Skip to content

Commit 006499b

Browse files
maltanarmaltanar
committed
[Docs] add versioning to all op docs, v2 and v1 for Trunc, overview
1 parent e044c63 commit 006499b

File tree

6 files changed

+148
-4
lines changed

6 files changed

+148
-4
lines changed

‎docs/qonnx-custom-ops/bipolar_quant_op.md‎ renamed to ‎docs/qonnx-custom-ops/bipolarquant_v1.md‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -5,7 +5,7 @@ Additionally, takes one float as input, which define the scaling.
55

66
#### Version
77

8-
This operator is not part of the ONNX standard and is not currently versioned.
8+
The description of this operator in this document corresponds to `qonnx.custom_ops.general` opset version 1.
99

1010
#### Attributes
1111

‎docs/qonnx-custom-ops/floatquant_op.md‎ renamed to ‎docs/qonnx-custom-ops/floatquant_v1.md‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -16,7 +16,7 @@ special (symbolic) values. This makes it nontrivial to infer the maximum represe
1616

1717
#### Version
1818

19-
This operator is not part of the ONNX standard and is not currently versioned.
19+
The description of this operator in this document corresponds to `qonnx.custom_ops.general` opset version 1.
2020

2121
#### Attributes
2222

‎docs/qonnx-custom-ops/intquant_op.md‎ renamed to ‎docs/qonnx-custom-ops/intquant_v1.md‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -9,11 +9,11 @@ rounding_mode defines how quantized values are rounded.
99

1010
Notes:
1111
* This operator was previously named `Quant` but is renamed to `IntQuant` to distinguish it from `FloatQuant`. For a transition period, qonnx will transparently handle `Quant` as `IntQuant` for backwards compatibility reasons, but only `IntQuant` should be used for new models.
12-
* This operator does not work for binary or bipolar quantization, for this purpose the simpler BipolarQuant node exists.
12+
* This operator does not work for binary or bipolar quantization, for this purpose the simpler `BipolarQuant` node exists.
1313

1414
#### Version
1515

16-
This operator is not part of the ONNX standard and is not currently versioned.
16+
The description of this operator in this document corresponds to `qonnx.custom_ops.general` opset version 1.
1717

1818
#### Attributes
1919

‎docs/qonnx-custom-ops/overview.md‎

Lines changed: 13 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,13 @@
1+
## Operator Schemas
2+
3+
This file lists the QONNX custom operators, similar to `Operators.md` for the ONNX standard.
4+
It is manually updated, since QONNX custom operators are relatively few in number.
5+
6+
### qonnx.custom_op.general
7+
8+
|**Operator**|**Since version**||
9+
|-|-|-|
10+
|<a href="bipolarquant_v1.md">BipolarQuant</a>|<a href="bipolarquant_v1.md">1</a>|
11+
|<a href="floatquant_v1.md">FloatQuant</a>|<a href="floatquant_v1.md">1</a>|
12+
|<a href="intquant_v1.md">IntQuant</a>|<a href="intquant_v1.md">1</a>|
13+
|<a href="trunc_v2.md">Trunc</a>|<a href="trunc_v2.md">2</a>, <a href="trunc_v1.md">1</a>|

‎docs/qonnx-custom-ops/trunc_v1.md‎

Lines changed: 131 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,131 @@
1+
### <a name="Trunc"></a><a name="abs">**Trunc**</a>
2+
3+
Truncates the values of one input data (Tensor<T>) at a specified bitwidth and produces one output data (Tensor<T>).
4+
Additionally, takes four float tensors as input, which define the scale, zero-point, input bit-width and output bit-width of the quantization.
5+
The attribute rounding_mode defines how truncated values are rounded.
6+
7+
#### Version
8+
9+
The description of this operator in this document corresponds to `qonnx.custom_ops.general` opset version 1.
10+
11+
#### Attributes
12+
13+
<dl>
14+
<dt><tt>rounding_mode</tt> : string (default is "FLOOR")</dt>
15+
<dd>Defines how rounding should be applied during truncation. Currently available modes are: "ROUND", "CEIL" and "FLOOR". Here "ROUND" implies a round-to-even operation. Lowercase variants for the rounding mode string are also supported: "round", "ceil", "floor".</dd>
16+
</dl>
17+
18+
#### Inputs
19+
20+
<dl>
21+
<dt><tt>X</tt> (differentiable) : tensor(float32)</dt>
22+
<dd>input tensor to truncate</dd>
23+
<dt><tt>scale</tt> : float32</dt>
24+
<dd>The scale factor</dd>
25+
<dt><tt>zeropt</tt> : float32</dt>
26+
<dd>The zero-point</dd>
27+
<dt><tt>in_bitwidth</tt> : int32</dt>
28+
<dd>The number of bits used at the input of the truncation</dd>
29+
<dt><tt>out_bitwidth</tt> : int32</dt>
30+
<dd>The number of bits used at the output of the truncation</dd>
31+
</dl>
32+
33+
34+
#### Outputs
35+
36+
<dl>
37+
<dt><tt>Y</tt> (differentiable) : tensor(float32)</dt>
38+
<dd>Output tensor</dd>
39+
</dl>
40+
41+
42+
#### Examples
43+
<details>
44+
<summary>Trunc</summary>
45+
46+
```python
47+
from onnx import helper
48+
import numpy as np
49+
50+
# Define node settings and input
51+
x = np.random.randn(100).astype(np.float32)*10.
52+
scale = np.array(1.)
53+
zeropt = np.array(0.)
54+
in_bitwidth = np.array(10)
55+
out_bitwidth = np.array(4)
56+
rounding_mode = "ROUND"
57+
58+
# Create node
59+
node = helper.make_node(
60+
'Trunc',
61+
domain='finn.custom_op.general',
62+
inputs=['x', 'scale', 'zeropt', 'in_bitwidth', 'out_bitwidth'],
63+
outputs=['y'],
64+
rounding_mode=rounding_mode,
65+
)
66+
67+
# Execute the same settings with the reference implementation (trunc)
68+
# See the sample implementation for more details on trunc.
69+
output_ref = trunc(inp_tensor, scale, zeropt, in_bitwidth, out_bitwidth, rounding_mode)
70+
71+
# Execute node and compare
72+
expect(node, inputs=[x, scale, zeropt, bitwidth], outputs=[output_ref], name='test_trunc')
73+
74+
```
75+
76+
</details>
77+
78+
79+
#### Sample Implementation
80+
81+
<details>
82+
<summary>Trunc</summary>
83+
84+
```python
85+
# SPDX-License-Identifier: Apache-2.0
86+
87+
from __future__ import absolute_import
88+
from __future__ import division
89+
from __future__ import print_function
90+
from __future__ import unicode_literals
91+
92+
import numpy as np
93+
94+
def trunc(inp_tensor, scale, zeropt, input_bit_width, output_bit_width, rounding_mode):
95+
# Port of TruncIntQuant class from Brevitas: https://bit.ly/3wzIpTR
96+
97+
# Scaling
98+
y = inp_tensor / scale
99+
y = y + zeropt
100+
# Rounding
101+
y = np.round(y)
102+
# Truncate
103+
trunc_bit_width = input_bit_width - output_bit_width
104+
trunc_scale = 2.0 ** trunc_bit_width
105+
y = y / trunc_scale
106+
107+
# To int
108+
rounding_fx = resolve_rounding_mode(rounding_mode)
109+
y = rounding_fx(y)
110+
111+
# Rescale
112+
y = y - zeropt
113+
y = y * scale
114+
115+
return y
116+
117+
def resolve_rounding_mode(mode_string):
118+
"""Resolve the rounding mode string of Quant and Trunc ops
119+
to the corresponding numpy functions."""
120+
if mode_string == "ROUND":
121+
return np.round
122+
elif mode_string == "CEIL":
123+
return np.ceil
124+
elif mode_string == "FLOOR":
125+
return np.floor
126+
else:
127+
raise ValueError(f"Could not resolve rounding mode called: {mode_string}")
128+
129+
```
130+
131+
</details>

‎docs/qonnx-custom-ops/trunc_op.md‎ renamed to ‎docs/qonnx-custom-ops/trunc_v2.md‎

File renamed without changes.

0 commit comments

Comments
 (0)