Preprocessing data using a network interface
Methods and systems for obtaining data having a first format, converting the data to a second format, storing the converted data in memory accessible by at least one parallel processing unit, and processing the converted data stored in the memory using the at least one parallel processing unit.
1 . A system comprising:
first memory;
at least one graphics processing unit (GPU);
a network interface to obtain input data having a first format, the network interface to convert the input data to obtain converted input data having a multi-dimensional second format compatible with the at least one GPU, and store the converted input data directly to the first memory, the at least one GPU to be separate from the network interface, to access the converted input data stored in the first memory by the network interface, and to perform at least one operation on the converted input data; and
one or more central processing units that are different from the at least one GPU and the network interface.
2 . The system of claim 1 , wherein the at least one operation comprises at least one transformation operation.
3 . The system of claim 2 , wherein the at least one transformation operation comprises at least one of normalizing intensity values, normalizing spatial dimensions, scaling pixel size from non-isotropic to isotropic, or performing histogram normalization.
4 . The system of claim 1 , wherein the at least one operation comprises one or more inference operations.
5 . The system of claim 1 , further comprising:
an imaging device to capture imaging data and transmit the imaging data to the network interface, the network interface obtaining the input data from the imaging data.
6 . The system of claim 5 , wherein the imaging device is a medical imaging device and the first format is specified by a Digital Imaging and Communications in Medicine (“DICOM”) standard.
7 . The system of claim 6 , wherein the at least one operation comprises one or more inference operations to obtain a diagnostic result.
8 . The system of claim 1 , further comprising:
second memory, the network interface to obtain the input data by loading the input data from the second memory.
9 . The system of claim 1 , wherein the network interface comprises one or more data processing units to convert the input data,
the one or more central processing units are different from the one or more data processing units, and
the one or more central processing units are not to process the input data or the converted input data before the at least one GPU performs the at least one operation on the converted input data.
10 . The system of claim 1 , wherein converting the input data to obtain the converted input data comprises performing an alignment transformation operation on the input data.
11 . A method comprising:
obtaining input data having a first format;
converting with at least one network interface the input data to obtain converted input data having a multi-dimensional second format compatible with at least one graphics processing unit (GPU), the multi-dimensional second format being different from the first format;
using the at least one network interface to store the converted input data directly to memory accessible by the at least one GPU that is separate from the at least one network interface and one or more central processing units that are different from the network interface; and
processing the converted input data stored in the memory using the at least one GPU.
12 . The method of claim 11 , wherein processing the converted input data comprises performing at least one transformation operation on the converted input data to produce transformed information, and performing at least one inference operation on the transformed information.
13 . The method of claim 12 , wherein the at least one transformation operation comprises at least one of normalizing intensity values, normalizing spatial dimensions, scaling pixel size from non-isotropic to isotropic, or performing histogram normalization.
14 . The method of claim 12 , wherein the input data comprises imaging data.
15 . The method of claim 14 , wherein the imaging data comprises a sinogram.
16 . The method of claim 14 , wherein the imaging data is medical data having a format specified by a Digital Imaging and Communications in Medicine (“DICOM”) standard.
17 . The method of claim 11 , wherein the memory is a first memory,
obtaining the input data comprises loading the input data from a second memory, and
the first memory is different from the second memory.
18 . The method of claim 11 , wherein obtaining the input data comprises at least one of receiving the input data from an Ethernet network, receiving the input data from an Infiniband network, or reading the input data from an emulated storage device.
19 . The method of claim 11 , wherein processing the converted input data comprises performing at least one transformation operation on the converted input data to produce transformed information, constructing a volume from the transformed information, and performing at least one inference operation on the volume.
20 . A network interface comprising:
at least one processor comprising one or more circuits to:
obtain input data having a first format;
convert the input data by the one or more circuits of the at least one processor of the network interface into a multi-dimensional second format compatible with at least one graphics processing unit (GPU) that is different from one or more central processing units, which are different from the network interface; and
store the converted input data in a location accessible by the at least one GPU that is separate from the network interface and operable to perform at least one operation on the converted input data.
21 . The network interface of claim 20 , wherein the one or more circuits are to obtain the input data from an imaging device operable to capture imaging data and transmit the imaging data to the network interface.
22 . The network interface of claim 21 , wherein the imaging device is medical imaging device and the first format is specified by a Digital Imaging and Communications in Medicine (“DICOM”) standard.
23 . The network interface of claim 20 , wherein the one or more circuits are to obtain the input data by reading the input data from a memory.
24 . The network interface of claim 20 , wherein converting the input data comprises performing an alignment transformation operation on the input data.
25 . The network interface of claim 20 , wherein the multi-dimensional second format is a tensor format, a TensorFlow tensor format, a PyTorch tensor format, or a CuPy array format.
26 . The system of claim 1 , wherein the one or more central processing units are not to process the input data or the converted input data before the at least one GPU performs the at least one operation on the converted input data.
27 . The system of claim 1 , wherein the multi-dimensional second format is a tensor format, a TensorFlow tensor format, a PyTorch tensor format, or a CuPy array format.
28 . The method of claim 11 , wherein the multi-dimensional second format is a tensor format, a TensorFlow tensor format, a PyTorch tensor format, or a CuPy array format.