Open-Set Multi-Source Multi-Target Domain Adaptation

How to run

Dataset directory

Click to see full directory tree

   data
    ├── domain_net
    │   ├── clipart
    │   ├── clipart.txt
    │   ├── infograph
    │   ├── infograph.txt
    │   ├── painting
    │   ├── painting.txt
    │   ├── quickdraw
    │   ├── quickdraw.txt
    │   ├── real
    │   ├── real.txt
    │   ├── sketch
    │   └── sketch.txt
    ├── office
    │   ├── amazon
    │   ├── amazon.txt
    │   ├── dslr
    │   ├── dslr.txt
    │   ├── webcam
    │   └── webcam.txt
    ├── office-home
    │   ├── Art
    │   ├── Art.txt
    │   ├── Clipart
    │   ├── Clipart.txt
    │   ├── Product
    │   ├── Product.txt
    │   ├── Real_World
    │   └── RealWorld.txt
    ├── office_home_mixed
    │   ├── Art_Clipart_Product
    │   ├── Art_Clipart_Product.txt
    │   ├── Art_Clipart_Real_World
    │   ├── Art_Clipart_Real_World.txt
    │   ├── Art_Product_Real_World
    │   ├── Art_Product_Real_World.txt
    │   ├── Clipart_Product_Real_World
    │   └── Clipart_Product_Real_World.txt
    └── pacs
        ├── art_painting
        ├── art_painting.txt
        ├── cartoon
        ├── cartoon.txt
        ├── __MACOSX
        ├── photo
        ├── photo.txt
        ├── sketch
        └── sketch.txt

Install the dependencies and run scripts.

Prerequisites:

• See requirements.txt
• Install dependencies using pip3 install -r requirements.txt

Prepare pretrain model

We choose R50-ViT-B_16 as our backbone.

# Download pretrained R50-ViT-B_16
wget https://storage.googleapis.com/vit_models/imagenet21k/R50+ViT-B_16.npz 
mkdir -p ./model/vit_checkpoint/imagenet21k 
mv R50+ViT-B_16.npz ./model/vit_checkpoint/imagenet21k/R50+ViT-B_16.npz

Download the dataset

• Run sh source_only.sh
• Suppose you want to donwload Office-Home, just uncomment any line of Office-Home source training and it will download

CODE

• Run sh source_only.sh for multi-src training.
• Run scripts/warmup.sh for calculating centroids
• Run scripts/adapt.sh for adaptation

Codes

1. degaa_new.py

   • Passes whole data through LOF not batchwise
   • The LOF is applied every kth epoch of training (k=20)
   • Output Dir: ./adapt/run2
   • Pseudo Code:

     for epoch in epochs:
     
       if epoch % num_episodes = 0:
         for i in iters_per_epoch:
           Feat_S = concat(Feat_S, netB(netF(batch_s) || DE))
           Feat_T = concat(Feat_T, netB(netF(batch_t) || DE))
         
         Y_preds = LOF(Feat_T) # Array of [1 and -1]
         Label_T.where( Y_preds == -1 ) = 25 
     
         known = Feat_T.where(Y_preds == 1)
         Label_T.where( Y_preds == 1) = KNN(known, centroids)
     
         TempData = [Feat_s, Label_S, Feat_T, Label_T]
     
       for batch in TempData:
         F_s, F_t = GAA(Feat_S, Feat_T)
         Y_s, Y_t = softmax(netC(F_s)), softmax(netC(F_t))
     
         loss = CE(Y_s, Label_s) + CE(Y_t, Label_t)
     
         loss.backward()

This code is taken from Transfer-Learning-Library. This will be boiler-plate code. Do checkout some of the examples.