4.8 KiB

Исходник Постоянная ссылка Ответственный История

Multiverso Torch Binding API

init(sync)

Initialize mutliverso.

This should be called only once before training at the beginning of the whole project.

If sync is true, a sync server will be created. Otherwise an async server will be created.

If a sync server is created, you must make sure every process call add and get in the same order and for the same times. Otherwise some processes will be blocked. In sync server mode, all get method will return exactly the same results. If a async server is created, there won't be limitations like a sync server. But we can't make sure get method will return the same results. If you want to get the same results in async server mode, you should use barrier and get with the argument sync set to true to sync the processes.

barrier()

Set a barrier for all workers to wait.

Workers will wait until all workers reach a specific barrier.

shutdown()

Shutdown multiverso.

This should be called only once after finishing training at the end of the whole project.

num_workers()

Return the total number of workers.

worker_id()

Return the id (zero-based index) for current worker.

TableHandler

TableHandler is an interface to sync different kinds of values.

In most cases, you are supposed to sync models (for initialization) and gradients (during training) so as to let multiverso help you manage the models in distributed environments. Currently, two types of TableHandler are supported, namely ArrayTableHandler and MatrixTableHandler.

ArrayTableHandler

ArrayTableHandler is used to sync array-like (one-dimensional) value.

Although the model tends to be a matrix, when using torch.nn package we can get the flattened parameters and gradients with module.getParameters(). So in most cases, we should use ArrayTableHandler instead of MatrixTableHandler we will introduce soon.

ArrayTableHandler:new(size)

Create a ArrayTableHandler for syncing array-like (one-dimensional) value.

The size should be a number equal to the size of value we want to sync.

If init_value is nil, zeros will be used to initialize the table, otherwise the table will be initialized as the init_value. Notice: Only the init_value from the master will be used!

ArrayTableHandler:add(data, sync)

Add a array-like (one-dimensional) data to the server.

The data should be a torch.Tensor or Lua table. During training process, the data should be the gradients (delta value). The size of data must be equal to the size specified in initialization.

sync should be a boolean value. The default value is false. If sync is true, this call will blocked by IO until the call finish. Otherwise it will return immediately

ArrayTableHandler:get()

Get the array-like (one-dimensional) value from the server.

The value we get will be a torch.Tensor. Usually, we are supposed to use Tensor:copy() to assign the value to desired destination.

MatrixTableHandler

MatrixTableHandler is used to sync matrix-like (two-dimensional) value.

MatrixTableHandler:New(num_row, num_col, init_value)

Create a MatrixTableHandler for syncing matrix-like (two-dimensional) value.

The num_row should be the number of rows and the num_col should be the number of columns. Both of them should be a number equal to the exact size of value we want to sync.

If init_value is nil, zeros will be used to initialize the table, otherwise the table will be initialized as the init_value. Notice: if the init_value is different in different processes, the average of them will be used.

MatrixTableHandler:add(data, row_ids, sync)

Add a matrix-like (two-dimensional) data to the server.

Same as the clarification in ArrayTableHandler, the data should be a torch.Tensor or Lua table and we should pass the gradients (delta value) not the exact value to it. The row_ids is an optional parameter and it should be an array of 'row_id' numbers when specified. If specified, multiverso will only update the value in specific rows and the size of data should be equal to the size of value we want to update.