Did you use Drop Out?

[rlaengud123]

In the paper(https://arxiv.org/pdf/1411.4389.pdf, 5page)

They said "Additionally, when training the LRCN network end-to-end, we found that aggressive dropout (0.9) was needed to avoid overfitting"

But, I cant find Drop Out.

Did you use Drup Out In your code?

Thx

[doronharitan]

Hi, I didn't use dropout. I stop the training of the model before the overfitting (used the train vs validation learning curve do decide when to do so) Also, as you can see I tested the model on new data (not even the test data) and its performance was over 91% so I figure that it doesn't overfit.

[rlaengud123]

Thank to your answer. I have some questions..

1. I used your code well (thank you). but, end the test, i got acc is 55%. I think my parameter was wrong. can you share the parameter?

2. I want to use Multi-GPU. I have 8 GPUs, but when i use train.py, workstation used only device:0 2GB(i think in your code "cuda:device:0") In torch, offer DataParallel, Can I use it? or in your code offer Multi-GPU? How can i use Mulit-GPU? Thx -----Original Message----- From: "doronharitan"notifications@github.com To: "doronharitan/human_activity_recognition_LRCN"human_activity_recognition_LRCN@noreply.github.com; Cc: "rlaengud123"enguddkdlel@naver.com; "Author"author@noreply.github.com; Sent: 2020-02-06 (목) 00:45:25 (GMT+09:00) Subject: Re: [doronharitan/human_activity_recognition_LRCN] Did you use Drop Out? (#2)

Hi, I didn't use dropout. I stop the training of the model before the overfitting (used the train vs validation learning curve do decide when to do so) Also, as you can see I tested the model on new data (not even the test data) and its performance was over 91% so I figure that it doesn't overfit. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[doronharitan]

Hi,

1. as I mentioned in the readme file I trained and tested my model on 55 categories from the UCF101 possible categories. My params were: Preprocessing step: --sampling_rate 10
   --num_frames_to_extract 5 --ucf101_fps 25

train.py params --epochs 40 --batch-size 256 --lr 5e-4 --split_size 0.2 --seed 42 --latent_dim 512 --hidden_size 256 --lstm_layers 2 --bidirectional True --number_of_classes 55

test mode params: --num_frames_video 5 --seed 42 --latent_dim 512 --hidden_size 256 --lstm_layers 2 --bidirectional True --number_of_classes 55 --model_dir model_dir
--model_name model_name

2. I don't have experience with multi-GPU training (except GCP). I trained my model on Google Cloud Platform using K80 GPU, took ~ 25 hours to reach ~93% accuracy on validation set

Hope this helps. Good luck Doron

[rlaengud123]

Sorry, I have another question.

Are there any rules for trainlist (1,2,3) and testlist (1,2,3)?

Is it simply random?

However, I see some repeating sections. And also regularity in the file.

When you run the code Are all three sets used?

How should I organize my files?

Thank you, sir. ᐧ

2020년 2월 6일 (목) 오전 1:16, doronharitan notifications@github.com님이 작성:

Hi,

1. as I mentioned in the readme file I trained and tested my model on 55 categories from the UCF101 possible categories. My params were: Preprocessing step: --sampling_rate 10 --num_frames_to_extract 5 --ucf101_fps 25

train.py params --epochs 40 --batch-size 64 --lr 5e-4 --split_size 0.2 --seed 42 --latent_dim 512 --hidden_size 256 --lstm_layers 2 --bidirectional True --number_of_classes 55

test mode params: --num_frames_video 5 --seed 42 --latent_dim 512 --hidden_size 256 --lstm_layers 2 --bidirectional True --number_of_classes 55 --model_dir model_dir --model_name model_name

1. I don't have experience with multi-GPU training (except GCP). I trained my model on Google Cloud Platform using K80 GPU, took ~ 25 hours to reach ~93% accuracy on validation set

Hope this helps. Good luck Doron

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/doronharitan/human_activity_recognition_LRCN/issues/2?email_source=notifications&email_token=AMYBZFBF7N4DBOTV55VB7STRBLQ4BA5CNFSM4KQNABIKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEK4AMQA#issuecomment-582485568, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMYBZFAGPNVB3YNSVKWF4M3RBLQ4BANCNFSM4KQNABIA .

[yangyuan6956873]

hi,I have the same question. If I want to train two categories from the UCF101 or other datasets. How can i change the trainlist and the testlist. Thank you very much.