Skip to content

README.md

Latest commit

 

History

History
40 lines (23 loc) · 2.4 KB

README.md

File metadata and controls

40 lines (23 loc) · 2.4 KB

PULP-TrainLib - Training Library sources

PULP-TrainLib is organized as follows:

  • include/: folder containing the header files of the library (together with all of the definitions of the structures and the function primitives)
  • sources/: folder contatining all the implementations

To use PULP-TrainLib in your project, just include pulp_train.h.

Structure definitions and data types

PULP-TrainLib is designed to work with different data types. For each data format, a set of source files is designed (for example, pulp_xxx_fp32.h refers to a header file with functions and data in fp32).

Many of the functions inside PULP-TrainLib require specific data structures to work, passing the operands implicitly (as void* args) for compatibility with PULP PMSIS. These data structures, for each data type, are defined inside pulp_train_utils_fpxx.h.

Training primitives

With the same naming convention, each DNN layer has its own Forward and Backward functions, for each of the available data types. E.g: for a CONV2D layer in fp16, its training primitives are defined inside pulp_conv2d_fp16.h and pulp_conv2d_fp16.c. Inside each of these files, you can find both the forward and the backward functions. The same holds for the activation functions.

Define the fp16 format

The fp16 format refers to PULP native data formats. To select the format you need, please refer to pulp_train_defines.h, in which the format to be used is defined (float16 refers to the FP16 Sign-Exponent-Mantissa 1-5-10 format, while float16alt to the Bfloat16 1-8-7 format).

Other general defines

pulp_train_defines.h contains different kinds of definitions, as the available macros, the data type definitions etc.

Building a DNN training workload

To build the code to run a DNN training on a PULP-based platform with PULP-TrainLib, you need to:

  • define the arrays which will contain the data of the tensors
  • define the sizes of each tensor
  • define a blob structure to wrap the tensor and link the data array and the tensor sizes to it
  • define the training algorithm (loss, optimizer, etc)
  • run the computation on the PULP cluster, since each of the training functions is able to parallelize the execution on the NUM_CORES parameter.

Please refer to the tests/ folder for examples.

To speed up the design process, you can make use of the TrainLib Deployer.