❓ How to train a model with GrayScale images

[wenwu2021]

Your Question

D:\anaconda3\envs\ultralytics\python.exe D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\PatchcoreTrain.py
D:\anaconda3\envs\ultralytics\lib\site-packages\openvino\runtime_init_.py:10: DeprecationWarning: The openvino.runtime module is deprecated and will be removed in the 2026.0 release. Please replace openvino.runtime with openvino.
warnings.warn(
D:\anaconda3\envs\ultralytics\lib\site-packages\timm\models\layers_init_.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
warnings.warn(f"Importing from {name} is deprecated, please import via timm.layers", FutureWarning)
D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\utilities\parsing.py:209: Attribute 'pre_processor' is an instance of nn.Module and is already saved during checkpointing. It is recommended to ignore them using self.save_hyperparameters(ignore=['pre_processor']).
GPU available: False, used: False
TPU available: False, using: 0 TPU cores
HPU available: False, using: 0 HPUs
D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\core\optimizer.py:183: LightningModule.configure_optimizers returned None, this fit will run with no optimizer

| Name | Type | Params | Mode

0 | pre_processor | PreProcessor | 0 | train
1 | post_processor | PostProcessor | 0 | train
2 | evaluator | Evaluator | 0 | train
3 | model | PatchcoreModel | 24.9 M | train

24.9 M Trainable params
0 Non-trainable params
24.9 M Total params
99.450 Total estimated model params size (MB)
19 Modules in train mode
174 Modules in eval mode
D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\trainer\connectors\data_connector.py:425: The 'train_dataloader' does not have many workers which may be a bottleneck. Consider increasing the value of the num_workers argumenttonum_workers=15in theDataLoaderto improve performance. D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\trainer\connectors\data_connector.py:425: The 'val_dataloader' does not have many workers which may be a bottleneck. Consider increasing the value of thenum_workers argument to num_workers=15 in the DataLoader to improve performance.
Epoch 0: 0%| | 0/14 [00:00<?, ?it/s] Traceback (most recent call last):
File "D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\PatchcoreTrain.py", line 158, in
train()
File "D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\PatchcoreTrain.py", line 25, in train
engine.fit(datamodule=datamodule, model=model)
File "D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\src\anomalib\engine\engine.py", line 416, in fit
self.trainer.fit(model, train_dataloaders, val_dataloaders, datamodule, ckpt_path)
File "D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\trainer\trainer.py", line 561, in fit
call._call_and_handle_interrupt(
File "D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\trainer\call.py", line 48, in _call_and_handle_interrupt
return trainer_fn(*args, **kwargs)
File "D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\trainer\trainer.py", line 599, in _fit_impl
self._run(model, ckpt_path=ckpt_path)
File "D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\trainer\trainer.py", line 1012, in _run
results = self._run_stage()
File "D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\trainer\trainer.py", line 1056, in _run_stage
self.fit_loop.run()
File "D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\loops\fit_loop.py", line 216, in run
self.advance()
File "D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\loops\fit_loop.py", line 455, in advance
self.epoch_loop.run(self._data_fetcher)
File "D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\loops\training_epoch_loop.py", line 150, in run
self.advance(data_fetcher)
File "D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\loops\training_epoch_loop.py", line 303, in advance
call._call_callback_hooks(trainer, "on_train_batch_start", batch, batch_idx)
File "D:\anaconda3\envs\ultralytics\lib\site-packages\lightning\pytorch\trainer\call.py", line 227, in _call_callback_hooks
fn(trainer, trainer.lightning_module, *args, **kwargs)
File "D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\src\anomalib\pre_processing\pre_processor.py", line 91, in on_train_batch_start
batch.image, batch.gt_mask = self.transform(batch.image, batch.gt_mask)
File "D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\src\anomalib\data\dataclasses\generic.py", line 128, in set
value = validator(value)
File "D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\src\anomalib\data\validators\torch\image.py", line 518, in validate_image
raise ValueError(msg)
ValueError: Image must have 3 channels, got 1.
Epoch 0: 0%| | 0/14 [00:00<?, ?it/s]

Process finished with exit code 1

Forum Check

• I have searched the discussions forum for an answer to my question.

[github-actions[bot]]

Thanks for submitting this issue! It has been added to our triage queue. A maintainer will review it shortly.

[alfieroddan]

Can you show the script you used?

Also in markdown you can format your code: https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/creating-and-highlighting-code-blocks

Try it!

[abc-125]

Hello, for PatchCore you can change number of channels from default 3 to 1 here by adding in_chans=1 parameter:

anomalib/src/anomalib/models/components/feature_extractors/timm.py

Lines 120 to 127 in 90e1192

	self.feature_extractor = timm.create_model(
	backbone,
	pretrained=pre_trained,
	pretrained_cfg=None,
	features_only=True,
	exportable=True,
	out_indices=self.idx,
	)

[wenwu2021]

def trainGray():
datamodule = MVTecAD(train_batch_size=16,
eval_batch_size=16, num_workers=0)
# datamodule.setup()
transform = v2.Compose([
v2.Resize(size=(256, 256)),
v2.Grayscale(),
])
pre_processor = PreProcessor(transform=transform)
model = Patchcore(pre_processor =pre_processor)
engine = Engine()
engine.fit(datamodule=datamodule, model=model)

v2.Grayscale(),This method is useless

File "D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\src\anomalib\data\validators\torch\image.py", line 518, in validate_image
raise ValueError(msg)
ValueError: Image must have 3 channels, got 1.

[alfieroddan]

Ah I see, you are wanting to train with Grayscale images.

So your v2.Grayscale() converts the RGB back to grayscale.

The feature extractor of the model must therefore accept grayscale images, as @abc-125 suggested:

def trainGray(): datamodule = MVTecAD(train_batch_size=16, eval_batch_size=16, num_workers=0) # datamodule.setup() transform = v2.Compose([ v2.Resize(size=(256, 256)), v2.Grayscale(), ]) pre_processor = PreProcessor(transform=transform) model = Patchcore(pre_processor =pre_processor) engine = Engine() engine.fit(datamodule=datamodule, model=model)

v2.Grayscale(),This method is useless

File "D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\src\anomalib\data\validators\torch\image.py", line 518, in validate_image raise ValueError(msg) ValueError: Image must have 3 channels, got 1.

[samet-akcay]

ValueError: Image must have 3 channels, got 1.

This is because of the data validations I guess.

anomalib/src/anomalib/data/validators/torch/image.py

Lines 516 to 518 in 90e1192

	if image.shape[1] != 3:
	msg = f"Image must have 3 channels, got {image.shape[1]}."
	raise ValueError(msg)

Maybe we could remove this validation?

[alfieroddan]

anomalib/src/anomalib/data/validators/torch/image.py

Lines 479 to 486 in 90e1192

	def validate_image(image: torch.Tensor) -> Image:
	"""Validate the image for a batch.
	This method validates batches of images stored as PyTorch tensors. It handles:
	- Single images and batches
	- RGB images only
	- Channel-first format [N, C, H, W]
	- Type conversion to float32

The whole function expects RGB Images, does the rest of anomalib depend on the shape [B, C=3, H, W]?

If not, then as you say don't validate for RGB?

If yes, then you can also do v2.Grayscale(num_output_channels=3).

[samet-akcay]

does the rest of anomalib depend on the shape [B, C=3, H, W]?

We might need to validate this :)

But you are right, v2.Grayscale(num_output_channels=3) would be the best bet for now

[wenwu2021]

Ah I see, you are wanting to train with Grayscale images.

So your v2.Grayscale() converts the RGB back to grayscale.

The feature extractor of the model must therefore accept grayscale images, as @abc-125 suggested:

def trainGray(): datamodule = MVTecAD(train_batch_size=16, eval_batch_size=16, num_workers=0) # datamodule.setup() transform = v2.Compose([ v2.Resize(size=(256, 256)), v2.Grayscale(), ]) pre_processor = PreProcessor(transform=transform) model = Patchcore(pre_processor =pre_processor) engine = Engine() engine.fit(datamodule=datamodule, model=model)
v2.Grayscale(),This method is useless
File "D:\BaiduNetdiskDownload\anomalib\anomalib-2.0.0\src\anomalib\data\validators\torch\image.py", line 518, in validate_image raise ValueError(msg) ValueError: Image must have 3 channels, got 1.

the sanme error

[wenwu2021]

ValueError: Image must have 3 channels, got 1.

This is because of the data validations I guess.

anomalib/src/anomalib/data/validators/torch/image.py

Lines 516 to 518 in 90e1192

if image.shape[1] != 3:
msg = f"Image must have 3 channels, got {image.shape[1]}."
raise ValueError(msg)
Maybe we could remove this validation?

no

[samet-akcay]

ValueError: Image must have 3 channels, got 1.

This is because of the data validations I guess.
anomalib/src/anomalib/data/validators/torch/image.py
Lines 516 to 518 in 90e1192
if image.shape[1] != 3:
msg = f"Image must have 3 channels, got {image.shape[1]}."
raise ValueError(msg)
Maybe we could remove this validation?

no

what do you mean by no? can you elaborate ?

[wenwu2021]

ValueError: Image must have 3 channels, got 1.

This is because of the data validations I guess.
anomalib/src/anomalib/data/validators/torch/image.py
Lines 516 to 518 in 90e1192
if image.shape[1] != 3:
msg = f"Image must have 3 channels, got {image.shape[1]}."
raise ValueError(msg)
Maybe we could remove this validation?

no

what do you mean by no? can you elaborate ?

Not only this one needs to be changed, there are other things that need to be changed

[samet-akcay]

Have you tried this option v2.Grayscale(num_output_channels=3)?

[lucianchauvin]

What if a model explicitly requires 1 channel? Is there any reason this validation explicitly requires 3?

[samet-akcay]

@lucianchauvin is this an existing anomalib model, or an external model? If there are models expecting 1 channel, we could drop this validation I think

[lucianchauvin]

@lucianchauvin is this an existing anomalib model, or an external model? If there are models expecting 1 channel, we could drop this validation I think

An external model. I am trying to integrate an existing model into Anomalib. (speaking of, all the documentation for doing this disappeared in 2.0.0 or some point before :C)

[wenwu2021]

Have you tried this option ?v2.Grayscale(num_output_channels=3)

(model): PatchcoreModel(
(feature_extractor): TimmFeatureExtractor(
(feature_extractor): FeatureListNet(
(conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
The model still has three channels

[samet-akcay]

Yes, but it should be gray scale, not rgb? Have you checked the channels? They should be the same

[wenwu2021]

Not what I wanted, single channel prediction, less time and computational resources. So if it's a 4-channel graph, then the model is also a 3-channel one? That doesn't reflect the characteristics of Channel 4

[alexriedel1]

Not what I wanted, single channel prediction, less time and computational resources. So if it's a 4-channel graph, then the model is also a 3-channel one? That doesn't reflect the characteristics of Channel 4

The effect of 3 vs. 1 channels in the first model layer is negligibly small. On top of that, the model is pre-trained on 3 channel images, so you would need to train a new backbone model for patchcore with a 1 channel input on imagenet.

If you have a 4 channel input, you would need to adjust the model accordingly or find another way to aggregate the channels