Why is the result different when using TilerConfiguration callback and not using TilerConfiguration callback？

[wenwu2021]

def predict(a):
transform = Compose([
Resize(size=(256, 256)),

])
dataset = PredictDataset(
    path=r"D:\BaiduNetdiskDownload\Zbar\AnomalibTensorrtAlg\x64\Debug\result\033.png",
    transform=transform,
)
pre_processor = PreProcessor(transform=transform)

# model = Patchcore()
# # prepare tiling configuration callback
model = Patchcore(pre_processor=pre_processor)
print(a)
if a==1:

    tiler_config_callback = TilerConfigurationCallback(enable=True, tile_size=[128, 128], stride=64)
    engine = Engine(callbacks=[tiler_config_callback])
else:
    engine = Engine()
# train the model (tiling is seamlessly utilized in the background)
predictions=engine.predict(dataset=dataset,
               model=model,
               ckpt_path=r"D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\results\Patchcore\MVTecAD\tile\v4\weights\lightning\model.ckpt")
# 5. Access the results
if predictions is not None:
    for prediction in predictions:
        image_path = prediction.image_path
        print(image_path)
        anomaly_map = prediction.anomaly_map  # Pixel-level anomaly heatmap
        pred_label = prediction.pred_label  # Image-level label (0: normal, 1: anomalous)
        pred_score = prediction.pred_score  # Image-level anomaly score
        print(pred_score)

a=1
tensor([0.3981])

a=2
tensor([0.4398])

[alfieroddan]

Tiling will naturally produce different results for different models as image features are computed at different scales.

A Tiler splits a large image into smaller tiles (patches), usually with some overlap, and processes each tile independently (often with the same model). After inference, the predictions from each tile are reassembled into a full-size output.

[wenwu2021]

tiler_config_callback = TilerConfigurationCallback(enable=True, tile_size=[128, 128], stride=64)
engine = Engine(callbacks=[tiler_config_callback])

I trained in this way,So what method is used for deep learning inference？

[alfieroddan]

Callbacks currently are 'non-critical' meaning they are not needed for the model to run in prediction. I suggest keeping the same code for inference for now.

Something like this:

# 1. Import required modules
from pathlib import Path
from anomalib.data import PredictDataset
from anomalib.engine import Engine
from anomalib.models import Patchcore
from anomalib.callbacks import TilerConfigurationCallback


# 2. Initialize the model and load weights
model = Patchcore()
tiler_config_callback = TilerConfigurationCallback(enable=True, tile_size=[128, 128], stride=128)
engine = Engine(callbacks=[tiler_config_callback])

# 3. Prepare test data
# You can use a single image or a folder of images
dataset = PredictDataset(
    path=Path("datasets/MVTecAD/bottle/test/broken_large/000.png"),
)

# 4. Get predictions
predictions = engine.predict(
    model=model,
    dataset=dataset,
    ckpt_path="results/Patchcore/MVTecAD/bottle/v6/weights/lightning/model.ckpt",
)

# 5. Access the results
if predictions is not None:
    for prediction in predictions:
        image_path = prediction.image_path
        anomaly_map = prediction.anomaly_map  # Pixel-level anomaly heatmap
        pred_label = prediction.pred_label  # Image-level label (0: normal, 1: anomalous)
        pred_score = prediction.pred_score  # Image-level anomaly score
        print(anomaly_map.shape)

@samet-akcay is the plan down the line to load callbacks like TilerConfigurationCallback for inference or in the model graph like Pre and Post Processing?

Currently the TilerConfigurationCallback is not saved in the checkpoint and is not used at inference.

[wenwu2021]

1. Import required modules

from pathlib import Path
from anomalib.data import PredictDataset
from anomalib.engine import Engine
from anomalib.models import Patchcore
from anomalib.callbacks import TilerConfigurationCallback

2. Initialize the model and load weights

model = Patchcore()
tiler_config_callback = TilerConfigurationCallback(enable=True, tile_size=[128, 128], stride=128)
engine = Engine(callbacks=[tiler_config_callback])

3. Prepare test data

You can use a single image or a folder of images

dataset = PredictDataset(
path=Path("datasets/MVTecAD/bottle/test/broken_large/000.png"),
)

4. Get predictions

predictions = engine.predict(
model=model,
dataset=dataset,
ckpt_path="results/Patchcore/MVTecAD/bottle/v6/weights/lightning/model.ckpt",
)

5. Access the results

if predictions is not None:
for prediction in predictions:
image_path = prediction.image_path
anomaly_map = prediction.anomaly_map # Pixel-level anomaly heatmap
pred_label = prediction.pred_label # Image-level label (0: normal, 1: anomalous)
pred_score = prediction.pred_score # Image-level anomaly score
print(anomaly_map.shape)

tensor([0.4241])

[samet-akcay]

Callbacks currently are 'non-critical' meaning they are not needed for the model to run in prediction. I suggest keeping the same code for inference for now.

Something like this:

1. Import required modules

from pathlib import Path
from anomalib.data import PredictDataset
from anomalib.engine import Engine
from anomalib.models import Patchcore
from anomalib.callbacks import TilerConfigurationCallback

2. Initialize the model and load weights

model = Patchcore()
tiler_config_callback = TilerConfigurationCallback(enable=True, tile_size=[128, 128], stride=128)
engine = Engine(callbacks=[tiler_config_callback])

3. Prepare test data

You can use a single image or a folder of images

dataset = PredictDataset(
path=Path("datasets/MVTecAD/bottle/test/broken_large/000.png"),
)

4. Get predictions

predictions = engine.predict(
model=model,
dataset=dataset,
ckpt_path="results/Patchcore/MVTecAD/bottle/v6/weights/lightning/model.ckpt",
)

5. Access the results

if predictions is not None:
for prediction in predictions:
image_path = prediction.image_path
anomaly_map = prediction.anomaly_map # Pixel-level anomaly heatmap
pred_label = prediction.pred_label # Image-level label (0: normal, 1: anomalous)
pred_score = prediction.pred_score # Image-level anomaly score
print(anomaly_map.shape)
@samet-akcay is the plan down the line to load callbacks like TilerConfigurationCallback for inference or in the model graph like Pre and Post Processing?

Currently the TilerConfigurationCallback is not saved in the checkpoint and is not used at inference.

@alfieroddan, yeah, ideally Tiler can be a preprocessor or within an existing pre-processor. --we can discuss the design choices

In this case, we would not even need TilerConfigurationCallback, which was implemented in v0.*

[wenwu2021]

transform = Compose([
    Resize(size=(256, 256)),
  
])
dataset = PredictDataset(
    path="datasets/MVTecAD/bottle/test/broken_large/000.png",
    transform=transform,
)
pre_processor = PreProcessor(transform=transform)

model = Patchcore(pre_processor=pre_processor)
engine = Engine()
predictions=engine.predict(dataset=dataset,
               model=model,
               ckpt_path=r"results\Patchcore\MVTecAD\tile\v4\weights\lightning\model.ckpt")
# 5. Access the results
if predictions is not None:
    for prediction in predictions:
        image_path = prediction.image_path
        print(image_path)
        anomaly_map = prediction.anomaly_map  # Pixel-level anomaly heatmap
        pred_label = prediction.pred_label  # Image-level label (0: normal, 1: anomalous)
        pred_score = prediction.pred_score  # Image-level anomaly score
        print(pred_score)

is this ?