From 8a6fb6ff00b975f5af0f4dbc493d974d337b381c Mon Sep 17 00:00:00 2001
From: Hendrik Borras <hendrikborras@web.de>
Date: Tue, 19 Oct 2021 19:17:15 +0100
Subject: [PATCH] Enabled non-scalar scales for the Trunc node and made bit
width readout more robust.
---
.../qonnx/infer_quant_avg_pool_2d.py | 25 ++++++++-----------
1 file changed, 11 insertions(+), 14 deletions(-)
diff --git a/src/finn/transformation/qonnx/infer_quant_avg_pool_2d.py b/src/finn/transformation/qonnx/infer_quant_avg_pool_2d.py
index 70d97946a..3de317b9e 100644
--- a/src/finn/transformation/qonnx/infer_quant_avg_pool_2d.py
+++ b/src/finn/transformation/qonnx/infer_quant_avg_pool_2d.py
@@ -127,32 +127,28 @@ class AvgPoolAndTruncToQuantAvgPool(Transformation):
f"except the first, must be statically "
f"initialized. However, {inp} is not."
)
- scale = model.get_initializer(t_node.input[1])
- if len(scale.shape) != 0:
- raise ValueError(
- f"Finn only supports scalar scales of zero dimension "
- f"for the Trunc node, it currently is {scale}."
- )
zero_pt = model.get_initializer(t_node.input[2])
if len(zero_pt.shape) != 0 or zero_pt != 0:
raise ValueError(
f"Finn only supports 0 as the zero point for "
f"the Trunc node, it currently is {zero_pt}."
)
- trunc_in_bits = model.get_initializer(t_node.input[3])
- trunc_out_bits = model.get_initializer(t_node.input[4])
+ trunc_in_bits = model.get_initializer(t_node.input[3]).flatten()
+ trunc_out_bits = model.get_initializer(
+ t_node.input[4]
+ ).flatten()
if (
- len(trunc_in_bits.shape) != 0
- or len(trunc_out_bits.shape) != 0
+ len(trunc_in_bits.shape) != 1
+ or len(trunc_out_bits.shape) != 1
):
raise ValueError(
- f"Finn only supports scalar bit widths of zero "
- f"dimension for the Trunc node. The input bit width "
+ f"Finn only supports scalar bit widths "
+ f"for the Trunc node. The input bit width "
f"currently is: {trunc_in_bits}, "
f"while the output bit width is: {trunc_out_bits}."
)
- trunc_in_bits = int(trunc_in_bits)
- trunc_out_bits = int(trunc_out_bits)
+ trunc_in_bits = int(trunc_in_bits[0])
+ trunc_out_bits = int(trunc_out_bits[0])
# Calculate parameters for the QuantAvgPool2d node,
# Calculate input bit width. Basically this backwards:
@@ -170,6 +166,7 @@ class AvgPoolAndTruncToQuantAvgPool(Transformation):
data_layout = "NCHW"
# Insert scale nodes, QuantAvgPool2d node and required tensors
+ scale = model.get_initializer(t_node.input[1])
scale_div_tensor = helper.make_tensor_value_info(
model.make_new_valueinfo_name(),
TensorProto.FLOAT,
--
GitLab