parsa.haghighi
/
CLIPSampler
Watch 1
Star 0
Fork 0
Code Issues 0 Pull Requests 0 Releases 0 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
1 Commit
1 Branch
Tree: bcb96dfebb
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'bcb96dfebb'
${ noResults }
CLIPSampler/config.py

config.py 6.8KB
Raw Blame History

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889 
  1. import argparse

  2. import os

  3. import torch

  4. 

  5. media_path_root = "path/to/media"

  6. 

  7. class Config:

  8.     def __init__(self):

  9.         self.parser = argparse.ArgumentParser(description='Train a student network.')

  10.         self.add_arguments()

  11.         self.args = self.parse()

  12.         self.update_paths()

  13. 

  14.     def add_arguments(self):

  15.         # General settings

  16.         self.parser.add_argument('--seed', type=int, default=12345, help='Init Seed')

  17.         self.parser.add_argument('--device', type=int, default=0, help='Device number to use for training')

  18.         self.parser.add_argument("--do_train", type=int, default=1, help="Whether to run training.")

  19.         self.parser.add_argument("--do_eval", type=int, default=1, help="Whether to run evaluation.")

  20.         self.parser.add_argument('--dataset', type=str, default='activity-net', choices=['activity-net',  'msrvtt'], help='Dataset to use')

  21.         self.parser.add_argument('--teacher_model', type=str, default='CLIP', choices=['CLIP', 'DiffusionRet',  'EMCL',  'HBI', 'DiCoSA'], help='Teacher model to use')

  22.         self.parser.add_argument('--workers', type=int, default=12, help='Number of workers')

  23. 

  24.         # Model configurations

  25.         self.parser.add_argument('--dropout_ratio', type=float, default=0.3, help='Dropout ratio for the model')

  26.         self.parser.add_argument('--fc1_size', type=int, default=512, help='First FC layer Hidden size')

  27.         self.parser.add_argument('--fc2_size', type=int, default=256, help='Second FC layer Hidden size')

  28.         self.parser.add_argument('--fc3_size', type=int, default=128, help='Third FC layer Hidden size')

  29.         self.parser.add_argument('--backbone', type=str, default='resnet18', choices=['resnet18', 'efficientnet_b0', 'mobilenetv2_100', 'mobilenetv3_large_100'], help='Backbone model')

  30.         self.parser.add_argument('--n_trainable_backbone_layers', type=int, default=1, help='Number of trainable backbone layers')

  31.         self.parser.add_argument('--activation', type=str, default='relu', choices=['relu', 'leakyrelu', 'gelu'], help='Activation function')

  32.         self.parser.add_argument('--normalization_layer', type=str, default='sigmoid', choices=['sigmoid', 'softmax', 'gumbel', 'k_top'], help='Normalization layer type (for text-to-video retireval pipeline)')

  33.         self.parser.add_argument('--init_tau', default=5.0, type=float, help="annealing init temperature")

  34.         self.parser.add_argument('--min_tau', default=0.5, type=float, help="min temperature to anneal to")

  35.         self.parser.add_argument('--decay_factor', default=0.045, type=float, help="exp decay factor per epoch")

  36. 

  37.         self.parser.add_argument('--store_complete_model', type=int, default=0, help='Store the compelte model')

  38.         self.parser.add_argument('--load_complete_model',  type=int, default=0, help='1: Load')

  39.         self.parser.add_argument('--trained_model_epoch', type=int, default=100, help='Load the trained model from which epoch')

  40.         

  41.         self.parser.add_argument('--store_backbone', type=int, default=0, help='Store just backbone')

  42.         self.parser.add_argument('--load_backbone',  type=int, default=1, help='1: Load')

  43.         self.parser.add_argument('--trained_backbone_epoch', type=int, default=100, help='Load a trained backbone model from which epoch')

  44.         

  45.         # Dataloader configurations

  46.         self.parser.add_argument('--normalization', type=str, default='binary', choices=['raw', 'min_max', 'binary'], help='Normalization type for dataloader')

  47.         self.parser.add_argument('--binary_percentage', type=float, default=0.1, help='Binary percentage for dataloader')

  48. 

  49.         # Training configurations

  50.         self.parser.add_argument('--batch_size', type=int, default=16, help='Batch size for training')

  51.         self.parser.add_argument('--epochs', type=int, default=20, help='Number of epochs for training')

  52.         self.parser.add_argument('--similarity_matrix_loss', type=str, default='kl', choices=['mse','kl'], help='frame2frame loss')

  53.         self.parser.add_argument('--use_weighted_loss', type=int, default=0, help='Whether use weighted BCE, to handle unbalanced classes')

  54.         self.parser.add_argument('--alpha', type=float, default=1, help='Weight for f2f loss')

  55.         self.parser.add_argument('--beta', type=float, default=1, help='Weight for f2t loss')

  56.         self.parser.add_argument('--gamma', type=float, default=1, help='Weight for v2t loss')

  57.         self.parser.add_argument('--lr', type=float, default=0.0001, help='Learning rate')

  58.         self.parser.add_argument('--use_scheduler', type=int, default=1, help='Use a learning rate scheduler')

  59.         self.parser.add_argument('--scheduler_step_size', type=float, default=5, help='After how many epochs the learning rate will be adjusted')

  60.         self.parser.add_argument('--scheduler_gamma', type=float, default=0.4, help='The factor by which the learning rate will be reduced at each step.')

  61.         self.parser.add_argument('--weight_decay', type=float, default=0.001, help='Weight decay for the optimizer. Set 0 to disable it')

  62.         self.parser.add_argument('--do_cluster', type=int, default=1, help='Calculate results with cluster as well or not.')

  63.         

  64.         

  65.     def parse(self):

  66.         return self.parser.parse_args()

  67. 

  68.     def update_paths(self):

  69.         self.args.device = torch.device(f'cuda:{self.args.device}' if torch.cuda.is_available() else "cpu")

  70. 

  71.         self.args.raw_backbone_path = f"{media_path_root}/saved_models/student_pretrained_backbones/{self.args.backbone}_pretrained.pth"

  72.         self.args.st_scores_path = f"{media_path_root}/{self.args.dataset}/{self.args.teacher_model}/st_scores/st_scores.json"

  73.         self.args.trained_backbone_path = f"{media_path_root}/experiments/saved_models/{self.args.backbone}/{self.args.similarity_matrix_loss}"

  74.         self.args.trained_model_path = f"{self.args.trained_backbone_path}/complete_model/"

  75.         self.args.log_dir = f'logs/{self.args.project_name}/'

  76. 

  77.         os.makedirs(self.args.st_scores_path, exist_ok=True)

  78.         os.makedirs(self.args.trained_backbone_path, exist_ok=True)

  79.         os.makedirs(self.args.trained_model_path, exist_ok=True)

  80.         os.makedirs(self.args.log_dir, exist_ok=True)

  81. 

  82.         self.args.paths = {'train':{}, 'valid':{}}

  83. 

  84.         for data_split in ['train', 'valid']:

  85.             split = 'val_1' if data_split=='valid' else data_split

  86.             self.args.paths[data_split]['frames'] = f"{media_path_root}/{self.args.dataset}/sampled_frames/{split}"

  87.             self.args.paths[data_split]['frame_features'] = f"{media_path_root}/{self.args.dataset}/{self.args.teacher_model}/embeddings/frames/{split}"

  88.             self.args.paths[data_split]['text_features'] =  f"{media_path_root}/{self.args.dataset}/{self.args.teacher_model}/embeddings/description/{split}"

  89.             self.args.paths[data_split]['gt_scores'] = f"{media_path_root}/{self.args.dataset}/{self.args.teacher_model}/gt_scores/gt_{split}.json"