renjieChao

I need the version number as detailed as possible! like Cython==0.27.3,but how about python,pytorch ?I don't know! thanks!!!

cpu only

1

I want to use it with pytorch-cuponly ,what should I do? thanks!!

I need the version number as detailed as possible! like Cython==0.27.3,but how about python,pytorch ?I don't know! thanks!!!

单阶段目标检测算法中的 box jittering

3

如果我想把预测网络换成单阶段网络,如yolo系列,在box jittering部分如何做的呢?请问有什么思路吗

同样的错误loss_cls?

4

我训练的时候同样遇到了KeyError: 'loss_cls' ,我因此在报错前面print了loss,发现在报错之前是正常的,但是报错的时候,loss少了loss_cls. 这是我print的地方:ssod/models/soft_teacher.py:243-245 loss = self.student.roi_head.bbox_head.loss( bbox_results["cls_score"], bbox_results["bbox_pred"], rois, *bbox_targets, reduction_override="none", ) print("~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~`") print(loss) print("~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~`") loss["loss_cls"] = loss["loss_cls"].sum() / max(bbox_targets[1].sum(), 1.0) loss["loss_bbox"] = loss["loss_bbox"].sum() / max( bbox_targets[1].size()[0], 1.0 )...

for循环的作用?

1

for fold in 1 2 3 4 5; do bash ... 这个指令执行了五次bash也就是后面的半监督训练了五次,这个作用是什么呢?,我可以只训练一次么,因为时间太久了 for FOLD in 1 2 3 4 5; do bash tools/dist_train_partially.sh semi ${FOLD} 10 1; done 改成 for...

我想在纯cpu上训练推理，怎么样设置呢？ 或者把代码改成cpu上的思路可以大致说一下么

rknn_inputs_set, param input size(1228800) < model input size(4915200)

2

`在rknn_inputs_set处出现错误，`ret = rknn_inputs_set(ctx, io_num.n_input, `inputs);` 128800是640,640,3的tensor，而4915200是12800,12800,3的tensor,我的rknn模型在连板调试中使用Python推理正确，且输入确实是640的，但是这里提醒我模型期望输入却是4915200(1280,1280,3)，不知道为什么，当我讲图片resize成1280后，又出现了新的错误，如下图，我想知道函数rknn_inputs_set中返回的模型期望输入是从哪里计算出来的？因为即使 ret = rknn_query(ctx, RKNN_QUERY_INPUT_ATTR, &(input_attrs[i]), sizeof(rknn_tensor_attr));查询模型输入，返回的也依旧是640,640,3  

why onnx.load("resnet50_v1.pth") when I use pytorch backend?

11

is not onnx.load() with args "xxx.onnx"? vision/classification_and_detection/python/backend_pytorch.py line 34:self.model = onnx.load(model_path) when I run "./run_local.sh pytorch resnet50 cpu" the backend_pytorch.py is called,but model_path is "resnet50_v1.pth" and it caused bug onnx...

[QUESTION] IS zero bubble pp support flash-atten?

1

I SEE zero-bubble-pipeline-parallelism disabled FusdLayerNorm，Is it because of the fused op can not split backward of w and x？