保存推理模型时被错误裁剪掉导致，可以修改成以下代码：
- 直接使用Ks，不用Ks.inverse()
- 推理的时候，模型输入的Ks是相机内参的逆矩阵
def decode_location_without_transmat(self,
points,
points_offset,
depths,
Ks,
down_ratios=None):
"""
retrieve objects location in camera coordinate based on projected points
Args:
points: projected points on feature map in (x, y)
points_offset: project points offset in (delata_x, delta_y)
depths: object depth z
Ks: camera intrinsic matrix, shape = [N, 3, 3]
trans_mats: transformation matrix from image to feature map, shape = [N, 3, 3]
Returns:
locations: objects location, shape = [N, 3]
"""
if down_ratios is None:
down_ratios = [(1, 1)]
# number of points
N = points_offset.shape[0]
# batch size
N_batch = Ks.shape[0]
#batch_id = paddle.arange(N_batch).unsqueeze(1)
batch_id = paddle.arange(N_batch).reshape((N_batch, 1))
# obj_id = batch_id.repeat(1, N // N_batch).flatten()
obj_id = batch_id.tile([1, N // N_batch]).flatten()
Ks_inv = Ks[obj_id]
down_ratio = down_ratios[0]
points = paddle.reshape(points, (numel_t(points) // 2, 2))
proj_points = points + points_offset
# trans point from heatmap to ori image, down_sample * resize_scale
proj_points[:, 0] = down_ratio[0] * proj_points[:, 0]
proj_points[:, 1] = down_ratio[1] * proj_points[:, 1]
# transform project points in homogeneous form.
proj_points_extend = paddle.concat(
[proj_points, paddle.ones((N, 1))], axis=1)
# expand project points as [N, 3, 1]
proj_points_extend = proj_points_extend.unsqueeze(-1)
# with depth
proj_points_img = proj_points_extend * paddle.reshape(
depths, (N, numel_t(depths) // N, 1))
# transform image coordinates back to object locations
locations = paddle.matmul(Ks_inv, proj_points_img)
return locations.squeeze(2)
他为什么多了这么多不支持的算子吖？ paddle2onnx --model_dir exported_model --model_filename inference.pdmodel --params_filename inference.pdiparams --save_file ./outputs/model.onnx --enable_dev_version True [Paddle2ONNX] Start to parse PaddlePaddle model... [Paddle2ONNX] Model file path: exported_model/inference.pdmodel [Paddle2ONNX] Paramters file path: exported_model/inference.pdiparams [Paddle2ONNX] Start to parsing Paddle model... [Paddle2ONNX] Oops, there are some operators not supported yet, including elementwise_floordiv,gather_nd,group_norm,index_select,inverse,p_norm, [ERROR] Due to the unsupported operators, the conversion is aborted. Aborted (core dumped)
