Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    Piu
    @swingbalance
    Yap, I follow the instructions in tools/deploy/README.md of Caffe deploy part,
    but error of verification results is large
    Piu
    @swingbalance
    Once I convert .pth model to .onnx model and visualize it by Netron,
    I noticed preprocessing operation(Mean and std) was contained in bottom of model,
    so I'm wondering caffe model also contain preprocessing part in converted model ?
    If so, how could I do
    thanks
    Piu
    @swingbalance
    Is it possible to visualize heatmap(CAM) by uncomment function(get_actmap) in last few line of fastreid/utils/visualizer.py ?
    • fastreid/engine/utils/visualizer.py
    Xingyu Liao
    @L1aoXingyu
    Actually, the preprocessing op can be converted in onnx, but cannot in caffe.
    So I add this preprocess when use caffe model

    Good news,
    I retrained MobileNetV2 after modify relu6 layer to general relu layer, and follow your deploy instructions,
    verify results of caffe and pytorch on tools/deploy/test_data/0022_c6s1_002976_01.jpg,
    cosine similarity is 1.000 and L1 distance is 0.000014,
    other images in tools/deploy/test_data were similar as well

    According to this, that caffe convertor cannot convert relu6 layer. If you replace relu6 with relu, will performance drop?

    Piu
    @swingbalance
    with relu layer, it slightly dropped 1% of rank1 in pytorch model
    I'm working on evaluate by caffe model,
    I want to know the difference of evaluation results(rank1 and mAP) on pytorch and caffe.
    Piu
    @swingbalance
    Is there a faster way to do evaluation if currently I have features npy file which were inferenced by caffe model ?
    You can evaluate the results with distmat, pids and camids using this function directly.
    Piu
    @swingbalance
    image.png
    result seems great
    Xingyu Liao
    @L1aoXingyu
    That's great!
    bixiaopeng
    @bixiaopeng0
    tql
    gzhhong
    @gzhhong
    Hello, I want to try the fast-reid in my project, but I don't have powerful machines to do any training on the model. I want to run the pre-trained model in a docker image and start the docker on my Mac, then develop a small application to send picture or video to fast-reid service and get some result. Is it possible? Is there any existing guidelines for this kind of work? Thanks.
    Xingyu Liao
    @L1aoXingyu
    Yes, this is possible. We will put an example soon.
    gzhhong
    @gzhhong
    That's great, thanks
    gzhhong
    @gzhhong
    @L1aoXingyu , when the docker sample will be ready? I try to build the docker image but it is very hard as the dependency of libraries are very complicated. Thanks for the support!
    Xingyu Liao
    @L1aoXingyu
    Next week we will put the dockerfile on github.
    gzhhong
    @gzhhong
    Thanks a million :)
    gzhhong
    @gzhhong
    Hello, I download the source code and install the related libraries, then I run "bash ./demo/run_demo.sh" from the root folder of the source tree, I get the error "No such file or directory: 'logs/dukemtmc/mgn_R50-ibn/config.yaml'", from the run_demo.sh I can see there is a config "--config-file logs/dukemtmc/mgn_R50-ibn/config.yaml", how can I fix this issue?
    Xingyu Liao
    @L1aoXingyu
    you need to train your model first, then use demo/run_demo.sh to get the features.
    Madhu Ameneni
    @madhuameneni
    This message was deleted
    1 reply
    Levi Pereira
    @levipereira
    First of all, I would like to congratulate you on your wonderful work.
    I am currently using YOLO + Deepsort to perform vehicle tracking. I want to move from Deepsort to FAST-REID.
    I read that FAST-REID has already built in some Vehicle Tracking models.
    I'm studying FAST-REID and from what I understand I don't need to do training to use veri-wild.
    How can I use FAST-REID using the VERI-Wild Baseline model?
    I need an example of how I can implement it.
    Madhu Ameneni
    @madhuameneni

    Hello, I just installed all the packages using docker. When I run the training script "python3 tools/train_net.py --config-file ./configs/VehicleID/bagtricks_R50-ibn.yml MODEL.DEVICE "cuda:0"" and the training freezes at --------> "Selected optimization level O1: Insert automatic casts around Pytorch functions and Tensor methods.

    Defaults for this optimization level are:
    enabled : True
    opt_level : O1
    cast_model_type : None
    patch_torch_functions : True
    keep_batchnorm_fp32 : None
    master_weights : None
    loss_scale : dynamic
    Processing user overrides (additional kwargs that are not None)...
    After processing overrides, optimization options are:
    enabled : True
    opt_level : O1
    cast_model_type : None
    patch_torch_functions : True
    keep_batchnorm_fp32 : None
    master_weights : None
    loss_scale : dynamic
    " ---------------------------------------- Can I know how to solve the issue and what is happening at background?

    Xingyu Liao
    @L1aoXingyu
    @madhuameneni I will provide a dockerfile later and you can try it again.
    @levipereira You can export the pytorch model to pt, onnx or trt runtime, then you can just load the weights file to do inference without fastreid dependency.
    Levi Pereira
    @levipereira

    @levipereira You can export the pytorch model to pt, onnx or trt runtime, then you can just load the weights file to do inference without fastreid dependency.

    Thank you and sorry for previous dumb question. Now I get it, I'm trying to move from Cosine Metric Learning to Fast-REID. DeepSort is a framework that use Cosine Metric Learning models to perform object tracking. So, I have to build a Framework to use FAST-REID models and implement into my project to perfom object tracking.

    Madhu Ameneni
    @madhuameneni
    @L1aoXingyu Thanks for the dockerfile. I think there are some changes to be done after the update given yesterday in the bagtricks_R50-ibn.yml file POOL_LAYER: gempool to replaced with GeneralizedMeanPooling. Sorry if i am wrong.
    gzhhong
    @gzhhong
    @L1aoXingyu thanks for the docker file, that's a big achievement
    wangzhiyuanking
    @wangzhiyuanking
    大家好,请问有对veri-wild数据集处理的吗?处理后想veri-776数据集那样?
    Xingyu Liao
    @L1aoXingyu

    @L1aoXingyu Thanks for the dockerfile. I think there are some changes to be done after the update given yesterday in the bagtricks_R50-ibn.yml file POOL_LAYER: gempool to replaced with GeneralizedMeanPooling. Sorry if i am wrong.

    Yes, u are right.

    Xingyu Liao
    @L1aoXingyu
    @levipereira I know that deepsort is a framework for mot. Actually, u can just replace the feature model with fastreid model, then maintain others unchanged.
    gzhhong
    @gzhhong
    @wangzhiyuanking , 请问您想如何处理,达到什么效果?
    pazyork
    @pazyork
    怎么部署呀
    Madhu Ameneni
    @madhuameneni
    @L1aoXingyu Hello, I am actually training the model with custom data for VeRi, i am facing a issue during evaluation phase where the input["target"] is not getting converted into tensor. So I am facing a issue called :
    File "./fastreid/engine/defaults.py", line 457, in test
    results_i = inference_on_dataset(model, data_loader, evaluator, flip_test=cfg.TEST.FLIP_ENABLED)
    File "./fastreid/evaluation/evaluator.py", line 127, in inference_on_dataset
    evaluator.process(inputs, outputs)
    File "./fastreid/evaluation/reid_evaluation.py", line 45, in process
    'pids': inputs['targets'].to(self._cpu_device) --------------------------------------------------------- my training data is just 3500
    Can you please help me out what might be the issue. Thanks in Advance.
    Xingyu Liao
    @L1aoXingyu
    @madhuameneni When you create your own custom dataset, you need to set the target integer, please refer to this JDAI-CV/fast-reid#220
    @pazyork 部署将 model 转成 trt,然后放到自己的 pipeline 里面
    wangzhiyuanking
    @wangzhiyuanking
    @gzhhong veri-wild数据集所有的图片是包含是images文件夹里面的,格式是000001.jpg,需要根据train_test_split文件对images里面的图片进行划分和重新命名成00006_c078_00000048.jpg,新建文件夹query3000_images 、test3000_images、query5000_images 、test5000_images、query10000_images 、test10000_images,并将重新命名后的图片放到对应新生成的文件夹中,我是reid刚入门,对这个train_test_split处理脚本还有点懵,不知道大佬有veri-wild数据集的处理脚本吗?有的话可以共享一下不
    Madhu Ameneni
    @madhuameneni
    @L1aoXingyu Thanks for the Info.
    Piu
    @swingbalance

    Hello, for CAM visualization, is there any code to visualize features map ?

    In general CAM procedure, it needs fc layer to extract weights for specific category(person ID),
    is there any way to visualize feature map without fc layer ?

    thanks

    Xingyu Liao
    @L1aoXingyu
    @swingbalance Actually, we can just visualize the activation map not CAM.
    gengwb
    @gengwb
    using caffe_export.py I got a caffemodel and prototxt, but the caffe output shape is 1x2048x24x8,the onnx output is 2048x1, I want to know what is the difference bettwen caffe and onnx, am I got the right caffemodel ?
    Levi Pereira
    @levipereira

    @levipereira I know that deepsort is a framework for mot. Actually, u can just replace the feature model with fastreid model, then maintain others unchanged.

    I found this https://github.com/GeekAlexis/FastMOT they are using FastREID

    ViokingTung
    @ViokingTung
    请问最后的demo/demo.py推理结果是npy文件,需要自己怎么解析一下吗?
    Xingyu Liao
    @L1aoXingyu
    使用 np.load('xx.npy') 就可以获得 feature
    ViokingTung
    @ViokingTung
    好的,谢谢