update

README.md

@@ -11,27 +11,61 @@ Nearest neighbors are traditionally used to compute final decisions, e.g., in Su
 
 Download pretrained models for CUB-200, Cars-196, and Dogs-120 at [this link](https://drive.google.com/drive/folders/1pC_5bEi5DryDZCaKb51dzCE984r8EnqW?usp=sharing).
 
-## Sampling → Training → Binary classify
+## Training image comparator network
+
 
 For CUB-200,
-> sh cub_run.sh
+
+Step 1: Set `global_training_type = 'CUB'` and `self.set = 'train'` in `params.py`
+
+Step 2:
+> sh train_cub.sh
 
 For Cars-196,
-> sh cars_run.sh
+
+Step 1: Set `global_training_type = 'CARS'` and `self.set = 'train'` in `params.py`
+
+Step 2:
+> sh train_cars.sh
+
 
 For Dogs-120,
-> sh dogs_run.sh
 
-Notes: Set `global_training_type` to  = `'CUB'` or `'CARS'` or `'DOGS'`
+Step 1: Set `global_training_type = 'DOGS'` and `self.set = 'train'` in `params.py`
+
+Step 2:
+> sh train_dogs.sh
+
+## Testing (Binary classification and Reranking)
+
+For CUB-200,
+
+Step 1: Set `global_training_type = 'CUB'` and `self.set = 'test'` in `params.py`
+
+Step 2:
+> sh test_cub.sh
+
+For Cars-196,
+
+Step 1: Set `global_training_type = 'CARS'` and `self.set = 'test'` in `params.py`
 
+Step 2:
+> sh test_cars.sh
 
-## Reranking
+For Dogs-120,
 
-> python cub_extract_feature_for_reranking.py
+Step 1: Set `global_training_type = 'DOGS'` and `self.set = 'test'` in `params.py`
 
-> python cub_reranking.py
+Step 2:
+> sh test_dogs.sh
 
 ## How to visualize qualitative figures
+For CUB-120,
+
+Step 1: Set `global_training_type = 'CUB'` and `self.set = 'test'` in `params.py`
+
+Step 2:
+
 1. Corrections of S
 > python cub_visualize_corrections.py
 2. Training pairs of S
@@ -40,20 +74,3 @@ Notes: Set `global_training_type` to  = `'CUB'` or `'CARS'` or `'DOGS'`
 > python cub_infer.py
 
 Same steps for Cars-196 and Dogs-120.
-
-[//]: # ()
-[//]: # (### RN50 image comparator improving NTSNet)
-
-[//]: # (Set NTSNET in params.py to True. Then set MODEL1_RESNET in cub_extract_feature_adv_process.py to False then run:)
-
-[//]: # (> python cub_extract_feature_adv_process.py)
-
-[//]: # ()
-[//]: # (Then set MODEL1_RESNET in cub_advising_process.py to False then run:)
-
-[//]: # (> python cub_advising_process.py)
-
-[//]: # ()
-[//]: # (You can also set PRODUCT_OF_EXPERTS in cub_advising_process.py to True/False to use/not use the product of experts.)
-
-[//]: # ()

cars-196/car_image_comparator_training.py

@@ -288,11 +288,10 @@ def train_model(model, loss_func, optimizer, scheduler, num_epochs=25):
     config=config,
 )
 
-wandb.save(os.path.basename(__file__), policy='now')
-wandb.save('../params.py', policy='now')
-wandb.save('../datasets.py', policy='now')
-wandb.save('car_image_comparator_training.py', policy='now')
-wandb.save('../transformer.py', policy='now')
+wandb.save(f'{RunningParams.prj_dir}/params.py')
+wandb.save(f'{RunningParams.prj_dir}/datasets.py')
+wandb.save('car_image_comparator_training.py')
+wandb.save(f'{RunningParams.prj_dir}/transformer.py')
 
 _, best_acc = train_model(
     MODEL2,

cub-200/cub_image_comparator_training.py

@@ -275,11 +275,10 @@ def train_model(model, loss_func, optimizer, scheduler, num_epochs=25):
     config=config,
 )
 
-wandb.save(os.path.basename(__file__), policy='now')
-wandb.save('../params.py', policy='now')
-wandb.save('../datasets.py', policy='now')
-wandb.save('cub_image_comparator_training.py', policy='now')
-wandb.save('../transformer.py', policy='now')
+wandb.save(f'{RunningParams.prj_dir}/params.py')
+wandb.save(f'{RunningParams.prj_dir}/datasets.py')
+wandb.save('cub_image_comparator_training.py')
+wandb.save(f'{RunningParams.prj_dir}/transformer.py')
 
 if RunningParams.VisionTransformer is True:
     _, best_acc = train_model(

params.py

@@ -17,7 +17,7 @@ def __init__(self, training_type=None):
         # Retrieving NNs and sample positive and negative pairs
         # Set it when you extract the NNs. data_dir is the folder containing query images for KNN retrieval
         # Set it when you run train/test
-        self.set = 'test'
+        self.set = 'train'
 
         self.parent_dir = '/home/giang/Downloads'
         self.prj_dir = '/home/giang/Downloads/advising_network'

stanford-dogs/dogs_image_comparator_training.py

@@ -283,11 +283,10 @@ def train_model(model, loss_func, optimizer, scheduler, num_epochs=25):
     config=config,
 )
 
-wandb.save(os.path.basename(__file__), policy='now')
-wandb.save(f'{RunningParams.prj_dir}/params.py', policy='now')
-wandb.save(f'{RunningParams.prj_dir}/datasets.py', policy='now')
-wandb.save('dogs_image_comparator_training.py', policy='now')
-wandb.save(f'{RunningParams.prj_dir}/transformer.py', policy='now')
+wandb.save(f'{RunningParams.prj_dir}/params.py')
+wandb.save(f'{RunningParams.prj_dir}/datasets.py')
+wandb.save('dogs_image_comparator_training.py')
+wandb.save(f'{RunningParams.prj_dir}/transformer.py')
 
 if RunningParams.VisionTransformer is True:
     _, best_acc = train_model(

test_cars.sh

@@ -0,0 +1,8 @@
+#!/bin/sh
+
+CUDA_VISIBLE_DEVICES=4 python cars-196/car_extract_feature.py  # sampling for test set
+python cars-196/augment_images.py
+CUDA_VISIBLE_DEVICES=4 python cars-196/car_binary_classify.py  # binary classification
+
+CUDA_VISIBLE_DEVICES=4 python cars-196/car_extract_feature_for_reranking.py  # sampling for reranking
+CUDA_VISIBLE_DEVICES=4,5 python cars-196/car_reranking.py  # reranking

test_cub.sh

@@ -0,0 +1,8 @@
+#!/bin/sh
+
+CUDA_VISIBLE_DEVICES=4 python cub-200/cub_extract_feature.py  # sampling for test set
+python cub-200/augment_images.py
+CUDA_VISIBLE_DEVICES=4 python cub-200/cub_binary_classify.py  # binary classification
+
+CUDA_VISIBLE_DEVICES=4 python cub-200/cub_extract_feature_for_reranking.py  # sampling for reranking
+CUDA_VISIBLE_DEVICES=4,5 python cub-200/cub_reranking.py  # reranking

test_dogs.sh

@@ -0,0 +1,8 @@
+#!/bin/sh
+
+CUDA_VISIBLE_DEVICES=4 python stanford-dogs/dogs_extract_feature.py  # sampling for test set
+python stanford-dogs/augment_images.py
+CUDA_VISIBLE_DEVICES=4 python stanford-dogs/dogs_binary_classify.py  # binary classification
+
+CUDA_VISIBLE_DEVICES=4 python stanford-dogs/dogs_extract_feature_for_reranking.py  # sampling for reranking
+CUDA_VISIBLE_DEVICES=4,5 python stanford-dogs/dogs_reranking.py  # reranking

cars_run.sh → train_cars.sh

@@ -2,7 +2,6 @@
 
 rm -rf /home/giang/Downloads/Cars/Stanford-Cars-dataset/train/
 cp -r /home/giang/Downloads/Cars/Stanford-Cars-dataset/BACKUP/train /home/giang/Downloads/Cars/Stanford-Cars-dataset/train/
-python CUDA_VISIBLE_DEVICES=4 cars-196/car_extract_feature.py
+CUDA_VISIBLE_DEVICES=4 python cars-196/car_extract_feature.py
 python cars-196/augment_images.py
-python CUDA_VISIBLE_DEVICES=4,5 cars-196/car_image_comparator_training.py  # train
-python CUDA_VISIBLE_DEVICES=4 cars-196/car_binary_classify.py  # test
+CUDA_VISIBLE_DEVICES=4,5 python cars-196/car_image_comparator_training.py  # train

train_cub.sh

@@ -0,0 +1,5 @@
+#!/bin/sh
+
+CUDA_VISIBLE_DEVICES=4 python cub-200/cub_extract_feature.py  # sampling
+python cub-200/augment_images.py
+CUDA_VISIBLE_DEVICES=4,5 python cub-200/cub_image_comparator_training.py  # train

train_dogs.sh

@@ -0,0 +1,5 @@
+#!/bin/sh
+
+CUDA_VISIBLE_DEVICES=4 python stanford-dogs/dogs_extract_feature.py
+python stanford-dogs/augment_images.py
+CUDA_VISIBLE_DEVICES=4,5 python stanford-dogs/dogs_image_comparator_training.py  # train