thinc_mnist_linear_model_relu_softmax_training.py

1from thinc.api import chain, Relu, Softmax, Adam, fix_random_seed
2import numpy
3
4# Set a random seed for reproducibility
5fix_random_seed(0)
6
7def train_model(data, n_iter=10, n_hidden=32, learn_rate=0.001):
8    (train_X, train_Y), (test_X, test_Y) = data
9    
10    # 1. Define the model
11    model = chain(
12        Relu(n_hidden, dropout=0.2), 
13        Relu(n_hidden, dropout=0.2), 
14        Softmax()
15    )
16    
17    # 2. Initialize the model with a batch of data
18    model.initialize(X=train_X[:5], Y=train_Y[:5])
19    
20    # 3. Create the optimizer
21    optimizer = Adam(learn_rate)
22    
23    # 4. Training loop
24    for i in range(n_iter):
25        # Get loss and gradient of the loss with respect to the output
26        Yh, backprop = model.begin_update(train_X)
27        loss = ((Yh - train_Y) ** 2).sum()
28        d_loss = Yh - train_Y
29        
30        # Backpropagate the gradient and update weights
31        backprop(d_loss)
32        model.finish_update(optimizer)
33        
34        # Evaluate on test data
35        score = model.predict(test_X).argmax(axis=1) == test_Y.argmax(axis=1)
36        print(f"Iter {i}: loss {loss:.3f}, accuracy {score.mean():.3f}")
37
38if __name__ == "__main__":
39    # Example: Create dummy MNIST-like data
40    def get_dummy_data(samples, input_dim, output_dim):
41        X = numpy.random.uniform(size=(samples, input_dim)).astype("float32")
42        Y = numpy.zeros((samples, output_dim), dtype="float32")
43        for i in range(samples):
44            Y[i, numpy.random.randint(0, output_dim)] = 1.0
45        return X, Y
46
47    train_data = get_dummy_data(100, 784, 10)
48    test_data = get_dummy_data(20, 784, 10)
49    train_model((train_data, test_data))