Blob is the fundamental data representation in Mocha. It is used as both data (e.g. mini-batch of data samples) and parameters (e.g. filters of a convolution layer). Conceptually, a blob is a N-dimensional tensor.
For example, in vision, a data blob is usually a 4D-tensor. Following the vision (and Caffe) convention, the four dimensions are called width, height, channels and num. The fastest changing dimension is width and slowest changing dimension is num.
The memory layout of a blob in Mocha is compatible with Caffe’s blob. So a blob (e.g. layer parameters) in Mocha can be saved to HDF5 and load it from Caffe without doing any dimension permutation, and vise versa. However, since Julia use column-major convention for tensor and matrix, and Caffe use row-major convention, in Mocha API, the order of the four dimensions is width, height, channels, and num, while in Caffe API, it is num, channels, height, width.
Each backend has its own blob implementation, as a subtype of Blob. For example, a blob in the CPU backend is a shallow wrapper of a Julia Array object, while a blob in the GPU backend references to a piece of GPU memory.
Constructors and Destructors¶
A backend-dependent blob can be created with the following function:
- make_blob(backend, data_type, dims)¶
dims is a NTuple, specifying the dimensions of the blob to be created. Currently data_type should be either Float32 or Float64.
Several helper functions are also provided:
- make_blob(backend, data_type, dims...)
Spell out the dimensions explicitly.
- make_blob(backend, array)
array is a Julia AbstractArray. This makes a blob with the same data type and shape as array and initialize the blob contents with array.
- make_zero_blob(backend, data_type, dims)¶
Create a blob and initialize with zeros.
- reshape_blob(backend, blob, new_dims)¶
Create a reference to an existing blob with a possiblely different shape. The behavior is the same as Julia’s reshape function on an array: the new blob shares data with the existing one.
The resources of a blob could be released by calling
Note the resources need to be released explicitly. A Julia blob object being GC-ed does not release the underlying resource automatically.
Accessing Properties of a Blob¶
The blob implements some simple API for a Julia array:
Get the element type of the blob.
Get the tensor dimension of the blob. The same as length(size(blob)).
Get the shape of the blob. The return value is a NTuple.
- size(blob, dim)
Get the size at a particular dimension. dim could be negative. For example, size(blob, -1) is the same as size(blob)[end]. For convenience, if dim exceeds ndims(blob), the function returns 1 instead of firing an error.
Get the total number of elements in a blob.
The same as size(blob, 1).
The same as size(blob, 2).
The same as size(blob, -1).
The the feature size in a blob, which is the same as prod(size(blob)[1:end-1]).
The wrappers get_chann is removed from v0.0.5 when Mocha upgrade from 4D-tensor to general ND-tensor, because the channel dimension is usually ambiguous for general ND-tensors.
Accessing Data of a Blob¶
Because accessing GPU memory is costly, a blob does not has interface to do element-wise accessing. The data could either be manipulated in a backend-dependent manner, relying on the underlying implementation details; or in a backend-independent way by copying the contents back and to a Julia array.
- copy!(dst, src)¶
Copy the contents of src to dst. src and dst could be either a blob or a Julia array.
The following utilities could be used to initialize the contents of a blob
- fill!(blob, value)¶
Fill every element of blob with value.
Fill blob with zeros. Depending on the implementation, erase!(blob) might be more efficient than fill!(blob, 0).