https://github.com/dcarpintero/fastai-deeplearning
Learning notes and projects built for the Deep Learning course by fast.ai
https://github.com/dcarpintero/fastai-deeplearning
data-augmentation deep-learning fastai resnet-50
Last synced: over 1 year ago
JSON representation
Learning notes and projects built for the Deep Learning course by fast.ai
- Host: GitHub
- URL: https://github.com/dcarpintero/fastai-deeplearning
- Owner: dcarpintero
- License: mit
- Created: 2024-02-07T16:45:10.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2024-03-10T23:56:28.000Z (over 2 years ago)
- Last Synced: 2024-11-20T15:56:55.531Z (over 1 year ago)
- Topics: data-augmentation, deep-learning, fastai, resnet-50
- Language: Jupyter Notebook
- Homepage:
- Size: 5.94 MB
- Stars: 0
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Practical Deep Learning @ fast.ai
Learning notes and projects built for the Deep Learning course by [fast.ai](https://course.fast.ai/).
## 01. Getting Started - Bird Classifier
[Notebook](./course2024/lesson_01.ipynb) |
[Summary](./course2024/lesson_01.summary.md) |
[Quiz](./course2024/lesson_01.quiz.md)
## 02. Deployment - Astronomy Classifier
[Notebook](./course2024/lesson_02.astronomy.ipynb) |
[Summary](./course2024/lesson_01.summary.md) |
[Quiz](./course2024/lesson_02.quiz.md) |
[Model](https://huggingface.co/dcarpintero/fastai-interstellar-class) |
[Try in HuggingFace Spaces](https://huggingface.co/spaces/dcarpintero/interstellar)
`[deep-learning]` `[data-augmentation]` `[ResNet-50]` `[transfer-learning]`
Visual learner to classify images of astronomical objects using ResNet and transfer learning (1 + 3 epochs). We provide two versions: `class model`, and `object model`.
In the `class model`, the possible labels are: `galaxy`, `nebula`, `comet`, `asteroid`, `quasar`, and `star cluster`.
In the `object model` the labels are specific astronomy objects: `m31 andromeda galaxy`, `m33 triangulum galaxy`, `m81 bode galaxy`, `m82 cigar galaxy`, `ngc 1300 galaxy`, `m104 sombrero galaxy`, `m51 whirlpool galaxy`, `m42 orion nebula`, `m17 omega nebula`, and `m45 pleiades star cluster`.
The datasets have been created using [Bing Search API](https://www.microsoft.com/en-us/bing/apis/bing-web-search-api), 150 images per class with augmentation.
The model reaches 84% accuracy on class level:
and 91% accuracy at object level:
## 03. Neural Net Foundations - Digit Classifier w/ Multi-Layer Perceptron
[Summary](./course2024/lesson_03.summary.md) |
[Quiz](./course2024/lesson_03.quiz.md) |
[Digit Classifier w/ Multi-Layer Perceptron](./course2024/lesson_03.full.mnist.mlp.ipynb) |
[Try in HuggingFace Spaces](https://huggingface.co/spaces/dcarpintero/mlp-digit-classifier)
`[deep-learning]` `[perceptron]` `[backpropagation]` `[gradient-descend]` `[linear-layer]` `[relu]` `[optimizer]` `[mnist]` `[multi-class]` `[universal-approximation-theorem]`
[Annotated Multi-Layer Perceptron](./course2024/lesson_03.full.mnist.mlp.md) trained on the MNIST dataset to classify handwritten digits. It implements from scratch the following modules: **linear layer**, **relu-activation-function**, **sequential-layer**, **flatten-layer**, **basic optimizer**, and **learner**.
We define a hidden layer of size `256, 64` wherein the input tensors (28 x 28) are flattened in the height, and width dimensions. The model achieves 95.6% accuracy with `15 training epochs` and `batch size = 64`.
