Hello, I change the dataset for my own dataset, and the test.py file run well, but i dont kown how this file work, could you tell me what does this file do? How to use a trained model? Thanks in advance!

Hi Sabokrou,

Thank you for the source code of this good work!
I would like to explore more about your approach. Could you please provide the frame-level scores of Ped2 you obtained in your experiment (which you used to estimate the EER in Table 2)?

Thank you!

Hello there,

In your "test.py" file the variable "nd_patch_step" is not defined. If I track down the variable it should be equal to "nStride", but with such setting the test does not seem to be working correct.

Am I missing some thing?

Thanks

Hi there, I want to extend your idea for adversarial training. Is there anyone which has time for helping me to use it for that purpose?

Is the output of the discriminator an anomaly score during the test of anomaly detection? Why is the output of the discriminator a very small value when I am testing, on the order of e-9 to e-20?

(What the title says)

In my experiment, the output of the discriminator is about 0.49 after GAN convergence. How could you make the discriminator output probability larger than 0.5?

Hi,
I'm trying to repeat your experiments on MNIST. But to get the F1-score, may I know how you set the likelihood threshold to distinguish the inliers and outliers of the test samples?

Thanks!

when I reproduct this code on my own application, I can easily get d_loss with 0 and g_loss with 20.
How could I fix this up, your any advices would be helpful, thinks!

It seems that your code doesn't converge on MNIST dataset. Please check it out, thanks.

getting error for UCSD dataset:
File "test.py", line 168, in process_frame
frame_patches = nd_patch.transpose([1,0,2,3])
ValueError: axes don't match array

Hi, khalooei:
Thanks for sharing your code, it's interesting.
I've a little confusion about the model, could you please explain it?
It seems the ALOCC can model generic single class(e.g. penguins) and others class(e.g. dogs, cats...) very well.
Will it works for these scenarios?
a. generic train-set multi-classes class and others class? i.e. R models a distribution for not only one explicit class but a complex distribution for all classes in train-set
b. for fine-grained dataset, one explicit class as base class and others as novelty classes? i.e. base class and novelty class may have more similar distributions than generic class dataset
c. fine-grained multi-classes as base class and others as novelty class?
Thanks.

I find the function of sampler and generator actually are the same? I wonder if there is any mistake.

hello~
when I trained and tested on my own dataset, I got some weird output from results_d: there exists negative value and values which is bigger than 1.0, such as:

results d: [[-0.06518985]
 [ 0.01245594]
 [ 0.18250445]
 [ 0.14557755]
 [-0.12947509]
 [-0.32424578]
 [ 0.01147144]
 [-0.20546311]
 [ 0.38732064]
 [ 0.45803282]
 [ 0.22594012]
 [ 0.02803989]
 [ 0.43993357]
 [ 0.81624144]
 [ 0.5465872 ]
 [ 1.1168824 ]
 [ 0.4104824 ]
 [ 0.60788006]
 [ 0.68482363]
 [ 1.3565235 ]
 [ 1.1945145 ]
 [ 1.1945145 ]
 [ 1.1945145 ]
 [ 1.1945145 ]
 [ 1.1945145 ]
 [-0.06518985]
 [ 0.01245594]
 [ 0.18250445]
 [ 0.14557755]
 [-0.12947509]
 [-0.32424578]
 [ 0.01147144]
 [-0.20546311]
 [ 0.38732064]
 [ 0.45803282]
 [ 0.22594012]
 [ 0.02803989]
 [ 0.43993357]
 [ 0.81624144]
 [ 0.5465872 ]
 [ 1.1168824 ]
 [ 0.4104824 ]
 [ 0.60788006]
 [ 0.68482363]
 [ 1.3565235 ]
 [ 1.1945145 ]
 [ 1.1945145 ]
 [ 1.1945145 ]
 [ 1.1945145 ]
 [ 1.1945149 ]]

Did I miss something during training or testing process? Your any advices would be helpful, thanks!

Hi, the output values of generator G is the one of tanh and thus it takes [-1,1] values.
But the true input takes [0,1] values (at least, MNIST case).
Thus, it will be too easy for discriminator D to classify True or False, by just checking if minus values being contained. The output of G should take sigmoid instead of tanh?

Issue: No output sample is generated for “mnist” or “ucsd” (even when the pre-trained model provided in github is used)

We used following commands (separately) to run the test program on the dataset provided and using the pre-trained model provided as per the instructions. Neither "mnist" or "ucsd" output images are generated. The code fragment is copy-pasted below.

for test on UCSD and patch_size 45*45 and some specific dir like ['Test004'], etc. We prefer to open test.py file and edit every thing that you want

python test.py --dataset UCSD --dataset_address ./dataset/UCSD_Anomaly_Dataset.v1p2/UCSDped2/Train --input_height 45 --output_height 45

for test on MNIST

python test.py --dataset mnist --dataset_address ./dataset/mnist/ --input_height 28 --output_height 28

Following is the copy-paste of the code in test.py file which is executed (after some bug fixing)

if FLAGS.dataset=='mnist':
mnist = input_data.read_data_sets(FLAGS.dataset_address)

        specific_idx_anomaly = np.where(mnist.train.labels != 6)[0]
        specific_idx = np.where(mnist.train.labels == 6)[0]
        ten_precent_anomaly = [specific_idx_anomaly[x] for x in
                               random.sample(range(0, len(specific_idx_anomaly)), len(specific_idx) // 40)]

        data = mnist.train.images[specific_idx].reshape(-1, 28, 28, 1)
        tmp_data = mnist.train.images[ten_precent_anomaly].reshape(-1, 28, 28, 1)
        data = np.append(data, tmp_data).reshape(-1, 28, 28, 1)

        lst_prob = tmp_ALOCC_model.f_test_frozen_model(data[0:FLAGS.batch_size])
        print('check is ok')
        exit()

else in UCDS (depends on infrustructure)

    for s_image_dirs in sorted(glob(os.path.join(FLAGS.dataset_address, 'Test[0-9][0-9][0-9]'))):
        tmp_lst_image_paths = []
        
        if os.path.basename(s_image_dirs) not in ['Test004']:
           print('Skip ',os.path.basename(s_image_dirs))
           continue
        for s_image_dir_files in sorted(glob(os.path.join(s_image_dirs + '/*'))):
            if os.path.basename(s_image_dir_files) not in ['068.tif']:
                print('Skip ', os.path.basename(s_image_dir_files))
                continue
            tmp_lst_image_paths.append(s_image_dir_files)

        #random
        #lst_image_paths = [tmp_lst_image_paths[x] for x in random.sample(range(0, len(tmp_lst_image_paths)), n_fetch_data)]
        lst_image_paths = tmp_lst_image_paths
        #images =read_lst_images(lst_image_paths,nd_patch_size,nd_patch_step,b_work_on_patch=False)
        
        nd_patch_step = 0              # we added this line (from the github to avoid an error)
       images = read_lst_images_w_noise2(lst_image_paths, nd_patch_size, nd_patch_step)

        lst_prob = process_frame(os.path.basename(s_image_dirs),images,tmp_ALOCC_model)
        print('pseudocode test is finished')

Dear Sir:
I am very interested in your paper. But some problems occurred when I repeated MNIST experiment. I wrote the test code and adopted the same LOSS function as the training code. Visualize the LOSS value of G and D, but the normal samples and abnormal samples may not be distinguished.
For example, In the training,the number 1 is treated as a normal sample; in the test, the number 0,1,2 and 3 are visualized; the number 1 is red, the number 0 is blue, the number 2 is green, and the number 3 is yellow. The parameters are all default values. The results look good.

But if the number 2 is treated as a normal sample; in the test, the number 0,1,2 and 3 are visualized; the number 2 is red, the number 0 is blue, the number 1 is green, and the number 3 is yellow. The parameters are all default values. The results turned out badly.

I tried adding noise to the test sample and adjusted r_alpha, but none of the results improved significantly.
This question has been bothering me for a few days and How should the LOSS of the test sample be selected，Could you give me some advice. Thank you!
Best wishes!

Dear Sir,
In the paper refinement loss is defined as euclidean loss but in the tensorflow code it is defined as crossentropy loss why is it different?
Also did you try L1 loss ?
Thanks

For example，line 83 “ FLAGS.nStride ” nStride is not difined, line 161 "nd_patch_step" is not difined,
using glob.glob() instead of glob() may be better and so on.

test.py doesnot output any meaningful result? the code is not finished? how about AUC?

Will this also be applicable to structure data - continuous or categorical columns.

I ran training, but hit this error... Anyone getting this?

python train.py --dataset UCSD --dataset_address ./dataset/UCSD_Anomaly_Dataset.v1p2/UCSDped2/Train --input_height 45 --output_height 45
{'attention_label': 1,
 'batch_size': 128,
 'beta1': 0.5,
 'checkpoint_dir': 'checkpoint',
 'dataset': 'UCSD',
 'dataset_address': './dataset/UCSD_Anomaly_Dataset.v1p2/UCSDped2/Train',
 'epoch': 40,
 'input_fname_pattern': '*',
 'input_height': 45,
 'input_width': None,
 'learning_rate': 0.002,
 'log_dir': 'log',
 'output_height': 45,
 'output_width': None,
 'r_alpha': 0.2,
 'sample_dir': 'samples',
 'train': True,
 'train_size': inf}
2018-12-22 15:03:46.625140: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\platform\cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2
2018-12-22 15:03:46.948538: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\gpu\gpu_device.cc:1030] Found device 0 with properties:
name: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.6705
pciBusID: 0000:01:00.0
totalMemory: 6.00GiB freeMemory: 4.97GiB
2018-12-22 15:03:46.965900: I C:\tf_jenkins\home\workspace\rel-win\M\windows-gpu\PY\35\tensorflow\core\common_runtime\gpu\gpu_device.cc:1120] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: GeForce GTX 1060, pci bus id: 0000:01:00.0, compute capability: 6.1)
Program is on Train Mode
libpng error: Write Error

I can't reproduce the results in your paper,that is,D(G(z)), in your paper,all of them are between 0 and 1,I tried to print the variable 'D_',it turned out that all of them are negative,and so ,the results of variable 'D_logits_' are all too close to zero,which means the algorithm didn't go very well because all the test patches are abnormal,I was stuck for a long time,and I hope you can help me.