Timeline and files in implementing DNN from scratch
- 00- v1.0 - Data preparation file
- 01- v1.1 - Logistic Regression and 2 Layer Shallow Neural Network (sigmoid -> sigmoid)
- 02- Data augmentation (increasing dataset, thus creating new dataset)
- 03- v1.1.1 - v1.1 with new dataset
- 04- v1.2 - shallow nn implementation (relu -> sigmoid) on original dataset
- 05- v1.2.1 - v1.2 with new dataset (relu -> sigmoid)
- 06- v1.3 - deep nn with original dataset ((n-1)relu -> sigmoid)
- 07- v1.3.1 - deep nn with new dataset
- 08- v1.3.2 - v1.3.1 with He/Xavier initialization and Dropout regularization (failed)
- 09- v1.3.3 - v1.3.1 with He/Xavier Initialization
- 10- v1.4 - DNN with He/Xavier initialization and Adam optimizer
- 11- v_1.1 - 3 classes instead of 2
- 12- v_1.1.1 - softmax function changed
- 13- v_1.2 - uses adam optimizer
- 14- v_1.2.1 - now 7 classes,with adam optimizer
File No. | Training Accuracy | Testing Accuracy |
---|---|---|
01 (logistic) | 100 % | 73.77 % |
01 (shallow) | 100 % | 78 % |
03 (logistic) | 100 % | 68.59 % |
03 (shallow) | 100 % | 67 % |
04 | 100 % | 75.41 % |
05 | 99.99 % | 81.81 % |
06 | 100 % | 73.77 % |
07 | 99.9 % | 74.38 % |
08 (for 2 layers) | 98.76 % | 76.86 % |
08 (for 4 layers,failed) | -- | -- |
09 | 99.22 % | 74.38 % |
10 | 99.05 % | 79.12 % |
11 | 38.9 % | 35.64 % |
12 | 95.52 % | 61.69 % |
13 | 64.63 % | 53.23 % |
14 | 100 % | 32.11 % |
Dataset from - Dog Dataset
Code Snippets from
Inspired by - Andrew Ng
Project partner- Rahul Lamge