Increase calibration samples and tolerance for flaky quantized op tests

Author: GregoryComer

backends/xnnpack/test/ops/test_add.py

@@ -7,7 +7,7 @@
 import unittest
 
 import torch
-from executorch.backends.xnnpack.test.tester import Tester
+from executorch.backends.xnnpack.test.tester import Quantize, Tester
 
 
 class TestAdd(unittest.TestCase):
@@ -136,9 +136,12 @@ def test_qs8_add2(self):
 
     def test_qs8_add3(self):
         inputs = (torch.randn(1, 1, 4, 4), torch.randn(1, 1, 4, 1))
+        calibration_samples = [
+            (torch.randn(1, 1, 4, 4), torch.randn(1, 1, 4, 1)) for _ in range(100)
+        ]
         (
             Tester(self.Add(), inputs)
-            .quantize()
+            .quantize(Quantize(calibration_samples=calibration_samples))
             .export()
             .check_count({"torch.ops.aten.add.Tensor": 4})
             .check(["torch.ops.quantized_decomposed"])
@@ -152,7 +155,7 @@ def test_qs8_add3(self):
             )
             .to_executorch()
             .serialize()
-            .run_method_and_compare_outputs()
+            .run_method_and_compare_outputs(num_runs=10, atol=0.02, rtol=0.02)
         )
 
     class AddRelu(torch.nn.Module):

backends/xnnpack/test/ops/test_conv1d.py

@@ -13,7 +13,7 @@
 from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
 from executorch.backends.xnnpack.test.test_xnnpack_utils import randomize_bn
 
-from executorch.backends.xnnpack.test.tester import RunPasses, Tester
+from executorch.backends.xnnpack.test.tester import Quantize, RunPasses, Tester
 from executorch.backends.xnnpack.test.tester.tester import ToEdgeTransformAndLower
 from executorch.exir.passes.constant_prop_pass import constant_prop_pass
 
@@ -98,9 +98,17 @@ def _test_conv1d(
         stage=None,
         skip_to_executorch=False,
     ):
+        calibration_samples = (
+            [tuple(torch.randn_like(inputs[i]) for i in range(len(inputs)))]
+            if quantized
+            else None
+        )
+
         tester = (
             (
-                Tester(module, inputs, dynamic_shape).quantize()
+                Tester(module, inputs, dynamic_shape).quantize(
+                    Quantize(calibration_samples=calibration_samples)
+                )
                 if quantized
                 else Tester(module, inputs)
             )
@@ -114,7 +122,9 @@ def _test_conv1d(
         # For some tests we want to skip to_executorch because otherwise it will require the
         # quantized operators to be loaded and we don't want to do that in the test.
         if not skip_to_executorch:
-            tester.to_executorch().serialize().run_method_and_compare_outputs()
+            tester.to_executorch().serialize().run_method_and_compare_outputs(
+                num_runs=10, atol=0.01, rtol=0.01
+            )
 
     def test_fp16_conv1d(self):
         inputs = (torch.randn(2, 2, 4).to(torch.float16),)

backends/xnnpack/test/tester/tester.py

@@ -12,7 +12,7 @@
 import sys
 from abc import ABC, abstractmethod
 from collections import Counter, OrderedDict
-from typing import Any, Callable, Dict, List, Optional, Tuple, Type, Union
+from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Type, Union
 
 import torch
 from executorch.backends.xnnpack._passes import XNNPACKPassManager
@@ -146,12 +146,14 @@ def __init__(
         quantizer: Optional[Quantizer] = None,
         quantization_config: Optional[QuantizationConfig] = None,
         calibrate: bool = True,
+        calibration_samples: Optional[Sequence[Any]] = None,
     ):
         self.quantizer = quantizer or XNNPACKQuantizer()
         self.quantization_config = (
             quantization_config or get_symmetric_quantization_config()
         )
         self.calibrate = calibrate
+        self.calibration_samples = calibration_samples
 
         self.quantizer.set_global(self.quantization_config)
 
@@ -168,7 +170,11 @@ def run(
 
         if self.calibrate:
             # Calibrate prepared model to provide data to quantization observers.
-            prepared(*inputs)
+            if self.calibration_samples is not None:
+                for inp in self.calibration_samples:
+                    prepared(*inp)
+            else:
+                prepared(*inputs)
 
         converted = convert_pt2e(prepared)
         self.converted_graph = converted