Specify Range-4 operation (#1180)

diff --git a/docs/ops/generation/Range_1.md b/docs/ops/generation/Range_1.md
index e1097f232782583a4e402a2287adb30cb0b20960..06a16cc1899f6c1145efe4b2f21bcc5869fbc0b7 100644 (file)
--- a/docs/ops/generation/Range_1.md
+++ b/docs/ops/generation/Range_1.md
@@ -12,13 +12,13 @@ No attributes available.
 
 **Inputs**:
 
-* **1**: "start" - If a value is not given then *start* = 0. A scalar of type T. **Required.**
+* **1**: "start" - A scalar of type T. **Required.**
 * **2**: "stop" - A scalar of type T. **Required.**
-* **3**: "step" - If a value is not given then *step* = 1. A scalar of type T. **Required.**
+* **3**: "step" - A scalar of type T. **Required.**
 
 **Outputs**:
 
-* **1**: A tensor with type matching 2nd tensor.
+* **1**: A tensor of type T.
 
 **Types**
 
@@ -28,8 +28,22 @@ No attributes available.
 
 *Range* operation generates a sequence of numbers starting from the value in the first input (start) up to but not including the value in the second input (stop) with a step equal to the value in the third input, according to the following formula:
 
+For a positive `step`:
+
+\f[
+start<=val[i]<stop,
+\f]
+
+for a negative `step`:
+
+\f[
+start>=val[i]>stop,
+\f]
+
+where
+
 \f[
-[start, start + step, start + 2 * step, ..., start + K * step], where K is the maximal integer value that satisfies condition start + K*step < stop, then step is positive value and start + K*step > stop, then step is negative value.
+val[i]=start+i*step
 \f]
 
 **Examples**

diff --git a/docs/ops/generation/Range_4.md b/docs/ops/generation/Range_4.md
new file mode 100644 (file)
index 0000000..33e1b7b
--- /dev/null
+++ b/docs/ops/generation/Range_4.md
@@ -0,0 +1,127 @@
+## Range<a name="Range"></a>
+
+**Versioned name**: *Range-4*
+
+**Category**: Generation
+
+**Short description**: *Range* operation generates a sequence of numbers according input values [start, stop) with a step.
+
+**Attributes**:
+
+* *output_type*
+
+  * **Description**: the output tensor type
+  * **Range of values**: any numeric type
+  * **Type**: string
+  * **Default value**: None
+  * **Required**: *Yes*
+
+**Inputs**:
+
+* **1**: "start" - A scalar of type T1. **Required.**
+* **2**: "stop" - A scalar of type T2. **Required.**
+* **3**: "step" - A scalar of type T3. If `step` is equal to zero after casting to `output_type`, behavior is undefined. **Required.**
+
+**Outputs**:
+
+* **1**: A tensor with type specified by attribute *output_type*.
+
+**Types**
+
+* *T1*, *T2*, *T3*: any numeric type.
+
+**Detailed description**:
+
+*Range* operation generates a sequence of numbers starting from the value in the first input (`start`) up to but not including the value in the second input (`stop`) with a `step` equal to the value in the third input, according to the following formula:
+
+For a positive `step`:
+
+\f[
+start<=val[i]<stop,
+\f]
+
+for a negative `step`:
+
+\f[
+start>=val[i]>stop,
+\f]
+
+the i-th element is calculated by the following formula:
+
+\f[
+val[i+1]=val[i]+step.
+\f]
+
+The calculations are done after casting all values to `accumulate_type(output_type)`. `accumulate_type` is a type that have better or equal accuracy for accumulation than `output_type` on current hardware, e.g. `fp64` for `fp16`. The number of elements is calculated in the floating point type according to the following formula:
+
+\f[
+max(ceil((end − start) / step), 0)
+\f]
+
+This is aligned with PyTorch's operation `torch.arange`, to align with tensorflow operation `tf.range` all inputs must be casted to `output_type` before calling *Range*. The rounding for casting values are done towards zero.
+
+**Examples**
+
+*Example 1: positive step*
+
+```xml
+<layer ... type="Range">
+    <data output_type="i32">
+    <input>
+        <port id="0">  <!-- start value: 2 -->
+        </port>
+        <port id="1">  <!-- stop value: 23 -->
+        </port>
+        <port id="2">  <!-- step value: 3 -->
+        </port>
+    </input>
+    <output>
+        <port id="3">
+            <dim>7</dim> <!-- [ 2,  5,  8, 11, 14, 17, 20] -->
+        </port>
+    </output>
+</layer>
+```
+
+*Example 2: negative step*
+
+```xml
+<layer ... type="Range">
+    <data output_type="i32">
+    <input>
+        <port id="0">  <!-- start value: 23 -->
+        </port>
+        <port id="1">  <!-- stop value: 2 -->
+        </port>
+        <port id="2">  <!-- step value: -3 -->
+        </port>
+    </input>
+    <output>
+        <port id="3">
+            <dim>7</dim> <!-- [23, 20, 17, 14, 11, 8, 5] -->
+        </port>
+    </output>
+</layer>
+```
+
+*Example 3: floating point*
+
+```xml
+<layer ... type="Range">
+    <data output_type="f32">
+    <input>
+        <port id="0">  <!-- start value: 1 -->
+        </port>
+        <port id="1">  <!-- stop value: 2.5 -->
+        </port>
+        <port id="2">  <!-- step value: 0.5 -->
+        </port>
+    </input>
+    <output>
+        <port id="3">
+            <dim>3</dim> <!-- [ 1.0,  1.5,  2.0] -->
+        </port>
+    </output>
+</layer>
+```
+

diff --git a/docs/ops/opset4.md b/docs/ops/opset4.md
index 0b4f66423563b548c9a1b91b62ab1b6f78217a30..8f1852aa3a2254522caff02355b452afed33038e 100644 (file)
--- a/docs/ops/opset4.md
+++ b/docs/ops/opset4.md
@@ -96,7 +96,7 @@ declared in `namespace opset4`.
 * [PriorBox](detection/PriorBox_1.md)
 * [Proposal](detection/Proposal_1.md)
 * [PSROIPooling](detection/PSROIPooling_1.md)
-* [Range](generation/Range_1.md)
+* [Range](generation/Range_4.md)
 * [ReLU](activation/ReLU_1.md)
 * [ReadValue](infrastructure/ReadValue_3.md)
 * [ReduceLp](reduction/ReduceLp_4.md)