wangwei created SINGA-205:
-----------------------------
Summary: Enable slice and concatenate operations for Tensor objects
Key: SINGA-205
URL: https://issues.apache.org/jira/browse/SINGA-205
Project: Singa
Issue Type: New Feature
Reporter: wangwei
Assignee: wangwei
Slice and concatenate operations are typical array operations, and are
necessary to support some models, e.g., GoogleLeNet.
For this ticket, we are going to implement the two operations to support 1D and
2D tensors. The following functions would be implemented,
1. SliceRows(Tensor, start, end), the returned tensor shares the memory with
the origin tensor.
2. CopyRows(Tensor, start, end), CopyColumns(Tensor, start, end) the returned
tensor copies values from the original tensor
3. ConcatenateRows(vector<Tensor>) and ConcatenateColumns(vector<Tensor>); the
returned tensor copies values from the origin tensors.
To support SliceRows, i.e., memory sharing between two tensors of different
sizes, we need to update the Block class as
{code}
class Block {
Block(void* ptr, size_t size, size_t offset=0);
void* mutable_data() const { return data_+offset_;}
const void* data() const { return data_+offset_;}
private:
void *data_;
size_t offset_ = 0; // offset in terms of Bytes.
shared_ptr<atomic<int>> ref_count_;
};
{code}
The sub-tensors (i.e., generated from SliceRows) would have a different Block
instance as the original tensor, but would share the internal `ref_count_` and
`data_` fields.
`offset_` is 0 for the original tensors, but non-zero for sub-tensors . atomic
makes the inc and dec operations thread-safe. shared_ptr helps to free
`ref_count_` safely.
SliceColumns is not supported as singa::Tensor uses row major storage which is
difficult to support SliceColumns SliceRows(). Users can use CopyColumns().
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
