thinc_linear_relu_softmax_model_with_backprop_training.py

1from thinc.api import chain, Relu, Softmax, model_ops
2
3# 1. Define the model architecture
4# This creates a simple feed-forward network: Linear + ReLU -> Linear + Softmax
5model = chain(
6    Relu(nO=32, nI=16), 
7    Softmax(nO=10, nI=32)
8)
9
10# 2. Initialize the model with sample data to infer shapes
11import numpy
12X = numpy.zeros((128, 16), dtype="f")
13Y = numpy.zeros((128, 10), dtype="f")
14model.initialize(X=X, Y=Y)
15
16# 3. Predict
17Y_hat = model.predict(X)
18
19# 4. Training loop (simplified)
20# The model returns the output and a callback to complete the backprop
21Y_hat, backprop = model.begin_update(X)
22gradient = Y_hat - Y
23d_X = backprop(gradient)
24
25# 5. Update weights
26optimizer = model.make_optimizer()
27model.finish_update(optimizer)