Debug options to set up a given QuantizationDebugger.
tf
.
lite
.
experimental
.
QuantizationDebugOptions
(
layer_debug_metrics
:
Optional
[
Mapping
[
str
,
Callable
[[
np
.
ndarray
],
float
]]]
=
None
,
model_debug_metrics
:
Optional
[
Mapping
[
str
,
Callable
[[
Sequence
[
np
.
ndarray
],
Sequence
[
np
.
ndarray
]],
float
]]]
=
None
,
layer_direct_compare_metrics
:
Optional
[
Mapping
[
str
,
Callable
[[
Sequence
[
np
.
ndarray
],
Sequence
[
np
.
ndarray
],
float
,
int
],
float
]]]
=
None
,
denylisted_ops
:
Optional
[
List
[
str
]]
=
None
,
denylisted_nodes
:
Optional
[
List
[
str
]]
=
None
,
fully_quantize
:
bool
=
False
)
->
None
Used in the notebooks
Used in the tutorials
Args
a dict to specify layer debug functions
{function_name_str: function} where the function accepts result of
NumericVerify Op, which is value difference between float and
dequantized op results. The function returns single scalar value.
a dict to specify model debug functions
{function_name_str: function} where the function accepts outputs from
two models, and returns single scalar value for a metric. (e.g.
accuracy, IoU)
a dict to specify layer debug functions
{function_name_str: function}. The signature is different from that of
layer_debug_metrics
, and this one gets passed (original float value,
original quantized value, scale, zero point). The function's
implementation is responsible for correctly dequantize the quantized
value to compare. Use this one when comparing diff is not enough.
(Note) quantized value is passed as int8, so cast to int32 is needed.
