MobileNetv2 FP32 + QS8 Test

Summary:
Adding some CI for Mobilenetv2.

The test tests for FP32 model and QS8 Model via long term quantization flow.

Differential Revision: D48488928

fbshipit-source-id: 58c97896f01d0bd10938d025f08b3d2a069c01b6

Author: mcr229

backends/xnnpack/test/TARGETS

@@ -120,3 +120,16 @@ python_unittest(
         "//executorch/backends/xnnpack/test/tester:tester",
     ],
 )
+
+python_unittest(
+    name = "test_xnnpack_models",
+    srcs = glob([
+        "models/*.py",
+    ]),
+    deps = [
+        "//caffe2:torch",
+        "//executorch/backends/xnnpack/partition:xnnpack_partitioner",
+        "//executorch/backends/xnnpack/test/tester:tester",
+        "//pytorch/vision:torchvision",
+    ],
+)

backends/xnnpack/test/models/mobilenet_v2.py

@@ -0,0 +1,70 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import unittest
+
+import torch
+from executorch.backends.xnnpack.partition.xnnpack_partitioner import (
+    XnnpackQuantizedPartitioner2,
+)
+from executorch.backends.xnnpack.test.tester import Partition, Tester
+from torchvision import models
+from torchvision.models.mobilenetv2 import MobileNet_V2_Weights
+
+
+class TestXNNPACKMobileNetV2(unittest.TestCase):
+
+    mv2 = models.mobilenet_v2(weights=MobileNet_V2_Weights)
+    mv2 = mv2.eval()
+
+    all_operators = {
+        "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default",
+        "executorch_exir_dialects_edge__ops_aten_add_Tensor",
+        "executorch_exir_dialects_edge__ops_aten_permute_copy_default",
+        "executorch_exir_dialects_edge__ops_aten_addmm_default",
+        "executorch_exir_dialects_edge__ops_aten_mean_dim",
+        "executorch_exir_dialects_edge__ops_aten_hardtanh_default",
+        "executorch_exir_dialects_edge__ops_aten_convolution_default",
+    }
+
+    def test_fp32(self):
+        model_inputs = (torch.ones(1, 3, 224, 244),)
+
+        (
+            Tester(self.mv2, model_inputs)
+            .export()
+            .to_edge()
+            .check(list(self.all_operators))
+            .partition()
+            .check(["torch.ops.executorch_call_delegate"])
+            .check_not(list(self.all_operators))
+            .to_executorch()
+            .serialize()
+            .run_method()
+            .compare_outputs()
+        )
+
+    def test_qs8_pt2e(self):
+        model_inputs = (torch.ones(1, 3, 224, 244),)
+        # Quantization fuses away batchnorm, so it is no longer in the graph
+        ops_after_quantization = self.all_operators - {
+            "executorch_exir_dialects_edge__ops_aten__native_batch_norm_legit_no_training_default",
+        }
+
+        (
+            Tester(self.mv2, model_inputs)
+            .quantize2()
+            .export()
+            .to_edge()
+            .check(list(ops_after_quantization))
+            .partition(Partition(partitioner=XnnpackQuantizedPartitioner2))
+            .check(["torch.ops.executorch_call_delegate"])
+            .check_not(list(ops_after_quantization))
+            .to_executorch()
+            .serialize()
+            .run_method()
+            .compare_outputs()
+        )