maekawatoshiki commented on code in PR #14536:
URL: https://github.com/apache/tvm/pull/14536#discussion_r1193416091
##########
tests/python/relay/test_pass_fake_quantization_to_integer.py:
##########
@@ -1114,5 +1114,53 @@ def test_fake_quantize_take():
compare_fq_to_int(op, [x_np])
+def test_fake_quantize_softmax():
+ shape = [5, 10]
+ x_ = relay.var("x", shape=shape, dtype="int8")
+
+ is_sorted = lambda a: np.all(a[:-1] <= a[1:])
+
+ for scale in [1.0, 0.1, 0.01]:
+ x = relay.qnn.op.dequantize(x_, relay.const(scale), relay.const(0))
+ op = relay.op.nn.softmax(x, axis=1)
+ op = relay.qnn.op.quantize(
+ op, relay.const(1.0 / 256.0), relay.const(-128), out_dtype="int8"
+ )
+
+ x_np = np.random.randint(-128, 127, size=shape, dtype="int8")
+ x_np = np.sort(x_np)
+ args = [x_np]
+
+ mod = tvm.IRModule.from_expr(op)
+ mod = tvm.relay.transform.InferType()(mod)
+ mod_int = tvm.relay.transform.FakeQuantizationToInteger(
+ hard_fail=True, optional_qnn_ops=["nn.softmax"]
+ )(mod)
+ assert not tvm.ir.structural_equal(mod, mod_int)
+
+ result = (
+ relay.create_executor("vm", mod=mod, device=tvm.cpu(),
target="llvm")
+ .evaluate()(*args)
+ .numpy()
+ )
+ result_int = (
+ relay.create_executor("vm", mod=mod_int, device=tvm.cpu(),
target="llvm")
+ .evaluate()(*args)
+ .numpy()
+ )
+
+ # Check at least the softmax output is in ascending order,
+ # since it is difficult to use allclose due to not-so-good accuracy.
+ for qdq, qop in zip(result, result_int):
+ assert is_sorted(qdq)
+ assert is_sorted(qop)
Review Comment:
The input is sorted at L1131.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
