-
Notifications
You must be signed in to change notification settings - Fork 0
/
run.py
126 lines (95 loc) · 5.07 KB
/
run.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
"""Painter by Numbers.
"""
import tensorflow as tf
from core import boot
from core.utils import str2bool
boot.gpus_with_memory_growth()
from argparse import ArgumentParser
import numpy as np
parser = ArgumentParser()
parser.add_argument("--seed", default=0, type=int)
parser.add_argument("--override", default=False, type=str2bool)
parser.add_argument("--mixed_precision", default=False, type=str2bool)
parser.add_argument("--jit_compile", default=False, type=str2bool)
# Dataset
parser.add_argument("--data_dir", default="./data")
parser.add_argument("--batch_size", default=32, type=int)
parser.add_argument("--patch_size", default=299, type=int)
parser.add_argument("--data_split", default="frequent", choices=["original", "frequent"])
parser.add_argument("--data_frequent_test_size", default=0.1, type=float)
# Preprocess
parser.add_argument("--preprocess_max_size", default=6000, type=int)
parser.add_argument("--preprocess_workers", default=24, type=int)
# Features (Backbone)
parser.add_argument("--strategy", default="ce", choices=["ce", "supcon", "supcon_mh"])
parser.add_argument("--backbone_architecture", default="InceptionV3")
parser.add_argument("--backbone_features_layer", default="avg_pool")
parser.add_argument("--logs_dir", default="./experiments/logs/")
parser.add_argument("--weights_dir", default="./experiments/weights/")
parser.add_argument("--preds_dir", default="./experiments/predictions/")
## Features (Backbone) Training
parser.add_argument("--backbone_valid_split", type=float, default=0.1)
parser.add_argument("--backbone_valid_seed", type=int, default=581)
parser.add_argument("--backbone_train_workers", type=int, default=8)
parser.add_argument("--backbone_optimizer", default="momentum", type=str, choices=["sgd", "adam", "momentum"])
parser.add_argument("--backbone_train_epochs", default=5, type=int)
parser.add_argument("--backbone_train_lr", default=0.01, type=float)
parser.add_argument("--backbone_finetune_epochs", default=100, type=int)
parser.add_argument("--backbone_freezebn", default=False, type=str2bool)
parser.add_argument("--backbone_finetune_layers", default="all") # ResNet Arch => first: "conv1_pad" b4: "conv4_block1_preact_bn"
parser.add_argument("--backbone_finetune_lr", default=0.01, type=float)
parser.add_argument("--backbone_train_supcon_temperature", default=0.1, type=float)
## Features (Backbone) Inference
parser.add_argument("--patches_train", default=2, type=int) # 2 random crops are the default in SupCon (arxiv 2004.11362).
parser.add_argument("--positives_train", default=1, type=int) # Positives from different sources.
parser.add_argument("--patches_test", default=20, type=int) # 2 random crops are the default in SupCon (arxiv 2004.11362).
parser.add_argument("--batch_test", default=2, type=int) # Usually small because `patches_test` is large.
parser.add_argument("--features_parts", default=4, type=int) # default: 4
# Adversarial Feature Hallucination Network
parser.add_argument("--afhp_problem", default="painter", type=str, choices=["painter", "style", "genre"])
parser.add_argument("--afhp_lr", default=1e-5, type=float)
parser.add_argument("--afhp_epochs", default=100, type=int)
parser.add_argument("--afhp_persist", default=True, type=str2bool)
parser.add_argument("--afhp_gp_w", default=10.0, type=float)
parser.add_argument("--afhp_ac_w", default=1.0, type=float)
parser.add_argument("--afhp_cf_w", default=1.0, type=float)
# Steps
parser.add_argument("--step_preprocess_reduce", default=True, type=str2bool)
parser.add_argument("--step_features_train", default=True, type=str2bool)
parser.add_argument("--step_features_infer", default=True, type=str2bool)
parser.add_argument("--step_afhp_train", default=True, type=str2bool)
parser.add_argument("--step_afhp_test", default=True, type=str2bool)
def run(args):
import datasets.pbn
print(__doc__)
print("=" * 65)
print(*(f"{k:<20} = {v}" for k, v in vars(args).items()), sep="\n")
print("=" * 65)
dataset = datasets.pbn.get_dataset(args)
if args.step_preprocess_reduce:
import steps.preprocess_step
steps.preprocess_step.reduce_massive_images(dataset, args)
train_info, test_info = datasets.pbn.train_test_split(dataset, args)
if args.step_features_train or args.step_features_infer:
if args.strategy == "ce":
import steps.backbone_step
steps.backbone_step.vanilla(train_info, dataset, STRATEGY, args)
elif args.strategy == "supcon":
import steps.backbone_step
steps.backbone_step.supcon(train_info, dataset, STRATEGY, args)
else:
import steps.backbone_step
steps.backbone_step.supcon_mh(train_info, dataset, STRATEGY, args)
if args.step_afhp_train or args.step_afhp_test:
import steps.afhp_step
afhp_model, _, test_data = steps.afhp_step.run(train_info, dataset, STRATEGY, args)
if args.step_afhp_test:
import steps.fsl_step
steps.fsl_step.run(args, afhp_model, *test_data)
if __name__ == "__main__":
args = parser.parse_args()
np.random.seed(args.seed + 42)
STRATEGY = boot.appropriate_distributed_strategy()
if args.mixed_precision:
tf.keras.mixed_precision.set_global_policy('mixed_float16')
run(args)