Closed Tord-Zhang closed 6 years ago
Hi @mangdian , After clone the ECO-pytorch repo, you can download the models with the following command:
sh models/download_models.sh
sorry,there are some errors for using download_models.sh requests.exceptions.ConnectionError: HTTPSConnectionPool(host='docs.google.com', port=443): Max retries exceeded with url: /uc?export=download&id=1QffeXdoZYhPEEGXv4FT6Aicu0Hmi2o76 (Caused by NewConnectionError('<urllib3.connection.VerifiedHTTPSConnection object at 0x7f70f8d7d250>: Failed to establish a new connection: [Errno 101] Network is unreachable',))
Can you give me the models you have downloaded? thank you so much
Hi @shajie17 , You can download the pretrained models here: https://pan.baidu.com/s/1Hx52akJLR_ISfX406bkIog
hello @zhang-can , downloading the model from 'sh models/download_models.sh' gives a file named 'eco_lite_rgb_16F_kinetics_v1.pth.tar' , extracting that gives another zipped file.
I tried downloading the model from https://pan.baidu.com/s/1Hx52akJLR_ISfX406bkIog
It downloads a .dmg file - do I have to open that file? I can't tell because it's in chinese
what file extension is the trained model in? and how do we use it for inference on some wild videos?
thanks,
@zhang-can thank you for your response. I found that https://github.com/mzolfaghari/ECO-pytorch only provides the pretrained model on Kinetics. But I need the pretrained model on Something Something? Do you have the pytorch model on Something Something? I would really appreciate it.
thank you very much for your help @zhang-can ,but I want to use ECO_Lite_kinetics.caffemodel and ECO_Lite_UCF101.caffemodel in caffe to do online_recognition.py Can you give me these ,thank you
hello @mangdian , can you walk me through the steps you followed to download the pretrained model on Kinetics from https://github.com/mzolfaghari/ECO-efficient-video-understanding/issues/url.
I ran 'sh models/download_models.sh', but got a file named 'eco_lite_rgb_16F_kinetics_v1.pth.tar' that only unzips to further zipped files, am i doing something wrong?
@zhang-can As reported in the paper of ECO, pretrained 2D BNInception and 3D resnet-18 models on Kinetics dataset are not enough to get a good result, training the ECO model for another 10 epochs on Kinectics would promise a better result. However, I am really not able to train the model on Kinectics(GPU and memory limitation). Since you mentioned that you are training the models on Kinetics, would you please share the trained model? I am going to use the trained weights to initialize the model and train it on something something dataset. I can report the testing result and share it in this repository.
@sophia-wright-blue you do not need to unzip that downloaded file.
hello,@sophia-wright-blue Do you have ECO_Lite_kinetics.caffemodel and ECO_Lite_UCF101.caffemodel used in online_recognition.py
hi @shajie17 , I'm afraid I don't have ECO_Lite_kinetics.caffemodel and ECO_Lite_UCF101.caffemodel, I'm only interested in the PyTorch models. I guess you're gonna have to reach out to @zhang-can or @mzolfaghari for the PyTorch equivalent of
thank you for replying @mangdian , I'm a little confused as to how to use the downloaded trained model - 'eco_lite_rgb_16F_kinetics_v1.pth.tar'
I would like to use the trained model for inference on wild videos (mp4 files), or a live feed, so I guess this relates to the comment by @shajie17 and I would need a PyTorch equivalent of
@sophia-wright-blue After the trained model is downloaded, actually you do not need to do anything. The model is pretrained model and has not been trained well, I guess it can't be used for practical use yet. You can finetune the model on your dataset. The online version of ECO based on PyTorch has not been implemented yet.
@sophia-wright-blue eco_lite_rgb_16F_kinetics_v1.pth.tar you can see 60th lines in mian.py resume='./eco_lite_rgb_16F_kinetics_v1.pth.tar' if os.path.isfile(resume): print(("=> loading checkpoint '{}'".format(resume))) checkpoint = torch.load(resume)#加载模型 model.load_state_dict(checkpoint['state_dict'])#网络结构导入模型 print(("=> loaded checkpoint '{}' (epoch {})" .format(evaluate, checkpoint['epoch']))) else: print(("=> no checkpoint found at '{}'".format(resume)))
thank you so much for clarifying that @mangdian and @shajie17 ,
@mangdian @sophia-wright-blue sorry,I am very anxious to know how to change num_segments=16,I have done it some days .do you how to do it ?thank you very much
@shajie17 change it in .prototxt file
thank you @mangdian but how to do it in pytorch https://github.com/zhang-can/ECO-pytorch
@shajie17 change it in the corresponding .sh file in scripts directory. And maybe you should use the code in this repository https://github.com/mzolfaghari/ECO-pytorch
@mangdian thank you ,
Did you have the pytorch pretrained models on Kinetics? I would really appreciate it if you can share it