Results don't match for same videos in dataset about anomalydetectioncvpr2018

[WaqasSultani]

Hi,
I am not sure about the different feature vectors. Actually, I have downloaded C3D around 2 years ago. They might have changed something after that---not sure though. If that is the case, you might need to retrain the model.

[lrd8715923]

Dear sir,
@WaqasSultani,Thank you for your great work, it is really nice.I have been able to extract features using c3d (caffe2) and test results by use FC's model from you provide,i comparison the results of our results with SampleVideos from you provide.I find it is a little different.(For each images,the first results is SampleVideos from you provide,the sencond results is our)I was so confused,i use same FC model,so i guess is C3D'model made this difference.I want to get same results as SampleVideos from you provide. Could you please provide your C3D'model,please provide a download link or send to my email:[email protected]. Thank you very much.
And i have a question,i feel the 3D convolution's method of your paper is a little old.I want to replace C3D with newest method(such as R(2+1)D or Multi-Fiber Net) and retrain FC'layer to Improving AUC .What do you think of that?Could you give me some instructions.
Best greetings from Liruodao.

[Afladimir1]

Hi,
I am not sure about the different feature vectors. Actually, I have downloaded C3D around 2 years ago. They might have changed something after that---not sure though. If that is the case, you might need to retrain the model.

You are probably right, I'm also getting very different results. Can you provide us your version of C3D? Thanks in advance.

[junaidwahid]

@Afladimir1 same here, @WaqasSultani can you provide your C3D version?

[nerdykamil]

Thank you for the suggestion @WaqasSultani. When I try to retrain the model using the command python TrainingAnomalyDetector_public.py. It throws error on line: model.train_on_batch(inputs, targets). I searched the error but couldn't find anything helpful.

My Feature matrix has shape (256,4096)
The shape of target vector is (256,) - tried reshaping to (256,1), didn't help

The output of the train command is:

**Using Theano backend.
/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/pywt/init.py:16: RuntimeWarning: numpy.dtype size changed, may indicate binary incompatibility. Expected 96, got 88
from ._extensions._pywt import *
/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/pywt/_swt.py:8: RuntimeWarning: numpy.dtype size changed, may indicate binary incompatibility. Expected 96, got 88
from ._extensions._swt import swt_max_level, swt as _swt, swt_axis as _swt_axis
Create Model
Starting training...
number and name of folders
2
['Abnormal', 'Normal']
Loading Abnormal videos Features...
Abnormal Features loaded
Abnormal Features loaded
Abnormal Features loaded
Abnormal Features loaded
Loading Normal videos...
Features loaded
Traceback (most recent call last):
File "TrainingAnomalyDetector_public.py", line 309, in
batch_loss =model.train_on_batch(inputs, targets)
File "/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/keras/models.py", line 719, in train_on_batch
class_weight=class_weight)
File "/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/keras/engine/training.py", line 1239, in train_on_batch
outputs = self.train_function(ins)
File "/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/keras/backend/theano_backend.py", line 792, in call
return self.function(*inputs)
File "/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/theano/compile/function_module.py", line 917, in call
storage_map=getattr(self.fn, 'storage_map', None))
File "/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/theano/gof/link.py", line 325, in raise_with_op
reraise(exc_type, exc_value, exc_trace)
File "/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/theano/compile/function_module.py", line 903, in call
self.fn() if output_subset is None else
ValueError: Input of CAReduce{maximum} has zero-size on axis 0
Apply node that caused the error: Reduce{maximum}{0}(Subtensor{int64:int64:}.0)
Toposort index: 1218
Inputs types: [TensorType(float32, vector)]
Inputs shapes: [(0,)]
Inputs strides: [(4,)]
Inputs values: [array([], dtype=float32)]
Outputs clients: [[MakeVector{dtype='float32'}(max), InplaceDimShuffle{x}(max)]]

Backtrace when the node is created(use Theano flag traceback.limit=N to make it longer):
File "TrainingAnomalyDetector_public.py", line 278, in
model.compile(loss=custom_objective, optimizer=adagrad)
File "/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/keras/models.py", line 553, in compile
**kwargs)
File "/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/keras/engine/training.py", line 630, in compile
sample_weight, mask)
File "/home/lfd/anaconda3/envs/c3d_caffe2_try3/lib/python2.7/site-packages/keras/engine/training.py", line 332, in weighted
score_array = fn(y_true, y_pred)
File "TrainingAnomalyDetector_public.py", line 222, in custom_objective
sub_max = T.concatenate([sub_max, T.stack(T.max(Feat_Score))]) # Keep the maximum score of scores of all instances in a Bag (video)

HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node.**

[nerdykamil]

Keeping the batch_size 60 I was able to train the model and get the results. Thank you 👍

[DenisN03]

@nerdykamil, @lrd8715923. can you tell me how you prepared the features? What should the folder structure be for training?

[mrkstt]

Hi @nerdykamil , on your side how long the training process has done?. I'm using GTX 1050 Ti, and it's already two days, and the training is still ongoing.

[hali6779]

Hi @lrd8715923, may I have some guidance on how to replace C3D with other methods or do you have any sample code for that? Thanks in advance.