| 169 | |
| 170 | @unittest.skipIf(not singa_api.USE_CUDA, 'CUDA is not enabled') |
| 171 | def test_rnn_with_seq_lengths(self, dev=gpu_dev): |
| 172 | bs = 2 |
| 173 | seq_length = 3 |
| 174 | hidden_size = 2 |
| 175 | em_size = 2 |
| 176 | x_np = np.array([[[0.1, 0.1], [0.2, 0.2], [0.3, 0.3]], |
| 177 | [[0.3, 0.3], [0.4, 0.4], [0.0, |
| 178 | 0.0]]]).astype(np.float32) |
| 179 | y_np = np.array([[0.4, 0.4], [0.5, 0.5]]).astype(np.float32) |
| 180 | seq_lengths_np = np.array([3, 2]).astype(np.int32) |
| 181 | |
| 182 | x = tensor.from_numpy(x_np) |
| 183 | x.to_device(dev) |
| 184 | y = tensor.from_numpy(y_np) |
| 185 | y.to_device(dev) |
| 186 | seq_lengths = tensor.from_numpy(seq_lengths_np) |
| 187 | |
| 188 | m = LSTMModel3(hidden_size) |
| 189 | m.compile([x, seq_lengths], |
| 190 | is_train=True, |
| 191 | use_graph=False, |
| 192 | sequential=False) |
| 193 | m.train() |
| 194 | for i in range(10): |
| 195 | out = m.forward(x, seq_lengths) |
| 196 | loss = autograd.mse_loss(out, y) |
| 197 | print("train l:", tensor.to_numpy(loss)) |
| 198 | m.optimizer(loss) |
| 199 | m.eval() |
| 200 | out = m.forward(x, seq_lengths) |
| 201 | loss = autograd.mse_loss(out, y) |
| 202 | print(" eval l:", tensor.to_numpy(loss)) |
| 203 | |
| 204 | @unittest.skipIf(not singa_api.USE_CUDA, 'CUDA is not enabled') |
| 205 | def test_lstm_model(self, dev=gpu_dev): |