nelson1425/efficientad

Unofficial implementation of EfficientAD https://arxiv.org/abs/2303.14535

EfficientAD

Unofficial implementation of paper https://arxiv.org/abs/2303.14535

Results

Model	Dataset	Official Paper	efficientad.py
EfficientAD-M	Mvtec AD	99.1	99.1
EfficientAD-M	VisA	98.1	98.2
EfficientAD-M	Mvtec LOCO	90.7	90.1
EfficientAD-S	Mvtec AD	98.8	99.0
EfficientAD-S	VisA	97.5	97.6
EfficientAD-S	Mvtec LOCO	90.0	89.5

Benchmarks

Model	GPU	Official Paper	benchmark.py
EfficientAD-M	A6000	4.5 ms	4.4 ms
EfficientAD-M	A100	-	4.6 ms
EfficientAD-M	A5000	5.3 ms	5.3 ms

Setup

Packages

Python==3.10
torch==1.13.0
torchvision==0.14.0
tifffile==2021.7.30
tqdm==4.56.0
scikit-learn==1.2.2

Mvtec AD Dataset

For Mvtec evaluation code install:

numpy==1.18.5
Pillow==7.0.0
scipy==1.7.1
tabulate==0.8.7
tifffile==2021.7.30
tqdm==4.56.0

Download dataset (if you already have downloaded then set path to dataset (--mvtec_ad_path) when calling efficientad.py).

mkdir mvtec_anomaly_detection
cd mvtec_anomaly_detection
wget https://www.mydrive.ch/shares/38536/3830184030e49fe74747669442f0f282/download/420938113-1629952094/mvtec_anomaly_detection.tar.xz
tar -xvf mvtec_anomaly_detection.tar.xz
cd ..

Download evaluation code:

wget https://www.mydrive.ch/shares/60736/698155e0e6d0467c4ff6203b16a31dc9/download/439517473-1665667812/mvtec_ad_evaluation.tar.xz
tar -xvf mvtec_ad_evaluation.tar.xz
rm mvtec_ad_evaluation.tar.xz

efficientad.py

Training and inference:

python efficientad.py --dataset mvtec_ad --subdataset bottle

Evaluation with Mvtec evaluation code:

python mvtec_ad_evaluation/evaluate_experiment.py --dataset_base_dir './mvtec_anomaly_detection/' --anomaly_maps_dir './output/1/anomaly_maps/mvtec_ad/' --output_dir './output/1/metrics/mvtec_ad/' --evaluated_objects bottle

Reproduce paper results

Reproducing results from paper requires ImageNet stored somewhere. Download ImageNet training images from https://www.kaggle.com/competitions/imagenet-object-localization-challenge/data or set --imagenet_train_path of efficientad.py to other folder with general images in children folders for example downloaded https://drive.google.com/uc?id=1n6RF08sp7RDxzKYuUoMox4RM13hqB1Jo

Calls:

python efficientad.py --dataset mvtec_ad --subdataset bottle --model_size medium --weights models/teacher_medium.pth --imagenet_train_path ./ILSVRC/Data/CLS-LOC/train
python efficientad.py --dataset mvtec_ad --subdataset cable --model_size medium --weights models/teacher_medium.pth --imagenet_train_path ./ILSVRC/Data/CLS-LOC/train
python efficientad.py --dataset mvtec_ad --subdataset capsule --model_size medium --weights models/teacher_medium.pth --imagenet_train_path ./ILSVRC/Data/CLS-LOC/train
...

python efficientad.py --dataset mvtec_loco --subdataset breakfast_box --model_size medium --weights models/teacher_medium.pth --imagenet_train_path ./ILSVRC/Data/CLS-LOC/train
python efficientad.py --dataset mvtec_loco --subdataset juice_bottle --model_size medium --weights models/teacher_medium.pth --imagenet_train_path ./ILSVRC/Data/CLS-LOC/train
...

This produced the Mvtec AD results in results/mvtec_ad_medium.json.

Mvtec LOCO Dataset

Download dataset:

mkdir mvtec_loco_anomaly_detection
cd mvtec_loco_anomaly_detection
wget https://www.mydrive.ch/shares/48237/1b9106ccdfbb09a0c414bd49fe44a14a/download/430647091-1646842701/mvtec_loco_anomaly_detection.tar.xz
tar -xf mvtec_loco_anomaly_detection.tar.xz
cd ..

Download evaluation code:

wget https://www.mydrive.ch/shares/48245/a4e9922c5efa93f57b6a0ff9f5c6b969/download/430648014-1646847095/mvtec_loco_ad_evaluation.tar.xz
tar -xvf mvtec_loco_ad_evaluation.tar.xz
rm mvtec_loco_ad_evaluation.tar.xz

Install same packages as for Mvtec AD evaluation code, see above.

Training and inference for LOCO sub-dataset:

python efficientad.py --dataset mvtec_loco --subdataset breakfast_box

Evaluation with LOCO evaluation code:

python mvtec_loco_ad_evaluation/evaluate_experiment.py --dataset_base_dir './mvtec_loco_anomaly_detection/' --anomaly_maps_dir './output/1/anomaly_maps/mvtec_loco/' --output_dir './output/1/metrics/mvtec_loco/' --object_name breakfast_box

efficientad's Issues

The image_size

Hello, Thank you for your work.
During training, does the image_size of the data need to be 256?

Inference is too slow,is something wrong?

I wrote the interface for model inference following the logic of the predict function in the efficientad.py
However, in actual testing, it was found that the average time was close to 1 second (excluding the model loading time), and it was also very slow after converting to the ONNX model

Device information：
CPU：
Intel(R) Core(TM) i7-10700K CPU @ 3.80GHz
GPU：
NVIDIA GeForce RTX 3080 16G

This is the code tested with the pth model：

import time
import numpy as np
import torch
from tqdm import tqdm


def pth_predict(image, teacher_model, student_model, ae_model, teacher_mean, teacher_std, out_channels,
                q_st_start=None, q_st_end=None, q_ae_start=None, q_ae_end=None):
    teacher_output = teacher_model(image)
    teacher_output = (teacher_output - teacher_mean) / teacher_std
    student_output = student_model(image)
    autoencoder_output = ae_model(image)
    map_st = torch.mean((teacher_output - student_output[:, :out_channels]) ** 2,
                        dim=1, keepdim=True)
    map_ae = torch.mean((autoencoder_output -
                         student_output[:, out_channels:]) ** 2,
                        dim=1, keepdim=True)
    if q_st_start is not None:
        map_st = 0.1 * (map_st - q_st_start) / (q_st_end - q_st_start)
    if q_ae_start is not None:
        map_ae = 0.1 * (map_ae - q_ae_start) / (q_ae_end - q_ae_start)
    map_combined = 0.5 * map_st + 0.5 * map_ae
    return map_combined, map_st, map_ae


if __name__ == '__main__':
    # Load the PTH model
    device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
    teacher_net = torch.load('./output/ad_small/trainings/mvtec_ad/rain/teacher_final.pth', map_location=device)
    student_net = torch.load('./output/ad_small/trainings/mvtec_ad/rain/student_final.pth', map_location=device)
    ae_net = torch.load('./output/ad_small/trainings/mvtec_ad/rain/autoencoder_final.pth', map_location=device)

    # Construct the input data
    fake_img_tensor = torch.rand((1, 3, 256, 256))

    output_channels_num = 384
    # Model prediction
    teacher_mean_tensor = torch.rand((1, output_channels_num, 1, 1))
    teacher_std_tensor = torch.rand((1, output_channels_num, 1, 1))

    time_range = 100
    time_cost_list = []
    for i in tqdm(range(time_range)):
        s1 = time.time()
        pth_predict(fake_img_tensor, teacher_net, student_net, ae_net, teacher_mean_tensor, teacher_std_tensor,
                    output_channels_num,
                    q_st_start=None, q_st_end=None, q_ae_start=None, q_ae_end=None)
        s2 = time.time()
        time_cost_list.append(s2 - s1)
    print(f'average time cost:{np.mean(time_cost_list):.6f}s')

This is the test code for inference with the ONNX model：

import time
import numpy as np
import onnxruntime
from tqdm import tqdm


def onnx_predict(img_arr, teacher_session, student_session, ae_session, teacher_mean, teacher_std, out_channels,
                 q_st_start=None, q_st_end=None, q_ae_start=None, q_ae_end=None):
    ort_inputs1 = {teacher_session.get_inputs()[0].name: img_arr}
    teacher_output = teacher_session.run(None, ort_inputs1)
    teacher_output = teacher_output[0]
    teacher_output = (teacher_output - teacher_mean) / teacher_std

    ort_inputs2 = {student_session.get_inputs()[0].name: img_arr}
    student_output = student_session.run(None, ort_inputs2)
    student_output = student_output[0]

    ort_inputs3 = {ae_session.get_inputs()[0].name: img_arr}
    autoencoder_output = ae_session.run(None, ort_inputs3)
    autoencoder_output = autoencoder_output[0]

    map_st = np.mean((teacher_output - student_output[:, :out_channels]) ** 2, axis=1)
    map_ae = np.mean((autoencoder_output -
                      student_output[:, out_channels:]) ** 2, axis=1)
    if q_st_start is not None:
        map_st = 0.1 * (map_st - q_st_start) / (q_st_end - q_st_start)
    if q_ae_start is not None:
        map_ae = 0.1 * (map_ae - q_ae_start) / (q_ae_end - q_ae_start)
    map_combined = 0.5 * map_st + 0.5 * map_ae

    return map_combined, map_st, map_ae


if __name__ == '__main__':
    # Load the ONNX model
    teacher_ort_session = onnxruntime.InferenceSession('./output/onnx_path/teacher.onnx')
    student_ort_session = onnxruntime.InferenceSession('./output/onnx_path/student.onnx')
    ae_ort_session = onnxruntime.InferenceSession('./output/onnx_path/autoencoder.onnx')

    # Construct the input data
    fake_img_arr = np.random.rand(1, 3, 256, 256)
    fake_img_arr = fake_img_arr.astype(np.float32)
    output_channels_num = 384
    # Model prediction
    teacher_mean_arr = np.random.rand(1, output_channels_num, 1, 1)
    teacher_std_arr = np.random.rand(1, output_channels_num, 1, 1)

    time_range = 100
    time_cost_list = []
    for i in tqdm(range(time_range)):
        s1 = time.time()
        onnx_predict(fake_img_arr, teacher_ort_session, student_ort_session, ae_ort_session,
                     teacher_mean=teacher_mean_arr,
                     teacher_std=teacher_std_arr, out_channels=output_channels_num,
                     q_st_start=None, q_st_end=None, q_ae_start=None, q_ae_end=None)
        s2 = time.time()
        time_cost_list.append(s2 - s1)
    print(f'average time cost:{np.mean(time_cost_list):.6f}s')

This is a function of converting the PTH model to the ONNX model：

def convert_to_onnx_with_dynamic_img_shape(model, input_size, onnx_path):
    model.eval()

    dummy_input = torch.randn(1, *input_size, requires_grad=True)

    torch.onnx.export(model, 
                      dummy_input, 
                      onnx_path,  
                      export_params=True,  
                      opset_version=11,  
                      do_constant_folding=True,  
                      input_names=['modelInput'],  
                      output_names=['modelOutput'],  
                      dynamic_axes={'modelInput': {0: 'batch_size', 2: 'img_height', 3: 'img_weight'}, 
                                    'modelOutput': {0: 'batch_size', 2: 'img_height', 3: 'img_weight'}})
    print('Model has been converted to ONNX')


if __name__ == '__main__':
    out_channels = 384

    # teacher = get_pdn_small(out_channels)
    # student = get_pdn_small(2 * out_channels)
    # autoencoder = get_autoencoder(out_channels)

    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    teacher_model = torch.load('./output/ad_small/trainings/mvtec_ad/rain/teacher_final.pth',
                               map_location=device)
    student_model = torch.load('./output/ad_small/trainings/mvtec_ad/rain/student_final.pth',
                               map_location=device)
    autoencoder_model = torch.load('./output/ad_small/trainings/mvtec_ad/rain/autoencoder_final.pth',
                                   map_location=device)

    teacher_onnx_path = './output/onnx_path/teacher.onnx'
    convert_to_onnx_with_dynamic_img_shape(teacher_model, input_size=(3, 256, 256), onnx_path=teacher_onnx_path)

    student_onnx_path = './output/onnx_path/student.onnx'
    convert_to_onnx_with_dynamic_img_shape(student_model, input_size=(3, 256, 256), onnx_path=student_onnx_path)

    autoencoder_onnx_path = './output/onnx_path/autoencoder.onnx'
    convert_to_onnx_with_dynamic_img_shape(autoencoder_model, input_size=(3, 256, 256), onnx_path=autoencoder_onnx_path)

mvtec_ad_evaluation.tar.xz

mvtec_ad_evaluation.tar.xz
If I can't connect to this link, is there any other way to get this file

I encountered an error while evaluating

(EfficientAD) E:\PycharmProjects\EfficientAD-main>python mvtec_ad_evaluation/evaluate_experiment.py --dataset_base_dir E:\Dataset\anomaly_detection\mvtec_anomaly_detection\mvtec_anomaly_detection\ --anomaly_maps_dir E:\PycharmProjects\EfficientAD-main\output\1\anomaly_maps\mvtec_ad --output_dir E:\PycharmProjects\EfficientAD-main\output\1\metrics/mvtec_ad --
evaluated_objects bottle
=== Evaluate bottle ===
Parsed 83 ground truth image files.
Read ground truth files and corresponding predictions...
0%| | 0/83 [00:00<?, ?it/s]
Traceback (most recent call last):
File "E:\PycharmProjects\EfficientAD-main\mvtec_ad_evaluation\evaluate_experiment.py", line 247, in
main()
File "E:\PycharmProjects\EfficientAD-main\mvtec_ad_evaluation\evaluate_experiment.py", line 215, in main
calculate_au_pro_au_roc(
File "E:\PycharmProjects\EfficientAD-main\mvtec_ad_evaluation\evaluate_experiment.py", line 148, in calculate_au_pro_au_roc
prediction = util.read_tiff(pred_name)
File "E:\PycharmProjects\EfficientAD-main\mvtec_ad_evaluation\generic_util.py", line 105, in read_tiff
raise IOError('Found multiple files with a TIFF extension at'
OSError: Found multiple files with a TIFF extension at E:\PycharmProjects\EfficientAD-main\output\1\anomaly_maps\mvtec_ad\bottle\test\broken_large\000
Please specify which TIFF extension to use via the exts parameter of this function.

Error when Training and inference & evaluating with MVTec code

I used the python script the author gave as follow :

!python efficientad.py --dataset mvtec_ad --subdataset wood
!python mvtec_ad_evaluation/evaluate_experiment.py --dataset_base_dir './mvtec_anomaly_detection/' --anomaly_maps_dir './output/1/anomaly_maps/mvtec_ad/' --output_dir './output/1/metrics/mvtec_ad/' --evaluated_objects bottle

but the error says:
"python3: can't open file '/content/efficientad.py': [Errno 2] No such file or directory"

I am running from colab
What should i do next? Hope someone can point a way. Thanks

[BUG] In some cases the quantile may lead to NAN exception or divide zero exception .

the q_st_start and q_st_end may be equal in some cases, which will lead to divide zero error. So it is a bug.

good job bro!

Error when evaluate with MVTec code

I used the python script you gave as follow:

python mvtec_ad_evaluation/evaluate_experiment.py --dataset_base_dir './mvtec_anomaly_detection/' --anomaly_maps_dir './output/1/anomaly_maps/mvtec_ad/' --output_dir './output/1/metrics/mvtec_ad/' --evaluated_objects bottle

but there are errors:

`python mvtec_ad_evaluation/evaluate_experiment.py --dataset_base_dir './mvtec_anomaly_detection/' --anomaly_maps_dir './output/1/anomaly_maps/mvtec_ad/' --output_dir './output/1/metrics/mvtec_ad/' --evaluated_objects bottle
=== Evaluate bottle ===
Parsed 83 ground truth image files.
Read ground truth files and corresponding predictions...
0%| | 0/83 [00:00<?, ?it/s]
Traceback (most recent call last):
File "F:\implementation_anomalydetection\nelson1425_EfficientAD\nelson_data\mvtec_ad_evaluation\evaluate_experiment.py", line 247, in
main()
File "F:\implementation_anomalydetection\nelson1425_EfficientAD\nelson_data\mvtec_ad_evaluation\evaluate_experiment.py", line 215, in main
calculate_au_pro_au_roc(
File "F:\implementation_anomalydetection\nelson1425_EfficientAD\nelson_data\mvtec_ad_evaluation\evaluate_experiment.py", line 148, in calculate_au_pro_au_roc
prediction = util.read_tiff(pred_name)
^^^^^^^^^^^^^^^^^^^^^^^^^
File "F:\implementation_anomalydetection\nelson1425_EfficientAD\nelson_data\mvtec_ad_evaluation\generic_util.py", line 102, in read_tiff
raise FileNotFoundError('Could not find a file with a TIFF extension'
FileNotFoundError: Could not find a file with a TIFF extension at ./output/1/anomaly_maps/mvtec_ad/bottle\test\broken_large\000
(base) PS F:\implementation_anomalydetection\nelson1425_EfficientAD\nelson_data>

`

all the mvtec_ad_evaluation.tar.xz and mvtec_anomaly_detection.tar.xz are from what you share.
I am running from aconda powershell from windows 10

can anyone point a way ?

Convolutions' padding - student/teacher model

Hi! I want to congratulate with you for the implementation of EfficientAD, thank you for your work. I would like to better understand some parts of your implementations, specifically the student/teacher model. In the paper the model has a fixed value for the padding in the convolutions resulting in an output with the shape 384x64x64, instead in your implementation the padding is set by the boolean's variable and the output's shape might be different (384x56x56). Is there a reason behind it?

How inteprete the hyper parameter table？

qa, qb, and qhard is three hyper parameter. But the meaning of the table is vague. I think the combination of three hyper parameter determine the auroc, but this table give auroc values for each hyper parameter value.

how to inference

hello. thanks for the nice code

How do I just infer an image?

error happend when training teacher_small.pth by my self.

it said: [RuntimeError: The size of tensor a (64) must match the size of tensor b (128) at non-singleton dimension 3 at line loss = torch.mean((target - prediction)**2)], so @nelson1425 could please tell me how to solve it? Many thanks!

The source of teacher weights

The feature size not the same with the papar

As referred in the paper, the feature size is 64 x 64, but in this repo is 56 x 56

If padding=True, worse results but good segmentation precision. Why?

Hi Nelson, i hope this message finds you well. I'm curious on what is your take on this.
I am very satisfied with results i obtain with pad_maps = True and padding=False.
However, as you know, pad_maps creates a "frame" that pads the resulting anomaly map and it seems that anomalies contained in the padded region are not seen by the model.
Playing with these two pad_maps and padding parameters i realized that, regardless of pad_maps value, it's padding that determines if the padded region is ignored or not. So, if i set padding=True the model also sees anomalies close to the borders of the image. Also, if I set padding=True and pad_maps=False I obtain precise segmentations and i also see defects close to borders. To me it seems that pad_maps is needed only to correct the segmentations "translation" that is caused by padding=False.

So why i'm writing this message and i'm not simply using padding=True and pad_maps=False?

The problem is that, regardless of the pad_maps (so we can forget about it, because it's getting confusing 😁 ), padding=True significantly worsen the results. I mean, it seems that the model learns less things.
So the first thing that came to my mind is that maybe the Teacher has been pretrained with padding=False, so it maybe has a different architecture with respect to the Student with padding=True that i'm trying to use.
But based on the code in this repo, it seems you pretrained the Teacher using padding=True and then you use a Student with padding=False when you do the efficientad training. Is this true?

This image shows what i obtain with padding=True and pad_maps=False . It segments perfectly the fake defects i am using to do these tests. Unfortunately the overall performance on the real defects in the rest of the test test is significantly worse wrt the model trained with padding=False (you can see it also in this image: the model struggles to segment the real defect on the bottom).

What do you think? What could be the reason why padding=True jeopardizes my trainings?

about the validation of the MvTec-AD LOCO dataset

I use valid code that the author gives，but the there is something wrong that shows “ assert len(set(init_queries)) == len(init_queries) ”
I hope someone who meet this question can help me

Per category result for MVTec AD

Hi Nelson,
can oyu provide the results of EfficientAD-M per category on the MVTec AD dataset? For matching the results in the anomlib framework, I would like to have the single category data :)
Thanks!

result of inference on broken large bottle

hi nelson and thanks for your great code
i just executed efficientad.py with default setting for bottle with 70000 steps
below is the result i got for the first broken large image(000.png):

i just wanted to know if the result image is what you expected or not

Asking why use batch size =1 ......? Please...help

Why use batch size =1 ? What is for it?
train_loader = DataLoader(train_set, batch_size=1, shuffle=True,
num_workers=4, pin_memory=True)
train_loader_infinite = InfiniteDataloader(train_loader)
validation_loader = DataLoader(validation_set, batch_size=1)

IndexError: list index out of range

I use custom dataset.

(torch1) D:\code\git-DS\EfficientAD>python mvtec_ad_evaluation/evaluate_experiment.py --dataset_base_dir "\DS-NAS\ds\oilhole(crank)" --anomaly_maps_dir './output/1/anomaly_maps/oilhole(crank)/' --output_dir './output/1/metrics/oilhole(crank)/' --evaluated_objects type_i
=== Evaluate type_i ===
Parsed 0 ground truth image files.
Read ground truth files and corresponding predictions...
0it [00:00, ?it/s]
Compute PRO curve...
Traceback (most recent call last):
File "D:\code\git-DS\EfficientAD\mvtec_ad_evaluation\evaluate_experiment.py", line 247, in
main()
File "D:\code\git-DS\EfficientAD\mvtec_ad_evaluation\evaluate_experiment.py", line 215, in main
calculate_au_pro_au_roc(
File "D:\code\git-DS\EfficientAD\mvtec_ad_evaluation\evaluate_experiment.py", line 157, in calculate_au_pro_au_roc
pro_curve = compute_pro(
File "D:\code\git-DS\EfficientAD\mvtec_ad_evaluation\pro_curve_util.py", line 37, in compute_pro
anomaly_maps[0].shape[0],
IndexError: list index out of range

I've encounted this issue.

I made it all the way to train and got a final image AUC: 98.3421.
The tiff files are stored in the anomaly_maps folder, but I don't know if I'm having trouble reading them or why.

【EfficientAD】realize the visualization of heat map?

How to realize the visualization of heat map?

Pretrained weights

Hello, my hardware is quite old, and training takes a long time. Can you please share your pre training weights so that I can quickly check the effectiveness of this model?

Purpose of padding

https://github.com/nelson1425/EfficientAD/blob/fcab5146f84ae17597044ad5ddf1656ccf805401/efficientad.py#LL279C9-L279C75
I'm wondering why this padding is necessary. Can you help me?

The provided Google Drive URL is not available anymore

Hi, I just want to let you know that the Google Drive URL mentioned in the readme, i.e., https://drive.google.com/uc?id=1n6RF08sp7RDxzKYuUoMox4RM13hqB1Jo, is not accessible anymore. I have pretrained the teacher model using the ImageNet dataset instead.

Why change the output size to 56 * 56，instead of 64 * 64

Hi，Your work is greatly appreciated！I got a question, Why change the output size to 56 * 56，instead of 64 * 64 in paper.

How to obtain --imagenet_ train_ path

Hello，Thanks for your work.How can I obtain --imagenet_ train_ path？

RuntimeError: quantile() input tensor is too large

Traceback (most recent call last):
File "D:\code\git-DS\EfficientAD\efficientad.py", line 451, in
main()
File "D:\code\git-DS\EfficientAD\efficientad.py", line 268, in main
q_st_start, q_st_end, q_ae_start, q_ae_end = map_normalization(
File "C:\Users\2878045\AppData\Roaming\Python\Python39\site-packages\torch\autograd\grad_mode.py", line 27, in decorate_context
return func(*args, **kwargs)
File "D:\code\git-DS\EfficientAD\efficientad.py", line 374, in map_normalization
q_st_start = torch.quantile(maps_st, q=0.9)
RuntimeError: quantile() input tensor is too large

anyone solve this?

Pixelwise_AUC isn't evaluated?

How come pixelwise-AUC isn't evaluated?

Asking about deployment of onnx format models

Please ask! What is the final output obtained by converting the trained model into onnx format and importing it into onxxruntime for inference? Is it the input sample detection result (and whether it is an anomalous sample) and the corresponding confidence level? Or is it the anomaly_map with the same channel 1 as the input height and width?

Question about training steps used to achieve reported performance

For the performance reported in the readme, I wonder, did you use the same training steps (70,000) for all datasets (mvtec, mvtec loco, and visa)?

Get error when evaluate breakfast_box's output with LOCO's official evaluation code

Thanks for your great effort in contributing and sharing the reimplementation.

I have a problem here that needs your assistance. After training the MVTec LOCO's breakfast box category, I want to evaluate the output feature maps (in tiff format) using their official evaluation code, but I get this error:

File "[mydrive]/EfficientAD/mvtec_loco_ad_evaluation/src/aggregation.py", line 57, in binary_refinement
    assert len(set(init_queries)) == len(init_queries)

I debugged the issue, and found this error is caused by the fact that the output feature maps contain too many 0 values, such that when the evaluation code creates the initial_thresholds, a.k.a. init_queries (a list of 50 anomaly scores), it contains duplicate values 0. An example is like:

[2.83783058984375, 0.3397859036922455, 0.11129003018140793, 0.04379851371049881, 0.018422557041049004, 0.005928939674049616, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -0.00046889434452168643, -0.004845781251788139, -0.008791022002696991, -0.012286320328712463, -0.015344921499490738, -0.01806800439953804, -0.02048872597515583, -0.0226922407746315, -0.024705413728952408, -0.026554450392723083, -0.028272513300180435, -0.029879318550229073, -0.0313967764377594, -0.0328390970826149, -0.03421252220869064, -0.0355297289788723, -0.036807697266340256, -0.03804701194167137, -0.03924761712551117, -0.04042193666100502, -0.04157993942499161, -0.04271307215094566, -0.043838512152433395, -0.044960867613554, -0.046086572110652924, -0.04722655192017555, -0.048392221331596375, -0.049606941640377045, -0.05089380219578743, -0.052285678684711456, -0.05385420098900795, -0.055760398507118225, -0.05840958654880524, -0.07654058350419998]

I wonder if anyone encounters the same issue here. I appreciate if you can help me with solving the issue.

The output size of the network is inconsistent with the description in the paper

Thanks author for the awsome job. But the output size of the Network called PDN(patch descriptor network) is 384 × 56 ×56, while description the in the paper is 384 × 64 × 64. Is it a mistake or my misunderstanding?

Distilled teacher model

Sorry for my ignorance, i have small question about teacher model.
The teacher model pre-distilled with image dataset makes it smarter when training with other datasets?

Can you share the background of this rule. I feed there is no similarity between mvrec dataset and image dataset.

Thank you.

Train on ViSA dataset

Hello, Thank you for your work.
To reproduce the paper results. I want to train on mvtec and visa datasets. But there is no configuration for the visa dataset. However, you mentioned the score of EfficientAD on the Visa dataset in the table.
How we can train on Visa Dataset.
Secondly, while training the algorithm, you save the model after some interval without considering the auroc score. I'm looking forward to hearing back from you. Thank you

Can you share the teacher model pretrained from imagenet?

@nelson1425 Can you share the teacher model pretrained from imagenet? The database is too large . I can't train the teacher model .

Can the pre-trained model of the teacher network 'teacher_small.pth' be trained by myself？

Why pad 4 pixels before resizing instead of resizing directly？

Hi~, in the part of ouput anomaly map:

map_combined = torch.nn.functional.pad(map_combined, (4, 4, 4, 4))
map_combined = torch.nn.functional.interpolate(map_combined, (orig_height, orig_width), mode='bilinear')

why pad 4 pixel each side? what's the difference with resize to (orig_height, orig_width) directly?
🤔🤔🤔

Question about the loss between student output and ae output

Hi， I'm puzzled about the loss between student output and ae output，is it necessary?

A minor typo in distillation training

Before freezing the backbone model, feature_dimensions() is called with a forward pass, thus the BN parameters are affected and it is no longer the original pre-trained model. Though it should cause little differences.

Do I need to re-pretrain a teacher network when I train my own dataset?

Hardware / time needed for pretraining

Hi, first of all thank you for your work and for making it completely public. I'm curious on the hardware you used for the pretraining on ImageNet, and on how much time did it take. Good job!

缺陷检测交流微信群，互相交流学习进步

缺陷检测交流微信群，互相交流学习进步