| 81 | |
| 82 | # Train the network for one or more epochs, validating after each epoch. |
| 83 | def learn(self, num_epochs=2): |
| 84 | # Train the network for a single epoch |
| 85 | def train(epoch): |
| 86 | self.network.train() |
| 87 | optimizer = optim.SGD(self.network.parameters(), lr=self.learning_rate, momentum=self.sgd_momentum) |
| 88 | for batch, (data, target) in enumerate(self.train_loader): |
| 89 | data, target = Variable(data), Variable(target) |
| 90 | optimizer.zero_grad() |
| 91 | output = self.network(data) |
| 92 | loss = F.nll_loss(output, target) |
| 93 | loss.backward() |
| 94 | optimizer.step() |
| 95 | if batch % self.log_interval == 0: |
| 96 | print( |
| 97 | "Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}".format( |
| 98 | epoch, |
| 99 | batch * len(data), |
| 100 | len(self.train_loader.dataset), |
| 101 | 100.0 * batch / len(self.train_loader), |
| 102 | loss.data.item(), |
| 103 | ) |
| 104 | ) |
| 105 | |
| 106 | # Test the network |
| 107 | def test(epoch): |
| 108 | self.network.eval() |
| 109 | test_loss = 0 |
| 110 | correct = 0 |
| 111 | for data, target in self.test_loader: |
| 112 | with torch.no_grad(): |
| 113 | data, target = Variable(data), Variable(target) |
| 114 | output = self.network(data) |
| 115 | test_loss += F.nll_loss(output, target).data.item() |
| 116 | pred = output.data.max(1)[1] |
| 117 | correct += pred.eq(target.data).cpu().sum() |
| 118 | test_loss /= len(self.test_loader) |
| 119 | print( |
| 120 | "\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n".format( |
| 121 | test_loss, correct, len(self.test_loader.dataset), 100.0 * correct / len(self.test_loader.dataset) |
| 122 | ) |
| 123 | ) |
| 124 | |
| 125 | for e in range(num_epochs): |
| 126 | train(e + 1) |
| 127 | test(e + 1) |
| 128 | |
| 129 | def get_weights(self): |
| 130 | return self.network.state_dict() |