---
license: mit
language:
- en
metrics:
- accuracy
pipeline_tag: image-classification
tags:
- TensorFlow
- Keras
- Pokedex
- Image-Classification
datasets:
- RogerKoala/gen1-pokemon-images
base_model:
- keras/densenet_201_imagenet
library_name: keras
---

<style>
img {
 display: inline;
}
</style>

[![Google Colab Notebook](https://img.shields.io/badge/Google%20Colab-Notebook-blue#model-badge)](#google-colab-notebook)
| [![Demo](https://img.shields.io/badge/Live%20Demo-Hugging%20Face%20Space-red?style=flat)](#demo)

# Model Summary

* Architecture: DenseNet121.
* Accuracy: 96% on the test set.
* Framework: Tensorflow.


## Usage

```
import keras
from keras.src.utils import load_img
from keras.src.applications.densenet import preprocess_input
import numpy as np

pokedex = keras.saving.load_model("pokedex.keras")

image = load_img('image.png', target_size=(224, 224))

x = keras.utils.img_to_array(image)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)

preds = pokedex.predict(x)

# Important: You must ensure that the Pokémon are ordered alphabetically, as the model was trained using this sequence. You can obtain the .txt file from the following link: https://huggingface.co/spaces/RogerKoala/Pokedex/blob/main/Pokemons.txt

with open('Pokemons.txt', 'r') as f:
  class_labels = f.read().splitlines()

top_indices = preds[0].argsort()[-3:][::-1] 

for i in top_indices:
    print(f"{class_labels[i]}: {preds[0][i]*100:.2f}%")
```

## System

- Input reqs: 224×224×3 RGB, normalized.

- Downstream deps: Class index→Pokémon metadata lookup.

## Implementation requirements

- Training: T4 Google Colab.
- Duration: 1 hour and 15 minutes.

# Model Characteristics

## Model initialization

- Fine-tuned from ImageNet DenseNet121.

## Model stats

| Layer (type)                  | Output Shape         | Param #     |
|------------------------------|----------------------|-------------|
| densenet121 (Functional)     | (None, 7, 7, 1024)    | 7,037,504   |
| global_average_pooling2d (GlobalAveragePooling2D)      | (None, 1024)          | 0           |
| dense (Dense)                | (None, 128)           | 131,200     |
| dropout (Dropout)            | (None, 128)           | 0           |
| dense_1 (Dense)              | (None, 151)           | 19,479      |

**Total params:** 21,397,255 (81.62 MB) <br>
 **Trainable params:** 7,104,535 (27.10 MB) <br>
 **Non-trainable params:** 83,648 (326.75 KB) <br>
 **Optimizer params:** 14,209,072 (54.20 MB) <br>

## Training data

- Collected via scripts from public archives.

- Pre-processing: resize to 224×224, normalize.


## Evaluation data

 - Train: 2249 images.
 - Test: 840 images.

# Evaluation Results

## Summary

Test accuracy: 96%

## Usage limitations

- Only original 151.

- Fails on non-canonical art styles, low light, occlusion.

## Google Colab Notebook

Explore the model training and inference workflow in this interactive notebook:

* [Pokédex Colab Notebook](https://colab.research.google.com/drive/1IeGzndeOZ_9PnnWSnLjYgaS7QFn-4Voc?usp=sharing)

## Demo

Visit the live Space to try it out:

* [Hugging Face Space](https://huggingface.co/spaces/RogerKoala/Pokedex)