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 | } | ||