# Captum
In all this notebook, we will use the following data:
``` python
from fastai.vision.all import *
```
``` python
path = untar_data(URLs.PETS)/'images'
fnames = get_image_files(path)
def is_cat(x): return x[0].isupper()
dls = ImageDataLoaders.from_name_func(
path, fnames, valid_pct=0.2, seed=42,
label_func=is_cat, item_tfms=Resize(128))
```
``` python
from random import randint
```
``` python
learn = vision_learner(dls, resnet34, metrics=error_rate)
learn.fine_tune(1)
```
# Captum Interpretation
The Distill Article
[here](https://distill.pub/2020/attribution-baselines/) provides a good
overview of what baseline image to choose. We can try them one by one.
------------------------------------------------------------------------
source
### CaptumInterpretation
``` python
def CaptumInterpretation(
learn, cmap_name:str='custom blue', colors:NoneType=None, N:int=256,
methods:tuple=('original_image', 'heat_map'), signs:tuple=('all', 'positive'), outlier_perc:int=1
):
```
*Captum Interpretation for Resnet*
## Interpretation
``` python
captum=CaptumInterpretation(learn)
idx=randint(0,len(fnames))
captum.visualize(fnames[idx])
```
``` python
captum.visualize(fnames[idx],baseline_type='uniform')
```
``` python
captum.visualize(fnames[idx],baseline_type='gauss')
```
``` python
captum.visualize(fnames[idx],metric='NT',baseline_type='uniform')
```
``` python
captum.visualize(fnames[idx],metric='Occl',baseline_type='gauss')
```
## Captum Insights Callback
``` python
@patch
def _formatted_data_iter(x: CaptumInterpretation,dl,normalize_func):
dl_iter=iter(dl)
while True:
images,labels=next(dl_iter)
images=normalize_func.decode(images).to(dl.device)
yield Batch(inputs=images, labels=labels)
```
------------------------------------------------------------------------
source
### CaptumInterpretation.insights
``` python
def insights(
x:CaptumInterpretation, inp_data, debug:bool=True
):
```
*Call self as a function.*
``` python
captum=CaptumInterpretation(learn)
captum.insights(fnames)
```
CaptumInsights(insights_config={'classes': ['False', 'True'], 'methods': ['Deconvolution', 'Deep Lift', 'Guide…
Output()