deepmeta: Deep Meta Learning

This package implements deep meta learning algorithms.

Usage Example

import numpy as np
from tensorflow.keras import layers, models
from deepmeta import MAML

# Define a deep-network model
input_shape = (28, 28, 1)
category_count = 10
model = models.Sequential([
    layers.Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=input_shape),
    layers.MaxPooling2D(pool_size=(2, 2)),
    layers.Flatten(),
    layers.Dense(128, activation='relu'),
    layers.Dense(category_count, activation='softmax')
])

# Initialize MAML
maml = MAML(model, meta_lr=0.001, task_lr=0.01, inner_steps=5)

# Generate dummy data
train_sample_count = 100
test_sample_count = 50
train_features = np.random.rand(*([train_sample_count] + list(input_shape)))
train_targets = np.random.randint(0, category_count, size=(train_sample_count,))
test_features = np.random.rand(*([test_sample_count] + list(input_shape)))
test_targets = np.random.randint(0, category_count, size=(test_sample_count,))

# Split data into tasks
support_set = [(train_features[i:i+5], train_targets[i:i+5]) for i in range(0, 100, 5)]
query_set = [(test_features[i:i+5], test_targets[i:i+5]) for i in range(0, 50, 5)]

# Train MAML
for epoch in range(10):
    meta_loss = maml.train_step(support_set, query_set)
    print(f"epoch: {epoch}, meta loss: {meta_loss.numpy()}")