image_classification cifar-10 train (use sparse trainning)

[shenhuinuist]

When execute sparse training, we need to set sparse_update=True in network config. Where should I set the command when I train image_classification cifar-10?

[qingqing01]

@shenhuinuist the sparse_update is an attribute of ParamAttr. For example

 fc = fc_layer(input=data, size=128, param_attr=ParamAttr(sparse_update=True))

The sparse training is usually used to accelerate calculation when input is sparse data with highly dimension.

But the data for image classification of cifar-10 is dense type, sparse_update is not applicable to this case.

[shenhuinuist]

@qingqing01 Thank you so much . Now , I set sparse_update =True and sparse_remote_update = False in paddle.trainer_config_helpers.attr.py , the input data was still image classification of cifar-10, the error was as follows: TypeError: init( ) got an un expected keyword argument ‘sparse_update’. I have two questions. Where does paddle execute the command ? What's more, How does paddle judge the input data is sparse or not ?

[qingqing01]

@shenhuinuist I have mentioned above, the cifar-10 data is not sparse input. So the sparse_update is not applicable to this case. We will provide a demo to show how to use sparse training later.

The data type is defined in this line in the data provider file. The data provider defines the types of the input data and provides training or testing data to PaddlePaddle.

Thanks. If there is any question, we can continue to discuss :)

[shenhuinuist]

@qingqing01 Thank you!