camp-explain-ai / inputiba Goto Github PK

We can remove the run_tests.sh for the moment

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/aa728708eeedd43355168172b5f37377a27d0614/iba/models/bottlenecks/vision_feat_iba.py#L109

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/aa728708eeedd43355168172b5f37377a27d0614/iba/models/bottlenecks/vision_input_iba.py#L77

We'd better replace these letters with variable names like lamb, eps, etc.

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/aa728708eeedd43355168172b5f37377a27d0614/iba/models/bottlenecks/vision_input_iba.py#L80

As marked in the code, we need to consider the case where reverse_lambda = True in VisionInputIBA

Since smooth kernel is a 2D kernel.

Solution:
Remove smooth option completely since word are produced discretely (in LSTM)

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/775fad551e900907010ae8914fe727cb8e126d82/iba/models/utils.py#L129

This line uses skimage.transform.resize function. In order to reduce the dependencies, I suggest to replace it with cv2.resize.

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/775fad551e900907010ae8914fe727cb8e126d82/iba/models/attributors/vision_attributor.py#L60

When writing notification, please add a notification that explains why input_mask is sigmoid(alpha) while feat_mask is a KL-divergence.

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/67b4b5a1c2e5697fbacacaf39a0fbdc297763c70/iba/evaluation/insertion_deletion.py#L40

Currently we are using train and val for the datasets in configs. It would be better to use estimation and attribution instead.

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/215c1fed8717ae0f729677154ceeff4d93d99624/iba/models/attributors/nlp_attributor.py#L71

The term capacity is very misleading, though it might have special meaning in information theory. The IBA uses KL divergence to measure feature importance (see 3.2 in IBA's paper). Therefore, it would be better to re-name it to e.g. kl_divergence (take care of the possible name conflict with function kl_div).

In addition, BaseIBA has attributes loss_buffer, cls_loss_buffer, info_loss_buffer. These three attributes' names are in patter xxx_buffer, whichbuffer_capacity does not conform to. Also, these three buffers are list, while buffer_capacity is a Tensor. So, I suggest to change the name of buffer_capacity, and register it to nn.Module's buffer.

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/775fad551e900907010ae8914fe727cb8e126d82/iba/models/bottlenecks/vision_feat_iba.py#L105

Hi everyone,
Ild like to learn how long does it take to have explanations on cpu device. Its been half and hour and I only have plt.show() executed.

Hi,

This is excellent work, and thank you for sharing the well-designed official codes for a better understanding of the details of your NeurIPS paper.

I would like to understand the details of some equations, and I'd appreciate it if you could explain them to me.
In your paper, you proposed two assumptions for estimating the $I[I, Z]$,
which are $Z_I = \Lambda Z_{G} + (1 - \Lambda) \epsilon$ and $Z_G = \lambda_G Z_{I} + (1 - \lambda_G) \epsilon_{G}$,

1. the term $\epsilon$ is only mentioned again in the $Z^* = \lambda^* R + (1-\lambda^{*}) \epsilon$ for feature IBA,
   can I assume they are not the same?
2. in order to derive remark 2 (equation 9) in your paper, can I assume the $\epsilon$ for $Z_I$ is actually equal to $\epsilon_G$ , so that I can merge those two epsilon terms together?
3. based on the above, the $\mu_I$ and $\sigma_I$ are actually $\mu_G$ and $\sigma_G$ in remark 2 and equation 9
4. if my understanding above is wrong, could you please let me know how can I get the remark 2, as $\epsilon_G$ is a Gaussian parameterized by learnable parameters $\mu_G$ and $\sigma_G$, but the parameters of $Z_I$'s distribution is related with $\mu_I$ and $\epsilon_I.$

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/7e3a56e9150f7487345e89f675faa36503717948/iba/models/bottlenecks/base_feat_iba.py#L33

Rename these three attributes to xxx_history
https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/775fad551e900907010ae8914fe727cb8e126d82/iba/models/bottlenecks/base_iba.py#L31

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/7e3a56e9150f7487345e89f675faa36503717948/iba/models/bottlenecks/base_feat_iba.py#L106

Should function estimate be an abstract method for class base_feature_iba?

remove the notebooks folder. We can create a tutorials directory later, and put all the tutorials there

If the tutorials can successfully run on colab with the newest albumentations , remember to unpin its version

https://github.com/YaNgZhAnG-V5/informationbottleneck/blob/775fad551e900907010ae8914fe727cb8e126d82/iba/models/utils.py#L48

torchvision.transforms.GaussianBlur also supports Tensor input.