关于metric learning中的reader.py

[linshufei]

版本、环境信息： 1）PaddlePaddle版本：1.6 2）CPU：CPU 3）GPU：无 4）系统环境：Linux python2.7 metric learning中的reader.py源码如下：createreader中的keep_order ` def createreader(settings, mode): def metric_reader(): if mode == 'train': train_data, train_image_list = init_sop('train') loss_name = settings.loss_name if loss_name in ["softmax", "arcmargin"]: return arcmargin_iterator(train_image_list, settings)() elif loss_name == 'triplet': return triplet_iterator(train_data, settings)() else: return common_iterator(train_data, settings)() elif mode == 'val': val_data, val_image_list = init_sop('val') return image_iterator(val_image_list, 'val')() else: test_image_list = init_sop('test') return image_iterator(test_image_list, 'test')()

image_shape = settings.image_shape.split(',')
assert(image_shape[1] == image_shape[2])
image_size = int(image_shape[2])
keep_order = False if mode != 'train' or settings.loss_name in ['softmax', 'arcmargin'] else True
image_mapper = functools.partial(process_image,
        mode=mode, color_jitter=False, rotate=False, crop_size=image_size)
reader = paddle.reader.xmap_readers(
        image_mapper, metric_reader, 8, 1000, order=keep_order)
return reader                                                                              

` 问题描述： keep_order = False if mode != 'train' or settings.loss_name in ['softmax', 'arcmargin'] else True中当非训练模式置keep_order为False的作用是什么？

[kebinC]

版本、环境信息： 1）PaddlePaddle版本：1.6 2）CPU：CPU 3）GPU：无 4）系统环境：Linux python2.7 metric learning中的reader.py源码如下：createreader中的keep_order ` def createreader(settings, mode): def metric_reader(): if mode == 'train': train_data, train_image_list = init_sop('train') loss_name = settings.loss_name if loss_name in ["softmax", "arcmargin"]: return arcmargin_iterator(train_image_list, settings)() elif loss_name == 'triplet': return triplet_iterator(train_data, settings)() else: return common_iterator(train_data, settings)() elif mode == 'val': val_data, val_image_list = init_sop('val') return image_iterator(val_image_list, 'val')() else: test_image_list = init_sop('test') return image_iterator(test_image_list, 'test')()

image_shape = settings.image_shape.split(',')
assert(image_shape[1] == image_shape[2])
image_size = int(image_shape[2])
keep_order = False if mode != 'train' or settings.loss_name in ['softmax', 'arcmargin'] else True
image_mapper = functools.partial(process_image,
        mode=mode, color_jitter=False, rotate=False, crop_size=image_size)
reader = paddle.reader.xmap_readers(
        image_mapper, metric_reader, 8, 1000, order=keep_order)
return reader                                                                              

` 问题描述： keep_order = False if mode != 'train' or settings.loss_name in ['softmax', 'arcmargin'] else True中当非训练模式置keep_order为False的作用是什么？

reader里面keep_order的作用是： 设置为True, 在训triplet，quadruplet 等pair对的loss，保持一个batch里面pair对的顺序不变; 设置为False，经过reader的多线程处理后，一个batch里面的数据顺序会乱掉, 这对softmax，arcmargin以及test模型都不影响，它们都不依赖batch里数据的顺序。

[linshufei]

如果都设置为True，对训练有什么影响？

[linshufei]

版本、环境信息： 1）PaddlePaddle版本：1.6 2）CPU：CPU 3）GPU：无 4）系统环境：Linux python2.7 metric learning中的reader.py源码如下：createreader中的keep_order ` def createreader(settings, mode): def metric_reader(): if mode == 'train': train_data, train_image_list = init_sop('train') loss_name = settings.loss_name if loss_name in ["softmax", "arcmargin"]: return arcmargin_iterator(train_image_list, settings)() elif loss_name == 'triplet': return triplet_iterator(train_data, settings)() else: return common_iterator(train_data, settings)() elif mode == 'val': val_data, val_image_list = init_sop('val') return image_iterator(val_image_list, 'val')() else: test_image_list = init_sop('test') return image_iterator(test_image_list, 'test')()

image_shape = settings.image_shape.split(',')
assert(image_shape[1] == image_shape[2])
image_size = int(image_shape[2])
keep_order = False if mode != 'train' or settings.loss_name in ['softmax', 'arcmargin'] else True
image_mapper = functools.partial(process_image,
        mode=mode, color_jitter=False, rotate=False, crop_size=image_size)
reader = paddle.reader.xmap_readers(
        image_mapper, metric_reader, 8, 1000, order=keep_order)
return reader                                                                              

` 问题描述： keep_order = False if mode != 'train' or settings.loss_name in ['softmax', 'arcmargin'] else True中当非训练模式置keep_order为False的作用是什么？

reader里面keep_order的作用是： 设置为True, 在训triplet，quadruplet 等pair对的loss，保持一个batch里面pair对的顺序不变; 设置为False，经过reader的多线程处理后，一个batch里面的数据顺序会乱掉, 这对softmax，arcmargin以及test模型都不影响，它们都不依赖batch里数据的顺序。

如果都设置为True，对训练有什么影响？

[kebinC]

如果都设置为True，对训练有什么影响？

reader速度会慢点，训练效果没影响。