rel 1.13.1 cherry pick round 2 (microsoft#13390)

intel · Oct 21, 2022 · 95bd0e7 · 95bd0e7
1 parent 8042f72
commit 95bd0e7
Show file tree

Hide file tree

Showing 2 changed files with 21 additions and 6 deletions.
diff --git a/csharp/src/Microsoft.ML.OnnxRuntime/targets/netstandard/targets.xml b/csharp/src/Microsoft.ML.OnnxRuntime/targets/netstandard/targets.xml
@@ -6,12 +6,12 @@
     the PlatformTarget is empty, and you don't know until runtime (i.e. which dotnet.exe)
     what processor architecture will be used.
   -->
-    <Error Condition="('$(PlatformTarget)' != 'x64' AND '$(PlatformTarget)' != 'x86' AND '$(PlatformTarget)' != 'AnyCPU') AND
+    <Error Condition="('$(PlatformTarget)' != 'x64' AND '$(PlatformTarget)' != 'arm32' AND '$(PlatformTarget)' != 'arm64' AND '$(PlatformTarget)' != 'x86' AND '$(PlatformTarget)' != 'AnyCPU') AND
                       ('$(OutputType)' == 'Exe' OR '$(OutputType)'=='WinExe') AND
                       !('$(TargetFrameworkIdentifier)' == '.NETCoreApp' AND '$(PlatformTarget)' == '') AND
                       ('$(TargetFrameworkIdentifier)' != 'Xamarin.iOS' AND
                        $([MSBuild]::GetTargetPlatformIdentifier('$(TargetFramework)')) != 'ios') AND
                       '$(SuppressOnnxRuntimePlatformCompatibilityError)' != 'true'"
-           Text="Microsoft.ML.OnnxRuntime only supports the AnyCPU, x64, and x86 platforms at this time."/>
+           Text="Microsoft.ML.OnnxRuntime only supports the AnyCPU, x64, arm32, arm64 and x86 platforms at this time."/>
   </Target>
-</Project>
+</Project>
diff --git a/onnxruntime/python/tools/quantization/calibrate.py b/onnxruntime/python/tools/quantization/calibrate.py
@@ -570,14 +570,23 @@ def collect_absolute_value(self, name_to_arr):
         for tensor, data_arr in name_to_arr.items():
             data_arr = np.asarray(data_arr)
             data_arr = data_arr.flatten()
+            if data_arr.size > 0:
+                min_value = np.min(data_arr)
+                max_value = np.max(data_arr)
+            else:
+                min_value = 0
+                max_value = 0
+
             data_arr = np.absolute(data_arr)  # only consider absolute value
 
             if tensor not in self.histogram_dict:
                 # first time it uses num_bins to compute histogram.
                 hist, hist_edges = np.histogram(data_arr, bins=self.num_bins)
-                self.histogram_dict[tensor] = (hist, hist_edges)
+                self.histogram_dict[tensor] = (hist, hist_edges, min_value, max_value)
             else:
                 old_histogram = self.histogram_dict[tensor]
+                old_min = old_histogram[2]
+                old_max = old_histogram[3]
                 old_hist = old_histogram[0]
                 old_hist_edges = old_histogram[1]
                 temp_amax = np.max(data_arr)
@@ -589,7 +598,7 @@ def collect_absolute_value(self, name_to_arr):
                     old_hist_edges = np.hstack((old_hist_edges, new_bin_edges))
                 hist, hist_edges = np.histogram(data_arr, bins=old_hist_edges)
                 hist[: len(old_hist)] += old_hist
-                self.histogram_dict[tensor] = (hist, hist_edges)
+                self.histogram_dict[tensor] = (hist, hist_edges, min(old_min, min_value), max(old_max, max_value))
 
     def collect_value(self, name_to_arr):
         """
@@ -688,6 +697,7 @@ def compute_percentile(self):
             cdf = np.cumsum(hist / total)
             if self.symmetric:
                 idx_right = np.searchsorted(cdf, percentile / 100.0)
+
                 thresholds_dict[tensor] = (
                     -float(hist_edges[idx_right]),
                     float(hist_edges[idx_right]),
@@ -700,7 +710,12 @@ def compute_percentile(self):
                     float(hist_edges[idx_left]),
                     float(hist_edges[idx_right]),
                 )
-
+            min_value = histogram[2]
+            max_value = histogram[3]
+            if thresholds_dict[tensor][0] < min_value:
+                thresholds_dict[tensor] = (min_value, thresholds_dict[tensor][1])
+            if thresholds_dict[tensor][1] > max_value:
+                thresholds_dict[tensor] = (thresholds_dict[tensor][0], max_value)
             # Plot histogram for debug only
             if False:
                 apply_plot(hist, hist_edges)