❓ <title>Dinomaly heatmap issue #3035
Description
Your Question
i trained the Dinomaly model on my custiom dataset
this is my config ->
anomalib==2.3.0dev
seed_everything: true
trainer:
accelerator: auto
strategy: auto
devices: auto
num_nodes: 1
precision: null
logger: null
callbacks:
- class_path: lightning.pytorch.callbacks.EarlyStopping
init_args:
monitor: train_loss_step
min_delta: 0.0
patience: 20
verbose: false
mode: max
strict: true
check_finite: true
stopping_threshold: null
divergence_threshold: null
check_on_train_epoch_end: null
log_rank_zero_only: false
fast_dev_run: false
max_epochs: 60
min_epochs: null
max_steps: -1
min_steps: null
max_time: null
limit_train_batches: null
limit_val_batches: null
limit_test_batches: null
limit_predict_batches: null
overfit_batches: 0.0
val_check_interval: null
check_val_every_n_epoch: 1
num_sanity_val_steps: null
log_every_n_steps: null
enable_checkpointing: null
enable_progress_bar: null
enable_model_summary: null
accumulate_grad_batches: 1
gradient_clip_val: null
gradient_clip_algorithm: null
deterministic: null
benchmark: null
inference_mode: true
use_distributed_sampler: true
profiler: null
detect_anomaly: false
barebones: false
plugins: null
sync_batchnorm: false
reload_dataloaders_every_n_epochs: 0
model_registry: null
model:
class_path: anomalib.models.Dinomaly
init_args:
encoder_name: dinov2reg_vit_base_14
bottleneck_dropout: 0.2
decoder_depth: 8
target_layers: null
fuse_layer_encoder: null
fuse_layer_decoder: null
remove_class_token: false
pre_processor: true
post_processor: true
evaluator: true
visualizer: true
data:
class_path: anomalib.data.Folder
init_args:
name: rotated_with_mask_testing
normal_dir: train/good
root: /home/rotated
abnormal_dir: train/bad
normal_test_dir: test/good
mask_dir: train/ground_truth_binary
normal_split_ratio: 0.2
extensions: null
train_batch_size: 16
eval_batch_size: 16
num_workers: 8
train_augmentations: null
val_augmentations: null
test_augmentations: null
augmentations: null
test_split_mode: from_dir
test_split_ratio: 0.2
val_split_mode: same_as_test
val_split_ratio: 0.5
seed: 42
return_predictions: true
ckpt_path: /home/model.ckpt
logging:
log_graph: true
default_root_dir: results
i used the cli command to train
anomalib fit -c
(dinomaly_venv) orama@orama-H610M-K-DDR4:~/Desktop/prince/dinomaly_anomalib$ anomalib fit -c config.yaml
2025-10-17 12:41:28,276 - anomalib.models.components.base.anomalib_module - INFO - Initializing Dinomaly model.
[10/17/25 12:41:28] INFO Initializing Dinomaly model. anomalib_module.py:131
2025-10-17 12:41:29,212 - anomalib.models.image.dinomaly.components.dinov2_loader - INFO - Loaded model: dinov2reg_vit_base_14
[10/17/25 12:41:29] INFO Loaded model: dinov2reg_vit_base_14 dinov2_loader.py:71
2025-10-17 12:41:29,648 - anomalib.callbacks - INFO - Loading the callbacks
INFO Loading the callbacks init.py:114
2025-10-17 12:41:29,650 - anomalib.engine.engine - INFO - Overriding gradient_clip_val from None with 0.1 for Dinomaly
INFO Overriding gradient_clip_val from None with 0.1 for Dinomaly engine.py:102
2025-10-17 12:41:29,651 - anomalib.engine.engine - INFO - Overriding num_sanity_val_steps from None with 0 for Dinomaly
INFO Overriding num_sanity_val_steps from None with 0 for Dinomaly engine.py:102
GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
HPU available: False, using: 0 HPUs
You are using a CUDA device ('NVIDIA GeForce RTX 3060') that has Tensor Cores. To properly utilize them, you should set torch.set_float32_matmul_precision('medium' | 'high') which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
2025-10-17 12:41:29,750 - lightning.pytorch.accelerators.cuda - INFO - LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
[10/17/25 12:41:29] INFO LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0] cuda.py:61
2025-10-17 12:41:29,960 - lightning.pytorch.callbacks.model_summary - INFO -
| Name | Type | Params | Mode
0 | pre_processor | PreProcessor | 0 | train
1 | post_processor | PostProcessor | 0 | train
2 | evaluator | Evaluator | 0 | train
3 | model | DinomalyModel | 147 M | train
4 | trainable_modules | ModuleList | 61.4 M | train
61.4 M Trainable params
86.6 M Non-trainable params
147 M Total params
591.898 Total estimated model params size (MB)
385 Modules in train mode
0 Modules in eval mode
INFO model_summary.py:104
| Name | Type | Params | Mode
------------------------------------------------------------
0 | pre_processor | PreProcessor | 0 | train
1 | post_processor | PostProcessor | 0 | train
2 | evaluator | Evaluator | 0 | train
3 | model | DinomalyModel | 147 M | train
4 | trainable_modules | ModuleList | 61.4 M | train
------------------------------------------------------------
61.4 M Trainable params
86.6 M Non-trainable params
147 M Total params
591.898 Total estimated model params size (MB)
385 Modules in train mode
0 Modules in eval mode
Epoch 0: 0%| | 0/6 [00:00<?, ?it/s]/home/orama/Desktop/prince/dinomaly_anomalib/dinomaly_venv/lib/python3.10/site-packages/lightning/pytorch/utilities/data.py:79: Trying to infer the batch_size from an ambiguous collection. The batch size we found is 3. To avoid any miscalculations, use self.log(..., batch_size=batch_size).
Epoch 20: 100%|█████████████████████████████████████████████████████| 6/6 [00:14<00:00, 0.41it/s, train_loss_step=0.0647, train_loss_epoch=0.0629]
2025-10-17 12:48:39,322 - anomalib.callbacks.timer - INFO - Training took 429.37 seconds
[10/17/25 12:48:39] INFO Training took 429.37 seconds
After reviewing the results, I found that the model is able to detect defects in the data. However, the heatmaps generated are slightly shifted from the actual defect regions.
For example, in the image below, the defect is detected, but the attention in the heatmap highlights a slightly different location:
how to solve this issue ?
Forum Check
- I have searched the discussions forum for an answer to my question.