File size: 4,369 Bytes
ff82fe6 3163344 10f217b 3163344 8e068be 3163344 00a4b21 10f217b ff82fe6 416c7bb 855e240 01d9a57 0693434 bc0f7fc ff82fe6 fd9442f 3163344 fd9442f 3163344 65e514f 3163344 0693434 3163344 8364103 3163344 8d2e833 68dd12a 8364103 3163344 bc0f7fc fd9442f 3163344 68dd12a 01d9a57 855e240 3163344 c681b80 3163344 416c7bb 3163344 1750035 3163344 912d566 3163344 8d9c0ef 5c43f60 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 |
import argparse
import os
import torch
import pytorch_lightning as ptl
from pytorch_lightning.loggers import TensorBoardLogger
from detector.data import FontDataModule
from detector.model import *
from utils import get_current_tag
torch.set_float32_matmul_precision("high")
parser = argparse.ArgumentParser()
parser.add_argument(
"-d",
"--devices",
nargs="*",
type=int,
default=[0],
help="GPU devices to use (default: [0])",
)
parser.add_argument(
"-b",
"--single-batch-size",
type=int,
default=64,
help="Batch size of single device (default: 64)",
)
parser.add_argument(
"-c",
"--checkpoint",
type=str,
default=None,
help="Trainer checkpoint path (default: None)",
)
parser.add_argument(
"-m",
"--model",
type=str,
default="resnet18",
choices=["resnet18", "resnet34", "resnet50", "resnet101"],
help="Model to use (default: resnet18)",
)
parser.add_argument(
"-p",
"--pretrained",
action="store_true",
help="Use pretrained model for ResNet (default: False)",
)
parser.add_argument(
"-i",
"--crop-roi-bbox",
action="store_true",
help="Crop ROI bounding box (default: False)",
)
parser.add_argument(
"-a",
"--augmentation",
type=str,
default=None,
choices=["v1", "v2"],
help="Augmentation strategy to use (default: None)",
)
parser.add_argument(
"-l",
"--lr",
type=float,
default=0.0001,
help="Learning rate (default: 0.0001)",
)
parser.add_argument(
"-s",
"--datasets",
nargs="*",
type=str,
default=["./dataset/font_img"],
help="Datasets paths, seperated by space (default: ['./dataset/font_img'])",
)
args = parser.parse_args()
devices = args.devices
single_batch_size = args.single_batch_size
total_num_workers = os.cpu_count()
single_device_num_workers = total_num_workers // len(devices)
if os.name == "nt":
single_device_num_workers = 0
lr = args.lr
b1 = 0.9
b2 = 0.999
lambda_font = 2.0
lambda_direction = 0.5
lambda_regression = 1.0
regression_use_tanh = False
num_warmup_epochs = 5
num_epochs = 100
log_every_n_steps = 100
num_device = len(devices)
data_module = FontDataModule(
train_paths=[os.path.join(path, "train") for path in args.datasets],
val_paths=[os.path.join(path, "val") for path in args.datasets],
test_paths=[os.path.join(path, "test") for path in args.datasets],
batch_size=single_batch_size,
num_workers=single_device_num_workers,
pin_memory=True,
train_shuffle=True,
val_shuffle=False,
test_shuffle=False,
regression_use_tanh=regression_use_tanh,
train_transforms=args.augmentation,
crop_roi_bbox=args.crop_roi_bbox,
)
num_iters = data_module.get_train_num_iter(num_device) * num_epochs
num_warmup_iter = data_module.get_train_num_iter(num_device) * num_warmup_epochs
model_name = f"{get_current_tag()}"
logger_unconditioned = TensorBoardLogger(
save_dir=os.getcwd(), name="tensorboard", version=model_name
)
strategy = None if num_device == 1 else "ddp"
trainer = ptl.Trainer(
max_epochs=num_epochs,
logger=logger_unconditioned,
devices=devices,
accelerator="gpu",
enable_checkpointing=True,
log_every_n_steps=log_every_n_steps,
strategy=strategy,
deterministic=True,
)
if args.model == "resnet18":
model = ResNet18Regressor(
pretrained=args.pretrained, regression_use_tanh=regression_use_tanh
)
elif args.model == "resnet34":
model = ResNet34Regressor(
pretrained=args.pretrained, regression_use_tanh=regression_use_tanh
)
elif args.model == "resnet50":
model = ResNet50Regressor(
pretrained=args.pretrained, regression_use_tanh=regression_use_tanh
)
elif args.model == "resnet101":
model = ResNet101Regressor(
pretrained=args.pretrained, regression_use_tanh=regression_use_tanh
)
else:
raise NotImplementedError()
if torch.__version__ >= "2.0":
model = torch.compile(model)
detector = FontDetector(
model=model,
lambda_font=lambda_font,
lambda_direction=lambda_direction,
lambda_regression=lambda_regression,
lr=lr,
betas=(b1, b2),
num_warmup_iters=num_warmup_iter,
num_iters=num_iters,
num_epochs=num_epochs,
)
trainer.fit(detector, datamodule=data_module, ckpt_path=args.checkpoint)
trainer.test(detector, datamodule=data_module)
|