euntaik commented on a change in pull request #7093:
URL: https://github.com/apache/tvm/pull/7093#discussion_r542268005
##########
File path: python/tvm/relay/frontend/tflite.py
##########
@@ -1991,20 +1994,29 @@ def convert_conv(self, op, conv_type):
weight_tensor_type_str = self.get_tensor_type_str(weight_tensor_type)
in_expr = self.get_expr(input_tensor_idx)
- weight_value = self.get_tensor_value(weight_tensor)
-
- # TFLite kernel layout:
- # convolution:
- # OC KH KW IC, we require KH KW IC OC (HWIO)
- # depthwise convolution:
- # 1 KH KW C(input_c * depth_multiplier), we require
- # KH KW IC M (depth_multiplier) (HWOI)
- if is_depthwise_conv:
- weight_value = weight_value.reshape(kernel_h, kernel_w, input_c,
depth_multiplier)
+
+ if self.has_expr(weight_tensor.tensor_idx):
+ weight_expr = self.get_expr(weight_tensor.tensor_idx)
+ if is_depthwise_conv:
+ weight_expr = _op.reshape(
+ weight_expr, (kernel_h, kernel_w, input_c,
depth_multiplier)
+ )
+ else:
+ weight_expr = _op.transpose(weight_expr, axes=(1, 2, 3, 0))
Review comment:
Yes it is related to float16 support.
tflite converts 32bit models to 16bit models by introducing a Quantize op in
every op that contains a 32-bit values (weights, biases, and constants etc. ).
So for 16-bit converted tflite model has conv2d op receiving the weight as
an input tensor instead of a weight value attached to itself.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
