Skip to content

Commit 73a0669

Browse files
Start setup for improved W&B integration (#1948)
* Add helper functions for wandb and artifacts * cleanup * Reorganize files * Update wandb_utils.py * Update log_dataset.py We can remove this code, as the giou hyp has been deprecated for a while now. * Reorganize and update dataloader call * yaml.SafeLoader * PEP8 reformat * remove redundant checks * Add helper functions for wandb and artifacts * cleanup * Reorganize files * Update wandb_utils.py * Update log_dataset.py We can remove this code, as the giou hyp has been deprecated for a while now. * Reorganize and update dataloader call * yaml.SafeLoader * PEP8 reformat * remove redundant checks * Update util files * Update wandb_utils.py * Remove word size * Change path of labels.zip * remove unused imports * remove --rect * log_dataset.py cleanup * log_dataset.py cleanup2 * wandb_utils.py cleanup * remove redundant id_count * wandb_utils.py cleanup2 * rename cls * use pathlib for zip * rename dataloader to dataset * Change import order * Remove redundant code * remove unused import * remove unused imports Co-authored-by: Glenn Jocher <[email protected]>
1 parent 9646ca4 commit 73a0669

File tree

4 files changed

+186
-1
lines changed

4 files changed

+186
-1
lines changed

utils/datasets.py

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -348,7 +348,8 @@ def __init__(self, path, img_size=640, batch_size=16, augment=False, hyp=None, r
348348
self.mosaic = self.augment and not self.rect # load 4 images at a time into a mosaic (only during training)
349349
self.mosaic_border = [-img_size // 2, -img_size // 2]
350350
self.stride = stride
351-
351+
self.path = path
352+
352353
try:
353354
f = [] # image files
354355
for p in path if isinstance(path, list) else [path]:

utils/wandb_logging/__init__.py

Whitespace-only changes.

utils/wandb_logging/log_dataset.py

Lines changed: 39 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,39 @@
1+
import argparse
2+
from pathlib import Path
3+
4+
import yaml
5+
6+
from wandb_utils import WandbLogger
7+
from utils.datasets import LoadImagesAndLabels
8+
9+
WANDB_ARTIFACT_PREFIX = 'wandb-artifact://'
10+
11+
12+
def create_dataset_artifact(opt):
13+
with open(opt.data) as f:
14+
data = yaml.load(f, Loader=yaml.SafeLoader) # data dict
15+
logger = WandbLogger(opt, '', None, data, job_type='create_dataset')
16+
nc, names = (1, ['item']) if opt.single_cls else (int(data['nc']), data['names'])
17+
names = {k: v for k, v in enumerate(names)} # to index dictionary
18+
logger.log_dataset_artifact(LoadImagesAndLabels(data['train']), names, name='train') # trainset
19+
logger.log_dataset_artifact(LoadImagesAndLabels(data['val']), names, name='val') # valset
20+
21+
# Update data.yaml with artifact links
22+
data['train'] = WANDB_ARTIFACT_PREFIX + str(Path(opt.project) / 'train')
23+
data['val'] = WANDB_ARTIFACT_PREFIX + str(Path(opt.project) / 'val')
24+
path = opt.data if opt.overwrite_config else opt.data.replace('.', '_wandb.') # updated data.yaml path
25+
data.pop('download', None) # download via artifact instead of predefined field 'download:'
26+
with open(path, 'w') as f:
27+
yaml.dump(data, f)
28+
print("New Config file => ", path)
29+
30+
31+
if __name__ == '__main__':
32+
parser = argparse.ArgumentParser()
33+
parser.add_argument('--data', type=str, default='data/coco128.yaml', help='data.yaml path')
34+
parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset')
35+
parser.add_argument('--project', type=str, default='YOLOv5', help='name of W&B Project')
36+
parser.add_argument('--overwrite_config', action='store_true', help='overwrite data.yaml')
37+
opt = parser.parse_args()
38+
39+
create_dataset_artifact(opt)

utils/wandb_logging/wandb_utils.py

Lines changed: 145 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,145 @@
1+
import json
2+
import shutil
3+
import sys
4+
from datetime import datetime
5+
from pathlib import Path
6+
7+
import torch
8+
9+
sys.path.append(str(Path(__file__).parent.parent.parent)) # add utils/ to path
10+
from utils.general import colorstr, xywh2xyxy
11+
12+
try:
13+
import wandb
14+
except ImportError:
15+
wandb = None
16+
print(f"{colorstr('wandb: ')}Install Weights & Biases for YOLOv5 logging with 'pip install wandb' (recommended)")
17+
18+
WANDB_ARTIFACT_PREFIX = 'wandb-artifact://'
19+
20+
21+
def remove_prefix(from_string, prefix):
22+
return from_string[len(prefix):]
23+
24+
25+
class WandbLogger():
26+
def __init__(self, opt, name, run_id, data_dict, job_type='Training'):
27+
self.wandb = wandb
28+
self.wandb_run = wandb.init(config=opt, resume="allow",
29+
project='YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem,
30+
name=name,
31+
job_type=job_type,
32+
id=run_id) if self.wandb else None
33+
34+
if job_type == 'Training':
35+
self.setup_training(opt, data_dict)
36+
if opt.bbox_interval == -1:
37+
opt.bbox_interval = (opt.epochs // 10) if opt.epochs > 10 else opt.epochs
38+
if opt.save_period == -1:
39+
opt.save_period = (opt.epochs // 10) if opt.epochs > 10 else opt.epochs
40+
41+
def setup_training(self, opt, data_dict):
42+
self.log_dict = {}
43+
self.train_artifact_path, self.trainset_artifact = \
44+
self.download_dataset_artifact(data_dict['train'], opt.artifact_alias)
45+
self.test_artifact_path, self.testset_artifact = \
46+
self.download_dataset_artifact(data_dict['val'], opt.artifact_alias)
47+
self.result_artifact, self.result_table, self.weights = None, None, None
48+
if self.train_artifact_path is not None:
49+
train_path = Path(self.train_artifact_path) / 'data/images/'
50+
data_dict['train'] = str(train_path)
51+
if self.test_artifact_path is not None:
52+
test_path = Path(self.test_artifact_path) / 'data/images/'
53+
data_dict['val'] = str(test_path)
54+
self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation")
55+
self.result_table = wandb.Table(["epoch", "id", "prediction", "avg_confidence"])
56+
if opt.resume_from_artifact:
57+
modeldir, _ = self.download_model_artifact(opt.resume_from_artifact)
58+
if modeldir:
59+
self.weights = Path(modeldir) / "best.pt"
60+
opt.weights = self.weights
61+
62+
def download_dataset_artifact(self, path, alias):
63+
if path.startswith(WANDB_ARTIFACT_PREFIX):
64+
dataset_artifact = wandb.use_artifact(remove_prefix(path, WANDB_ARTIFACT_PREFIX) + ":" + alias)
65+
assert dataset_artifact is not None, "'Error: W&B dataset artifact doesn\'t exist'"
66+
datadir = dataset_artifact.download()
67+
labels_zip = Path(datadir) / "data/labels.zip"
68+
shutil.unpack_archive(labels_zip, Path(datadir) / 'data/labels', 'zip')
69+
print("Downloaded dataset to : ", datadir)
70+
return datadir, dataset_artifact
71+
return None, None
72+
73+
def download_model_artifact(self, name):
74+
model_artifact = wandb.use_artifact(name + ":latest")
75+
assert model_artifact is not None, 'Error: W&B model artifact doesn\'t exist'
76+
modeldir = model_artifact.download()
77+
print("Downloaded model to : ", modeldir)
78+
return modeldir, model_artifact
79+
80+
def log_model(self, path, opt, epoch):
81+
datetime_suffix = datetime.today().strftime('%Y-%m-%d-%H-%M-%S')
82+
model_artifact = wandb.Artifact('run_' + wandb.run.id + '_model', type='model', metadata={
83+
'original_url': str(path),
84+
'epoch': epoch + 1,
85+
'save period': opt.save_period,
86+
'project': opt.project,
87+
'datetime': datetime_suffix
88+
})
89+
model_artifact.add_file(str(path / 'last.pt'), name='last.pt')
90+
model_artifact.add_file(str(path / 'best.pt'), name='best.pt')
91+
wandb.log_artifact(model_artifact)
92+
print("Saving model artifact on epoch ", epoch + 1)
93+
94+
def log_dataset_artifact(self, dataset, class_to_id, name='dataset'):
95+
artifact = wandb.Artifact(name=name, type="dataset")
96+
image_path = dataset.path
97+
artifact.add_dir(image_path, name='data/images')
98+
table = wandb.Table(columns=["id", "train_image", "Classes"])
99+
class_set = wandb.Classes([{'id': id, 'name': name} for id, name in class_to_id.items()])
100+
for si, (img, labels, paths, shapes) in enumerate(dataset):
101+
height, width = shapes[0]
102+
labels[:, 2:] = (xywh2xyxy(labels[:, 2:].view(-1, 4)))
103+
labels[:, 2:] *= torch.Tensor([width, height, width, height])
104+
box_data = []
105+
img_classes = {}
106+
for cls, *xyxy in labels[:, 1:].tolist():
107+
cls = int(cls)
108+
box_data.append({"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]},
109+
"class_id": cls,
110+
"box_caption": "%s" % (class_to_id[cls]),
111+
"scores": {"acc": 1},
112+
"domain": "pixel"})
113+
img_classes[cls] = class_to_id[cls]
114+
boxes = {"ground_truth": {"box_data": box_data, "class_labels": class_to_id}} # inference-space
115+
table.add_data(si, wandb.Image(paths, classes=class_set, boxes=boxes), json.dumps(img_classes))
116+
artifact.add(table, name)
117+
labels_path = 'labels'.join(image_path.rsplit('images', 1))
118+
zip_path = Path(labels_path).parent / (name + '_labels.zip')
119+
if not zip_path.is_file(): # make_archive won't check if file exists
120+
shutil.make_archive(zip_path.with_suffix(''), 'zip', labels_path)
121+
artifact.add_file(str(zip_path), name='data/labels.zip')
122+
wandb.log_artifact(artifact)
123+
print("Saving data to W&B...")
124+
125+
def log(self, log_dict):
126+
if self.wandb_run:
127+
for key, value in log_dict.items():
128+
self.log_dict[key] = value
129+
130+
def end_epoch(self):
131+
if self.wandb_run and self.log_dict:
132+
wandb.log(self.log_dict)
133+
self.log_dict = {}
134+
135+
def finish_run(self):
136+
if self.wandb_run:
137+
if self.result_artifact:
138+
print("Add Training Progress Artifact")
139+
self.result_artifact.add(self.result_table, 'result')
140+
train_results = wandb.JoinedTable(self.testset_artifact.get("val"), self.result_table, "id")
141+
self.result_artifact.add(train_results, 'joined_result')
142+
wandb.log_artifact(self.result_artifact)
143+
if self.log_dict:
144+
wandb.log(self.log_dict)
145+
wandb.run.finish()

0 commit comments

Comments
 (0)