mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-11 06:55:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

Changes that should affect nothing: bug fixes etc.

Browse Source
This commit is contained in:
Daniel Povey 2023-06-18 04:00:43 +08:00
parent 70bd58c648
commit b656b0df36
2 changed files with 22 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
egs/librispeech/ASR/pruned_transducer_stateless7/scaling.py
View File

@ -296,7 +296,7 @@ class SoftmaxFunction(torch.autograd.Function):
dim = ctx.dim dim = ctx.dim
if dim < 0: if dim < 0:
dim = dim + ans.dim dim = dim + ans.ndim
split_dim = 0 if dim != 0 else 1 split_dim = 0 if dim != 0 else 1
# split_dim is the dimension we split up ans on. # split_dim is the dimension we split up ans on.
num_split = min(8, ans.shape[split_dim]) num_split = min(8, ans.shape[split_dim])

26
icefall/diagnostics.py
View File

@ -498,6 +498,22 @@ class ModelDiagnostic(object):
self.diagnostics[k].print_diagnostics() self.diagnostics[k].print_diagnostics()
def get_class_name(module: nn.Module):
ans = type(module).__name__
# we put the below in try blocks in case anyone is using a different version of these modules that
# might have different member names.
if ans == 'Balancer' or ans == 'ActivationBalancer':
try:
ans += f'[{module.min_positive},{module.max_positive},{module.min_abs},{module.max_abs}]'
except:
pass
elif ans == 'AbsValuePenalizer':
try:
ans += f'[{module.limit}]'
except:
pass
return ans
def attach_diagnostics( def attach_diagnostics(
model: nn.Module, opts: Optional[TensorDiagnosticOptions] = None model: nn.Module, opts: Optional[TensorDiagnosticOptions] = None
) -> ModelDiagnostic: ) -> ModelDiagnostic:
@ -537,12 +553,12 @@ def attach_diagnostics(
if isinstance(_output, Tensor) and _output.dtype in ( torch.float32, torch.float16, torch.float64 ): if isinstance(_output, Tensor) and _output.dtype in ( torch.float32, torch.float16, torch.float64 ):
_model_diagnostic[f"{_name}.output"].accumulate(_output, _model_diagnostic[f"{_name}.output"].accumulate(_output,
class_name=type(_module).__name__) class_name=get_class_name(_module))
elif isinstance(_output, tuple): elif isinstance(_output, tuple):
for i, o in enumerate(_output): for i, o in enumerate(_output):
if o.dtype in ( torch.float32, torch.float16, torch.float64 ): if o.dtype in ( torch.float32, torch.float16, torch.float64 ):
_model_diagnostic[f"{_name}.output[{i}]"].accumulate(o, _model_diagnostic[f"{_name}.output[{i}]"].accumulate(o,
class_name=type(_module).__name__) class_name=get_class_name(_module))
def backward_hook( def backward_hook(
_module, _input, _output, _model_diagnostic=ans, _name=name _module, _input, _output, _model_diagnostic=ans, _name=name
@ -551,12 +567,12 @@ def attach_diagnostics(
_output = _output[0] _output = _output[0]
if isinstance(_output, Tensor) and _output.dtype in ( torch.float32, torch.float16, torch.float64 ): if isinstance(_output, Tensor) and _output.dtype in ( torch.float32, torch.float16, torch.float64 ):
_model_diagnostic[f"{_name}.grad"].accumulate(_output, _model_diagnostic[f"{_name}.grad"].accumulate(_output,
class_name=type(_module).__name__) class_name=get_class_name(_module))
elif isinstance(_output, tuple): elif isinstance(_output, tuple):
for i, o in enumerate(_output): for i, o in enumerate(_output):
if o.dtype in ( torch.float32, torch.float16, torch.float64 ): if o.dtype in ( torch.float32, torch.float16, torch.float64 ):
_model_diagnostic[f"{_name}.grad[{i}]"].accumulate(o, _model_diagnostic[f"{_name}.grad[{i}]"].accumulate(o,
class_name=type(_module).__name__) class_name=get_class_name(_module))
module.register_forward_hook(forward_hook) module.register_forward_hook(forward_hook)
@ -574,7 +590,7 @@ def attach_diagnostics(
_input, = _input _input, = _input
assert isinstance(_input, Tensor) assert isinstance(_input, Tensor)
_model_diagnostic[f"{_name}.scalar"].accumulate_input(_input, _model_diagnostic[f"{_name}.scalar"].accumulate_input(_input,
class_name=type(_module).__name__) class_name=get_class_name(_module))
def scalar_backward_hook( def scalar_backward_hook(
_module, _input, _output, _model_diagnostic=ans, _name=name _module, _input, _output, _model_diagnostic=ans, _name=name