specification_of_config_json.md

Specification of config.json [中文]

The config.json file saves configurations used to quantize floating points in coefficient.npy.

Specification

Each item in config.json stands for the configuration of one layer. Take the following code as an example:

{
    "l1": {"/* the configuration of layer l1 */"},
    "l2": {"/* the configuration of layer l2 */"},
    "l3": {"/* the configuration of layer l3 */"},
    ...
}

The key of each item is the layer name. The convert tool convert.py searches for the corresponding .npy files according to the layer name. For example, if a layer is named "l1", the tool will search for l1's filter coefficients in "l1_filter.npy". The layer name in config.json should be consistent with the layer name in the name of .npy files.

The value of each item is the layer configuration. Please fill arguments about layer configuration listed in Table 1:

Table 1: Layer Configuration Arguments

Key	Type	Value
"operation"	string	- "conv2d" - "depthwise_conv2d" - "fully_connected"
"feature_type"	string	- "s16" for int16 quantization with element_width = 16 - "s8" for int8 quantization with element_width = 8
"filter_exponent"	integer	- If filled, filter is quantized according to the equation: value_float = value_int * 2^exponent1 - If dropped2, exponent is determined according to the equation: exponent = log2(max(abs(value_float)) / 2^(element_width - 1)), while filter is quantized according to the equation: value_float = value_int * 2^exponent
"bias"	string	- "True" for adding bias - "False" and dropped for no bias
"output_exponent"	integer	Both output and bias are quantized according to the equation: value_float = value_int * 2^exponent. For now, "output_exponent" is effective only for "bias" coefficient conversion. "output_exponent" must be provided when using per-tensor quantization. If there is no "bias" in a specific layer or using per-channel quantization, "output_exponent" could be dropped.
"input_exponent"	integer	When using per-channel quantization, the exponent of bias is related to input_exponent and filter_exponent. "input_exponent" must be provided for "bias" coefficient conversion. If there is no "bias" in a specific layer or using per-tensor quantization, "output_exponent" could be dropped.
"activation"	dict	- If filled, see details in Table 2 - If dropped, no activation

Table 2: Activation Configuration Arguments

Key	Type	Value
"type"	string	- "ReLU" - "LeakyReLU" - "PReLU"
"exponent"	integer	- If filled, activation is quantized according to the equation, value_float = value_int * 2^exponent - If dropped, exponent is determined according to the equation: exponent = log2(max(abs(value_float)) / 2^(element_width - 1))

1: exponent: the number of times the base is multiplied by itself for quantization. For better understanding, please refer to Quantization Specification.

2: dropped: to leave a specific argument empty.

Example

Assume that for a one-layer model:

1. using int16 per-tensor quantization:

• layer name: "mylayer"
• operation: Conv2D(input, filter) + bias
• output_exponent: -10, exponent for the result of operation
• feature_type: s16, which means int16 quantization
• type of activation: PReLU

The config.json file should be written as:

{
	"mylayer": {
		"operation": "conv2d",
		"feature_type": "s16",
        "bias": "True",
        "output_exponent": -10,
        "activation": {
            "type": "PReLU"
        }
	}
}

"filter_exponent" and "exponent" of "activation" are dropped.
must provide "output_exponent" for bias in this layer.

2. using int8 per-tensor quantization:

• layer name: "mylayer"
• operation: Conv2D(input, filter) + bias
• output_exponent: -7, exponent for the result of this layer
• feature_type: s8
• type of activation: PReLU

The config.json file should be written as:

{
	"mylayer": {
		"operation": "conv2d",
		"feature_type": "s8",
        "bias": "True",
        "output_exponent": -7,
        "activation": {
            "type": "PReLU"
        }
	}
}

must provide "output_exponent" for bias in this layer.

3. using int8 per-channel quantization:

• layer name: "mylayer"
• operation: Conv2D(input, filter) + bias
• input_exponent: -7, exponent for the input of this layer
• feature_type: s8
• type of activation: PReLU

The config.json file should be written as:

{
	"mylayer": {
		"operation": "conv2d",
		"feature_type": "s8",
        "bias": "True",
        "input_exponent": -7,
        "activation": {
            "type": "PReLU"
        }
	}
}

must provide "input_exponent" for bias in this layer.

Meanwhile, mylayer_filter.npy, mylayer_bias.npy and mylayer_activation.npy should be ready.