Accuracy of VisDA2017 only 70.76 about tvt HOT 27 CLOSED

Can you reproduce the results on other datasets?

from tvt.

I have follow your code, in VisDA2017, the accuracy only 70.76

BTW, can you show me the command you used?

from tvt.

python3 main.py --train_batch_size 32 --dataset visda17 --name visda --source_list /home/cy/data/visda/train/train_list.txt --target_list /home/cy/data/visda/validation/validation_list.txt --test_list /home/cy/data/visda/validation/validation_list.txt --num_classes 12 --model_type ViT-B_16 --pretrained_dir checkpoint/ViT-B_16.npz --num_steps 20000 --img_size 256 --beta 1.0 --gamma 0.01 --use_im

from tvt.

Can you reproduce the results on other datasets?

I have only run in VisDA

from tvt.

Thanks, let me check this.

from tvt.

Thanks, let me check this.

Is ok?

from tvt.

Thanks, let me check this.

Is ok?

I am not in campus now, so I tried to use VPN to access my machine. However, the link to download VPN is broken, so I submitted a ticket to the IT service, hope that they can fix this issue tomorrow.

from tvt.

when use 'Transferability Adaptation Module',

why only 'k' need 'Domain Adversarial Loss'?

from tvt.

I don't understand what does 'k' means, can you clarify it?

from tvt.

TVT framework(image.png in this project) in the paper,

the 'TAM' module (i.e., 'Transferable MSA'),

img-->patch token--> generous 'q', 'k', 'v',

why only 'k' use 'Domain Adversarial Discriminator'? i am not understand.

from tvt.

'q' is query of the [CLS] token, while 'k' is key of local patches. We use 'Domain Adversarial Discriminator' to identify transferable local patches.

BTW, I haven't received the response from IT service, maybe they are on their holidays.

from tvt.

'q' is query of the [CLS] token, while 'k' is key of local patches. We use 'Domain Adversarial Discriminator' to identify transferable local patches.

BTW, I haven't received the response from IT service, maybe they are on their holidays.

Is 'v' equal 'k' ? ( i.e., 'v' and 'k' both of local patches, and have same effect)

from tvt.

no, 'v' is value of local patches.

from tvt.

'v'(value) and 'k'(key) , they are all multiplied by a matrix,

why only 'k' use 'Domain Adversarial Discriminator', while 'v' not?

from tvt.

because we usually measure the 'similarity' between 'q' and 'k' in attention mechanism, then the similarity is used to aggregate 'v'. You can check equation 1 in Attention is all you need

from tvt.

Fine.

from tvt.

python3 main.py --train_batch_size 32 --dataset visda17 --name visda --source_list /home/cy/data/visda/train/train_list.txt --target_list /home/cy/data/visda/validation/validation_list.txt --test_list /home/cy/data/visda/validation/validation_list.txt --num_classes 12 --model_type ViT-B_16 --pretrained_dir checkpoint/ViT-B_16.npz --num_steps 20000 --img_size 256 --beta 1.0 --gamma 0.01 --use_im

This command mean train source model and adaption target model simultaneously?

(I do not find source only accuracy of 73.22% in visda2017)

from tvt.

Domain adaption setting, should train a source model, and use this trained source model to init target model, finally, train target model for adaption.

from tvt.

It trains a model by using both source data and target data. It is not source-only

from tvt.

Not necessary. We can directly train a model which leverages both source data and target data to learn domain-invariant features

from tvt.

In my experiment, I use pre-trained model (ViT-B_16.npz),

in visda 2017

the accuracy of source only just 65%, not get 73.22% in your paper

from tvt.

Sorry for the late reply. Did you try the pre-trained model on the source data?

from tvt.

I use pre-trained Vit-16-B, and get 71.3% (source only)

more important , I still get 'Accuracy of VisDA2017 only 70.76'.

from tvt.

I rerun the code and can reproduce the result. Can you provide the screenshot of the printing output?

from tvt.

from tvt.

This is what I get:

from tvt.

It seems like the validation loss in your experiment is quite large

from tvt.