Skip to content

Commit

Permalink
add node tests for quantized ops (onnx#1944)
Browse files Browse the repository at this point in the history 
* add node tests for quantized ops

* cosmetic changes

* more cosmetic changes

* typecheck error fixes

* update docs

* add node tests for qlinearconv
  • Loading branch information
@askhade @linkerzhang
askhade authored and linkerzhang committed Apr 18, 2019
1 parent bff40c8 commit b418295
Show file tree
Hide file tree
Showing 64 changed files with 1,016 additions and 19 deletions.
230 changes: 230 additions & 0 deletions docs/Operators.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2492,6 +2492,42 @@ This version of the operator has been available since version 10 of the default
</dl>


#### Examples

<details>
<summary>convinteger</summary>

```python

x = np.array([2, 3, 4, 5, 6, 7, 8, 9, 10]).astype(np.uint8).reshape((1, 1, 3, 3))
x_zero_point = np.array([1]).astype(np.uint8)
w = np.array([1, 1, 1, 1]).astype(np.uint8).reshape((1, 1, 2, 2))

y = np.array([12, 16, 24, 28]).astype(np.int32).reshape(1, 1, 2, 2)

# ConvInteger without padding
convinteger_node = onnx.helper.make_node('ConvInteger',
inputs=['x', 'w', 'x_zero_point'],
outputs=['y'])

expect(convinteger_node, inputs=[x, w, x_zero_point], outputs=[y],
name='test_basic_convinteger')

# ConvInteger with padding
y_with_padding = np.array([1, 3, 5, 3, 5, 12, 16, 9, 11, 24, 28, 15, 7, 15, 17, 9]).astype(np.int32).reshape((1, 1, 4, 4))

convinteger_node_with_padding = onnx.helper.make_node('ConvInteger',
inputs=['x', 'w', 'x_zero_point'],
outputs=['y'],
pads=[1, 1, 1, 1],)

expect(convinteger_node_with_padding, inputs=[x, w, x_zero_point], outputs=[y_with_padding],
name='test_convinteger_with_padding')
```

</details>


### <a name="ConvTranspose"></a><a name="convtranspose">**ConvTranspose**</a>

The convolution transpose operator consumes an input tensor and a filter,
Expand Down Expand Up @@ -3105,6 +3141,29 @@ This version of the operator has been available since version 10 of the default
</dl>


#### Examples

<details>
<summary>dequantizelinear</summary>

```python
node = onnx.helper.make_node('DequantizeLinear',
inputs=['x', 'x_scale', 'x_zero_point'],
outputs=['y'],)

# scalar zero point and scale
x = np.array([0, 3, 128, 255]).astype(np.uint8)
x_scale = np.array([2], dtype=np.float32)
x_zero_point = np.array([128], dtype=np.uint8)
y = np.array([-256, -250, 0, 254], dtype=np.float32)

expect(node, inputs=[x, x_scale, x_zero_point], outputs=[y],
name='test_dequantizelinear')
```

</details>


### <a name="Div"></a><a name="div">**Div**</a>

Performs element-wise binary division (with Numpy-style broadcasting support).
Expand Down Expand Up @@ -6319,6 +6378,41 @@ This version of the operator has been available since version 10 of the default
</dl>


#### Examples

<details>
<summary>matmulinteger</summary>

```python
node = onnx.helper.make_node('MatMulInteger',
inputs=['A', 'B', 'a_zero_point', 'b_zero_point'],
outputs=['Y'],)

A = np.array([[11, 7, 3],
[10, 6, 2],
[9, 5, 1],
[8, 4, 0], ], dtype=np.uint8)

a_zero_point = np.array([12], dtype=np.uint8)

B = np.array([[1, 4],
[2, 5],
[3, 6], ], dtype=np.uint8)

b_zero_point = np.array([0], dtype=np.uint8)

output = np.array([[-38, -83],
[-44, -98],
[-50, -113],
[-56, -128], ], dtype=np.int32)

expect(node, inputs=[A, B, a_zero_point, b_zero_point], outputs=[output],
name='test_matmulinteger')
```

</details>


### <a name="Max"></a><a name="max">**Max**</a>

Element-wise max of each of the input tensors (with Numpy-style broadcasting support).
Expand Down Expand Up @@ -8222,6 +8316,50 @@ This version of the operator has been available since version 10 of the default
</dl>


#### Examples

<details>
<summary>qlinearconv</summary>

```python
node = onnx.helper.make_node('QLinearConv',
inputs=['x', 'x_scale', 'x_zero_point', 'w', 'w_scale', 'w_zero_point', 'y_scale', 'y_zero_point'],
outputs=['y'],)

x = np.array([[255, 174, 162, 25, 203, 168, 58],
[15, 59, 237, 95, 129, 0, 64],
[56, 242, 153, 221, 168, 12, 166],
[232, 178, 186, 195, 237, 162, 237],
[188, 39, 124, 77, 80, 102, 43],
[127, 230, 21, 83, 41, 40, 134],
[255, 154, 92, 141, 42, 148, 247], ], dtype=np.uint8).reshape((1, 1, 7, 7))

x_scale = np.array([0.00369204697], dtype=np.float32)
x_zero_point = np.array([132], dtype=np.uint8)

w = np.array([0], dtype=np.uint8).reshape((1, 1, 1, 1))

w_scale = np.array([0.00172794575], dtype=np.float32)
w_zero_point = np.array([255], dtype=np.uint8)

y_scale = np.array([0.00162681262], dtype=np.float32)
y_zero_point = np.array([123], dtype=np.uint8)

output = np.array([[0, 81, 93, 230, 52, 87, 197],
[240, 196, 18, 160, 126, 255, 191],
[199, 13, 102, 34, 87, 243, 89],
[23, 77, 69, 60, 18, 93, 18],
[67, 216, 131, 178, 175, 153, 212],
[128, 25, 234, 172, 214, 215, 121],
[0, 101, 163, 114, 213, 107, 8], ], dtype=np.uint8).reshape((1, 1, 7, 7))

expect(node, inputs=[x, x_scale, x_zero_point, w, w_scale, w_zero_point, y_scale, y_zero_point], outputs=[output],
name='test_qlinearconv')
```

</details>


### <a name="QLinearMatMul"></a><a name="qlinearmatmul">**QLinearMatMul**</a>

Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html.
Expand Down Expand Up @@ -8277,6 +8415,76 @@ This version of the operator has been available since version 10 of the default
</dl>


#### Examples

<details>
<summary>qlinearmatmul</summary>

```python
node = onnx.helper.make_node('QLinearMatMul',
inputs=['a', 'a_scale', 'a_zero_point', 'b', 'b_scale', 'b_zero_point', 'y_scale', 'y_zero_point'],
outputs=['y'],)

#2D
a = np.array([[208, 236, 0, 238],
[3, 214, 255, 29], ], dtype=np.uint8)

a_scale = np.array([0.0066], dtype=np.float32)
a_zero_point = np.array([113], dtype=np.uint8)

b = np.array([[152, 51, 244],
[60, 26, 255],
[0, 127, 246],
[127, 254, 247]], dtype=np.uint8)

b_scale = np.array([0.00705], dtype=np.float32)
b_zero_point = np.array([114], dtype=np.uint8)

y_scale = np.array([0.0107], dtype=np.float32)
y_zero_point = np.array([118], dtype=np.uint8)

output = np.array([[168, 115, 255],
[1, 66, 151], ], dtype=np.uint8)

expect(node, inputs=[a, a_scale, a_zero_point, b, b_scale, b_zero_point, y_scale, y_zero_point], outputs=[output],
name='test_qlinearmatmul_2D')

#3D
a = np.array([[[208, 236, 0, 238],
[3, 214, 255, 29]],
[[208, 236, 0, 238],
[3, 214, 255, 29]]], dtype=np.uint8)

a_scale = np.array([0.0066], dtype=np.float32)
a_zero_point = np.array([113], dtype=np.uint8)

b = np.array([[[152, 51, 244],
[60, 26, 255],
[0, 127, 246],
[127, 254, 247]],
[[152, 51, 244],
[60, 26, 255],
[0, 127, 246],
[127, 254, 247]]], dtype=np.uint8)

b_scale = np.array([0.00705], dtype=np.float32)
b_zero_point = np.array([114], dtype=np.uint8)

y_scale = np.array([0.0107], dtype=np.float32)
y_zero_point = np.array([118], dtype=np.uint8)

output = np.array([[[168, 115, 255],
[1, 66, 151]],
[[168, 115, 255],
[1, 66, 151]]], dtype=np.uint8)

expect(node, inputs=[a, a_scale, a_zero_point, b, b_scale, b_zero_point, y_scale, y_zero_point], outputs=[output],
name='test_qlinearmatmul_3D')
```

</details>


### <a name="QuantizeLinear"></a><a name="quantizelinear">**QuantizeLinear**</a>

The linear per-tensor/layer quantization operator. It consumes a high precision tensor, a scale, a zero point to compute the low precision / quantized tensor.
Expand Down Expand Up @@ -8315,6 +8523,28 @@ This version of the operator has been available since version 10 of the default
</dl>


#### Examples

<details>
<summary>quantizelinear</summary>

```python
node = onnx.helper.make_node('QuantizeLinear',
inputs=['x', 'y_scale', 'y_zero_point'],
outputs=['y'],)

x = np.array([0, 2, 3, 1000, -254, -1000]).astype(np.float32)
y_scale = np.array([2], dtype=np.float32)
y_zero_point = np.array([128], dtype=np.uint8)
y = np.array([128, 129, 130, 255, 1, 0]).astype(np.uint8)

expect(node, inputs=[x, y_scale, y_zero_point], outputs=[y],
name='test_quantizelinear')
```

</details>


### <a name="RNN"></a><a name="rnn">**RNN**</a>

Computes an one-layer simple RNN. This operator is usually supported
Expand Down
Loading

0 comments on commit b418295

Please sign in to comment.