def before():
x = relay.var("x", shape=(1, 16, 16, 16), dtype="float32")
w = relay.var("w", shape=(32, 16, 3, 3), dtype="float32")
y = relay.nn.conv2d(x, w, padding=(1, 1))
y = relay.reshape(y, newshape=(1, 16, -1))
y = relay.reshape(y, newshape=(4, 8, -1, 16))
y = relay.reverse_reshape(y, newshape=(32, 0, -1))
return relay.Function([x, w], y)
def before():
x = relay.var("x", shape=(1, 16, 16, 16), dtype="float32")
w = relay.var("w", shape=(32, 16, 3, 3), dtype="float32")
y = relay.nn.conv2d(x, w, padding=(1, 1))
y = relay.reshape(y, newshape=(1, 16, -1))
y = relay.reshape(y, newshape=(4, 8, -1, 16))
y = relay.reverse_reshape(y, newshape=(32, 0, -1))
return tvm.IRModule.from_expr(y)
def symbolic():
b = tvm.te.size_var("b")
x = relay.var("x", shape=(b, 16, 16, 16), dtype="float32")
w = relay.var("w", shape=(32, 16, 3, 3), dtype="float32")
y = relay.nn.conv2d(x, w, padding=(1, 1))
y = relay.reshape(y, newshape=(1, 16, -1))
y = relay.reshape(y, newshape=(4, 8, -1, 16))
y = relay.reverse_reshape(y, newshape=(32, 0, -1))
return relay.Function([x, w], y)
def verify_reverse_reshape(shape, newshape, oshape):
x = relay.var("x", relay.TensorType(shape, "float32"))
z = relay.reverse_reshape(x, newshape=newshape)
zz = relay.ir_pass.infer_type(z)
assert "newshape=" in z.astext()
assert zz.checked_type == relay.ty.TensorType(oshape, "float32")
func = relay.Function([x], z)
x_data = np.random.uniform(low=-1, high=1, size=shape).astype("float32")
ref_res = np.reshape(x_data, oshape)
for target, ctx in ctx_list():
for kind in ["graph", "debug"]:
intrp = relay.create_executor(kind, ctx=ctx, target=target)
op_res = intrp.evaluate(func)(x_data)
tvm.testing.assert_allclose(op_res.asnumpy(), ref_res, rtol=1e-5)
def verify_reverse_reshape(shape, newshape, oshape):
x = relay.var("x", relay.TensorType(shape, "float32"))
z = relay.reverse_reshape(x, newshape=newshape)
zz = relay.ir_pass.infer_type(z)
assert "newshape=" in z.astext()
assert zz.checked_type == relay.ty.TensorType(oshape, "float32")
func = relay.Function([x], z)
x_data = np.random.uniform(low=-1, high=1, size=shape).astype("float32")
ref_res = np.reshape(x_data, oshape)
for target, ctx in ctx_list():
for kind in ["graph", "debug"]:
intrp = relay.create_executor(kind, ctx=ctx, target=target)
op_res = intrp.evaluate(func)(x_data)
tvm.testing.assert_allclose(op_res.asnumpy(), ref_res, rtol=1e-5)
def verify_reverse_reshape(executor_kind, shape, newshape, oshape):
x = relay.var("x", relay.TensorType(shape, "float32"))
z = relay.reverse_reshape(x, newshape=newshape)
zz = run_infer_type(z)
assert "newshape=" in z.astext()
assert zz.checked_type == relay.ty.TensorType(oshape, "float32")
func = relay.Function([x], z)
x_data = np.random.uniform(low=-1, high=1,
size=shape).astype("float32")
ref_res = np.reshape(x_data, oshape)
for target, dev in tvm.testing.enabled_targets():
op_res = relay.create_executor(
executor_kind, device=dev,
target=target).evaluate(func)(x_data)
tvm.testing.assert_allclose(op_res.numpy(), ref_res, rtol=1e-5)
def test_vm_reshape_tensor():
x_np = np.random.uniform(size=(8, 16)).astype("float32")
x = relay.var("x", shape=(8, 16), dtype="float32")
y = relay.reshape(x, [-1, 4, 8])
mod = tvm.IRModule()
mod["main"] = relay.Function([x], y)
with tvm.transform.PassContext(opt_level=3):
exec = relay.vm.compile(mod, "llvm")
assert "reshape_tensor" in exec.bytecode
check_result([x_np], x_np.reshape([4, 4, 8]), mod)
x = relay.var("x", shape=(8, 16), dtype="float32")
y = relay.reshape(x, [16, -1])
y = relay.reverse_reshape(y, [-1, 4, 0])
mod = tvm.IRModule()
mod["main"] = relay.Function([x], y)
with tvm.transform.PassContext(opt_level=3):
exec = relay.vm.compile(mod, "llvm")
assert exec.bytecode.count("reshape_tensor") == 1
check_result([x_np], x_np.reshape([4, 4, 8]), mod)
# reshape with symbolic/any shape
for n in [tvm.tir.Any(), tvm.te.size_var("n")]:
x = relay.var("x", shape=(n, 16), dtype="float32")
y = relay.reshape(x, [-1, 4])
y = relay.reshape(y, [0, 2, -1])
mod = tvm.IRModule()
mod["main"] = relay.Function([x], y)
with tvm.transform.PassContext(opt_level=3):
exec = relay.vm.compile(mod, "llvm")
assert exec.bytecode.count("reshape_tensor") == 1
check_result([x_np], x_np.reshape([32, 2, 2]), mod)
# dyn.reshape
x = relay.var("x", shape=(8, 16), dtype="float32")
y = relay.var("y", shape=(3, ), dtype="int32")
z = relay.reshape(x, [-1, 4, 8])
z = relay.reshape(z, y)
mod = tvm.IRModule()
mod["main"] = relay.Function([x, y], z)
with tvm.transform.PassContext(opt_level=3):
exec = relay.vm.compile(mod, "llvm")
assert exec.bytecode.count("reshape_tensor") == 2
assert "reshape_tensor" in exec.bytecode
y_np = np.array([8, 2, 8]).astype("int32")
check_result([x_np, y_np], x_np.reshape([8, 2, 8]), mod)
