mapextrackt 0.3

Pytorch Feature Map Extractor

MapExtrakt

Convolutional Neural Networks Are Beautiful

We all take our eyes for granted, we glance at an object for an instant and our brains can identify with ease. However distorted the information may be, we do a pretty good job at it.

Low light, obscured vision, poor eyesight... There are a myriad of situations where conditions are poor but still we manage to understand what an object it. Context helps, but we humans were created with sight in mind.

Computers have a harder time, but modern advances with convolutional neural networks are making this task a reality and have now surpassed human level accuracy.

Computers are beautiful, neural networks are beautiful. And the maps they create to determine what makes a cat a cat are beautiful.

---

MapExtrakt makes viewing feature maps a breeze.

Catch a glimpse of how a computer can see.

---

# load a model 
import torchvision
model = torchvision.models.vgg19(pretrained=True)

#import FeatureExtractor
from MapExtrackt import FeatureExtractor

#load the model and image
fe = FeatureExtractor(model)
fe.set_image("pug.jpg")

#gather maps
fe.display_from_map(layer_no=1)

View Single Cells At a Time

#gather cells
fe.display_from_map(layer_no=2, cell_no=4)

Slice the class to get a range of cells

Layer 2 Cells 0-9

fe[2,0:10]

Or Export Layers To Video

fe.write_video(out_size=(1920,1080), file_name="output.avi", 
               time_for_layer=60, transition_perc_layer=0.2)

---

More Examples

For LOTS more - view the jupyter notebook.

Examples

---

Installation

It's as easy as PyPI

pip install mapextrackt

or build from source in terminal

git clone https://github.com/lewis-morris/mapextrackt
cd mapextrackt
pip install -e .

---

---

Todo List

• Add the ability to slice the class i.e FeatureExtractor[1,3]
• Show parameters on the image
• Fix video generation
• Add video parameters such as duration in seconds.
• Clean up code

---

Author

Created by me, initially to view the outputs on demand.

open to suggestions/ collaborators.

If anyone has any suggestions or requests please send them over I'd be more than happy to consider

[e-mail]