|
|
--- |
|
|
license: mit |
|
|
language: en |
|
|
library_name: pytorch |
|
|
tags: |
|
|
- pytorch |
|
|
- tabular-classification |
|
|
- pokemon |
|
|
- finance |
|
|
- scikit-learn |
|
|
- shap |
|
|
--- |
|
|
|
|
|
# Pokémon TCG Price Predictor |
|
|
|
|
|
This repository contains a PyTorch model trained to analyze Pokemon card features to identify cards with potential for significant price increases. |
|
|
|
|
|
|
|
|
This model is the backend for the **[PokePrice Gradio Demo](https://huggingface.co/spaces/OffWorldTensor/PokePrice)**. |
|
|
|
|
|
## Model Description |
|
|
|
|
|
The model is a simple Multi-Layer Perceptron (MLP) implemented in PyTorch. It takes various features of a Pokémon card as input—such as its rarity, type, and historical price data—and outputs a single logit. A sigmoid function can be applied to this logit to get a probability score for the price rising. |
|
|
|
|
|
- **Model type:** Tabular Binary Classification |
|
|
- **Architecture:** `PricePredictor` (MLP) |
|
|
- **Framework:** PyTorch |
|
|
- **Training Data:** A custom dataset derived from the PokemonTCG/pokemon-tcg-data repository, augmented with pricing history. |
|
|
|
|
|
## How to Use |
|
|
|
|
|
To use this model, you will need `torch`, `scikit-learn`, `pandas`, and `huggingface_hub`. You can download the model artifacts directly from the Hub. |
|
|
|
|
|
First, ensure you have `network.py` (which defines the model class) in your working directory. |
|
|
|
|
|
```python |
|
|
import torch |
|
|
import joblib |
|
|
import json |
|
|
import pandas as pd |
|
|
from huggingface_hub import hf_hub_download |
|
|
from safetensors.torch import load_file |
|
|
|
|
|
# Make sure you have network.py in the same directory |
|
|
from network import PricePredictor |
|
|
|
|
|
REPO_ID = "your-username/pokemon-price-predictor" |
|
|
MODEL_FILENAME = "model.safetensors" |
|
|
CONFIG_FILENAME = "config.json" |
|
|
SCALER_FILENAME = "scaler.pkl" |
|
|
|
|
|
print("Downloading model files from the Hub...") |
|
|
model_path = hf_hub_download(repo_id=REPO_ID, filename=MODEL_FILENAME) |
|
|
config_path = hf_hub_download(repo_id=REPO_ID, filename=CONFIG_FILENAME) |
|
|
scaler_path = hf_hub_download(repo_id=REPO_ID, filename=SCALER_FILENAME) |
|
|
print("Downloads complete.") |
|
|
|
|
|
with open(config_path, "r") as f: |
|
|
config = json.load(f) |
|
|
|
|
|
feature_columns = config["feature_columns"] |
|
|
input_size = config["input_size"] |
|
|
|
|
|
model = PricePredictor(input_size=input_size) |
|
|
model.load_state_dict(load_file(model_path)) |
|
|
model.eval() |
|
|
|
|
|
scaler = joblib.load(scaler_path) |
|
|
|
|
|
data_to_predict = { |
|
|
'rawPrice': [10.0], 'gradedPriceTen': [100.0], 'gradedPriceNine': [50.0], |
|
|
} |
|
|
|
|
|
input_df = pd.DataFrame(data_to_predict) |
|
|
missing_cols = set(feature_columns) - set(input_df.columns) |
|
|
for c in missing_cols: |
|
|
input_df[c] = 0.0 |
|
|
input_df = input_df[feature_columns] |
|
|
|
|
|
|
|
|
input_scaled = scaler.transform(input_df.values) |
|
|
input_tensor = torch.tensor(input_scaled, dtype=torch.float32) |
|
|
|
|
|
with torch.no_grad(): |
|
|
logits = model(input_tensor) |
|
|
probability = torch.sigmoid(logits).item() |
|
|
|
|
|
print(f"\nPrediction for the input card:") |
|
|
print(f" - Probability of 30% price rise in 6 months: {probability:.4f}") |
|
|
|
|
|
if probability > 0.5: |
|
|
print(" - Prediction: Price WILL LIKELY rise.") |
|
|
else: |
|
|
print(" - Prediction: Price WILL LIKELY NOT rise.") |
|
|
``` |
|
|
|
|
|
|
|
|
## Model Explainability |
|
|
|
|
|
To understand the model's decisions, SHAP (SHapley Additive exPlanations) values were computed. |
|
|
|
|
|
### Global Feature Importance |
|
|
|
|
|
This plot shows the average impact of each feature on the model's output magnitude. Features at the top are most influential. |
|
|
|
|
|
 |
|
|
|
|
|
## Limitations and Bias |
|
|
|
|
|
- The model is trained on historical data and may not predict future trends accurately, especially in a volatile market. |
|
|
- The definition of "price rise" is fixed at 30% over 6 months. The model is not trained for other thresholds or timeframes. |
|
|
- The dataset may have inherent biases related to card popularity, set releases, or data collection artifacts. |
|
|
|
|
|
## Author |
|
|
|
|
|
Callum Anderson |
