https://github.com/yeyom/not_hotdog_model
In this notebook I will try to recreate a curious model used in an App shown in the HBO serie "Sillicon Valley".
https://github.com/yeyom/not_hotdog_model
keras mobilenetv2 not-hotdog sillicon-valey tensorflow
Last synced: 3 months ago
JSON representation
In this notebook I will try to recreate a curious model used in an App shown in the HBO serie "Sillicon Valley".
- Host: GitHub
- URL: https://github.com/yeyom/not_hotdog_model
- Owner: YeyoM
- License: apache-2.0
- Created: 2023-01-21T03:43:56.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2023-01-21T04:52:30.000Z (over 2 years ago)
- Last Synced: 2025-01-31T17:14:08.405Z (5 months ago)
- Topics: keras, mobilenetv2, not-hotdog, sillicon-valey, tensorflow
- Language: Jupyter Notebook
- Homepage:
- Size: 2.9 MB
- Stars: 1
- Watchers: 3
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Not_Hotdog_Model
In this notebook I will try to recreate a curious model used in an App shown in the HBO serie "Sillicon Valley".
For this model, I will use the hotdog-nothotdog dataset from Kaggle, and to build the model my way to go was to use Tensorflow using Keras layers to build a Convolutional Neural Network and a pre-trained model to make use of transfer learning from the mobileNet v2 model.
Along the notebook and the code blocks, I will try to explain as best as I can what the building process of this model works and at the end obviously the results I got.
Hopefully I can also build the App as in the serie with this model I built, which is the first model I built by my own.
A detailed explanation of how the model was built take a look at the notebook attached in this repo.
You can found the google colab notebook in the following link:
- https://colab.research.google.com/drive/1ScYoJsGa-C5tqyyR5ETeA7pAgXwtwCx1#scrollTo=R0za6i4K4hkg
Here are the results I got.
## Results
### Training
For the training process I run it for 20 epochs, with a batch size of 32 and a learning rate of 0.0001. Using the mobileNet v2 and some more layers including dropout and image augmentation to prevent overfitting. The results I got were the following:
[![training][training-screenshot]]()
### Evaluation
For the evaluation process I used the validation set of the dataset and the results I got were the following:
[![evaluate][evaluate-screenshot]]()
### Predictions
For the predictions I used the test set of the dataset and the results I got were the following:
[![predictions][predictions-screenshot]]()
[training-screenshot]: public/training.png
[evaluate-screenshot]: public/evaluate.png
[predictions-screenshot]: public/predictions.png