Spaces:
Configuration error
Configuration error
Commit
•
eb144e3
1
Parent(s):
fdddc61
Create train_video.py
Browse files- train_video.py +104 -0
train_video.py
ADDED
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import datetime
|
2 |
+
import random
|
3 |
+
from argparse import ArgumentParser
|
4 |
+
from pathlib import Path
|
5 |
+
|
6 |
+
import numpy as np
|
7 |
+
import torch
|
8 |
+
import yaml
|
9 |
+
from tqdm import tqdm
|
10 |
+
|
11 |
+
from datasets.video_dataset import AtlasDataset
|
12 |
+
from models.video_model import VideoModel
|
13 |
+
from util.atlas_loss import AtlasLoss
|
14 |
+
from util.util import get_optimizer
|
15 |
+
from util.video_logger import DataLogger
|
16 |
+
|
17 |
+
|
18 |
+
def train_model(config):
|
19 |
+
# set seed
|
20 |
+
seed = config["seed"]
|
21 |
+
if seed == -1:
|
22 |
+
seed = np.random.randint(2 ** 32)
|
23 |
+
random.seed(seed)
|
24 |
+
np.random.seed(seed)
|
25 |
+
torch.manual_seed(seed)
|
26 |
+
print(f"running with seed: {seed}.")
|
27 |
+
|
28 |
+
dataset = AtlasDataset(config)
|
29 |
+
model = VideoModel(config)
|
30 |
+
criterion = AtlasLoss(config)
|
31 |
+
optimizer = get_optimizer(config, model.parameters())
|
32 |
+
|
33 |
+
logger = DataLogger(config, dataset)
|
34 |
+
with tqdm(range(1, config["n_epochs"] + 1)) as tepoch:
|
35 |
+
for epoch in tepoch:
|
36 |
+
inputs = dataset[0]
|
37 |
+
optimizer.zero_grad()
|
38 |
+
outputs = model(inputs)
|
39 |
+
losses = criterion(outputs, inputs)
|
40 |
+
|
41 |
+
loss = 0.
|
42 |
+
if config["finetune_foreground"]:
|
43 |
+
loss += losses["foreground"]["loss"]
|
44 |
+
elif config["finetune_background"]:
|
45 |
+
loss += losses["background"]["loss"]
|
46 |
+
|
47 |
+
lr = optimizer.param_groups[0]["lr"]
|
48 |
+
log_data = logger.log_data(epoch, lr, losses, model, dataset)
|
49 |
+
|
50 |
+
loss.backward()
|
51 |
+
optimizer.step()
|
52 |
+
optimizer.param_groups[0]["lr"] = max(config["min_lr"], config["gamma"] * optimizer.param_groups[0]["lr"])
|
53 |
+
|
54 |
+
if config["use_wandb"]:
|
55 |
+
wandb.log(log_data)
|
56 |
+
else:
|
57 |
+
if epoch % config["log_images_freq"] == 0:
|
58 |
+
logger.save_locally(log_data)
|
59 |
+
|
60 |
+
tepoch.set_description(f"Epoch {epoch}")
|
61 |
+
tepoch.set_postfix(loss=loss.item())
|
62 |
+
|
63 |
+
|
64 |
+
if __name__ == "__main__":
|
65 |
+
parser = ArgumentParser()
|
66 |
+
parser.add_argument(
|
67 |
+
"--config",
|
68 |
+
default="./configs/video_config.yaml",
|
69 |
+
help="Config path",
|
70 |
+
)
|
71 |
+
parser.add_argument(
|
72 |
+
"--example_config",
|
73 |
+
default="car-turn_winter.yaml",
|
74 |
+
help="Example config name",
|
75 |
+
)
|
76 |
+
args = parser.parse_args()
|
77 |
+
config_path = args.config
|
78 |
+
|
79 |
+
with open(config_path, "r") as f:
|
80 |
+
config = yaml.safe_load(f)
|
81 |
+
with open(f"./configs/video_example_configs/{args.example_config}", "r") as f:
|
82 |
+
example_config = yaml.safe_load(f)
|
83 |
+
config["example_config"] = args.example_config
|
84 |
+
config.update(example_config)
|
85 |
+
|
86 |
+
run_name = f"-{config['checkpoint_path'].split('/')[-2]}"
|
87 |
+
if config["use_wandb"]:
|
88 |
+
import wandb
|
89 |
+
|
90 |
+
wandb.init(project=config["wandb_project"], entity=config["wandb_entity"], config=config, name=run_name)
|
91 |
+
wandb.run.name = str(wandb.run.id) + wandb.run.name
|
92 |
+
config = dict(wandb.config)
|
93 |
+
else:
|
94 |
+
now = datetime.datetime.now()
|
95 |
+
run_name = f"{now.strftime('%Y-%m-%d_%H-%M-%S')}{run_name}"
|
96 |
+
path = Path(f"{config['results_folder']}/{run_name}")
|
97 |
+
path.mkdir(parents=True, exist_ok=True)
|
98 |
+
with open(path / "config.yaml", "w") as f:
|
99 |
+
yaml.dump(config, f)
|
100 |
+
config["results_folder"] = str(path)
|
101 |
+
|
102 |
+
train_model(config)
|
103 |
+
if config["use_wandb"]:
|
104 |
+
wandb.finish()
|