quantized w2l (#1623)
Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/1623
With the features add prior to this diff, Quantized W2L should successfully work now
Reviewed By: digantdesai
Differential Revision: D52809537
fbshipit-source-id: 79e5bee58e0708a2f50008b074736c47e8028d99
diff --git a/backends/xnnpack/test/models/w2l.py b/backends/xnnpack/test/models/w2l.py
index 607ee4b..10d7ca1 100644
--- a/backends/xnnpack/test/models/w2l.py
+++ b/backends/xnnpack/test/models/w2l.py
@@ -15,7 +15,7 @@
batch_size = 10
input_frames = 700
vocab_size = 4096
- wav2letter = models.Wav2Letter(num_classes=vocab_size)
+ wav2letter = models.Wav2Letter(num_classes=vocab_size).eval()
model_inputs = (torch.randn(batch_size, 1, input_frames),)
@@ -37,3 +37,23 @@
.run_method()
.compare_outputs()
)
+
+ def test_qs8_w2l(self):
+ (
+ Tester(self.wav2letter.eval(), self.model_inputs)
+ .quantize()
+ .export()
+ .to_edge()
+ .partition()
+ .check_not(
+ [
+ "executorch_exir_dialectes_edge__ops_aten_convolution_default",
+ "executorch_exir_dialects_edge__ops_aten_relu_default",
+ ]
+ )
+ .check(["torch.ops.higher_order.executorch_call_delegate"])
+ .to_executorch()
+ .serialize()
+ .run_method()
+ .compare_outputs()
+ )
