1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
| void MLP() { auto sym_x = Symbol::Variable("X"); auto sym_label = Symbol::Variable("label");
const int nLayers = 2; vector<int> layerSizes({512, 10}); vector<Symbol> weights(nLayers); vector<Symbol> biases(nLayers); vector<Symbol> outputs(nLayers);
for (int i = 0; i < nLayers; i++) { string istr = to_string(i); weights[i] = Symbol::Variable(string("w") + istr); biases[i] = Symbol::Variable(string("b") + istr); Symbol fc = FullyConnected(string("fc") + istr, i == 0? sym_x : outputs[i-1], weights[i], biases[i], layerSizes[i]); outputs[i] = LeakyReLU(string("act") + istr, fc, LeakyReLUActType::kLeaky); } auto sym_out = SoftmaxOutput("softmax", outputs[nLayers - 1], sym_label);
Context ctx_dev(DeviceType::kCPU, 0);
NDArray array_x(Shape(128, 28), ctx_dev, false); NDArray array_y(Shape(128), ctx_dev, false);
mx_float* aptr_x = new mx_float[128 * 28]; mx_float* aptr_y = new mx_float[128];
for (int i = 0; i < 128; i++) { for (int j = 0; j < 28; j++) { aptr_x[i * 28 + j] = i % 10 * 1.0f; } aptr_y[i] = i % 10; }
array_x.SyncCopyFromCPU(aptr_x, 128 * 28); array_x.WaitToRead(); array_y.SyncCopyFromCPU(aptr_y, 128); array_y.WaitToRead();
NDArray array_w_1(Shape(512, 28), ctx_dev, false); NDArray array_b_1(Shape(512), ctx_dev, false); NDArray array_w_2(Shape(10, 512), ctx_dev, false); NDArray array_b_2(Shape(10), ctx_dev, false);
array_w_1 = 0.5f; array_b_1 = 0.0f; array_w_2 = 0.5f; array_b_2 = 0.0f;
NDArray array_w_1_g(Shape(512, 28), ctx_dev, false); NDArray array_b_1_g(Shape(512), ctx_dev, false); NDArray array_w_2_g(Shape(10, 512), ctx_dev, false); NDArray array_b_2_g(Shape(10), ctx_dev, false);
std::vector<NDArray> in_args; in_args.push_back(array_x); in_args.push_back(array_w_1); in_args.push_back(array_b_1); in_args.push_back(array_w_2); in_args.push_back(array_b_2); in_args.push_back(array_y); std::vector<NDArray> arg_grad_store; arg_grad_store.push_back(NDArray()); arg_grad_store.push_back(array_w_1_g); arg_grad_store.push_back(array_b_1_g); arg_grad_store.push_back(array_w_2_g); arg_grad_store.push_back(array_b_2_g); arg_grad_store.push_back(NDArray()); std::vector<OpReqType> grad_req_type; grad_req_type.push_back(kNullOp); grad_req_type.push_back(kWriteTo); grad_req_type.push_back(kWriteTo); grad_req_type.push_back(kWriteTo); grad_req_type.push_back(kWriteTo); grad_req_type.push_back(kNullOp);
std::vector<NDArray> aux_states;
cout << "make the Executor" << endl;
Executor* exe = new Executor(sym_out, ctx_dev, in_args, arg_grad_store, grad_req_type, aux_states);
cout << "Training" << endl; int max_iters = 20000; mx_float learning_rate = 0.0001; for (int iter = 0; iter < max_iters; ++iter) { exe->Forward(true);
if (iter % 100 == 0) { cout << "epoch " << iter << endl; std::vector<NDArray>& out = exe->outputs; float* cptr = new float[128 * 10]; out[0].SyncCopyToCPU(cptr, 128 * 10); NDArray::WaitAll(); OutputAccuracy(cptr, aptr_y); delete[] cptr; }
exe->Backward(); for (int i = 1; i < 5; ++i) { in_args[i] -= arg_grad_store[i] * learning_rate; } NDArray::WaitAll(); }
delete exe; delete[] aptr_x; delete[] aptr_y; }
|