config
Package constants used throughout pydrobert.torch
If this submodule is imported first in pydrobert.torch
and any values are
changed, the changes will propagate to any submodules.
This list is non-exhaustive; types or functions may have their own defaults.
- pydrobert.torch.config.DEFT_ALI_SUBDIR = 'ali'
Default subdirectory of a torch data directory containing alignments
- pydrobert.torch.config.DEFT_CHUNK_SIZE = 1000
Default number of units to process at once when performing multiprocessing
- pydrobert.torch.config.DEFT_CTM_CHANNEL = 'A'
Default channel to write to CTM files
- pydrobert.torch.config.DEFT_DEL_COST = 1.0
Default deletion cost in error rate/distance computations
- pydrobert.torch.config.DEFT_FEAT_SUBDIR = 'feat'
Default subdirectory of a torch data directory containing features
- pydrobert.torch.config.DEFT_FILE_PREFIX = ''
Default prefix of a torch data file
- pydrobert.torch.config.DEFT_FILE_SUFFIX = '.pt'
Default suffix of a torch data file
- pydrobert.torch.config.DEFT_FLOAT_PRINT_PRECISION = 3
Default precision to write floating point values to file with
- pydrobert.torch.config.DEFT_FRAME_SHIFT_MS = 10.0
The default frame shift in milliseconds for commands
- pydrobert.torch.config.DEFT_HYP_SUBDIR = 'hyp'
Default subdirectory of a torch data directory to write hypothesis tokens to
- pydrobert.torch.config.DEFT_INS_COST = 1.0
Default insertion cost in error rate/distance computations
- pydrobert.torch.config.DEFT_NUM_WORKERS = 2
Default number of workers when performing multiprocessing
- pydrobert.torch.config.DEFT_PAD_VALUE = 0.0
Default value to pad floating-point tensors with
- pydrobert.torch.config.DEFT_PDFS_SUBDIR = 'pdfs'
Default subdirectory of a torch data directory to write pdfs to
- pydrobert.torch.config.DEFT_REF_SUBDIR = 'ref'
Default subdirectory of a torch data directory containing reference tokens
- pydrobert.torch.config.DEFT_SUB_COST = 1.0
Default substitution cost in error rate/distance computations
- pydrobert.torch.config.DEFT_TEXTGRID_SUFFIX = '.TextGrid'
The default suffix indicating TextGrid files for commands
- pydrobert.torch.config.DEFT_TEXTGRID_TIER_ID = 0
Default TextGrid tier to read transcripts from
- pydrobert.torch.config.DEFT_TEXTGRID_TIER_NAME = 'transcript'
Default TextGrid tiear to write transcripts to
- pydrobert.torch.config.EPS_0 = -2.3841860752327643e-07
A large enough value in log space that exponentiating it is very close to 1
This number is sometimes used in place of 0 in log-space values to avoid masking. Decreasing it will decrease the accuracy of computations, but may avoid NaNs.
- pydrobert.torch.config.EPS_INF = 44.361419526034176
A large enough value in log space that exponentiating it is near infinity
This number is sometimes used in place of infinity in log-space values to avoid masking. Decreasing it will decrease the accuracy of computations, but may avoid NaNs.
- pydrobert.torch.config.EPS_NINF = -43.66827237527655
A small enough value in log space that exponentiating it is very close to zero
This number is sometimes used in place of -infinity in log-space values to avoid masking. Increasing it will decrease the accuracy of computations, but may avoid NaNs.
- pydrobert.torch.config.INDEX_PAD_VALUE = -100
The value to pad index-based tensors with
Batched operations often involve variable-width input. This value is used to right-pad indexed-based tensors with to indicate that this element should be ignored.
The default value (
-100
) was chosen to coincide with the PyTorch 1.0 default forignore_index
in the likelihood losses
- pydrobert.torch.config.TINY = 1.1754943508222875e-38
Smallest representable floating-point integer
- pydrobert.torch.config.USE_JIT = False
Whether to eagerly compile functions with JIT
If
True
,pydrobert.torch
compile all functions it can with JIT on import. Otherwise, if using PyTorch >= 1.8.0, relevant items will be decorated withtorch.jit.script_if_tracing()
. The default isTrue
if and only if the environment variablePYTORCH_JIT=1
.