def _validate_data(self, data):
"""Check whether data is valid, try to convert with best effort if not"""
if isinstance(data, pd.DataFrame):
# TODO(zhreshold): allow custom label column without this renaming trick
if self._label != 'label' and self._label in data.columns:
# data is deepcopied so it's okay to overwrite directly
data = data.rename(columns={
'label': '_unused_label',
self._label: 'label'
},
errors='ignore')
if not (hasattr(data, 'classes') and hasattr(data, 'to_mxnet')):
if isinstance(data, pd.DataFrame):
# raw dataframe, try to add metadata automatically
if 'label' in data.columns and 'image' in data.columns:
# check image relative/abs path is valid
sample = data.iloc[0]['image']
if not os.path.isfile(sample):
raise OSError(
f'Detected invalid image path `{sample}`, please ensure all image paths are absolute or you are using the right working directory.'
)
logger.log(
20,
'Converting raw DataFrame to ImagePredictor.Dataset...'
)
infer_classes = sorted(data.label.unique().tolist())
logger.log(
20,
f'Detected {len(infer_classes)} unique classes: {infer_classes}'
)
instruction = 'train_data = ImagePredictor.Dataset(train_data, classes=["foo", "bar"])'
logger.log(
20,
f'If you feel the `classes` is inaccurate, please construct the dataset explicitly, e.g. {instruction}'
)
data = _ImageClassification.Dataset(data,
classes=infer_classes)
else:
err_msg = 'Unable to convert raw DataFrame to ImagePredictor Dataset, ' + \
'`image` and `label` columns are required.' + \
'You may visit `https://auto.gluon.ai/stable/tutorials/image_prediction/dataset.html` ' + \
'for details.'
raise AttributeError(err_msg)
else:
raise TypeError(
f"Unable to process dataset of type: {type(data)}")
elif isinstance(data, _ImageClassification.Dataset):
assert 'label' in data.columns
assert hasattr(data, 'classes')
# check whether classes are outdated, no action required if all unique labels is subset of `classes`
unique_labels = sorted(data['label'].unique().tolist())
if not (all(ulabel in data.classes for ulabel in unique_labels)):
data = _ImageClassification.Dataset(data,
classes=unique_labels)
logger.log(20, f'Reset labels to {unique_labels}')
if len(data) < 1:
raise ValueError('Empty dataset.')
return data
def test_time_out_image_classification():
time_limit = 30
from gluoncv.auto.tasks import ImageClassification
task = ImageClassification({'num_trials': 1, 'epochs': 50})
tic = time.time()
classifier = task.fit(IMAGE_CLASS_DATASET, time_limit=time_limit)
# check time_limit with a little bit overhead
assert (time.time() - tic) < time_limit + 180
def test_torch_image_classification_custom_net():
from gluoncv.auto.tasks import ImageClassification
from timm import create_model
import torch.nn as nn
net = create_model('resnet18')
net.fc = nn.Linear(512, 4)
task = ImageClassification({'num_trials': 1, 'epochs': 1, 'custom_net': net, 'batch_size': 8})
classifier = task.fit(IMAGE_CLASS_DATASET)
assert task.fit_summary().get('valid_acc', 0) > 0
test_result = classifier.predict(IMAGE_CLASS_TEST)
def test_image_classification():
from gluoncv.auto.tasks import ImageClassification
task = ImageClassification({
'model': 'resnet18_v1',
'num_trials': 1,
'epochs': 1,
'batch_size': 8
})
classifier = task.fit(IMAGE_CLASS_DATASET)
assert task.fit_summary().get('valid_acc', 0) > 0
test_result = classifier.predict(IMAGE_CLASS_TEST)
def test_image_classification_custom_net():
from gluoncv.auto.tasks import ImageClassification
from gluoncv.model_zoo import get_model
net = get_model('resnet18_v1')
task = ImageClassification({
'num_trials': 1,
'epochs': 1,
'custom_net': net
})
classifier = task.fit(IMAGE_CLASS_DATASET)
assert task.fit_summary().get('valid_acc', 0) > 0
test_result = classifier.predict(IMAGE_CLASS_TEST)
def _validate_data(self, data):
"""Check whether data is valid, try to convert with best effort if not"""
if not (hasattr(data, 'classes') and hasattr(data, 'to_mxnet')):
if isinstance(data, pd.DataFrame):
# raw dataframe, try to add metadata automatically
if 'label' in data.columns and 'image' in data.columns:
# check image relative/abs path is valid
sample = data.iloc[0]['image']
if not os.path.isfile(sample):
raise OSError(f'Detected invalid image path `{sample}`, please ensure all image paths are absolute or you are using the right working directory.')
logger.log(20, 'Converting raw DataFrame to ImagePredictor.Dataset...')
infer_classes = list(data.label.unique().tolist())
logger.log(20, f'Detected {len(infer_classes)} unique classes: {infer_classes}')
instruction = 'train_data = ImagePredictor.Dataset(train_data, classes=["foo", "bar"])'
logger.log(20, f'If you feel the `classes` is inaccurate, please construct the dataset explicitly, e.g. {instruction}')
data = _ImageClassification.Dataset(data, classes=infer_classes)
else:
err_msg = 'Unable to convert raw DataFrame to ImagePredictor Dataset, ' + \
'`image` and `label` columns are required.' + \
'You may visit `https://auto.gluon.ai/stable/tutorials/image_prediction/dataset.html` ' + \
'for details.'
raise AttributeError(err_msg)
if len(data) < 1:
raise ValueError('Empty dataset.')
return data
def test_image_classification():
from gluoncv.auto.tasks import ImageClassification
task = ImageClassification({'num_trials': 1})
classifier = task.fit(IMAGE_CLASS_DATASET)
assert task.fit_summary.get('valid_acc', 0) > 0
test_result = classifier.predict(IMAGE_CLASS_TEST)
'batch_size':
ag.Int(4, 7), # [16, 32, 64, 128]
'momentum':
ag.Real(0.85, 0.95),
'wd':
ag.Real(1e-6, 1e-2, log=True),
'epochs':
15,
'num_trials':
args.num_trials,
'search_strategy':
'bayesopt'
}
# specify learning task
task = ImageClassification(config)
# specify dataset
dataset = Dataset.get(args.dataset)
train_data, valid_data = dataset.split(0.8)
# fit auto estimator
classifier = task.fit(train_data, valid_data)
# evaluate auto estimator
top1, top5 = classifier.evaluate(valid_data)
logging.info('evaluation: top1={}, top5={}'.format(top1, top5))
# save and load auto estimator
classifier.save('classifier.pkl')
classifier = ImageClassification.load('classifier.pkl')
