Python CuDevice Examples

Example #1

 def run_tests(self):
     from kaldi.cudamatrix import cuda_available
     if cuda_available():
         from kaldi.cudamatrix import CuDevice
         CuDevice.instantiate().set_debug_stride_mode(True)
         CuDevice.instantiate().select_gpu_id("yes")
         super(test_cuda, self).run_tests()
         CuDevice.instantiate().print_profile()
     else:
         print("CUDA not available. Running tests on CPU.")
         super(test_cuda, self).run_tests()

Example #2

File: cu-vector-test.py Project: ruchirtravadi/pykaldi

    def testCuVectorGetItem(self):
        v = CuVector()
        with self.assertRaises(IndexError):
            v[0]

        v = CuVector.new([3, 5, 7, 11, 13])
        self.assertAlmostEqual(3.0, v[0])
        self.assertAlmostEqual(7.0, v[2])
        self.assertAlmostEqual(13.0, v[4])

        with self.assertRaises(IndexError):
            v[5]


if __name__ == '__main__':
    if cuda_available():
        from kaldi.cudamatrix import CuDevice

        for i in range(2):
            CuDevice.instantiate().set_debug_stride_mode(True)
            if i == 0:
                CuDevice.instantiate().select_gpu_id("no")
            else:
                CuDevice.instantiate().select_gpu_id("yes")

            unittest.main()

            CuDevice.instantiate().print_profile()
    else:
        unittest.main()

Example #3

File: nnet3-batch-recognizer.py Project: prismdata/kor_asr

#!/usr/bin/env python

from __future__ import print_function

from kaldi.asr import NnetLatticeFasterBatchRecognizer
from kaldi.decoder import LatticeFasterDecoderOptions
from kaldi.nnet3 import NnetBatchComputerOptions
from kaldi.util.table import SequentialMatrixReader

from kaldi.cudamatrix import cuda_available
if cuda_available():
    from kaldi.cudamatrix import CuDevice
    CuDevice.instantiate().select_gpu_id('yes')
    CuDevice.instantiate().allow_multithreading()

# Construct recognizer
decoder_opts = LatticeFasterDecoderOptions()
decoder_opts.beam = 13
decoder_opts.max_active = 7000
compute_opts = NnetBatchComputerOptions()
compute_opts.acoustic_scale = 1.0
compute_opts.frame_subsampling_factor = 3
compute_opts.frames_per_chunk = 150
asr = NnetLatticeFasterBatchRecognizer.from_files(
    "final.mdl", "HCLG.fst", "words.txt",
    decoder_opts=decoder_opts, compute_opts=compute_opts, num_threads=4)

# Define feature pipelines as Kaldi rspecifiers
feats_rspec = "ark:compute-mfcc-feats --config=mfcc.conf scp:wav.scp ark:- |"
ivectors_rspec = (
    "ark:compute-mfcc-feats --config=mfcc.conf scp:wav.scp ark:-"

Example #4

            matrices[matrix].resize(0, 0)

    gflops = num_floats_processed / (tim.elapsed() * 1.0e9)
    print(
        "For CuMatrix::Reize float, for size multiple {}, speed was {} gigaflops"
        .format(size_multiple, gflops))


def testCudaMatrixResize():
    sizes = [1, 2, 4, 8, 16, 32]
    for s in sizes:
        aux(s)


if __name__ == '__main__':
    if cuda_available():
        from kaldi.cudamatrix import CuDevice

        for loop in range(2):
            CuDevice.Instantiate().SetDebugStrideMode(True)
            if loop == 0:
                CuDevice.Instantiate().SelectGpuId("no")
            else:
                CuDevice.Instantiate().SelectGpuId("yes")

            testCudaMatrixResize()

            CuDevice.Instantiate().PrintProfile()

    else:
        testCudaMatrixResize()