r1c7 / slowfastnetworks Goto Github PK
View Code? Open in Web Editor NEWPyTorch implementation of "SlowFast Networks for Video Recognition".
PyTorch implementation of "SlowFast Networks for Video Recognition".
I trained the model in ucf101, after 70 epoch, the top 1 accuracy is 73.3%, the top 5 accuracy is 95.7% in training data, is this a normal result?
如题
non
can you give me an example code about how to detect a single video, and get its result?
thank you very much!
Is the video in train and validation the same?
Thanks for your work, i got a question about dataset split. UCF101 and HMDB51 has 3 offical train/test splits but in you code there is a train / validation in training , how do you to split the dataset?
hi, thanks for building this project! I want to finetune my own dataset. Can you have the pretrained weights trained in Kinetics dataset?
ucf101的数据集该如何划分train,val
thanks for your work.
When i train it on my datasets,after 100 epoch,it only get 35% top1-acc.My datasets has 7 classses,and i use the default config as u provided,What are the possible causes?
Thank you for sharing your implementation, have you evaluated the net in the UCF101? how about the result?
I am studing your code.But I have some problem.In the article, the video of input size is 64224224.In your code,it is 8112112.This is just the the simple implementation about the article
Excuse me,could I ask you about that only verification method, how to test on ucf101?
Hi,
Great work with implementing the paper here. I'm trying to replicate the results on Kinetics-400 and so far it looks really promising!
I wanted to ask about the temporal convolutions in the slow path of the model -
In the paper they apply non-degenerate temporal convolutions in residual layers 4 and 5 of the slow path (kernel size>1 for depth dimension). Is it something you chose not to apply here? did it hurt performance somehow? I just want to know whether it's something worth attempting.
Thanks!
Thanks for your great implementation. When I read your code, I found you manually select the first 2 frames and the first 16 frames as the input of the fast path and the slow path , respectively.
# lib/slowfastnet.py
92 fast, lateral = self.FastPath(input[:, :, ::2, :, :])
93 slow = self.SlowPath(input[:, :, ::16, :, :])
Could you please explain why you choose the first 2 frames rather than the last 2 frames or 2 frames in the middle? Will the selection change the final performance?
Hi, thanks for the great code! I was thinking to experiment with it but not sure what the accuracy and speed is?
I saw people trained on UCF but got mixed results, @r1ch88 did you have a chance to validate with the results from the paper on Kinetics or AVA or any other dataset?
In line
https://github.com/RI-CH/SlowFastNetworks/blob/master/lib/slowfastnet.py#L89
self.fc = nn.Linear(self.fast_inplanes+2048, class_num, bias=False)
here self.fast_inplanes is 8, but i think it should be 256 after the pooing in fast path, is it right?
Hi, I want to create a confusion matrix to see which actions could be related in UCF dataset. I know it sounds trivial, but I am trying to get the two matrices and pass them through sklearn, I just cant seem to do it though
Can someone please help??
Hi! thanks for your great works, I use two K40C each has 12GB memory, crop_size=112, clip_len=32, I only set batch_size=2 can run it, when I set batch_size=4 it tell me out of memory. But the Memory Usage is just about 1/10, Volatile GPU-Util is nearly 100%, I don't kown why.
where dose it do in code
Which version of the residual block is used in the code, and why is it so designed? Thanks
Thanks for sharing the code of your implementation, but i have a question about follow line, why should we do this operation in slow path?
Thanks for your reply!
sincerly!
i have trained the slowfastNetwork on ucf101 dataset,but the accuracy was just 40+% after 80 epochs
I set lr=0.001, other are default
I train it in UCF101
the top1 accuracy of train is around 73%
the top1 accuracy of test is around 40%
Shocking overfitting!!!!
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.