Why Predicted Heat maps trained using efficient_ad models look strange.

[Xuyilongs]

This is the result picture I got when I trained the efficient_ad model using the default data set format, why the predicted heat map here is strange, all green.

Why is the anomaly_maps data always the same value 0.4975?

But on the same data set, using other models such as fastflow, this is what the predicted heat map looks like.

[blaz-r]

Hi, this is indeed quite unusual. I don't remember if I've ever visualized the EfficientAD anomaly maps, so we'll need to check that.

Since the binary segmentation map is still valid, I think that the value difference between anomalous and normal region is just so small in this case, making it impossible to see. This is probably due invalid normalization values or approach. Are all the masks like this?

It is possible that since the map normalization is part of the model, adaptive threshold value that is calculated on maps causes minmax to scale the image in this way (since this threshold is also later used in minmax normalization).

[abc-125]

Hello, I am not sure if this is the case, but please check that you have normalization: none in your config file. EfficientAD handles normalization inside the model, so using the default config from a different model might be the problem.

[blaz-r]

Yeah, that might be a bit problematic then 😅 as the default normalization method for EfficientAD is set to minmax in config:
https://github.com/openvinotoolkit/anomalib/blob/2e09314a2d5ab19321806d0789fc97d369254582/src/anomalib/models/efficient_ad/config.yaml#L21-L30

[abc-125]

I meant normalization for the dataset, it can be set to imagenet, which would be a problem in the case of EfficientAD 😅
https://github.com/openvinotoolkit/anomalib/blob/2e09314a2d5ab19321806d0789fc97d369254582/src/anomalib/models/efficient_ad/config.yaml#L12

[blaz-r]

Oh, that is correct. But since the image above seems to have correct binary mask I think that the model actually learns fine, but the visualization is not right. Can you confirm this @Xuyilongs or is the model not learning at all?

[abc-125]

@Xuyilongs can you also please share your config.yaml file?

[Xuyilongs]

Thanks for your answer, here is my config.yam

dataset:
name: mvtec
format: mvtec
path: ./datasets/MVTec
category: bottle
task: segmentation # 可选：classification，detection，segmentation
train_batch_size: 1
eval_batch_size: 16
num_workers: 8
image_size: 256 # dimensions to which images are resized (mandatory)
center_crop: null # dimensions to which images are center-cropped after resizing (optional)
normalization: none # data distribution to which the images will be normalized: [none, imagenet]
transform_config:
train: null
eval: null
test_split_mode: from_dir # options: [from_dir, synthetic]
test_split_ratio: 0.2 # fraction of train images held out testing (usage depends on test_split_mode)
val_split_mode: same_as_test # options: [same_as_test, from_test, synthetic]
val_split_ratio: 0.5 # fraction of train/test images held out for validation (usage depends on val_split_mode)

model:
name: efficient_ad
teacher_out_channels: 384
model_size: small # options: [small, medium]
lr: 0.0001
weight_decay: 0.00001
padding: false
pad_maps: true # relevant for "padding: false", see EfficientAd in lightning_model.py

generic params

normalization_method: min_max # options: [null, min_max, cdf]

metrics:
image:
- F1Score
- AUROC
pixel:
- F1Score
- AUROC
threshold:
method: adaptive #options: [adaptive, manual]
manual_image: null
manual_pixel: null

visualization:
show_images: False # show images on the screen
save_images: True # save images to the file system
log_images: False # log images to the available loggers (if any)
image_save_path: null # path to which images will be saved
mode: full # options: ["full", "simple"]

project:
seed: 42
path: ./results

logging:
logger: [] # options: [comet, tensorboard, wandb, csv] or combinations.
log_graph: false # Logs the model graph to respective logger.

optimization:
export_mode: null # options: torch, onnx, openvino

PL Trainer Args. Don't add extra parameter here.

trainer:
enable_checkpointing: true
default_root_dir: null
gradient_clip_val: 0
gradient_clip_algorithm: norm
num_nodes: 1
devices: 1
enable_progress_bar: true
overfit_batches: 0.0
track_grad_norm: -1
check_val_every_n_epoch: 1
fast_dev_run: false
accumulate_grad_batches: 1
max_epochs: 20
min_epochs: null
max_steps: 70000
min_steps: null
max_time: null
limit_train_batches: 1.0
limit_val_batches: 1.0
limit_test_batches: 1.0
limit_predict_batches: 1.0
val_check_interval: 1.0
log_every_n_steps: 50
accelerator: auto # <"cpu", "gpu", "tpu", "ipu", "hpu", "auto">
strategy: null
sync_batchnorm: false
precision: 32
enable_model_summary: true
num_sanity_val_steps: 0
profiler: null
benchmark: false
deterministic: false
reload_dataloaders_every_n_epochs: 0
auto_lr_find: false
replace_sampler_ddp: true
detect_anomaly: false
auto_scale_batch_size: false
plugins: null
move_metrics_to_cpu: false
multiple_trainloader_mode: max_size_cycle

[Xuyilongs]

@blaz-r @abc-125 Thank you for your help. You guys have been friendly.

[SilverStarCoder]

Hi, my heatmaps are also looking like yours, even after I set normalization to none. Is there anything else you changed?

[abc-125]

It looks like this PR should fix it, you need this part in particular:

def on_validation_epoch_start(self, trainer: Trainer, pl_module: AnomalyModule) -> None:
  """Call when the validation epoch begins."""
  del trainer  # `trainer` variable is not used.
  
  if hasattr(pl_module, "normalization_metrics"):
      pl_module.normalization_metrics.reset()

[watertianyi]

Which file is normalization_method in? How to set normalization_method=none in TorchInferencer during inference?