Skip to content

[TFLite] Cast operator adapted for MLIR-based convertor#7639

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
anijain2305 merged 2 commits into from
Mar 19, 2021
Merged

[TFLite] Cast operator adapted for MLIR-based convertor #7639

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
98f078c
[TFLite] Cast operator adapted for MLIR-based convertor
d-smirnov Mar 9, 2021
3699924
Comment added
d-smirnov Mar 16, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 12 additions & 5 deletions python/tvm/relay/frontend/tflite.py
View file
Open in desktop
Original file line numberDiff line numberDiff line change
Expand Up@@ -2336,11 +2336,18 @@ def convert_cast(self, op):
input_tensor = input_tensors[0]
in_expr = self.get_expr(input_tensor.tensor_idx)

assert op.BuiltinOptionsType() == BuiltinOptions.CastOptions
op_options = op.BuiltinOptions()
cast_options = CastOptions()
cast_options.Init(op_options.Bytes, op_options.Pos)
cast_dtype = cast_options.OutDataType()
# MLIR-based converter outputs no BuiltinOptions for Cast operator. In this
# case the output type can be derived from the Cast operator output tensor.
# When TOCO converter is used there will be "normal" BuiltinOptions.CastOptions
# with output type.
if op.BuiltinOptions() is not None:
assert op.BuiltinOptionsType() == BuiltinOptions.CastOptions
op_options = op.BuiltinOptions()
cast_options = CastOptions()
cast_options.Init(op_options.Bytes, op_options.Pos)
cast_dtype = cast_options.OutDataType()
else:
cast_dtype = self.get_output_tensors(op)[0].tensor.Type()

@u99127u99127Mar 12, 2021

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this because the cast operator can be in 2 forms, one with the type specified by BuiltinOptions.CastOptions and the other type is where there is no output data type specified by BuiltinOptions.CastOptions but we can derive the output type of the cast operator based on the type of the output tensor for the Cast operator itself ?

Is that worth stating in the commit message or in a comment ?

Ramana


out = _op.cast(in_expr, self.get_tensor_type_str(cast_dtype))

Expand Down
19 changes: 14 additions & 5 deletions tests/python/frontend/tflite/test_forward.py
View file
Open in desktop
Original file line numberDiff line numberDiff line change
Expand Up@@ -647,19 +647,28 @@ def test_forward_transpose():
# ----


def _test_cast(data, cast_dtype):
def _test_cast(data, cast_dtype, use_mlir=False):
""" One iteration of CAST """
with tf.Graph().as_default():
in_data = array_ops.placeholder(shape=data.shape, dtype=data.dtype)
out = math_ops.cast(in_data, cast_dtype)
compare_tflite_with_tvm(data, "Placeholder:0", [in_data], [out])
compare_tflite_with_tvm(
data, "Placeholder:0", [in_data], [out], experimental_new_converter=use_mlir
)


def test_forward_cast():
""" CAST """
_test_cast(np.arange(6.0, dtype=np.float32).reshape((1, 6)), cast_dtype=tf.int32)
_test_cast(np.arange(6.0, dtype=np.float32).reshape((1, 6)), cast_dtype=tf.uint8)
_test_cast(np.arange(6.0, dtype=np.int32).reshape((1, 6)), cast_dtype=tf.int64)
for use_mlir in [False, True]:
_test_cast(
np.arange(6.0, dtype=np.float32).reshape((1, 6)), cast_dtype=tf.int32, use_mlir=use_mlir
)
_test_cast(
np.arange(6.0, dtype=np.float32).reshape((1, 6)), cast_dtype=tf.uint8, use_mlir=use_mlir
)
_test_cast(
np.arange(6.0, dtype=np.int32).reshape((1, 6)), cast_dtype=tf.int64, use_mlir=use_mlir
)


#######################################################################
Expand Down