Merging Methods

Merge methods (layers) are use to merge 2 or more layer's output using the methods list below.

These methods are also layers which return a layer instance. However, they normally take two or more layer's output and "merge" them into one output. These layer instance must be passed to either model.merge(method, in1, in2)or model.mergex(method, num_of_input, in1, in2, 1n3 ...). An example will be to concat the Inception structure.

---

Concat()

nnom_layer_t* Concat(int8_t axis);

Concatenate mutiple input on the selected axis.

Arguments

• axis: the axis number to concatenate in HWC format. The axis could be nagative, such as '-1' indicate the last one axis which is 'Channel'.

Return

• The concat layer instance

Notes

The concatenated axis can be different in all input layers passed to this method. Other axes must be same.

---

Mult()

nnom_layer_t* Mult(int32_t oshift);

Element wise mutiplication in all the inputs.

This layer cannot use to merge more than 2 layer, which might cause overflowing problem. 2 Mult() must be used separately if willing to multiply 3 layer's output. The output shift individually of the 2 steps must be identify individually. Please check the example below for more than 2 input.

Arguments

• oshift: the output shift of this layer.

Return

• The mult layer instance

Notes

All input layers passed to this method must have same output shape.

---

Add()

nnom_layer_t* Add(int32_t oshift);

Element wise addition in all the inputs.

This layer cannot use to merge more than 2 layer, which might cause overflowing problem. Please refer to Mult()

Arguments

• oshift: the output shift of this layer.

Return

• The add layer instance

Notes

All input layers passed to this method must have same output shape.

---

Sub()

nnom_layer_t* Sub(int32_t oshift);

Element wise substraction in all the inputs.

This layer cannot use to merge more than 2 layer, which might cause overflowing problem. Please refer to Mult()

Arguments

• oshift: the output shift of this layer.

Return

• The sub layer instance

Notes

All input layers passed to this method must have same output shape.

---

Example

Channelwise concat for Inception

    input_layer = Input(shape(INPUT_HIGHT, INPUT_WIDTH, INPUT_CH), nnom_input_data);

    // conv2d - 1 - inception
    x1 = model.hook(Conv2D(16, kernel(1, 5), stride(1, 1), PADDING_SAME, &c2_w, &c2_b), x);
    x1 = model.hook(MaxPool(kernel(1, 2), stride(1, 2), PADDING_VALID), x1);

    // conv2d - 2 - inception
    x2 = model.hook(Conv2D(16, kernel(1, 3), stride(1, 1), PADDING_SAME, &c3_w, &c3_b), x);
    x2 = model.hook(MaxPool(kernel(1, 2), stride(1, 2), PADDING_VALID), x2);

    // maxpool - 3 - inception
    x3 = model.hook(MaxPool(kernel(1, 2), stride(1, 2), PADDING_VALID), x);

    // concatenate 
    x = model.mergex(Concat(-1), 3, x1, x2, x3);

    // flatten
    x = model.hook(Flatten(), x);
    ...

Mult for 3 input (or Add, Sub)

In Keras

    #instead of 
    x = multiply([x1,x2,x3]) 

    # you must use this instead to allow sript to calculate oshift individually. 
    x = multiply([x1,x2])
    x = multiply([x,x3])

Then in NNoM

    // x = x1 * x2 * x3
    x = model.merge(Mult(oshift_1), x1, x2); 
    x = model.merge(Mult(oshift_2), x, x3);