https://github.com/nathancooperjones/meezer
Supervised Siamese networks using hard negative examples.
https://github.com/nathancooperjones/meezer
Last synced: 2 months ago
JSON representation
Supervised Siamese networks using hard negative examples.
- Host: GitHub
- URL: https://github.com/nathancooperjones/meezer
- Owner: nathancooperjones
- Created: 2020-06-18T15:36:03.000Z (about 5 years ago)
- Default Branch: master
- Last Pushed: 2023-03-24T23:50:51.000Z (about 2 years ago)
- Last Synced: 2025-02-10T22:29:20.807Z (4 months ago)
- Language: Jupyter Notebook
- Size: 8.97 MB
- Stars: 0
- Watchers: 2
- Forks: 0
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- Changelog: CHANGELOG.md
Awesome Lists containing this project
README
# meezer
Supervised Siamese-ish networks using hard negative examples.
A good deal of this code, especially files dealing with building and extracting information from the `Annoy` index, is based off of [Ivis](https://github.com/beringresearch/ivis). The `Img2Vec` code is based off of [this repo](https://github.com/jaredwinick/img2vec-keras).
## Usage
```python
import matplotlib.pyplot as plt
import numpy as np
from sklearn.datasets import load_digits
from sklearn.preprocessing import MinMaxScaler
from meezer import Meezer
# GET DATA
digits = load_digits()
X, labels = digits.data, digits.target
labels = labels.astype(int)
# SCALE DATA
X_scaled = MinMaxScaler().fit_transform(X)
# TRAIN A MODEL
model = Meezer(embedding_dims=2,
k=25,
batch_size=512,
epochs=30,
sub_epochs=10)
embeddings = model.fit_transform(X=X_scaled, Y=labels)
# VISUALIZE EMBEDDINGS
for i, mnist_num in enumerate(set(labels)):
mnist_num_idxs = np.argwhere(labels == mnist_num)
plt.scatter(embeddings[mnist_num_idxs, 0],
embeddings[mnist_num_idxs, 1],
color=plt.cm.gist_rainbow(i / len(set(labels))),
s=4,
label=mnist_num)
plt.xlabel('embedding dim 1')
plt.ylabel('embedding dim 2')
plt.legend(bbox_to_anchor=(1.04,1), loc='upper left')
plt.show()
```
### Development
Begin by installing [Docker](https://docs.docker.com/install/) if you have not already. Once Docker is running, run development from within the Docker container:
```bash
# build the Docker image
docker build -t meezer .
# run the Docker container in interactive mode.
docker run \
-it \
--rm \
-v "${PWD}:/meezer" \
-p 8888:8888 \
meezer /bin/bash
# then run JupyterLab and begin development
jupyter lab --ip 0.0.0.0 --no-browser --allow-root --NotebookApp.token='' --NotebookApp.password=''
```
## TODOs
- [ ] Exponential data `hard_mode` scheduler to introduce hard negative examples at a better rate than linearly.
- [ ] Clean up notebooks to show examples a bit better and with more available data sources.
- [ ] Add in traditional Siamese network using the triplet generator, but now feeding inputs to models two at a time, with `y` being a label of `1` if the pair are the same class or `0` otherwise. Using hard negative examples will help performance, most likely.
- [ ] Better support for callbacks (early stopping, learning rate scheduler, etc.).