Comments (11)
Hi ,
Thanks for your feedback , I'll try it next week , here is my guess to solve this problem :
- to get a good pre-train weights
- solver type change to sgd
from mobilenet-yolo.
I first try to train original model and using pretrain coco weights (conver from original website , tutorial ), and mAP was 0.72 , deploy model just uploaded , there are some reasons mAP be lower than original yolo :
- Original concat layer fixed shape when layer setup , so training can not use adaptive image size, like 320~608
- I can't convert voc weights to caffemodel , I think there are some different between these frameworks
- etc ..
And if have time and machine , I'll try to optimize it.
from mobilenet-yolo.
Hi,
Thanks for your reply.
I have tried some modifications on the darknet19+yolov2 architecture.
Can you briefly describe how to test the mAP after training?
I used the following command but only returned some weird information...
build/tools/caffe test --model yolo_darknet_test.prototxt --weights yolo_darknet_solver_iter_80000.caffemodel --iterations 4952 --gpu 0
from mobilenet-yolo.
I usually use to set solver max_iter = 0 , and start training . It should start to test
Note : do not modify batch size of test prototxt , currently , the yolo eval function do not support batch size > 1
from mobilenet-yolo.
Actually, I have tried so...
But got no response...
from mobilenet-yolo.
Below is my solver setting , and it works ,
I think if it do not show mAP , maybe eval layer didn't receive any detection results . It happened when used mismatch prototxt and model
Note : I update the darknet model today
train_net: "models/darknet/darknet19_train.prototxt"
test_net: "models/darknet/darknet19_test.prototxt"
test_iter: 4952
test_interval: 1000
base_lr: 0.001
display: 10
max_iter: 0
lr_policy: "multistep"
gamma: 0.1
weight_decay: 0.00005
snapshot: 1000
snapshot_prefix: "models/darknet/yolov2_deploy"
solver_mode: GPU
debug_info: false
snapshot_after_train: true
test_initialization: false
average_loss: 10
stepvalue: 60000
stepvalue: 80000
iter_size: 2
type: "SGD"
eval_type: "detection"
ap_version: "11point"
show_per_class_result: true
script :
./tools/caffe train --solver models/darknet/darknet19_solver.prototxt --weights models/darknet/yolov2_deploy_iter_50000.caffemodel --gpu=0 2>&1 | tee $LOG
from mobilenet-yolo.
Hi,
I trained the darknet-yolov2 with the pre-trained weights(darknet-19, 448 in imagenet, not coco) released here, and convert to caffe format with this repo.
I have tried many kinds of parameter setting but the best mAP that I can reach is less than 60%, which is much lower than 76%. I have sent you the model to you via your gmail, can you help inspect what is going wrong in my training procedure?
from mobilenet-yolo.
I'm not sure , but I don't have machine can try to training now , and below is my guess :
-
Transform param in test.prototxt was
scale: 0.007843
mean_value: 127.5
mean_value: 127.5
mean_value: 127.5But in the train.prototxt was gone in your prototxt , and I suggest scale set to 1/255.5 and mean set to
zero -
I try to convert darknet.cov23 to caffemodel yet , but count was dismatch, I didn't solve this problem ,
and I can not ensure the caffemodel was correct
from mobilenet-yolo.
Hi,
Sorry to bother you again...
-
I found that I sent you the wrong train.prototxt. It does have transform param settings in my experiments, actually. The darknet pre-processing scales the pixel value into [-1,+1], so I think it is right to set the transform param as:
scale: 0.007843
mean_value: 127.5
mean_value: 127.5
mean_value: 127.5 -
I think the converted weights with this repo is correct for I have tried the converted weights for other architectures and it works.
Thank you
from mobilenet-yolo.
Ok , Maybe I'll need train this model and try to find the problems .
Thanks for your feedback
from mobilenet-yolo.
I update a new mode training from darknet19.conv , can only get 0.71
https://github.com/eric612/Caffe-YOLOv2-Windows/tree/master/models/darknet
from mobilenet-yolo.
Related Issues (20)
- About deepwise_conv_layer.cu,when bias_term = false,somting is wrong
- How do I load images instead of LMDB? HOT 8
- how to do inference on multiple images as one batch? HOT 1
- Best weight in training
- cuda版本的配置在哪个文件?Check failed: error == cudaSuccess (803 vs. 0) system has unsupported display driver / cuda driver combination HOT 2
- *** Error in `./build/tools/caffe': double free or corruption (!prev): 0x00007f5e15685ae0 *** HOT 1
- mobilenet_yolov3_lite does not converge HOT 1
- 关于Pelee + SSD + driverable maps
- How to test my own model ? HOT 5
- mobilenetv2_yolov3 caffe2ncnn转出的模型不能运行 HOT 4
- 多个gpu训练,卡住不动,gpu占用率为0 HOT 2
- 希望可以分享一下darknet-yolov3的模型,感谢! HOT 7
- 合并BN后,在GPU下循环测试同一张图片,输出不一致!!! HOT 1
- 能否添加clip层 HOT 3
- 您好,能否给一下您的联系方式,有问题想请教您
- question about add parameter yolo_data_type in batch_sampler?
- YOLOv3-quadrangle
- I think some codes in yolov3_layer.cpp is unreasonable. Should revise it?
- Question: finetune 13 class with custom dataset
- base_conv_layer.cpp:121] Check failed: channels_ % group_ == 0 (1 vs. 0) HOT 1
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from mobilenet-yolo.