darkonhub / darkon Goto Github PK

source

• dataset: cifar10
• model: ResNet-110

limits

• one test-set
• origin train-set without augmentation

result

• simple example
• visualization

test_batch() function required in InfluenceFeeder for large test data feeding

I am using tensorflow 1.6, when import darkon, it turns out that

/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/__init__.py in <module>()
     20 
     21 # Add projects here, they will show up under tf.contrib.
---> 22 from tensorflow.contrib import batching
     23 from tensorflow.contrib import bayesflow
     24 from tensorflow.contrib import cloud

ImportError: cannot import name batching

Can you run and get the most influential training examples of a very large dataset?

[1] Pang Wei Koh and Percy Liang "Understanding Black-box Predictions via Influence Functions" ICML2017

I tried the code they provided but it never worked for big datasets, other than easy model and MNIST

Thanks

Implement parallel processing of _grad_diffs() and _grad_diffs_all for faster execution
Needs to modify Tensorflow gradient functions

Using

• opencv for image processing
• scikit-image for 1-D interpolation

Hi,
I am litte confused over the calculation of cur_estimate in _get_inverse_hvp_lissa()

Since we are assigning test_grad_loss to cur_estimate for every ihvp_config['num_repeats']
The cur_estimate computed would be same in every loop.Why is it assigned test_grad_loss inside the ihvp_config['num_repeats'] loop ?

Thanks.

using Travis ci or Circle ci

Hi,

I came across your implementation of the Influence Function and am interested in separately grabbing the Hessian that is needed for the calculations. Is there a way for me to grab it without having to break apart the code?

thank you very much.

I tired to use darkon influence to find the influential example for wrong predicted examples.
However, one example took about 45 mins to complete.
Is it that slow?

When influence instance is created and called upweight function many times, the cpu memory will be increased.

Hello,
nice work, i really appreciate it.
I would like to know if it's possible to apply the idea of gradCam to a network such: pre-trained CNN for feature extraction + LSTM for sequence classification.
Thanks in avdance for the help!

There are many pre-trained nets in Slim, can darkon to visualize other nets, like inception?