search

yjxiong / tsn-pytorch

Temporal Segment Networks (TSN) in PyTorch
BSD 2-Clause "Simplified" License
1.07k stars 308 forks source link

Getting Very low accuracy ( 38%) with HMDB-51 #63

Closed Mohamad73 closed 6 years ago

Mohamad73 commented 6 years ago

I am getting around 38% with RGB modality on the HMDB-51. I am running it as shown below. Am I doing anything wrong? python main.py hmdb51 RGB hmdb51_rgb_train_list hmdb51_rgb_val_list --arch BNInception --num_segments 3 --gd 20 --lr 0.001 --lr_steps 30 60 --epochs 80 --b 100 -j 2 --dropout 0.8

yjxiong commented 6 years ago

HMDB needs more epochs. Please calculate the parameters based on the caffe TSN repo's solver files.

Mohamad73 commented 6 years ago

First of all thank you for your replies Even when I run it on 160 epochs, it gives me the same result. I have only two gpus. I read that in Caffe the effective batchsize equals to batch_sizeiter_sizen_gpu has to be 128. How do I accommodate for that in pytorch since I dont have access to iter_size!

Mohamad73 commented 6 years ago

An answer will be appreciated it!!

Mkarami3 commented 3 years ago

Hi Mohamad73, Did you find the solution to your question? Kindly can you please share it with me? Thank you