You can deploy this model and start earning money today!

mysticdan/
mlp-mnist

Name: mlp-mnist
Availability: InStock

mit

Views

No views yet

MNIST-MLP

A lightweight MNIST digit classification model implemented from scratch in JAX.

Model Details

Architecture: 5-layer MLP (784–512–256–128–64–10) with ReLU activations and dropout (0.3).
Initialization: He normal (stddev = $\sqrt{2/m}$).
Loss: Cross-entropy with L1+L2 regularization (λ=1e-4).
Optimizer: SGD with exponential decay (initial lr=0.005, decay rate=0.015).
Early Stopping: Patience=20 epochs, min_delta=1e-4.
Final Test Accuracy: ~98% (Pending further tuning).

Files in this repository

mlp_model.py — Model definition (MLP class, feedforward, dropout logic).
mlp_mnist_model.pkl — Serialized model parameters and layer activations.
main_mnist_predict.py — Streamlit app (streamlit_drawable_canvas) for interactive inference.
requirements.txt — Minimal dependencies for inference:

Quick Start

pip install -r requirements.txt

streamlit run main_mnist_predict.py

Open http://localhost:8501 in your browser, draw a digit, and press Predict.

Loading the Model in Python

python

from huggingface_hub import hf_hub_download
import pickle
import jax
from mlp_model import MLP, activation_map

# Download parameters
path = hf_hub_download(repo_id="mysticdan/mlp-mnist", filename="mlp_mnist_model.pkl")
with open(path, "rb") as f:
    data = pickle.load(f)

# Reconstruct MLP
mlp = MLP.__new__(MLP)
mlp.params = data["params"]
mlp.activation = tuple(
    activation_map["identity"] if name == "<lambda>" else activation_map[name]
    for name in data["activation_names"]
)
mlp.key = jax.random.key(0)

# Inference
import jax.numpy as jnp
# assume `sample` is a (784,) float32 array normalized [0,1]
logits = mlp.predict(sample[None, :])
pred = int(jnp.argmax(logits, axis=1)[0])
print("Predicted digit:", pred)

mysticdan/mlp-mnist