- Given the stats of a Pokemon as the input, predict which type they belong to (their primary type). An example is, Bulbasaur -> Grass. The task, although it seems rather simple, is not that easy since a lot of the stats are very similar in a lot of different types. Pokemon that have two types make it even harder since their stats would be shared.
- There are 19 types in total, so the network performs a 19 class classification
- Implemented a simple 4 layered neural network (2 hidden layers) with a softmax layer at the end
- Uses adam optimization to perform updation
- Computes a top-5 match in the accuracy (ie. If any one of the top 5 classes match the correct output, we consider it as correct). Also performed a few experiments with top-3 and top-1 matches as well
Pokemon Type Prediction challenge by @Sirajology on Youtube
- Tensorflow
- Numpy
Run python main.py and it would train the network and then run it on a randomly subsampled test dataset (not included in the training) and print the accuracy.
| ID | Top-K | Network-Shape | Iterations | Accuracy.avg |
|---|---|---|---|---|
| 1 | 5 | (7,128,256,19) | 100 | 58.7499976158 |
| 2 | 5 | (7,128,256,19) | 500 | 60.2500001921 |
| 3 | 5 | (7,512,256,19) | 100 | 63.7499988079 |
| 4 | 5 | (7,512,256,19) | 500 | 65.9999976158 |
| 5 | 3 | (7,128,256,19) | 100 | 41.2499999049 |
| 7 | 3 | (7,512,256,19) | 100 | 43.7500000011 |