[jira] [Updated] (SINGA-505) Buffer Operators / Change the Autograd operators to be bufferable

["Chris Yeung (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/SINGA-505?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Chris Yeung updated SINGA-505:
------------------------------
    Description: 
We can buffer the operators, so that we can extract all the operators in autograd to build a graph after scheduling, where the simplest scheduling can use the FIFO principle from the buffered operators. A more complex scheduleing algorithm could be implemented which consider the dependency of operators that could make it parallel. One more clear advantage is that when we run the graph we only need to run the buffered operators, then there will be no need to run the autograd python code again throughout the training iterations.

So this ticket uses for two purpose:

1. Change the core components (e.g. tensor,device) to support buffering.

2. Change all the autograd operator to be bufferable, i.e. the input and output should be inside the block. For example, the SoftMax backward cannot be buffered because it is not doing the operations through the block, and it was using numpy:
    def backward(self, dy):
        # calculations are made on numpy array
        if self.axis == 1:
            dy = singa.DefaultTranspose(dy)
        grad = ctensor2numpy(dy)
        output = ctensor2numpy(self.output)
        out_1 = np.einsum("ki,ki->ki", grad, output)
        medium_out = np.einsum("ki,kj->kij", output, output)
        out_2 = np.einsum("kij,kj->ki", medium_out, grad)
        out = out_1 - out_2
        dx = CTensor(out_1.shape)
        dx.CopyFloatDataFromHostPtr(out.flatten())
        if self.axis == 0:
            return dx
        elif self.axis == 1:
            return singa.DefaultTranspose(dx)

  was:
We can buffer the operators, so that we can extract all the operators in autograd to build a graph after schedule, where the simplest scheduling can use the FIFO principle from the buffered operators. A more complex scheduleing algorithm could be implemented which consider the dependency of operators that could make it parallel. One more clear advantage is that when we run the graph we only need to run the buffered operators called by the autograd function, then there will be no need to run the autograd python code again throughout the training process.

So this ticket uses for two purpose:

1. Change the core components (e.g. tensor,device) to support buffering.

2. Change all the autograd operator to be bufferable, i.e. the input and output should be inside the block. For example, the SoftMax backward cannot be buffered because it is not doing the operating through the block, and it was using numpy:
    def backward(self, dy):
        # calculations are made on numpy array
        if self.axis == 1:
            dy = singa.DefaultTranspose(dy)
        grad = ctensor2numpy(dy)
        output = ctensor2numpy(self.output)
        out_1 = np.einsum("ki,ki->ki", grad, output)
        medium_out = np.einsum("ki,kj->kij", output, output)
        out_2 = np.einsum("kij,kj->ki", medium_out, grad)
        out = out_1 - out_2
        dx = CTensor(out_1.shape)
        dx.CopyFloatDataFromHostPtr(out.flatten())
        if self.axis == 0:
            return dx
        elif self.axis == 1:
            return singa.DefaultTranspose(dx)


> Buffer Operators / Change the Autograd operators to be bufferable
> -----------------------------------------------------------------
>
>                 Key: SINGA-505
>                 URL: https://issues.apache.org/jira/browse/SINGA-505
>             Project: Singa
>          Issue Type: Improvement
>          Components: Core
>            Reporter: Chris Yeung
>            Priority: Major
>
> We can buffer the operators, so that we can extract all the operators in autograd to build a graph after scheduling, where the simplest scheduling can use the FIFO principle from the buffered operators. A more complex scheduleing algorithm could be implemented which consider the dependency of operators that could make it parallel. One more clear advantage is that when we run the graph we only need to run the buffered operators, then there will be no need to run the autograd python code again throughout the training iterations.
> So this ticket uses for two purpose:
> 1. Change the core components (e.g. tensor,device) to support buffering.
> 2. Change all the autograd operator to be bufferable, i.e. the input and output should be inside the block. For example, the SoftMax backward cannot be buffered because it is not doing the operations through the block, and it was using numpy:
>     def backward(self, dy):
>         # calculations are made on numpy array
>         if self.axis == 1:
>             dy = singa.DefaultTranspose(dy)
>         grad = ctensor2numpy(dy)
>         output = ctensor2numpy(self.output)
>         out_1 = np.einsum("ki,ki->ki", grad, output)
>         medium_out = np.einsum("ki,kj->kij", output, output)
>         out_2 = np.einsum("kij,kj->ki", medium_out, grad)
>         out = out_1 - out_2
>         dx = CTensor(out_1.shape)
>         dx.CopyFloatDataFromHostPtr(out.flatten())
>         if self.axis == 0:
>             return dx
>         elif self.axis == 1:
>             return singa.DefaultTranspose(dx)



--
This message was sent by Atlassian Jira
(v8.3.4#803005)