commit 4f728c04bcc90742d9d57b0e253be68a7251984f
author Dwight Lidman <dwight.lidman@arm.com> Thu Dec 17 15:14:45 2020 +0100
committer Louis Verhaard <louis.verhaard@arm.com> Thu Feb 25 17:27:27 2021 +0100
tree 7e205e9fc90a36606d1939af406af20fa10b448d
parent 0cbb166cd032c779bd4681afef8097f0831ac8be

MLBEDSW-1499: Add MEAN operator

This commit adds support for the MEAN operator,
with some caveats.

Signed-off-by: Dwight Lidman <dwight.lidman@arm.com>
Change-Id: I165cb26cb5aefd68e70d2cfc68291ccf7b778921

ethosu/vela/weight_compressor.py

diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py
index b291dce..bb7cd67 100644
--- a/ethosu/vela/weight_compressor.py
+++ b/ethosu/vela/weight_compressor.py
@@ -426,13 +426,13 @@
 
     first_consumer_op = tens.consumer_list[0]
     ifm_dtype = first_consumer_op.inputs[0].dtype
-    ifm_scale = first_consumer_op.inputs[0].quantization.scale_f32
+    ifm_scale = first_consumer_op.get_input_quantization().scale_f32
     ofm_scale = first_consumer_op.get_output_quantization().scale_f32
     weight_scales = first_consumer_op.inputs[1].quantization.scale_f32
 
     # biases can have multiple consumers for rnn cells. if so, then check that they are all the same
     for op in tens.consumer_list[1:]:
-        assert ifm_scale == op.inputs[0].quantization.scale_f32
+        assert ifm_scale == op.get_input_quantization().scale_f32
         assert ofm_scale == op.get_output_quantization().scale_f32
         assert weight_scales == op.inputs[1].quantization.scale_f32
 
@@ -445,7 +445,14 @@
     # TensorFlow Lite casts the scales slightly differently for uint8 and int8
     if not rescale_for_faf:
         if ifm_dtype == DataType.uint8:
-            scales = [np.double(ifm_scale * weight_scale) / np.double(ofm_scale) for weight_scale in weight_scales]
+            # for some cases of the Mean operator, the scale must be calculated differently to match reference
+            if first_consumer_op.low_precision_scaling:
+                scales = [
+                    np.double(np.single(ifm_scale) / (np.single(weight_scale) * np.single(ofm_scale)))
+                    for weight_scale in weight_scales
+                ]
+            else:
+                scales = [np.double(ifm_scale * weight_scale) / np.double(ofm_scale) for weight_scale in weight_scales]
         elif ifm_dtype == DataType.int8 or ifm_dtype == DataType.int16:
             scales = [
                 (np.double(ifm_scale) * np.double(weight_scale)) / np.double(ofm_scale)