Showing the changes to example/tinytorch_example.cpp (1 of 21 files). Show all
Unified · Split
| @@ -0,0 +1,149 @@ | ||
| 1 | +#include <cstddef> | |
| 2 | +#include <iostream> | |
| 3 | +#include <vector> | |
| 4 | + | |
| 5 | +#include "../tinytorch.hpp" | |
| 6 | + | |
| 7 | +namespace tt = tinytorch; | |
| 8 | +using namespace std; // NOLINT | |
| 9 | + | |
| 10 | +// let's create a neural network to approximate the following function | |
| 11 | +int classify_fruit(float color, float size, float weight) { | |
| 12 | + // color in [0, 1] where 0 represents green and 1 represents red | |
| 13 | + // size in [0, 1] where 0 represents small and 1 represents large | |
| 14 | + // weight in [0, 1] where 0 represents light and 1 represents heavy | |
| 15 | + // y in [0, 3] representing three types of fruits: apple, orange, and banana | |
| 16 | + | |
| 17 | + double y = 0.8 * color + 1.0 * size + 1.0 * weight; | |
| 18 | + | |
| 19 | + if (y < 0.9) { | |
| 20 | + return 0; // Apple | |
| 21 | + } else if (y < 1.8) { | |
| 22 | + return 1; // Orange | |
| 23 | + } else { | |
| 24 | + return 2; // Banana | |
| 25 | + } | |
| 26 | +} | |
| 27 | + | |
| 28 | +tuple<vector<float>, int> generate_case() { | |
| 29 | + float x0 = (float)rand() / RAND_MAX; | |
| 30 | + float x1 = (float)rand() / RAND_MAX; | |
| 31 | + float x2 = (float)rand() / RAND_MAX; | |
| 32 | + int y = classify_fruit(x0, x1, x2); | |
| 33 | + return {{x0, x1, x2}, y}; | |
| 34 | +} | |
| 35 | + | |
| 36 | +tuple<vector<vector<float>>, vector<int>, vector<vector<float>>, vector<int>> generate_dataset( | |
| 37 | + int train_size, int val_size) { | |
| 38 | + vector<vector<float>> train_inputs, val_inputs; | |
| 39 | + vector<int> train_targets, val_targets; | |
| 40 | + for (int i = 0; i < train_size; i++) { | |
| 41 | + auto [x, y] = generate_case(); | |
| 42 | + train_inputs.push_back(x); | |
| 43 | + train_targets.push_back(y); | |
| 44 | + } | |
| 45 | + for (int i = 0; i < val_size; i++) { | |
| 46 | + auto [x, y] = generate_case(); | |
| 47 | + val_inputs.push_back(x); | |
| 48 | + val_targets.push_back(y); | |
| 49 | + } | |
| 50 | + return {train_inputs, train_targets, val_inputs, val_targets}; | |
| 51 | +} | |
| 52 | + | |
| 53 | +int main() { | |
| 54 | + // create a context to hold all the tensors | |
| 55 | + tt::TensorContext ctx((size_t)1024 * 1024); | |
| 56 | + | |
| 57 | + // generate some random data | |
| 58 | + int train_size = 10000, val_size = 100; | |
| 59 | + auto [train_inputs, train_targets, val_inputs, val_targets] = | |
| 60 | + generate_dataset(train_size, val_size); | |
| 61 | + | |
| 62 | + // create a simple one hidden layer neural network | |
| 63 | + int variables = 3; // number of input variables | |
| 64 | + int neurons = 32; // number of neurons in the hidden layer | |
| 65 | + int classes = 3; // number of classes | |
| 66 | + | |
| 67 | + // input and target tensors | |
| 68 | + auto &x = *ctx.NewTensor({1, variables}); | |
| 69 | + auto &y = *ctx.NewTensor({1}, tt::kI32); | |
| 70 | + // parameters | |
| 71 | + auto &W1 = *ctx.NewTensor({neurons, variables})->RandomNorm(); | |
| 72 | + auto &b1 = *ctx.NewTensor({neurons})->RandomNorm(); | |
| 73 | + auto &W2 = *ctx.NewTensor({classes, neurons})->RandomNorm(); | |
| 74 | + auto &b2 = *ctx.NewTensor({classes})->RandomNorm(); | |
| 75 | + vector<tt::Tensor *> params = {&W1, &b1, &W2, &b2}; | |
| 76 | + // create the computation graph | |
| 77 | + auto &hidden = (x.MatMul(W1) + b1).Gelu(); | |
| 78 | + auto &logits = hidden.MatMul(W2) + b2; | |
| 79 | + auto &probs = logits.Softmax(); | |
| 80 | + auto &loss = probs.CrossEntropy(y); | |
| 81 | + | |
| 82 | + float learning_rate = 0.001; | |
| 83 | + float training_loss = 0; | |
| 84 | + | |
| 85 | + // train the model | |
| 86 | + for (int step = 0; step < 30000; step++) { | |
| 87 | + if (step % 1000 == 0) { | |
| 88 | + // evaluate the model on the validation set | |
| 89 | + float val_loss = 0; | |
| 90 | + for (int i = 0; i < val_size; i++) { | |
| 91 | + x.Fill(val_inputs[i]); | |
| 92 | + y.Fill(val_targets[i]); | |
| 93 | + loss.Forward(); | |
| 94 | + val_loss += loss.Flatten()[0]; | |
| 95 | + } | |
| 96 | + std::cout << "step " << step << ", val_loss: " << val_loss / val_size << std::endl; | |
| 97 | + } | |
| 98 | + | |
| 99 | + // pick a random training case | |
| 100 | + int casei = rand() % train_size; | |
| 101 | + x.Fill(train_inputs[casei]); | |
| 102 | + y.Fill(train_targets[casei]); | |
| 103 | + | |
| 104 | + // forward and backward | |
| 105 | + loss.Forward(); | |
| 106 | + loss.ZeroGrad(); | |
| 107 | + loss.Backward(); | |
| 108 | + | |
| 109 | + // update the parameters | |
| 110 | + for (auto param : params) { | |
| 111 | + auto weights = (float *)param->data(); | |
| 112 | + auto grads = (float *)param->grad()->data(); | |
| 113 | + for (int i = 0; i < param->NumElements(); i++) { | |
| 114 | + weights[i] -= learning_rate * grads[i]; | |
| 115 | + } | |
| 116 | + } | |
| 117 | + | |
| 118 | + training_loss = loss.Flatten()[0]; | |
| 119 | + } | |
| 120 | + | |
| 121 | + std::cout << "training loss: " << training_loss << std::endl; | |
| 122 | + | |
| 123 | + // test the model | |
| 124 | + for (int i = 0; i < 30; i++) { | |
| 125 | + auto [testx, y_truth] = generate_case(); | |
| 126 | + x.Fill(testx); | |
| 127 | + probs.Forward(); | |
| 128 | + | |
| 129 | + // find the class with the highest probability | |
| 130 | + auto probs_data = probs.Flatten(); | |
| 131 | + int y_pred = 0; | |
| 132 | + for (int j = 0; j < classes; j++) { | |
| 133 | + if (probs_data[j] > probs_data[y_pred]) { | |
| 134 | + y_pred = j; | |
| 135 | + } | |
| 136 | + } | |
| 137 | + | |
| 138 | + std::cout << "test case " << i << ", x: " << testx[0] << ", " << testx[1] << ", " | |
| 139 | + << testx[2] << ", y_truth: " << y_truth << ", y_pred: " << y_pred; | |
| 140 | + if (y_truth == y_pred) { | |
| 141 | + std::cout << " (correct)"; | |
| 142 | + } else { | |
| 143 | + std::cout << " (wrong)"; | |
| 144 | + } | |
| 145 | + std::cout << endl; | |
| 146 | + } | |
| 147 | + | |
| 148 | + return 0; | |
| 149 | +} | |