d0352276 / csl-yolo Goto Github PK

您好，因为自己是非科班，老师也没经验，自己研究进度缓慢。最近一直在优化yolov4-tiny，关于网络设计方面，想请教几个问题：
（1）如何快速验证网络结构是否更优？我一般用VOC2007训练看结果，但也需要训练五六个小时，感觉试错成本太高
（2）网络在CPU和GPU上运行是有区别的，在CPU上更关注低Flpos和少分支结构，这个理解对吗？
（3）因为网络backbone得到了大量简化，所以yolo-head要采用更多的scale，而不是原始yolov4-tiny的2层scale，这个理解对吗？
（4）yolov4-tiny用最简单的结构实现了很好的效果，而我们人工设计的结构感觉都是在反复试错，您认为呢？
Sincerely, Best wishes

I am running on Jetson Nano 2GB, Jetpack 4.6.1, tensorflow-2.7.0+nv22.1
Here's error message:
[ WARN:0] global /home/nvidia/host/build_opencv/nv_opencv/modules/videoio/src/cap_gstreamer.cpp (933) open OpenCV | GStreamer warning: Cannot query video position: status=0, value=-1, duration=-1
2022-04-17 14:04:31.889463: I tensorflow/stream_executor/cuda/cuda_dnn.cc:377] Loaded cuDNN version 8201
2022-04-17 14:04:33.712375: W tensorflow/core/common_runtime/bfc_allocator.cc:275] Allocator (GPU_0_bfc) ran out of memory trying to allocate 16.07MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2022-04-17 14:04:49.653386: W tensorflow/stream_executor/gpu/asm_compiler.cc:111] *** WARNING *** You are using ptxas 10.2.300, which is older than 11.1. ptxas before 11.1 is known to miscompile XLA code, leading to incorrect results or invalid-address errors.

You may not need to update to CUDA 11.1; cherry-picking the ptxas binary is often sufficient.
2022-04-17 14:04:50.134582: W tensorflow/core/common_runtime/bfc_allocator.cc:275] Allocator (GPU_0_bfc) ran out of memory trying to allocate 16.00MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2022-04-17 14:04:50.184189: W tensorflow/core/common_runtime/bfc_allocator.cc:275] Allocator (GPU_0_bfc) ran out of memory trying to allocate 16.07MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2022-04-17 14:04:50.184357: W tensorflow/core/common_runtime/bfc_allocator.cc:275] Allocator (GPU_0_bfc) ran out of memory trying to allocate 16.08MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2022-04-17 14:05:19.685884: W tensorflow/core/common_runtime/bfc_allocator.cc:275] Allocator (GPU_0_bfc) ran out of memory trying to allocate 16.00MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2022-04-17 14:05:19.776200: W tensorflow/core/common_runtime/bfc_allocator.cc:275] Allocator (GPU_0_bfc) ran out of memory trying to allocate 16.00MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2022-04-17 14:05:19.821245: W tensorflow/core/common_runtime/bfc_allocator.cc:275] Allocator (GPU_0_bfc) ran out of memory trying to allocate 16.00MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2022-04-17 14:05:20.067985: W tensorflow/core/common_runtime/bfc_allocator.cc:275] Allocator (GPU_0_bfc) ran out of memory trying to allocate 16.00MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2022-04-17 14:05:20.114925: W tensorflow/core/common_runtime/bfc_allocator.cc:275] Allocator (GPU_0_bfc) ran out of memory trying to allocate 16.00MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2022-04-17 14:05:20.685095: W tensorflow/core/common_runtime/bfc_allocator.cc:275] Allocator (GPU_0_bfc) ran out of memory trying to allocate 16.00MiB with freed_by_count=0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.
2022-04-17 14:05:26.275044: F tensorflow/core/kernels/image/resize_bilinear_op_gpu.cu.cc:445] Non-OK-status: GpuLaunchKernel(kernel, config.block_count, config.thread_per_block, 0, d.stream(), config.virtual_thread_count, images.data(), height_scale, width_scale, batch, in_height, in_width, channels, out_height, out_width, output.data()) status: INTERNAL: too many resources requested for launch
Aborted (core dumped)

作者你好，我在rtx3090上測試了一下fps，發現只能跑到二三十幀，yolo4tiny能達到一百五六十幀，是不是我設置方面有錯誤？
還有精度，我訓練了自己的數據集，用224224的初始模型，訓練了200個epoch，每個epoch是 batch=16，step=46，因爲訓練集就700多張圖片，但是最後map只有66%，我用tiny能達到94%，後來我用512512的跑了50個epoch準確率直接為0。。。標簽方面位置信息是雙精度的，這一點可能和你給的有差別，其他應該都差不多

Hello, this is solid work.
Can you please provide the code related to the anchor constraint before Kmeans? I can't find it in the project.

我当前有训练集和验证集，具体格式如下图所示，如何转成CSP-YOLO可训练的数据格式？

其中，txt格式如下所示
class_id cx/w cy/h box_w/w box_h/h
0 0.510814878287011 0.33199194073677063 0.04678068331669093 0.12072434276342392

Traceback (most recent call last):
File "/usr/lib/python3.6/threading.py", line 916, in _bootstrap_inner
self.run()
File "/usr/lib/python3.6/threading.py", line 864, in run
self._target(*self._args, **self._kwargs)
File "/project/train/src_repo/CSL-YOLO/tools/thread_pool.py", line 22, in _Thread
if(parameter==None):result=function()
File "/project/train/src_repo/CSL-YOLO/data_generator.py", line 180, in _ReadFunction
batch_data=self._data_gen()
File "/project/train/src_repo/CSL-YOLO/data_generator.py", line 31, in call
return self.Read()
File "/project/train/src_repo/CSL-YOLO/data_generator.py", line 142, in Read
img_1,bboxes_1=self._GetImgAndBboxes(file_name_1)
File "/project/train/src_repo/CSL-YOLO/data_generator.py", line 46, in _GetImgAndBboxes
bboxes=JSON2Bboxes(json_path)
File "/project/train/src_repo/CSL-YOLO/tools/bboxes_io.py", line 4, in JSON2Bboxes
with open(json_path,"r") as json_fin:
FileNotFoundError: [Errno 2] No such file or directory: './dataset/coco/train/json/.json'

By the way, the description about how to train is wrong

hello,you say your data format is different from the official format of MS_COCO，so how can I get JSON files like you?
thanks!