diff --git a/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py b/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
index e5a89f4cb60a92f4fe75bd060a1f3b30e9d1c34f..3ed76db2982e411b711be5bd78e39dd866332714 100644
--- a/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
+++ b/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
@@ -51,7 +51,7 @@ from . import templates
def get_smallest_possible(vals):
"""Returns smallest (fewest bits) possible DataType that can represent
value. Prefers unsigned integers where possible."""
- vals = np.array(vals)
+ vals = np.array(vals, dtype=np.float64)
for v in vals:
assert int(v) == v, "Error float value"
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index 3f42564f968f433688e942519f74251d4b2c34b8..9d70ca46906af3e299600669ec91090e56492bb8 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -1259,7 +1259,7 @@ class InferChannelwiseLinearLayer(Transformation):
def get_smallest_possible(self, vals):
"""Returns smallest (fewest bits) possible DataType that can represent
value. Prefers unsigned integers where possible."""
- vals = np.array(vals)
+ vals = np.array(vals, dtype=np.float64)
for v in vals:
assert int(v) == v, "Error float value"
@@ -1535,7 +1535,9 @@ class InferGlobalAccPoolLayer(Transformation):
model.make_new_valueinfo_name(), TensorProto.FLOAT, [1]
)
model.graph.value_info.append(mul_value)
- model.set_initializer(mul_value.name, np.array(1 / (vecs[1] * vecs[2])))
+ model.set_initializer(
+ mul_value.name, np.array(1 / (vecs[1] * vecs[2]), dtype=np.float32)
+ )
new_mul = helper.make_node(
"Mul",
[pool_out, mul_value.name],
diff --git a/src/finn/transformation/qonnx/qonnx_activation_handlers.py b/src/finn/transformation/qonnx/qonnx_activation_handlers.py
index 406f7c0f50ea649890e079a79ff9c52a5f95764d..a50a5850779cadf7ab21b9c1c4dfdbb36232af42 100644
--- a/src/finn/transformation/qonnx/qonnx_activation_handlers.py
+++ b/src/finn/transformation/qonnx/qonnx_activation_handlers.py
@@ -32,6 +32,8 @@ from onnx import TensorProto, helper
from qonnx.core.modelwrapper import ModelWrapper
from qonnx.custom_op.registry import getCustomOp
+np_default_dtype = np.float32
+
class QuantActBaseHandler(ABC):
"""Base class for converting quantized activation expressed in the QONNX dialect
@@ -163,17 +165,16 @@ class QuantActBaseHandler(ABC):
if scale_scalar and bias_scalar and self._q_node.op_type == "BipolarQuant":
# Get Quant parameters
mul_scale = np.atleast_1d(mul_scale)
- # ONNX only accepts 64bit floats as attributes
- mul_scale = mul_scale.astype(dtype=np.float64)
adder_bias = np.atleast_1d(adder_bias)
- adder_bias = adder_bias.astype(dtype=np.float64)
# Set Bias and scale
- mt_inst.set_nodeattr("out_scale", mul_scale[0])
+ # note calls to .item() to get Python float instead of numpy float
+ # ONNX attribute setting fails otherwise
+ mt_inst.set_nodeattr("out_scale", mul_scale[0].item())
# FINN applies scale first then bias,
# which is the other way around in Brevitas,
# we thus need to adjust the bias in the MultiThreshold node
- finn_bias = adder_bias[0] * mul_scale[0]
+ finn_bias = adder_bias[0].item() * mul_scale[0].item()
mt_inst.set_nodeattr("out_bias", finn_bias)
# Set the output data type
@@ -189,8 +190,7 @@ class QuantActBaseHandler(ABC):
zero_bias = False
if bias_scalar:
adder_bias = np.atleast_1d(adder_bias)
- # ONNX only accepts 64bit floats as attributes
- adder_bias = adder_bias.astype(dtype=np.float64)[0]
+ adder_bias = adder_bias[0]
add_shape = tuple()
if adder_bias == 0.0:
zero_bias = True
@@ -233,7 +233,7 @@ class QuantActBaseHandler(ABC):
unity_scale = False
if scale_scalar:
mul_scale = np.atleast_1d(mul_scale)
- mul_scale = mul_scale.astype(dtype=np.float64)[0]
+ mul_scale = mul_scale[0]
mul_shape = tuple()
if mul_scale == 1.0:
unity_scale = True
@@ -312,7 +312,7 @@ class QuantReluHandler(QuantActBaseHandler):
# No bias allowed for Relu activations, see: https://github.com/Xilinx/
# brevitas/blob/a5bfd6dc5e030f0047ac1ee47932b60e8e873e17/src/brevitas/
# export/onnx/finn/handler/act.py#L48
- bias = np.array([0.0])
+ bias = np.array([0.0], dtype=np_default_dtype)
return bias
def _calculate_thresholds(self):
@@ -338,7 +338,9 @@ class QuantReluHandler(QuantActBaseHandler):
num_scale_channels = flat_scale.shape[0]
step = np.abs(flat_scale).astype(np.float32)
min_threshold = step / 2
- thresholds = np.empty((num_scale_channels, num_thresholds)).astype(np.float32)
+ thresholds = np.empty(
+ (num_scale_channels, num_thresholds), dtype=np_default_dtype
+ )
for c in range(num_scale_channels):
for t in range(num_thresholds):
thresholds[c][t] = min_threshold[c] + step[c] * t
@@ -437,13 +439,13 @@ class QuantIdentityHandler(QuantActBaseHandler):
# a5bfd6dc5e030f0047ac1ee47932b60e8e873e17/src/brevitas/export/
# onnx/finn/handler/act.py#L64
if bit_width == 1.0:
- bias = np.array([-0.5])
+ bias = np.array([-0.5], dtype=np_default_dtype)
else:
if narrow:
min_non_scaled_val = -(2 ** (bit_width - 1) - 1)
else:
min_non_scaled_val = -(2 ** (bit_width - 1))
- bias = np.array([min_non_scaled_val])
+ bias = np.array([min_non_scaled_val], dtype=np_default_dtype)
return bias
def _calculate_thresholds(self):
@@ -462,7 +464,7 @@ class QuantIdentityHandler(QuantActBaseHandler):
# blob/a5bfd6dc5e030f0047ac1ee47932b60e8e873e17/src/brevitas/
# export/onnx/finn/handler/act.py#L76
if bit_width == 1.0:
- thresholds = np.empty([1, 1])
+ thresholds = np.empty([1, 1], dtype=np_default_dtype)
thresholds[0] = 0
return thresholds
else:
@@ -476,7 +478,9 @@ class QuantIdentityHandler(QuantActBaseHandler):
num_scale_channels = flat_scale.shape[0]
step = np.abs(flat_scale)
half_step = step / 2.0
- thresholds = np.empty((num_scale_channels, num_thresholds))
+ thresholds = np.empty(
+ (num_scale_channels, num_thresholds), dtype=np_default_dtype
+ )
# compute the value of the smallest threshold, we'll neg-bias all
# generated thresholds by this much
min_threshold = -half_step - step * ((num_thresholds // 2) - 1)
diff --git a/tests/fpgadataflow/test_convert_to_hls_conv_fc_transition.py b/tests/fpgadataflow/test_convert_to_hls_conv_fc_transition.py
index 6eb8b2b0056877f23c7fc7237b05581983f56579..001ffc08b6d2400c3047206f28adefd3b39f64f2 100755
--- a/tests/fpgadataflow/test_convert_to_hls_conv_fc_transition.py
+++ b/tests/fpgadataflow/test_convert_to_hls_conv_fc_transition.py
@@ -202,7 +202,7 @@ def test_convert_to_hls_conv_fc_transition(conv_config, depthwise, use_reshape):
model.set_initializer(
"matmul_param", gen_finn_dt_tensor(fc_weight_dt, fc_param_shape)
)
- model.set_initializer("reshape_shape", np.array([1, -1]))
+ model.set_initializer("reshape_shape", np.array([1, -1], dtype=np.int64))
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
diff --git a/tests/fpgadataflow/test_convert_to_hls_layers_synthetic.py b/tests/fpgadataflow/test_convert_to_hls_layers_synthetic.py
index 155a124f416da1f3ce800cfd2be484a226af473f..35f082f55cb99c4fb6b63020f3adfb351e5effe2 100644
--- a/tests/fpgadataflow/test_convert_to_hls_layers_synthetic.py
+++ b/tests/fpgadataflow/test_convert_to_hls_layers_synthetic.py
@@ -127,12 +127,12 @@ def make_model(ch, ifmdim):
model = ModelWrapper(model)
# set initializers for scalar add/mul nodes
- model.set_initializer(add0_node.input[1], np.array([0.0]))
- model.set_initializer(add1_node.input[1], np.array([7.0]))
- model.set_initializer(add2_node.input[1], np.array([8.0]))
- model.set_initializer(mul1_node.input[1], np.array([2.0]))
- model.set_initializer(mul2_node.input[1], np.array([2.0]))
- model.set_initializer(reshape_node.input[1], np.array([1, -1]))
+ model.set_initializer(add0_node.input[1], np.array([0.0], dtype=np.float32))
+ model.set_initializer(add1_node.input[1], np.array([7.0], dtype=np.float32))
+ model.set_initializer(add2_node.input[1], np.array([8.0], dtype=np.float32))
+ model.set_initializer(mul1_node.input[1], np.array([2.0], dtype=np.float32))
+ model.set_initializer(mul2_node.input[1], np.array([2.0], dtype=np.float32))
+ model.set_initializer(reshape_node.input[1], np.array([1, -1], dtype=np.int64))
return model
diff --git a/tests/transformation/streamline/test_linear_past_eltwise.py b/tests/transformation/streamline/test_linear_past_eltwise.py
index 5655dbb5c9333241f671957780f3ef0970de860b..12633d750bb405757efca0c028dece92b289b472 100644
--- a/tests/transformation/streamline/test_linear_past_eltwise.py
+++ b/tests/transformation/streamline/test_linear_past_eltwise.py
@@ -40,6 +40,7 @@ import finn.core.onnx_exec as oxe
from finn.transformation.streamline.reorder import MoveLinearPastEltwiseAdd
export_onnx_path = "test_linear_past_eltwise.onnx"
+np_default_dtype = np.float32
# construct a synthetic graph to test:
# topk insertion, topk conversion to hls, add conversion to hls
@@ -81,10 +82,10 @@ def make_model(shape):
model = ModelWrapper(model)
# set initializers for scalar add/mul nodes
- model.set_initializer(add1_node.input[1], np.array([7.0]))
- model.set_initializer(add2_node.input[1], np.array([8.0]))
- model.set_initializer(mul1_node.input[1], np.array([3.0]))
- model.set_initializer(mul2_node.input[1], np.array([3.0]))
+ model.set_initializer(add1_node.input[1], np.array([7.0], dtype=np_default_dtype))
+ model.set_initializer(add2_node.input[1], np.array([8.0], dtype=np_default_dtype))
+ model.set_initializer(mul1_node.input[1], np.array([3.0], dtype=np_default_dtype))
+ model.set_initializer(mul2_node.input[1], np.array([3.0], dtype=np_default_dtype))
return model
diff --git a/tests/transformation/streamline/test_move_maxpool_past_multithreshold.py b/tests/transformation/streamline/test_move_maxpool_past_multithreshold.py
index d1c1daabb56aa3a5cb39221784c8547f4bcc1261..bf25eee9e685d2536faf5bd25bc7b1aa36700463 100644
--- a/tests/transformation/streamline/test_move_maxpool_past_multithreshold.py
+++ b/tests/transformation/streamline/test_move_maxpool_past_multithreshold.py
@@ -112,7 +112,7 @@ def test_move_maxpool_past_multithreshold():
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
- model.set_initializer("thres1", np.array([[0]]))
+ model.set_initializer("thres1", np.array([[0]], dtype=np.float32))
model.set_initializer(
"thres2", get_multithreshold_rand_params(*thres2_shape, seed=0)
)