Skip to content
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

Jump to bottom

[frontend][tflite] float16 quant support #7736

Merged
merged 9 commits into from
Apr 17, 2021
Merged

[frontend][tflite] float16 quant support #7736

Show file tree
Hide file tree
Changes from 7 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
3d4e491
Merge pull request #1 from apache/main
euntaik Feb 3, 2021
b50cb67
Merge branch 'main' of https://github.com/apache/tvm into main
euntaik Feb 22, 2021
97efd21
Merge branch 'main' of https://github.com/apache/tvm into main
euntaik Mar 15, 2021
3888dd2
Merge branch 'main' of https://github.com/apache/tvm into main
euntaik Mar 16, 2021
3d3561a
Merge branch 'main' of https://github.com/apache/tvm into main
euntaik Mar 24, 2021
782c29e
[frontend][tflite] float16 quant support
euntaik Mar 24, 2021
9c28870
remove skip conditions in tests
euntaik Apr 15, 2021
e9b13d8
Merge branch 'main' of https://github.com/apache/tvm into fix-tflite-…
Apr 16, 2021
35069e4
Merge branch 'main' of https://github.com/apache/tvm into fix-tflite-…
Apr 17, 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
35 changes: 24 additions & 11 deletions python/tvm/relay/frontend/tflite.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1882,9 +1882,12 @@ def convert_fully_connected(self, op):
# bias tensor type should be INT32 (quantization) or FLOAT32
assert bias_tensor_type in (TensorType.INT32, TensorType.FLOAT32)
bias_tensor_type_str = self.get_tensor_type_str(bias_tensor_type)
bias_expr = self.exp_tab.new_const(
self.get_tensor_value(bias_tensor), dtype=bias_tensor_type_str
)
if self.has_expr(bias_tensor.tensor_idx):
bias_expr = self.get_expr(bias_tensor.tensor_idx)
else:
bias_expr = self.exp_tab.new_const(
self.get_tensor_value(bias_tensor), dtype=bias_tensor_type_str
)
out = _op.nn.bias_add(out, bias_expr)

# Finally if the dense is quantized. Add a requantize at the end.
Expand Down Expand Up @@ -2852,11 +2855,18 @@ def convert_transpose_conv(self, op):
# weights tensor type should be UINT8 (quantization) or FLOAT32
assert weights_tensor_type in (TensorType.INT8, TensorType.UINT8, TensorType.FLOAT32)
anijain2305 marked this conversation as resolved.
Show resolved Hide resolved
weight_tensor_type_str = self.get_tensor_type_str(weights_tensor_type)
weight_value_ohwi = self.get_tensor_value(weights_tensor)
# Relay kernel_layout should be OIHW
# Relay weights layout should be different from kernel_layout - it should be IOHW
weight_value_iohw = np.transpose(weight_value_ohwi, (3, 0, 1, 2))
weight_expr_iohw = self.exp_tab.new_const(weight_value_iohw, dtype=weight_tensor_type_str)

if self.has_expr(weights_tensor.tensor_idx):
weight_expr_iohw = self.get_expr(weights_tensor.tensor_idx)
weight_expr_iohw = _op.transpose(weight_expr_iohw, axes=(3, 0, 1, 2))
else:
weight_value_ohwi = self.get_tensor_value(weights_tensor)
# Relay kernel_layout should be OIHW
# Relay weights layout should be different from kernel_layout - it should be IOHW
weight_value_iohw = np.transpose(weight_value_ohwi, (3, 0, 1, 2))
weight_expr_iohw = self.exp_tab.new_const(
weight_value_iohw, dtype=weight_tensor_type_str
)

# Output shape value
output_shape_value = self.get_tensor_value(output_shape_tensor)
Expand Down Expand Up @@ -2912,9 +2922,12 @@ def convert_transpose_conv(self, op):
# bias tensor type should be INT32 (quantization) or FLOAT32
assert bias_tensor_type in (TensorType.INT32, TensorType.FLOAT32)
bias_tensor_type_str = self.get_tensor_type_str(bias_tensor_type)
bias_expr = self.exp_tab.new_const(
self.get_tensor_value(bias_tensor), dtype=bias_tensor_type_str
)
if self.has_expr(bias_tensor.tensor_idx):
bias_expr = self.get_expr(bias_tensor.tensor_idx)
else:
bias_expr = self.exp_tab.new_const(
self.get_tensor_value(bias_tensor), dtype=bias_tensor_type_str
)
channel_axis = 3
out = _op.nn.bias_add(out, bias_expr, axis=channel_axis)

Expand Down
Loading