Refine SE-ResNeXt model and use ParallelExecutor. #816
Conversation
fluid/image_classification/train.py Outdated
| momentum=0.9, | ||
| regularization=fluid.regularizer.L2Decay(1e-4)) | ||
| | ||
| opts = optimizer.minimize(avg_cost) |
There was a problem hiding this comment.
opts = optimizer.minimize(avg_cost) should be placed in line 84. Because minimize will add the backward ops.
| train_loss = np.array(train_info[0]).mean() | ||
| train_acc1 = np.array(train_info[1]).mean() | ||
| train_acc5 = np.array(train_info[2]).mean() | ||
| for data in test_reader(): |
There was a problem hiding this comment.
Testing phase can be written as a function. Just advise, this is my own habit.
fluid/image_classification/train.py Outdated
| regularization=fluid.regularizer.L2Decay(1e-4)) | ||
| | ||
| opts = optimizer.minimize(avg_cost) | ||
| fluid.memory_optimize(fluid.default_main_program()) |
There was a problem hiding this comment.
test_program's memory also need to be optimized.
There was a problem hiding this comment.
Done. But the memory_optimize has never been tested in ParallelExecutor.
fluid/image_classification/train.py Outdated
| opts = optimizer.minimize(avg_cost) | ||
| fluid.memory_optimize(fluid.default_main_program()) | ||
| | ||
| inference_program = fluid.default_main_program().clone() |
There was a problem hiding this comment.
inference_program should only contain the forward ops. So line 80 should be placed in line62.
There was a problem hiding this comment.
Done. But the original code is also right.
| loss, acc1, acc5 = exe.run( | ||
| fluid.default_main_program(), | ||
| feed=feeder.feed(data), | ||
| fetch_list=[avg_cost, acc_top1, acc_top5]) |
There was a problem hiding this comment.
fetching operation is time-consuming, if possible, we can fetch data every 10 iterations.
2 participants
No description provided.