About Multi Task Learning

[DuckJ]

Great work!!! Can you give some examples for multi-task. I think the examples in 'experiments' are all single tasks. Is there anything wrong with my understanding?

[XiaohangZhan]

Yes, they are all single tasks. See TODO at the end of README.

[DuckJ]

Soga. If I change the config.yaml, add the batchsize,data_root ..., multi-task experiments should also be started . Is right?

[XiaohangZhan]

For example: train: rand_seed: 0 batch_size: [59, 5] # ratio ~ ms1m : webface loss_weight: [0.92, 0.08] # ratio ~ ms1m : webface base_lr: 0.3 lr_decay_scale: 0.1 lr_decay_steps: [8, 12] # need to increase a little bit because of more data max_epoch: 14 # need to increase a little bit because of more data momentum: 0.9 weight_decay: 0.0001 print_freq: 20 average_stats: 100 data_root: - 'data/ms1m/images' - 'data/webface/images' data_list: - 'data/ms1m/list.txt' - 'data/webface/list.txt'

I've not tried. Tell me if it does not work.

[DuckJ]

Ok. I will try it. Another question，I noticed that the input size of the model is 224. Did you try to compare the input performance of 112 dimensions?

[XiaohangZhan]

No. Pls have a try and tell me your observation. Many thanks.

[XiaohangZhan]

You may also try to change the feature dimension from 256 to 512.

[DuckJ]

For example: train: rand_seed: 0 batch_size: [59, 5] # ratio ~ ms1m : webface loss_weight: [0.92, 0.08] # ratio ~ ms1m : webface base_lr: 0.3 lr_decay_scale: 0.1 lr_decay_steps: [8, 12] # need to increase a little bit because of more data max_epoch: 14 # need to increase a little bit because of more data momentum: 0.9 weight_decay: 0.0001 print_freq: 20 average_stats: 100 data_root:

• 'data/ms1m/images'
• 'data/webface/images' data_list:
• 'data/ms1m/list.txt'
• 'data/webface/list.txt'

I've not tried. Tell me if it does not work.

I use the MS1M and a private dataset(2000 images), it works. But when I use the multi - gpus, error occurred. Even when I run single task, the multi-gpu process can not be started.

[DuckJ]

I have solved it. pytorch vision change to 0.3.1 can run