Skip to content

Quantization tool: Use nanmin, nanmax, nanmean in calibrator #23749

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jambayk merged 2 commits into from
Feb 19, 2025
Merged

Quantization tool: Use nanmin, nanmax, nanmean in calibrator #23749

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a4d69b4
use nan min/max/mean
jambayk Feb 19, 2025
d240abc
scale issue typo
jambayk Feb 19, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 10 additions & 10 deletions onnxruntime/python/tools/quantization/calibrate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -496,14 +496,14 @@ def compute_data(self) -> TensorsData:
pairs = []
for i in range(0, len(added_output_names), 2):
if self.moving_average:
min_value_array = np.mean(merged_added_output_dict[added_output_names[i]], axis=0)
max_value_array = np.mean(merged_added_output_dict[added_output_names[i + 1]], axis=0)
min_value_array = np.nanmean(merged_added_output_dict[added_output_names[i]], axis=0)
max_value_array = np.nanmean(merged_added_output_dict[added_output_names[i + 1]], axis=0)
else:
min_value_array = np.min(merged_added_output_dict[added_output_names[i]], axis=0)
max_value_array = np.max(merged_added_output_dict[added_output_names[i + 1]], axis=0)
min_value_array = np.nanmin(merged_added_output_dict[added_output_names[i]], axis=0)
max_value_array = np.nanmax(merged_added_output_dict[added_output_names[i + 1]], axis=0)

if self.symmetric:
max_absolute_value = np.max([np.abs(min_value_array), np.abs(max_value_array)], axis=0)
max_absolute_value = np.nanmax([np.abs(min_value_array), np.abs(max_value_array)], axis=0)
pairs.append((-max_absolute_value, max_absolute_value))
else:
pairs.append((min_value_array, max_value_array))
Expand Down Expand Up @@ -834,8 +834,8 @@ def collect_absolute_value(self, name_to_arr):
data_arr_np = data_arr
data_arr_np = data_arr_np.flatten()
if data_arr_np.size > 0:
min_value = np.min(data_arr_np)
max_value = np.max(data_arr_np)
min_value = np.nanmin(data_arr_np)
max_value = np.nanmax(data_arr_np)
else:
min_value = np.array(0, dtype=data_arr_np.dtype)
max_value = np.array(0, dtype=data_arr_np.dtype)
Expand All @@ -858,7 +858,7 @@ def collect_absolute_value(self, name_to_arr):
assert hasattr(old_max, "dtype"), f"old_min should be a numpy array but is {type(old_max)}"
old_hist = old_histogram[0]
old_hist_edges = old_histogram[1]
temp_amax = np.max(data_arr_np)
temp_amax = np.nanmax(data_arr_np)
if temp_amax > old_hist_edges[-1]:
# increase the number of bins
width = old_hist_edges[1] - old_hist_edges[0]
Expand All @@ -882,8 +882,8 @@ def collect_value(self, name_to_arr):
data_arr = data_arr.flatten() # noqa: PLW2901

if data_arr.size > 0:
min_value = np.min(data_arr)
max_value = np.max(data_arr)
min_value = np.nanmin(data_arr)
max_value = np.nanmax(data_arr)
else:
min_value = np.array(0, dtype=data_arr.dtype)
max_value = np.array(0, dtype=data_arr.dtype)
Expand Down
2 changes: 1 addition & 1 deletion onnxruntime/python/tools/quantization/quant_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -290,7 +290,7 @@ def compute_scale_zp(rmin, rmax, qmin, qmax, symmetric=False, min_real_range=Non
dr = numpy.array(rmax - rmin, dtype=numpy.float64)
dq = numpy.array(qmax, dtype=numpy.float64) - numpy.array(qmin, dtype=numpy.float64)
scale = numpy.array(dr / dq)
assert scale >= 0, "scale isse"
assert scale >= 0, "scale issue"
if scale < numpy.finfo(rmax.dtype).tiny:
scale = numpy.array(1.0, dtype=rmax.dtype)
zero_point = numpy.array(0, dtype=qmin.dtype)
Expand Down