fix test_retinanet_detection_output #59813
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
There was a problem hiding this comment.
op_item->num_results() == output_defs.size(), but received op_item->num_results():1 != output_defs.size():0.
这种bug可能有两个原因，这里是因为算子的kernel通过STRUCT kernel方式注册到phi算子体系中，导致phi kernel的解析有问题。解决方法是在paddle/fluid/pir/dialect/operator/utils/utils.cc的LegacyOpList添加新定义的算子名。
在采取了上面的解决方案后，如果依然报相同错误，那就是另一个原因，选择了错误的kernel key，可以参考comment里提及的interfaces的配置，尝试解决。
| @@ -109,6 +109,7 @@ test_elementwise_mul_op | |||
| test_elementwise_pow_op | |||
| test_erf_op | |||
| test_erfinv_op | |||
| test_etinanet_detection_output | |||
There was a problem hiding this comment.
| test_etinanet_detection_output | |
| test_retinanet_detection_output |
| - op: retinanet_detection_output | ||
| args: (Tensor[] bboxes, Tensor[] scores, Tensor[] anchors, Tensor iminfo, float score_threshold, int64_t nms_top_k, float nms_threshold, float nms_eta, int64_t keep_top_k) | ||
| output: Tensor(out) | ||
| infer_meta: | ||
| func: RetinanetDetectionOutputInferMeta | ||
| param: [bboxes, scores, anchors, iminfo] | ||
| kernel: | ||
| func: retinanet_detection_output | ||
| optional: bboxes, scores, anchors |
There was a problem hiding this comment.
| - op: retinanet_detection_output | |
| args: (Tensor[] bboxes, Tensor[] scores, Tensor[] anchors, Tensor iminfo, float score_threshold, int64_t nms_top_k, float nms_threshold, float nms_eta, int64_t keep_top_k) | |
| output: Tensor(out) | |
| infer_meta: | |
| func: RetinanetDetectionOutputInferMeta | |
| param: [bboxes, scores, anchors, iminfo] | |
| kernel: | |
| func: retinanet_detection_output | |
| optional: bboxes, scores, anchors | |
| - op: retinanet_detection_output | |
| args: (Tensor[] bboxes, Tensor[] scores, Tensor[] anchors, Tensor iminfo, float score_threshold, int64_t nms_top_k, float nms_threshold, float nms_eta, int64_t keep_top_k) | |
| output: Tensor(out) | |
| infer_meta: | |
| func: RetinanetDetectionOutputInferMeta | |
| param: [bboxes, scores, anchors, iminfo] | |
| kernel: | |
| func: retinanet_detection_output | |
| data_type : scores | |
| interfaces: RetinanetDetectionOutputOpParseKernelKey |
这里有两个问题跟你确认下：
- 为什么要添加optional字段呢？我看retinanet_detection_output_op.cc里面对这些输入声明了AsDuplicable，但是并没有声明AsDispensable
- interfaces字段的添加，这个是为了迁移原算子体系下RetinanetDetectionOutputOp::GetExpectedKernelType引入的。除了在yaml配置字段，还需要在
paddle/fluid/pir/dialect/operator/interface/interface.cc里定义一个RetinanetDetectionOutputOpParseKernelKey函数，可以参考 [PIR] Add parse kernel key interface #59124 定义的UniqueOpParseKernelKey。
| const MetaTensor& iminfo, | ||
| MetaTensor* out, | ||
| MetaConfig config) { | ||
| std::vector<decltype(bboxes[0]->dims())> bboxes_dims; |
There was a problem hiding this comment.
需要先PADDLE_ENFORCE( bboxes.size() > 0)，原infershape中的PADDLE_ENFORCE也需要迁移，判断输入输出是否存在的除外。
|
Sorry to inform you that 33fb40a's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
luotao1
changed the title
PR types
Others
PR changes
Others
Description
PIR Op单测修复
修复单测 test_retinanet_detection_output 否
第一次，不太会，麻烦大佬看看~
@xingmingyyj @kangguangli
修复后打开FLAGS_enable_pir_in_executor单测是否通过：未测试
（多余的修改，是分支切错了，最后会复原）