ailayer_dense_default.h File Reference

Default implementation of the Dense layer . More...

Go to the source code of this file.

Typedefs
typedef struct ailayer_dense	ailayer_dense_f32_t
typedef struct ailayer_dense	ailayer_dense_q31_t
typedef struct ailayer_dense	ailayer_dense_q7_t

Functions
ailayer_t *	ailayer_dense_f32_default (ailayer_dense_f32_t *layer, ailayer_t *input_layer)
	Initializes and connect a Dense layer with the F32 default implementation. More...
ailayer_t *	ailayer_dense_wt_f32_default (ailayer_dense_f32_t *layer, ailayer_t *input_layer)
	Initializes and connect a Dense layer with the F32 default implementation. More...
ailayer_t *	ailayer_dense_q31_default (ailayer_dense_q31_t *layer, ailayer_t *input_layer)
	Initializes and connect a Dense layer with the Q31 default implementation. More...
ailayer_t *	ailayer_dense_q7_default (ailayer_dense_q7_t *layer, ailayer_t *input_layer)
	Initializes and connect a Dense layer with the Q7 default implementation. More...
ailayer_t *	ailayer_dense_wt_q7_default (ailayer_dense_q7_t *layer, ailayer_t *input_layer)
	Initializes and connect a Dense layer with the Q7 default implementation for transposed weights tensor. More...
void	ailayer_dense_init_params_f32_default (ailayer_t *self)
	F32 default implementation of the ailayer.init_params function for the Conv2D layer More...
void	ailayer_dense_init_params_q31_default (ailayer_t *self)
	Q31 default implementation of the ailayer.init_params function for the Conv2D layer More...
void	ailayer_dense_quantize_q7_from_f32 (ailayer_dense_f32_t *f32_layer_ptr, ailayer_dense_q7_t *q7_layer_ptr)
	Convert a F32 dense layer to a Q7 representation. More...

Detailed Description

Default implementation of the Dense layer .

Version

2.2.0

Copyright

Copyright (C) 2020-2023 Fraunhofer Institute for Microelectronic Circuits and Systems. All rights reserved.

AIfES is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more details.

You should have received a copy of the GNU Affero General Public License along with this program. If not, see https://www.gnu.org/licenses/.

Hardware independent implementations of the Dense layer in F32 , Q31 and Q7 data-type. For more information about the Dense layer refer to ailayer_dense.h.

Function Documentation

ailayer_dense_f32_default()

ailayer_t* ailayer_dense_f32_default	(	ailayer_dense_f32_t *	layer,
		ailayer_t *	input_layer
	)

Initializes and connect a Dense layer with the F32 default implementation.

Example: Create the layer structure with pretrained weights:
In C:

// Use constant data only for inference. For training remove the const qualifier!!
const float weights_data_dense[] = {-10.1164f, -8.4212f, 5.4396f, 7.297f, -7.6482f, -9.0155f};
const float bias_data_dense[] = {-2.9653f,  2.3677f, -1.5968f};
ailayer_dense_f32_t dense_layer = {
    .neurons = 3,
    .weights.data = (float *) weights_data_dense,
    .bias.data = (float *) bias_data_dense
};

In C, C++ and on Arduino:

// Use constant data only for inference. For training remove the const qualifier!!
const float weights_data_dense[] = {-10.1164f, -8.4212f, 5.4396f, 7.297f, -7.6482f, -9.0155f};
const float bias_data_dense[] = {-2.9653f,  2.3677f, -1.5968f};
ailayer_dense_f32_t dense_layer = AILAYER_DENSE_F32_M(3, weights_data_dense, bias_data_dense);

Example: Create the layer structure for automatic parameter distribution:
In C:

ailayer_dense_f32_t dense_layer = {
    .neurons = 3
};

In C, C++ and on Arduino:

ailayer_dense_f32_t dense_layer = AILAYER_DENSE_F32_A(3);

Example: Initialize and connect the layer:

x = ailayer_dense_f32_default(&dense_layer, x);

Parameters

*layer	The layer structure to initialize.
*input_layer	The prior layer.

Returns

The (successfully) initialized layer structure.

ailayer_dense_init_params_f32_default()

void ailayer_dense_init_params_f32_default

(

ailayer_t *

self

)

F32 default implementation of the ailayer.init_params function for the Conv2D layer

Implementation of ailayer.init_params.

The function will initialize the weights and bias depending on the following activation function.

Activation function	Weights-init	Bias-init
None, tanh, logistic, softmax	Glorot	Zeros
ReLu and variants	He	Zeros

Parameters

*self

The layer structure

ailayer_dense_init_params_q31_default()

void ailayer_dense_init_params_q31_default

(

ailayer_t *

self

)

Q31 default implementation of the ailayer.init_params function for the Conv2D layer

Implementation of ailayer.init_params.

The function will initialize the weights and bias depending on the following activation function.

Activation function	Weights-init	Bias-init
None, tanh, logistic, softmax	Glorot	Zeros
ReLu and variants	He	Zeros

Parameters

*self

The layer structure

ailayer_dense_q31_default()

ailayer_t* ailayer_dense_q31_default	(	ailayer_dense_q31_t *	layer,
		ailayer_t *	input_layer
	)

Initializes and connect a Dense layer with the Q31 default implementation.

Example: Create the layer structure with pretrained weights:
In C:

// Use constant data only for inference. For training remove the const qualifier!!
// Weights (32 bit quantized)
const aimath_q31_params_t weights_q_params_dense = { .shift = 26, .zero_point = 0 };
const int32_t weights_data_dense[] = {-484203965, -294377834, -415073361, 234042792,  514922145, -408068817};
// Bias (32 bit quantized)
const aimath_q31_params_t bias_q_params_dense = { .shift = 26, .zero_point = 0 };
const int32_t bias_data_dense[] = {-91841604, 133375416, 185587399};
// Result (32 bit quantized)
const aimath_q31_params_t result_q_params_dense = { .shift = 24, .zero_point = 0 };
ailayer_dense_q31_t dense_layer = {
    .neurons = 3,
    .weights = {
        .tensor_params = (aimath_q31_params_t *) &weights_q_params_dense,
        .data = (int32_t *) weights_data_dense
    },
    .bias = {
        .tensor_params = (aimath_q31_params_t *) &bias_q_params_dense,
        .data = (int32_t *) bias_data_dense
    },
    .base.result.tensor_params = (aimath_q31_params_t *) &result_q_params_dense
};

In C, C++ and on Arduino:

// Use constant data only for inference. For training remove the const qualifier!!
// Weights (32 bit quantized)
const aimath_q31_params_t weights_q_params_dense = { .shift = 26, .zero_point = 0 };
const int32_t weights_data_dense[] = {-484203965, -294377834, -415073361, 234042792,  514922145, -408068817};
// Bias (32 bit quantized)
const aimath_q31_params_t bias_q_params_dense = { .shift = 26, .zero_point = 0 };
const int32_t bias_data_dense[] = {-91841604, 133375416, 185587399};
// Result (32 bit quantized)
const aimath_q31_params_t result_q_params_dense = { .shift = 24, .zero_point = 0 };
ailayer_dense_q31_t dense_layer = AILAYER_DENSE_Q31_M(3,
                                                              weights_data_dense, &weights_q_params_dense,
                                                              bias_data_dense, &bias_q_params_dense,
                                                              &result_q_params_dense);

Example: Create the layer structure for automatic parameter distribution:
In C:

ailayer_dense_q31_t dense_layer = {
    .neurons = 3
};

In C, C++ and on Arduino:

ailayer_dense_q31_t dense_layer = AILAYER_DENSE_Q31_A(3);

Example: Initialize and connect the layer:

x = ailayer_dense_q31_default(&dense_layer, x);

Parameters

*layer	The layer structure to initialize.
*input_layer	The prior layer.

Returns

The (successfully) initialized layer structure.

ailayer_dense_q7_default()

ailayer_t* ailayer_dense_q7_default	(	ailayer_dense_q7_t *	layer,
		ailayer_t *	input_layer
	)

Initializes and connect a Dense layer with the Q7 default implementation.

Example: Create the layer structure with pretrained weights:
In C:

// Use constant data only for inference. For training remove the const qualifier!!
// Weights (8 bit quantized)
const aimath_q7_params_t weights_q_params_dense = { .shift = 3, .zero_point = 0 };
const int8_t weights_data_dense[] = {-81, -67, 44, 58, -61, -72};
// Bias (32 bit quantized)
const aimath_q31_params_t bias_q_params_dense = { .shift = 10, .zero_point = 0 };
const int32_t bias_data_dense[] = {-3036,  2425, -1635};
// Result (8 bit quantized)
const aimath_q7_params_t result_q_params_dense = { .shift = 3, .zero_point = 41 };
ailayer_dense_q7_t dense_layer = {
    .neurons = 3,
    .weights = {
        .tensor_params = (aimath_q7_params_t *) &weights_q_params_dense,
        .data = (int8_t *) weights_data_dense
    },
    .bias = {
        .tensor_params = (aimath_q31_params_t *) &bias_q_params_dense,
        .data = (int32_t *) bias_data_dense
    },
    .base.result.tensor_params = (aimath_q7_params_t *) &result_q_params_dense
};

In C, C++ and on Arduino:

// Use constant data only for inference. For training remove the const qualifier!!
// Weights (8 bit quantized)
const aimath_q7_params_t weights_q_params_dense = { .shift = 3, .zero_point = 0 };
const int8_t weights_data_dense[] = {-81, -67, 44, 58, -61, -72};
// Bias (32 bit quantized)
const aimath_q31_params_t bias_q_params_dense = { .shift = 10, .zero_point = 0 };
const int32_t bias_data_dense[] = {-3036,  2425, -1635};
// Result (8 bit quantized)
const aimath_q7_params_t result_q_params_dense = { .shift = 3, .zero_point = 41 };
ailayer_dense_q7_t dense_layer = AILAYER_DENSE_Q7_M(3,
                                                    weights_data_dense, &weights_q_params_dense,
                                                    bias_data_dense, &bias_q_params_dense,
                                                    &result_q_params_dense);

Example: Create the layer structure for automatic parameter distribution:
In C:

ailayer_dense_q7_t dense_layer = {
    .neurons = 3
};

In C, C++ and on Arduino:

ailayer_dense_q7_t dense_layer = AILAYER_DENSE_Q7_A(3);

Example: Initialize and connect the layer:

x = ailayer_dense_q7_default(&dense_layer, x);

Parameters

*layer	The layer structure to initialize.
*input_layer	The prior layer.

Returns

The (successfully) initialized layer structure.

ailayer_dense_quantize_q7_from_f32()

void ailayer_dense_quantize_q7_from_f32	(	ailayer_dense_f32_t *	f32_layer_ptr,
		ailayer_dense_q7_t *	q7_layer_ptr
	)

Convert a F32 dense layer to a Q7 representation.

The weights get 8 bit quantied and the bias gets 32 bit quantized for optimal results.

Quantizsation parameters for the previous layer need to be calculated before calling this function.

Parameters

*f32_layer_ptr	The source layer structure.
*q7_layer_ptr	The destination layer structure.

ailayer_dense_wt_f32_default()

ailayer_t* ailayer_dense_wt_f32_default	(	ailayer_dense_f32_t *	layer,
		ailayer_t *	input_layer
	)

Initializes and connect a Dense layer with the F32 default implementation.

This implementation is the same as ailayer_dense_f32_default() but with a transposed weights matrix/tensor. This may lead to a speedup in runtime and makes hardware accelerator usage easier.

Example: Create the layer structure with pretrained weights:
In C:

// Use constant data only for inference. For training remove the const qualifier!!
const float weights_data_dense[] = {-10.1164f, 7.297f, -8.4212f, -7.6482f, 5.4396f, -9.0155f};
const float bias_data_dense[] = {-2.9653f,  2.3677f, -1.5968f};
ailayer_dense_f32_t dense_layer = {
    .neurons = 3,
    .weights.data = (float *) weights_data_dense,
    .bias.data = (float *) bias_data_dense
};

In C, C++ and on Arduino:

// Use constant data only for inference. For training remove the const qualifier!!
const float weights_data_dense[] = {-10.1164f, 7.297f, -8.4212f, -7.6482f, 5.4396f, -9.0155f};
const float bias_data_dense[] = {-2.9653f,  2.3677f, -1.5968f};
ailayer_dense_f32_t dense_layer = AILAYER_DENSE_F32_M(3, weights_data_dense, bias_data_dense);

Example: Create the layer structure for automatic parameter distribution:
In C:

ailayer_dense_f32_t dense_layer = {
    .neurons = 3
};

In C, C++ and on Arduino:

ailayer_dense_f32_t dense_layer = AILAYER_DENSE_F32_A(3);

Example: Initialize and connect the layer:

x = ailayer_dense_f32_wt_default(&dense_layer, x);

Parameters

*layer	The layer structure to initialize.
*input_layer	The prior layer.

Returns

The (successfully) initialized layer structure.

ailayer_dense_wt_q7_default()

ailayer_t* ailayer_dense_wt_q7_default	(	ailayer_dense_q7_t *	layer,
		ailayer_t *	input_layer
	)

Initializes and connect a Dense layer with the Q7 default implementation for transposed weights tensor.

This implementation is the same as ailayer_dense_q7_default() but with a transposed weights matrix/tensor. This may lead to a speedup in runtime and makes hardware accelerator usage easier.

Example: Create the layer structure with pretrained weights:
In C:

// Use constant data only for inference. For training remove the const qualifier!!
// Weights (8 bit quantized)
const aimath_q7_params_t weights_q_params_dense = { .shift = 3, .zero_point = 0 };
const int8_t weights_data_dense[] = {-81, 58, -67, -61, 44, -72};
// Bias (32 bit quantized)
const aimath_q31_params_t bias_q_params_dense = { .shift = 10, .zero_point = 0 };
const int32_t bias_data_dense[] = {-3036,  2425, -1635};
// Result (8 bit quantized)
const aimath_q7_params_t result_q_params_dense = { .shift = 3, .zero_point = 41 };
ailayer_dense_q7_t dense_layer = {
    .neurons = 3,
    .weights = {
        .tensor_params = (aimath_q7_params_t *) &weights_q_params_dense,
        .data = (int8_t *) weights_data_dense
    },
    .bias = {
        .tensor_params = (aimath_q31_params_t *) &bias_q_params_dense,
        .data = (int32_t *) bias_data_dense
    },
    .base.result.tensor_params = (aimath_q7_params_t *) &result_q_params_dense
};

In C, C++ and on Arduino:

// Use constant data only for inference. For training remove the const qualifier!!
// Weights (8 bit quantized)
const aimath_q7_params_t weights_q_params_dense = { .shift = 3, .zero_point = 0 };
const int8_t weights_data_dense[] = {-81, 58, -67, -61, 44, -72};
// Bias (32 bit quantized)
const aimath_q31_params_t bias_q_params_dense = { .shift = 10, .zero_point = 0 };
const int32_t bias_data_dense[] = {-3036,  2425, -1635};
// Result (8 bit quantized)
const aimath_q7_params_t result_q_params_dense = { .shift = 3, .zero_point = 41 };
ailayer_dense_q7_t dense_layer = AILAYER_DENSE_Q7_M(3,
                                                    weights_data_dense, &weights_q_params_dense,
                                                    bias_data_dense, &bias_q_params_dense,
                                                    &result_q_params_dense);

Example: Create the layer structure for automatic parameter distribution:
In C:

ailayer_dense_q7_t dense_layer = {
    .neurons = 3
};

In C, C++ and on Arduino:

ailayer_dense_q7_t dense_layer = AILAYER_DENSE_Q7_A(3);

Example: Initialize and connect the layer:

x = ailayer_dense_wt_q7_default(&dense_layer, x);

Parameters

*layer	The layer structure to initialize.
*input_layer	The prior layer.

Returns

The (successfully) initialized layer structure.