Taxabind/TaxaBind/SatBind/config.py

from easydict import EasyDict as edict

config = edict()

# Pixel level CLIP training
config.img_dir = '/mnt/hdd/inat2021_ds/inat21' 
config.imo_dir = '/mnt/hdd/inat2021_ds/sat_train_jpg_512px'   # sat_train_jpg_256px, sat_train_jpg_512px
config.imo_dir_val = '/mnt/hdd/inat2021_ds/sat_test_jpg_512px'    # sat_test_jpg_256px, sat_test_jpg_512px
config.train_json_path = '/mnt/hdd/inat2021_ds/inat21_train.json' 
config.val_json_path = '/mnt/hdd/inat2021_ds/inat21_val.json' # no filter needed
# config.sat_to_img_ids_train_json_path = '/mnt/hdd/inat2021_ds/filtered_mapping_sat_to_img_ids_train.json' 
config.sat_to_img_ids_train_json_path = '/mnt/hdd/inat2021_ds/target_search_ds/NEW_3-20_Targets/FINAL_COUNTS_COMBINED_top4000_pos_top55000_neg/SEARCH_DS_filtered_mapping_sat_to_img_ids_train_3-20counts.json' 
config.sat_to_img_ids_val_json_path = '/mnt/hdd/inat2021_ds/filtered_mapping_sat_to_img_ids_val.json' 
# config.filtered_train_json_path = '/mnt/hdd/inat2021_ds/inat21_filtered_pixel_clip_v2_train.json' 
# config.filtered_val_json_path = '/mnt/hdd/inat2021_ds/inat21_filtered_pixel_clip_val.json' # no filter needed

# # Pixel level CLIP training (Toy example)
# config.img_dir = '../scripts/preprocess/expt/fake_img_dataset' 
# config.imo_dir = '../scripts/preprocess/expt/sat_train_dataset' 
# config.imo_dir_val = '../scripts/preprocess/expt/sat_val_dataset' 
# config.train_json_path = '../scripts/preprocess/expt/inat2021_train_micro.json' 
# config.val_json_path = '../scripts/preprocess/expt/inat2021_val_micro.json' 
# config.sat_to_img_ids_train_json_path = '../scripts/preprocess/expt/sat_filtered/filtered_mapping_sat_to_img_ids_train_micro.json' 
# config.sat_to_img_ids_val_json_path = '../scripts/preprocess/expt/sat_filtered/filtered_mapping_sat_to_img_ids_val_micro.json' 
# # config.filtered_train_json_path = '../scripts/preprocess/expt/inat2021_train_micro.json' 
# # config.filtered_val_json_path = '../scripts/preprocess/expt/inat2021_val_micro.json' # no filter needed

# # Image level CLIP training
# config.img_dir = '/mnt/hdd/inat2021_ds/inat21' 
# config.imo_dir = '/mnt/hdd/inat2021_ds/sat_train_jpg' 
# config.imo_dir_val = '/mnt/hdd/inat2021_ds/sat_test_jpg' 
# config.train_json_path = '/mnt/hdd/inat2021_ds/inat21_filtered_img_clip_train.json' 
# config.val_json_path = '/mnt/hdd/inat2021_ds/inat21_val.json' # no filter needed

# # Image level CLIP training (Toy example)
# config.img_dir = '../scripts/preprocess/expt/fake_img_dataset' 
# config.imo_dir = '../scripts/preprocess/expt/sat_train_dataset' 
# config.imo_dir_val = '../scripts/preprocess/expt/sat_val_dataset' 
# config.train_json_path = '../scripts/preprocess/expt/inat2021_train_micro.json' 
# config.val_json_path = '../scripts/preprocess/expt/inat2021_val_micro.json' 

# batch_size * accumulate_grad_batches * devices = MUST BE CONSTANT (i.e. 256 * 8 * 2 = 4096)
config.batch_size = 32 # 256
config.lr = 1e-4    # 1e-4
config.accumulate_grad_batches = 64
config.max_epochs = 20
config.num_workers = 16
config.devices = 2 # 2
config.val_check_interval = 0.5
config.sat_encoder = 'openai/clip-vit-large-patch14-336' # openai/clip-vit-base-patch16, openai/clip-vit-large-patch14-336
config.patch_size = 14

config.save_dir = 'checkpoints'
config.filename = 'satbind-{epoch:02d}-{val_loss:.2f}'

config.locked_tuning = True

config.resume_from_checkpoint = False
config.resume_checkpoint_name = 'satbind-resume'

print("config: \n", config)