mirror of
https://github.com/0xFEEDC0DE64/arduino-esp32.git
synced 2025-07-27 09:17:16 +02:00
Update IDF to v3.2 977854975 (#2771)
* Update IDF to v3.2 977854975 * Update app_httpd.cpp
This commit is contained in:
@ -1,336 +0,0 @@
|
||||
#ifndef DL_LIB_H
|
||||
#define DL_LIB_H
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
#include "dl_lib_matrix.h"
|
||||
#include "dl_lib_matrixq.h"
|
||||
#include "dl_lib_matrix3d.h"
|
||||
#include "dl_lib_matrix3dq.h"
|
||||
|
||||
typedef int padding_state;
|
||||
/**
|
||||
* @brief Does a fast version of the exp() operation on a floating point number.
|
||||
*
|
||||
* As described in https://codingforspeed.com/using-faster-exponential-approximation/
|
||||
* Should be good til an input of 5 or so with a steps factor of 8.
|
||||
*
|
||||
* @param in Floating point input
|
||||
* @param steps Approximation steps. More is more precise. 8 or 10 should be good enough for most purposes.
|
||||
* @return Exp()'ed output
|
||||
*/
|
||||
fptp_t fast_exp(double x, int steps);
|
||||
|
||||
/**
|
||||
* @brief Does a softmax operation on a matrix.
|
||||
*
|
||||
* @param in Input matrix
|
||||
* @param out Output matrix. Can be the same as the input matrix; if so,
|
||||
output results overwrite the input.
|
||||
*/
|
||||
void dl_softmax(const dl_matrix2d_t *in,
|
||||
dl_matrix2d_t *out);
|
||||
|
||||
/**
|
||||
* @brief Does a softmax operation on a quantized matrix.
|
||||
*
|
||||
* @param in Input matrix
|
||||
* @param out Output matrix. Can be the same as the input matrix; if so, output results overwrite the input.
|
||||
*/
|
||||
void dl_softmax_q(const dl_matrix2dq_t *in, dl_matrix2dq_t *out);
|
||||
|
||||
/**
|
||||
* @brief Does a sigmoid operation on a floating point number
|
||||
*
|
||||
* @param in Floating point input
|
||||
* @return Sigmoid output
|
||||
*/
|
||||
fptp_t dl_sigmoid_op(fptp_t in);
|
||||
|
||||
/**
|
||||
* @brief Does a sigmoid operation on a matrix.
|
||||
*
|
||||
* @param in Input matrix
|
||||
* @param out Output matrix. Can be the same as the input matrix; if so, output results overwrite the input.
|
||||
*/
|
||||
void dl_sigmoid(const dl_matrix2d_t *in, dl_matrix2d_t *out);
|
||||
|
||||
/**
|
||||
* @brief Does a tanh operation on a floating point number
|
||||
*
|
||||
* @param in Floating point input number
|
||||
* @return Tanh value
|
||||
*/
|
||||
fptp_t dl_tanh_op(fptp_t v);
|
||||
|
||||
/**
|
||||
* @brief Does a tanh operation on a matrix.
|
||||
*
|
||||
* @param in Input matrix
|
||||
* @param out Output matrix. Can be the same as the input matrix; if so, output results overwrite the input.
|
||||
*/
|
||||
void dl_tanh(const dl_matrix2d_t *in, dl_matrix2d_t *out);
|
||||
|
||||
/**
|
||||
* @brief Does a relu (Rectifier Linear Unit) operation on a floating point number
|
||||
*
|
||||
* @param in Floating point input
|
||||
* @param clip If value is higher than this, it will be clipped to this value
|
||||
* @return Relu output
|
||||
*/
|
||||
fptp_t dl_relu_op(fptp_t in, fptp_t clip);
|
||||
|
||||
/**
|
||||
* @brief Does a ReLu operation on a matrix.
|
||||
*
|
||||
* @param in Input matrix
|
||||
* @param clip If values are higher than this, they will be clipped to this value
|
||||
* @param out Output matrix. Can be the same as the input matrix; if so, output results overwrite the input.
|
||||
*/
|
||||
void dl_relu(const dl_matrix2d_t *in, fptp_t clip, dl_matrix2d_t *out);
|
||||
|
||||
/**
|
||||
* @brief Fully connected layer operation
|
||||
*
|
||||
* @param in Input vector
|
||||
* @param weight Weights of the neurons
|
||||
* @param bias Biases for the neurons. Can be NULL if a bias of 0 is required.
|
||||
* @param out Output array. Outputs are placed here. Needs to be an initialized, weight->w by in->h in size, matrix.
|
||||
*/
|
||||
void dl_fully_connect_layer(const dl_matrix2d_t *in,
|
||||
const dl_matrix2d_t *weight,
|
||||
const dl_matrix2d_t *bias,
|
||||
dl_matrix2d_t *out);
|
||||
|
||||
/**
|
||||
* @brief Pre-calculate the sqrtvari variable for the batch_normalize function.
|
||||
* The sqrtvari matrix depends on the variance and epsilon values, which normally are constant. Hence,
|
||||
* this matrix only needs to be calculated once. This function does that.
|
||||
*
|
||||
* @param
|
||||
* @return
|
||||
*/
|
||||
void dl_batch_normalize_get_sqrtvar(const dl_matrix2d_t *variance,
|
||||
fptp_t epsilon,
|
||||
dl_matrix2d_t *out);
|
||||
|
||||
/**
|
||||
* @brief Batch-normalize a matrix
|
||||
*
|
||||
* @param m The matrix to normalize
|
||||
* @param offset Offset matrix
|
||||
* @param scale Scale matrix
|
||||
* @param mean Mean matrix
|
||||
* @param sqrtvari Matrix precalculated using dl_batch_normalize_get_sqrtvar
|
||||
* @return
|
||||
*/
|
||||
void dl_batch_normalize(dl_matrix2d_t *m,
|
||||
const dl_matrix2d_t *offset,
|
||||
const dl_matrix2d_t *scale,
|
||||
const dl_matrix2d_t *mean,
|
||||
const dl_matrix2d_t *sqrtvari);
|
||||
|
||||
/**
|
||||
* @brief Do a basic LSTM layer pass.
|
||||
*
|
||||
* @warning Returns state_h pointer, so do not free result.
|
||||
|
||||
* @param in Input vector
|
||||
* @param state_c Internal state of the LSTM network
|
||||
* @param state_h Internal state (previous output values) of the LSTM network
|
||||
* @param weights Weights for the neurons
|
||||
* @param bias Bias for the neurons. Can be NULL if no bias is required
|
||||
* @return Output values of the neurons
|
||||
*/
|
||||
dl_matrix2d_t *dl_basic_lstm_layer(const dl_matrix2d_t *in,
|
||||
dl_matrix2d_t *state_c,
|
||||
dl_matrix2d_t *state_h,
|
||||
const dl_matrix2d_t *weight,
|
||||
const dl_matrix2d_t *bias);
|
||||
|
||||
/**
|
||||
* @brief Do a basic LSTM layer pass, partial quantized version.
|
||||
* This LSTM function accepts 16-bit fixed-point weights and 32-bit float-point bias.
|
||||
*
|
||||
* @warning Returns state_h pointer, so do not free result.
|
||||
|
||||
* @param in Input vector
|
||||
* @param state_c Internal state of the LSTM network
|
||||
* @param state_h Internal state (previous output values) of the LSTM network
|
||||
* @param weights Weights for the neurons, need to be quantised
|
||||
* @param bias Bias for the neurons. Can be NULL if no bias is required
|
||||
* @return Output values of the neurons
|
||||
*/
|
||||
dl_matrix2d_t *dl_basic_lstm_layer_quantised_weights(const dl_matrix2d_t *in,
|
||||
dl_matrix2d_t *state_c,
|
||||
dl_matrix2d_t *state_h,
|
||||
const dl_matrix2dq_t *weight,
|
||||
const dl_matrix2d_t *bias);
|
||||
|
||||
/**
|
||||
* @brief Do a fully-connected layer pass, fully-quantized version.
|
||||
*
|
||||
* @param in Input vector
|
||||
* @param weight Weights of the neurons
|
||||
* @param bias Bias values of the neurons. Can be NULL if no bias is needed.
|
||||
* @param shift Number of bits to shift the result back by. See dl_lib_matrixq.h for more info
|
||||
* @return Output values of the neurons
|
||||
*/
|
||||
void dl_fully_connect_layer_q(const dl_matrix2dq_t *in,
|
||||
const dl_matrix2dq_t *weight,
|
||||
const dl_matrix2dq_t *bias,
|
||||
dl_matrix2dq_t *out,
|
||||
int shift);
|
||||
|
||||
/**
|
||||
* @brief Do a basic LSTM layer pass, fully-quantized version
|
||||
*
|
||||
* @warning Returns state_h pointer, so do not free result.
|
||||
|
||||
* @param in Input vector
|
||||
* @param state_c Internal state of the LSTM network
|
||||
* @param state_h Internal state (previous output values) of the LSTM network
|
||||
* @param weights Weights for the neurons
|
||||
* @param bias Bias for the neurons. Can be NULL if no bias is required
|
||||
* @param shift Number of bits to shift the result back by. See dl_lib_matrixq.h for more info
|
||||
* @return Output values of the neurons
|
||||
*/
|
||||
dl_matrix2dq_t *dl_basic_lstm_layer_q(const dl_matrix2dq_t *in,
|
||||
dl_matrix2dq_t *state_c,
|
||||
dl_matrix2dq_t *state_h,
|
||||
const dl_matrix2dq_t *weight,
|
||||
const dl_matrix2dq_t *bias,
|
||||
int shift);
|
||||
|
||||
/**
|
||||
* @brief Batch-normalize a matrix, fully-quantized version
|
||||
*
|
||||
* @param m The matrix to normalize
|
||||
* @param offset Offset matrix
|
||||
* @param scale Scale matrix
|
||||
* @param mean Mean matrix
|
||||
* @param sqrtvari Matrix precalculated using dl_batch_normalize_get_sqrtvar
|
||||
* @param shift Number of bits to shift the result back by. See dl_lib_matrixq.h for more info
|
||||
* @return
|
||||
*/
|
||||
void dl_batch_normalize_q(dl_matrix2dq_t *m,
|
||||
const dl_matrix2dq_t *offset,
|
||||
const dl_matrix2dq_t *scale,
|
||||
const dl_matrix2dq_t *mean,
|
||||
const dl_matrix2dq_t *sqrtvari,
|
||||
int shift);
|
||||
|
||||
/**
|
||||
* @brief Does a relu (Rectifier Linear Unit) operation on a fixed-point number
|
||||
* This accepts and returns fixed-point 32-bit number with the last 15 bits being the bits after the decimal
|
||||
* point. (Equivalent to a mantissa in a quantized matrix with exponent -15.)
|
||||
*
|
||||
* @param in Fixed-point input
|
||||
* @param clip If value is higher than this, it will be clipped to this value
|
||||
* @return Relu output
|
||||
*/
|
||||
qtp_t dl_relu_q_op(qtp_t in,
|
||||
qtp_t clip);
|
||||
|
||||
/**
|
||||
* @brief Does a ReLu operation on a matrix, quantized version
|
||||
*
|
||||
* @param in Input matrix
|
||||
* @param clip If values are higher than this, they will be clipped to this value
|
||||
* @param out Output matrix. Can be the same as the input matrix; if so, output results overwrite the input.
|
||||
*/
|
||||
void dl_relu_q(const dl_matrix2dq_t *in,
|
||||
fptp_t clip,
|
||||
dl_matrix2dq_t *out);
|
||||
|
||||
/**
|
||||
* @brief Does a sigmoid operation on a fixed-point number.
|
||||
* This accepts and returns a fixed-point 32-bit number with the last 15 bits being the bits after the decimal
|
||||
* point. (Equivalent to a mantissa in a quantized matrix with exponent -15.)
|
||||
*
|
||||
* @param in Fixed-point input
|
||||
* @return Sigmoid output
|
||||
*/
|
||||
int dl_sigmoid_op_q(const int in);
|
||||
|
||||
/**
|
||||
* @brief Does a sigmoid operation on a matrix, quantized version
|
||||
*
|
||||
* @param in Input matrix
|
||||
* @param out Output matrix. Can be the same as the input matrix; if so, output results overwrite the input.
|
||||
*/
|
||||
void dl_sigmoid_q(const dl_matrix2dq_t *in,
|
||||
dl_matrix2dq_t *out);
|
||||
|
||||
/**
|
||||
* @brief Does a tanh operation on a matrix, quantized version
|
||||
*
|
||||
* @param in Input matrix
|
||||
* @param out Output matrix. Can be the same as the input matrix; if so, output results overwrite the input.
|
||||
*/
|
||||
void dl_tanh_q(const dl_matrix2dq_t *in,
|
||||
dl_matrix2dq_t *out);
|
||||
|
||||
/**
|
||||
* @brief Do a basic CNN layer pass.
|
||||
*
|
||||
* @Warning This just supports the single channel input image, and the output is single row matrix.
|
||||
That is to say, the height of output is 1, and the weight of output is out_channels*out_image_width*out_image_height
|
||||
*
|
||||
* @param in Input single channel image
|
||||
* @param weight Weights of the neurons, weight->w = out_channels, weight->h = filter_width*filter_height
|
||||
* @param bias Bias for the CNN layer.
|
||||
* @param filter_height The height of convolution kernel
|
||||
* @param filter_width The width of convolution kernel
|
||||
* @param out_channels The number of output channels of convolution kernel
|
||||
* @param stride_x The step length of the convolution window in x(width) direction
|
||||
* @param stride_y The step length of the convolution window in y(height) direction
|
||||
* @param pad One of `"VALID"` or `"SAME"`, 0 is "VALID" and the other is "SAME"
|
||||
* @param out The result of CNN layer, out->h=1.
|
||||
* @return The result of CNN layer.
|
||||
*/
|
||||
dl_matrix2d_t *dl_basic_conv_layer(const dl_matrix2d_t *in,
|
||||
const dl_matrix2d_t *weight,
|
||||
const dl_matrix2d_t *bias,
|
||||
int filter_width,
|
||||
int filter_height,
|
||||
const int out_channels,
|
||||
const int stride_x,
|
||||
const int stride_y,
|
||||
padding_state pad,
|
||||
const dl_matrix2d_t *out);
|
||||
|
||||
/**
|
||||
* @brief Do a basic CNN layer pass, quantised wersion.
|
||||
*
|
||||
* @Warning This just supports the single channel input image, and the output is single row matrix.
|
||||
That is to say, the height of output is 1, and the weight of output is out_channels*out_image_width*out_image_height
|
||||
*
|
||||
* @param in Input single channel image
|
||||
* @param weight Weights of the neurons, weight->w = out_channels, weight->h = filter_width*filter_height,
|
||||
* @param bias Bias of the neurons.
|
||||
* @param filter_height The height of convolution kernel
|
||||
* @param filter_width The width of convolution kernel
|
||||
* @param out_channels The number of output channels of convolution kernel
|
||||
* @param stride_x The step length of the convolution window in x(width) direction
|
||||
* @param stride_y The step length of the convolution window in y(height) direction
|
||||
* @param pad One of `"VALID"` or `"SAME"`, 0 is "VALID" and the other is "SAME"
|
||||
* @param out The result of CNN layer, out->h=1
|
||||
* @return The result of CNN layer
|
||||
*/
|
||||
dl_matrix2d_t *dl_basic_conv_layer_quantised_weight(const dl_matrix2d_t *in,
|
||||
const dl_matrix2dq_t *weight,
|
||||
const dl_matrix2d_t *bias,
|
||||
int filter_width,
|
||||
int filter_height,
|
||||
const int out_channels,
|
||||
const int stride_x,
|
||||
const int stride_y,
|
||||
padding_state pad,
|
||||
const dl_matrix2d_t *out);
|
||||
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
#endif
|
||||
#endif
|
@ -1,47 +0,0 @@
|
||||
#ifndef DL_LIB_COEFGETTER_IF_H
|
||||
#define DL_LIB_COEFGETTER_IF_H
|
||||
|
||||
#include "dl_lib_matrix.h"
|
||||
#include "dl_lib_matrixq.h"
|
||||
#include "dl_lib_matrix3d.h"
|
||||
#include "dl_lib_matrix3dq.h"
|
||||
|
||||
//Set this if the coefficient requested is a batch-normalization popvar matrix which needs to be preprocessed by
|
||||
//dl_batch_normalize_get_sqrtvar first.
|
||||
#define COEF_GETTER_HINT_BNVAR (1<<0)
|
||||
|
||||
/*
|
||||
This struct describes the basic information of model data:
|
||||
word_num: the number of wake words or speech commands
|
||||
word_list: the name list of wake words or speech commands
|
||||
thres_list: the threshold list of wake words or speech commands
|
||||
info_str: the string used to reflect the version and information of model data
|
||||
which consist of the architecture of network, the version of model data, wake words and their threshold
|
||||
*/
|
||||
typedef struct {
|
||||
int word_num;
|
||||
char **word_list;
|
||||
int *win_list;
|
||||
float *thresh_list;
|
||||
char *info_str;
|
||||
} model_info_t;
|
||||
|
||||
/*
|
||||
This struct describes a generic coefficient getter: a way to get the constant coefficients needed for a neural network.
|
||||
For the two getters, the name describes the name of the coefficient matrix, usually the same as the Numpy filename the
|
||||
coefficient was originally stored in. The arg argument can be used to optionally pass an additional user-defined argument
|
||||
to the getter (e.g. the directory to look for files in the case of the Numpy file loader getter). The hint argument
|
||||
is a bitwise OR of the COEF_GETTER_HINT_* flags or 0 when none is needed. Use the free_f/free_q functions to release the
|
||||
memory for the returned matrices, when applicable.
|
||||
*/
|
||||
typedef struct {
|
||||
const dl_matrix2d_t* (*getter_f)(const char *name, void *arg, int hint);
|
||||
const dl_matrix2dq_t* (*getter_q)(const char *name, void *arg, int hint);
|
||||
const dl_matrix3d_t* (*getter_3d)(const char *name, void *arg, int hint);
|
||||
const dl_matrix3dq_t* (*getter_3dq)(const char *name, void *arg, int hint);
|
||||
void (*free_f)(const dl_matrix2d_t *m);
|
||||
void (*free_q)(const dl_matrix2dq_t *m);
|
||||
const model_info_t* (*getter_info)(void *arg);
|
||||
} model_coeff_getter_t;
|
||||
|
||||
#endif
|
@ -1,216 +0,0 @@
|
||||
#ifndef DL_LIB_MATRIX_H
|
||||
#define DL_LIB_MATRIX_H
|
||||
|
||||
typedef float fptp_t;
|
||||
|
||||
|
||||
//Flags for matrices
|
||||
#define DL_MF_FOREIGNDATA (1<<0) /*< Matrix *item data actually points to another matrix and should not be freed */
|
||||
|
||||
//'Normal' float matrix
|
||||
typedef struct {
|
||||
int w; /*< Width */
|
||||
int h; /*< Height */
|
||||
int stride; /*< Row stride, essentially how many items to skip to get to the same position in the next row */
|
||||
int flags; /*< Flags. OR of DL_MF_* values */
|
||||
fptp_t *item; /*< Pointer to item array */
|
||||
} dl_matrix2d_t;
|
||||
|
||||
//Macro to quickly access the raw items in a matrix
|
||||
#define DL_ITM(m, x, y) m->item[(x)+(y)*m->stride]
|
||||
|
||||
|
||||
//#define DL_ITM3D(m, n, x, y, z) (m)->item[(n) * (m)->stride * (m)->c + (z) * (m)->stride + (y) * (m)->w + (x)]
|
||||
|
||||
/**
|
||||
* @brief Allocate a matrix
|
||||
*
|
||||
* @param w Width of the matrix
|
||||
* @param h Height of the matrix
|
||||
* @return The matrix, or NULL if out of memory
|
||||
*/
|
||||
dl_matrix2d_t *dl_matrix_alloc(int w, int h);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Free a matrix
|
||||
* Frees the matrix structure and (if it doesn't have the DL_MF_FOREIGNDATA flag set) the m->items space as well.
|
||||
*
|
||||
* @param m Matrix to free
|
||||
*/
|
||||
void dl_matrix_free(dl_matrix2d_t *m);
|
||||
|
||||
/**
|
||||
* @brief Zero out the matrix
|
||||
* Sets all entries in the matrix to 0.
|
||||
*
|
||||
* @param m Matrix to zero
|
||||
*/
|
||||
void dl_matrix_zero(dl_matrix2d_t *m);
|
||||
|
||||
/**
|
||||
* @brief Generate a new matrix using a range of items from an existing matrix.
|
||||
* When using this, the data of the new matrix is not allocated/copied but it re-uses a pointer
|
||||
* to the existing data. Changing the data in the resulting matrix, as a result, will also change
|
||||
* the data in the existing matrix that has been sliced.
|
||||
*
|
||||
* @param x X-offset of the origin of the returned matrix within the sliced matrix
|
||||
* @param y Y-offset of the origin of the returned matrix within the sliced matrix
|
||||
* @param w Width of the resulting matrix
|
||||
* @param h Height of the resulting matrix
|
||||
* @param in Old matrix (with foreign data) to re-use. Passing NULL will allocate a new matrix.
|
||||
* @return The resulting slice matrix, or NULL if out of memory
|
||||
*/
|
||||
dl_matrix2d_t *dl_matrix_slice(const dl_matrix2d_t *src, int x, int y, int w, int h, dl_matrix2d_t *in);
|
||||
|
||||
/**
|
||||
* @brief select a range of items from an existing matrix and flatten them into one dimension.
|
||||
*
|
||||
* @Warning The results are flattened in row-major order.
|
||||
*
|
||||
* @param x X-offset of the origin of the returned matrix within the sliced matrix
|
||||
* @param y Y-offset of the origin of the returned matrix within the sliced matrix
|
||||
* @param w Width of the resulting matrix
|
||||
* @param h Height of the resulting matrix
|
||||
* @param in Old matrix to re-use. Passing NULL will allocate a new matrix.
|
||||
* @return The resulting flatten matrix, or NULL if out of memory
|
||||
*/
|
||||
dl_matrix2d_t *dl_matrix_flatten(const dl_matrix2d_t *src, int x, int y, int w, int h, dl_matrix2d_t *in);
|
||||
|
||||
/**
|
||||
* @brief Generate a matrix from existing floating-point data
|
||||
*
|
||||
* @param w Width of resulting matrix
|
||||
* @param h Height of resulting matrix
|
||||
* @param data Data to populate matrix with
|
||||
* @return A newaly allocated matrix populated with the given input data, or NULL if out of memory.
|
||||
*/
|
||||
dl_matrix2d_t *dl_matrix_from_data(int w, int h, int stride, const void *data);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Multiply a pair of matrices item-by-item: res=a*b
|
||||
*
|
||||
* @param a First multiplicand
|
||||
* @param b Second multiplicand
|
||||
* @param res Multiplicated data. Can be equal to a or b to overwrite that.
|
||||
*/
|
||||
void dl_matrix_mul(const dl_matrix2d_t *a, const dl_matrix2d_t *b, dl_matrix2d_t *res);
|
||||
|
||||
/**
|
||||
* @brief Do a dotproduct of two matrices : res=a.b
|
||||
*
|
||||
* @param a First multiplicand
|
||||
* @param b Second multiplicand
|
||||
* @param res Dotproduct data. *Must* be a *different* matrix from a or b!
|
||||
*/
|
||||
void dl_matrix_dot(const dl_matrix2d_t *a, const dl_matrix2d_t *b, dl_matrix2d_t *res);
|
||||
|
||||
/**
|
||||
* @brief Add a pair of matrices item-by-item: res=a-b
|
||||
*
|
||||
* @param a First matrix
|
||||
* @param b Second matrix
|
||||
* @param res Added data. Can be equal to a or b to overwrite that.
|
||||
*/
|
||||
void dl_matrix_add(const dl_matrix2d_t *a, const dl_matrix2d_t *b, dl_matrix2d_t *out);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Divide a pair of matrices item-by-item: res=a/b
|
||||
*
|
||||
* @param a First matrix
|
||||
* @param b Second matrix
|
||||
* @param res Divided data. Can be equal to a or b to overwrite that.
|
||||
*/
|
||||
void dl_matrix_div(const dl_matrix2d_t *a, const dl_matrix2d_t *b, dl_matrix2d_t *out);
|
||||
|
||||
/**
|
||||
* @brief Subtract a matrix from another, item-by-item: res=a-b
|
||||
*
|
||||
* @param a First matrix
|
||||
* @param b Second matrix
|
||||
* @param res Subtracted data. Can be equal to a or b to overwrite that.
|
||||
*/
|
||||
void dl_matrix_sub(const dl_matrix2d_t *a, const dl_matrix2d_t *b, dl_matrix2d_t *out);
|
||||
|
||||
/**
|
||||
* @brief Add a constant to every item of the matrix
|
||||
*
|
||||
* @param subj Matrix to add the constant to
|
||||
* @param add The constant
|
||||
*/
|
||||
void dl_matrix_add_const(dl_matrix2d_t *subj, const fptp_t add);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Concatenate the rows of two matrices into a new matrix
|
||||
*
|
||||
* @param a First matrix
|
||||
* @param b Second matrix
|
||||
* @return A newly allocated array with as avlues a|b
|
||||
*/
|
||||
dl_matrix2d_t *dl_matrix_concat(const dl_matrix2d_t *a, const dl_matrix2d_t *b);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Print the contents of a matrix to stdout. Used for debugging.
|
||||
*
|
||||
* @param a The matrix to print.
|
||||
*/
|
||||
void dl_printmatrix(const dl_matrix2d_t *a);
|
||||
|
||||
/**
|
||||
* @brief Return the average square error given a correct and a test matrix.
|
||||
*
|
||||
* ...Well, more or less. If anything, it gives an indication of the error between
|
||||
* the two. Check the code for the exact implementation.
|
||||
*
|
||||
* @param a First of the two matrices to compare
|
||||
* @param b Second of the two matrices to compare
|
||||
* @return value indicating the relative difference between matrices
|
||||
*/
|
||||
float dl_matrix_get_avg_sq_err(const dl_matrix2d_t *a, const dl_matrix2d_t *b);
|
||||
|
||||
|
||||
|
||||
/**
|
||||
* @brief Check if two matrices have the same shape, that is, the same amount of rows and columns
|
||||
*
|
||||
* @param a First of the two matrices to compare
|
||||
* @param b Second of the two matrices to compare
|
||||
* @return true if the two matrices are shaped the same, false otherwise.
|
||||
*/
|
||||
int dl_matrix_same_shape(const dl_matrix2d_t *a, const dl_matrix2d_t *b);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Get a specific item from the matrix
|
||||
*
|
||||
* Please use these for external matrix access instead of DL_ITM
|
||||
*
|
||||
* @param m Matrix to access
|
||||
* @param x Column address
|
||||
* @param y Row address
|
||||
* @return Value in that position
|
||||
*/
|
||||
inline static fptp_t dl_matrix_get(const dl_matrix2d_t *m, const int x, const int y) {
|
||||
return DL_ITM(m, x, y);
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Set a specific item in the matrix to the given value
|
||||
*
|
||||
* Please use these for external matrix access instead of DL_ITM
|
||||
*
|
||||
* @param m Matrix to access
|
||||
* @param x Column address
|
||||
* @param y Row address
|
||||
* @param val Value to write to that position
|
||||
*/
|
||||
inline static void dl_matrix_set(dl_matrix2d_t *m, const int x, const int y, fptp_t val) {
|
||||
DL_ITM(m, x, y)=val;
|
||||
}
|
||||
|
||||
#endif
|
||||
|
@ -1,5 +1,6 @@
|
||||
#pragma once
|
||||
|
||||
#include <stdint.h>
|
||||
typedef float fptp_t;
|
||||
typedef uint8_t uc_t;
|
||||
|
||||
@ -92,6 +93,16 @@ void dl_matrix3d_free(dl_matrix3d_t *m);
|
||||
*/
|
||||
void dl_matrix3du_free(dl_matrix3du_t *m);
|
||||
|
||||
|
||||
/*
|
||||
* @brief Dot product with a vector and matrix
|
||||
*
|
||||
* @param out Space to put the result
|
||||
* @param in input vector
|
||||
* @param f filter matrix
|
||||
*/
|
||||
void dl_matrix3d_dot_product(dl_matrix3d_t *out, dl_matrix3d_t *in, dl_matrix3d_t *f);
|
||||
|
||||
/**
|
||||
* @brief Do a relu (Rectifier Linear Unit) operation, update the input matrix3d
|
||||
*
|
||||
@ -162,6 +173,9 @@ void dl_matrix3du_slice_copy(dl_matrix3du_t *dst,
|
||||
int w,
|
||||
int h);
|
||||
|
||||
|
||||
void dl_matrix3d_conv_1x1 (dl_matrix3d_t *out, dl_matrix3d_t *in, dl_matrix3d_t *f);
|
||||
|
||||
/**
|
||||
* @brief Do a general CNN layer pass, dimension is (number, width, height, channel)
|
||||
*
|
||||
@ -183,6 +197,11 @@ dl_matrix3d_t *dl_matrix3d_conv(dl_matrix3d_t *in,
|
||||
int padding,
|
||||
int mode);
|
||||
|
||||
void dl_matrix3d_conv_3x3_normal (dl_matrix3d_t *out,
|
||||
dl_matrix3d_t *in,
|
||||
dl_matrix3d_t *f,
|
||||
int step_x,
|
||||
int step_y);
|
||||
/**
|
||||
* @brief Do a general CNN layer pass, dimension is (number, width, height, channel)
|
||||
*
|
||||
@ -223,6 +242,11 @@ dl_matrix3d_t *dl_matrix3d_depthwise_conv(dl_matrix3d_t *in,
|
||||
int padding,
|
||||
int mode);
|
||||
|
||||
void dl_matrix3d_depthwise_conv_3x3_normal(dl_matrix3d_t *out,
|
||||
dl_matrix3d_t *in,
|
||||
dl_matrix3d_t *f,
|
||||
int step_x,
|
||||
int step_y);
|
||||
/**
|
||||
* @brief Do a mobilenet block forward, dimension is (number, width, height, channel)
|
||||
*
|
||||
@ -418,3 +442,8 @@ void dl_matrix3d_print(dl_matrix3d_t *m, char *message);
|
||||
* @param message name of matrix
|
||||
*/
|
||||
void dl_matrix3du_print(dl_matrix3du_t *m, char *message);
|
||||
|
||||
|
||||
void dl_matrix3d_init_bias (dl_matrix3d_t *out, dl_matrix3d_t *bias);
|
||||
|
||||
void dl_matrix3d_multiply(dl_matrix3d_t *out, dl_matrix3d_t *in1, dl_matrix3d_t *in2);
|
||||
|
@ -85,6 +85,22 @@ dl_matrix3dq_t *dl_matrix3dq_conv (dl_matrix3dq_t *in, dl_matrix3dq_t *filter, d
|
||||
dl_matrix3dq_t *dl_matrix3dq_conv_normal (dl_matrix3dq_t *in, dl_matrix3dq_t *filter, dl_matrix3dq_t *bias,
|
||||
int stride_x, int stride_y, int padding, int exponent, int mode);
|
||||
|
||||
void dl_matrix3dq_conv_1x1 (dl_matrix3dq_t *out, dl_matrix3dq_t *in, dl_matrix3dq_t *f, dl_conv_mode mode);
|
||||
|
||||
void dl_matrix3dq_conv_3x3_normal (dl_matrix3dq_t *out,
|
||||
dl_matrix3dq_t *in,
|
||||
dl_matrix3dq_t *f,
|
||||
int step_x,
|
||||
int step_y);
|
||||
dl_matrix3dq_t *dl_matrix3dq_conv_3x3_with_bn (dl_matrix3dq_t *in,
|
||||
dl_matrix3dq_t *f,
|
||||
dl_matrix3dq_t *scale,
|
||||
dl_matrix3dq_t *offset,
|
||||
int step_x,
|
||||
int step_y,
|
||||
int padding,
|
||||
int exponent,
|
||||
int relu);
|
||||
/**
|
||||
* @brief Print the matrix3d items
|
||||
*
|
||||
@ -95,6 +111,15 @@ void dl_matrix3dq_print (dl_matrix3dq_t *m, char *message);
|
||||
|
||||
dl_matrix3dq_t *dl_matrix3dq_depthwise_conv (dl_matrix3dq_t *in, dl_matrix3dq_t *filter,
|
||||
int stride_x, int stride_y, int padding, int exponent, int mode);
|
||||
dl_matrix3dq_t *dl_matrix3dq_depthwise_conv_3x3_with_bn(dl_matrix3dq_t *in,
|
||||
dl_matrix3dq_t *f,
|
||||
dl_matrix3dq_t *scale,
|
||||
dl_matrix3dq_t *offset,
|
||||
int step_x,
|
||||
int step_y,
|
||||
int padding,
|
||||
int exponent,
|
||||
int relu);
|
||||
|
||||
void dl_matrix3dq_relu (dl_matrix3dq_t *m, fptp_t clip);
|
||||
|
||||
|
@ -1,359 +0,0 @@
|
||||
#ifndef DL_LIB_MATRIXQ_H
|
||||
#define DL_LIB_MATRIXQ_H
|
||||
|
||||
#include <stdint.h>
|
||||
#include "dl_lib_matrix.h"
|
||||
|
||||
typedef int16_t qtp_t;
|
||||
|
||||
//Quantized matrix. Uses fixed numbers and has the storage for the rows/columns inverted
|
||||
//for easy use as a multiplicand without stressing out the flash cache too much.
|
||||
typedef struct {
|
||||
int w;
|
||||
int h;
|
||||
int stride; //Normally equals h, not w!
|
||||
int flags;
|
||||
int exponent; //The values in items should be multiplied by pow(2,exponent) to get the real values.
|
||||
qtp_t *itemq;
|
||||
} dl_matrix2dq_t;
|
||||
|
||||
#define DL_QTP_SHIFT 15
|
||||
#define DL_QTP_RANGE ((1<<DL_QTP_SHIFT)-1)
|
||||
#define DL_ITMQ(m, x, y) m->itemq[(y)+(x)*m->stride]
|
||||
#define DL_QTP_EXP_NA 255 //non-applicable exponent because matrix is null
|
||||
|
||||
#define DL_SHIFT_AUTO 32
|
||||
|
||||
/**
|
||||
* @info About quantized matrices and shift values
|
||||
*
|
||||
* Grab a coffee (or tea, or hot water) and sit down when you read this for the first
|
||||
* time. Quantized matrices can speed up your operations, but come with some quirks, and
|
||||
* it's good to understand how they work before using them.
|
||||
*
|
||||
* The data in the quantized matrix type is stored similarily to floating-point types:
|
||||
* when storing a real value, the value is stored as a mantissa (base number) and an
|
||||
* exponent. The 'real' value that can be re-derived from those two numbers is something
|
||||
* similar to mantissa*2^exponent. Up to this point, there's not that much difference from
|
||||
* the standard floating point implementations like e.g. IEEE-754.
|
||||
*
|
||||
* The difference with respect to quantized matrices is that for a quantized matrix, it is
|
||||
* assumed all values stored have more-or-less the same order of magnitude. This allows the
|
||||
* matrix to only store all the mantissas, while the exponents are shared; there is only one
|
||||
* exponent for the entire matrix. This makes it quicker to handle matrix operations - the
|
||||
* logic to fix the exponents only needs to happen once, while the rest can be done in simple
|
||||
* integer arithmetic. It also nets us some memory savings - while normally a floating point
|
||||
* number is 32-bit, storing only 16-bit mantissas as the matrix items almost halves the
|
||||
* memory requirements.
|
||||
*
|
||||
* While most of the details of handling the intricacies of the quantized matrixes are done
|
||||
* transparently by the code in dl_lib_matrixq.c, some implementation details leak out,
|
||||
* specifically in places where addition/subtraction/division happens.
|
||||
*
|
||||
* The problem is that the routines do not know what the size of the resulting operation is. For
|
||||
* instance, when adding two matrices of numbers, the resulting numbers *could* be large enough
|
||||
* to overflow the mantissa of the result if the exponent is the same. However, if by default we
|
||||
* assume the mantissas needs to be scaled back, we may lose precision.
|
||||
*
|
||||
* In order to counter this, all operations that have this issue have a ``shift`` argument. If
|
||||
* the argument is zero, the routine will be conservative, that is, increase the exponent of
|
||||
* the result to such an extent it's mathematically impossible a value in the result will exceed
|
||||
* the maximum value that can be stored. However, when this argument is larger than zero, the
|
||||
* algorithm will hold back on this scaling by the indicated amount of bits, preserving precision
|
||||
* but increasing the chance of some of the calculated values not fitting in the mantissa anymore.
|
||||
* If this happens, the value will be clipped to the largest (or, for negative values, smallest)
|
||||
* value possible. (Neural networks usually are okay with this happening for a limited amount
|
||||
* of matrix indices).
|
||||
*
|
||||
* For deciding on these shift values, it is recommended to start with a shift value of one, then
|
||||
* use dl_matrixq_check_sanity on the result. If this indicates clipping, lower the shift value.
|
||||
* If it indicates bits are under-used, increase it. Note that for adding and subtraction, only
|
||||
* shift values of 0 or 1 make sense; these routines will error out if you try to do something
|
||||
* else.
|
||||
*
|
||||
* For neural networks and other noise-tolerant applications, note that even when
|
||||
* dl_matrixq_check_sanity does not indicate any problems, twiddling with the shift value may lead
|
||||
* to slightly improved precision. Feel free to experiment.
|
||||
**/
|
||||
|
||||
|
||||
/**
|
||||
* @brief Allocate a matrix
|
||||
*
|
||||
* @param w Width of the matrix
|
||||
* @param h Height of the matrix
|
||||
* @return The matrix, or NULL if out of memory
|
||||
*/
|
||||
dl_matrix2dq_t *dl_matrixq_alloc(int w, int h);
|
||||
|
||||
/**
|
||||
* @brief Convert a floating-point matrix to a quantized matrix
|
||||
*
|
||||
* @param m Floating-point matrix to convert
|
||||
* @param out Quantized matrix to re-use. If NULL, allocate a new one.
|
||||
* @Return The quantized version of the floating-point matrix
|
||||
*/
|
||||
dl_matrix2dq_t *dl_matrixq_from_matrix2d(const dl_matrix2d_t *m, dl_matrix2dq_t *out);
|
||||
|
||||
|
||||
/**
|
||||
* TODO: DESCRIBE THIS FUNCTION
|
||||
*/
|
||||
dl_matrix2dq_t *dl_matrixq_from_matrix2d_by_qmf(const dl_matrix2d_t *m, dl_matrix2dq_t *out, int m_bit, int f_bit);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Convert a quantized matrix to a floating-point one.
|
||||
*
|
||||
* @param m Floating-point matrix to convert
|
||||
* @param out Quantized matrix to re-use. If NULL, allocate a new one.
|
||||
* @Return The quantized version of the floating-point matrix
|
||||
**/
|
||||
dl_matrix2d_t *dl_matrix2d_from_matrixq(const dl_matrix2dq_t *m, dl_matrix2d_t *out);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Free a quantized matrix
|
||||
* Frees the matrix structure and (if it doesn't have the DL_MF_FOREIGNDATA flag set) the m->items space as well.
|
||||
*
|
||||
* @param m Matrix to free
|
||||
*/
|
||||
void dl_matrixq_free(dl_matrix2dq_t *m);
|
||||
|
||||
/**
|
||||
* @brief Zero out the matrix
|
||||
* Sets all entries in the matrix to 0.
|
||||
*
|
||||
* @param m Matrix to zero
|
||||
*/
|
||||
void dl_matrixq_zero(dl_matrix2dq_t *m);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Do a dotproduct of two quantized matrices : res=a.b, Result is a fixed-point matrix.
|
||||
*
|
||||
* @param a First multiplicand
|
||||
* @param b Second multiplicand
|
||||
* @param res Dotproduct data. *Must* be a *different* matrix from a or b!
|
||||
* @param shift Shift ratio
|
||||
*/
|
||||
void dl_matrixq_dot(const dl_matrix2dq_t *a, const dl_matrix2dq_t *b, dl_matrix2dq_t *res, int shift);
|
||||
|
||||
/**
|
||||
* @brief Do a dotproduct of two quantized matrices: res=a.b, Result is a floating-point matrix.
|
||||
*
|
||||
* @param a First multiplicand
|
||||
* @param b Second multiplicand
|
||||
* @param res Dotproduct data. *Must* be a *different* matrix from a or b!
|
||||
*/
|
||||
void dl_matrixq_dot_matrix_out(const dl_matrix2dq_t *a, const dl_matrix2dq_t *b, dl_matrix2d_t *res);
|
||||
|
||||
/**
|
||||
* @brief Do a dotproduct of two quantized matrices : res=a.b. This always uses the simple & stupid C algo for the dot product.
|
||||
*
|
||||
* Result is a fixed-point matrix.
|
||||
*
|
||||
* Use this only if you expect something is wrong with the accelerated routines that dl_matrixq_dot calls; this function can be
|
||||
* much slower than dl_matrixq_dot .
|
||||
*
|
||||
* @param a First multiplicand
|
||||
* @param b Second multiplicand
|
||||
* @param res Dotproduct data. *Must* be a *different* matrix from a or b!
|
||||
* @param shift Shift ratio
|
||||
*/
|
||||
void dl_matrixq_dot_c_impl(const dl_matrix2dq_t *a, const dl_matrix2dq_t *b, dl_matrix2dq_t *res, int shift);
|
||||
|
||||
/**
|
||||
* @brief Do a dotproduct of two quantized matrices : res=a.b. This always uses the simple & stupid C algo for the dot product.
|
||||
*
|
||||
* Result is a floating-point matrix.
|
||||
*
|
||||
* Use this only if you expect something is wrong with the accelerated routines that dl_matrixq_dot_matrix_out calls; this function can be
|
||||
* much slower than dl_matrixq_dot_matrix_out.
|
||||
*
|
||||
* @param a First multiplicand
|
||||
* @param b Second multiplicand
|
||||
* @param res Dotproduct data. *Must* be a *different* matrix from a or b!
|
||||
*/
|
||||
void dl_matrixq_dot_matrix_out_c_impl(const dl_matrix2dq_t *a, const dl_matrix2dq_t *b, dl_matrix2d_t *res);
|
||||
|
||||
/**
|
||||
* @brief Do a dotproduct of a floating point and a quantized matrix. Result is a floating-point matrix.
|
||||
*
|
||||
* @param a First multiplicand; float matrix
|
||||
* @param b Second multiplicand; quantized matrix
|
||||
* @param res Dotproduct data; float matrix. *Must* be a *different* matrix from a or b!
|
||||
*/
|
||||
void dl_matrix_matrixq_dot(const dl_matrix2d_t *a, const dl_matrix2dq_t *b, dl_matrix2d_t *res);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Print the contents of a quantized matrix to stdout. Used for debugging.
|
||||
*
|
||||
* @param a The matrix to print.
|
||||
*/
|
||||
void dl_printmatrixq(const dl_matrix2dq_t *a);
|
||||
|
||||
|
||||
/**
|
||||
* @brief Add a pair of quantizedmatrices item-by-item: res=a-b
|
||||
*
|
||||
* @param a First matrix
|
||||
* @param b Second matrix
|
||||
* @param res Added data. Can be equal to a or b to overwrite that.
|
||||
* @param shift Shift value. Only 0 or 1 makes sense here. <ToDo: check>
|
||||
*/
|
||||
void dl_matrixq_add(const dl_matrix2dq_t *a, const dl_matrix2dq_t *b, dl_matrix2dq_t *res, int shift);
|
||||
|
||||
/**
|
||||
* @brief Generate a new matrix using a range of items from an existing matrix.
|
||||
* When using this, the data of the new matrix is not allocated/copied but it re-uses a pointer
|
||||
* to the existing data. Changing the data in the resulting matrix, as a result, will also change
|
||||
* the data in the existing matrix that has been sliced.
|
||||
*
|
||||
* @Warning In contrast to the floating point equivalent of this function, the fixed-point version
|
||||
* of this has the issue that as soon as the output exponent of one of the slices changes, the data
|
||||
* in the sliced matrix gets corrupted (because the exponent of that matrix is still the same.) If you
|
||||
* use this function, either treat the slices as read-only, or assume the sliced matrix contains
|
||||
* garbage after modifying the data in one of the slices.
|
||||
*
|
||||
* @param x X-offset of the origin of the returned matrix within the sliced matrix
|
||||
* @param y Y-offset of the origin of the returned matrix within the sliced matrix
|
||||
* @param w Width of the resulting matrix
|
||||
* @param h Height of the resulting matrix
|
||||
* @param in Old matrix (with foreign data) to re-use. Passing NULL will allocate a new matrix.
|
||||
* @return The resulting slice matrix, or NULL if out of memory
|
||||
*/
|
||||
dl_matrix2dq_t *dl_matrixq_slice(const dl_matrix2dq_t *src, int x, int y, int w, int h, dl_matrix2dq_t *in);
|
||||
|
||||
/**
|
||||
* @brief select a range of items from an existing matrix and flatten them into one dimension.
|
||||
*
|
||||
* @Warning The results are flattened in row-major order.
|
||||
*
|
||||
* @param x X-offset of the origin of the returned matrix within the sliced matrix
|
||||
* @param y Y-offset of the origin of the returned matrix within the sliced matrix
|
||||
* @param w Width of the resulting matrix
|
||||
* @param h Height of the resulting matrix
|
||||
* @param in Old matrix to re-use. Passing NULL will allocate a new matrix.
|
||||
* @return The resulting flatten matrix, or NULL if out of memory
|
||||
*/
|
||||
dl_matrix2dq_t *dl_matrixq_flatten(const dl_matrix2dq_t *src, int x, int y, int w, int h, dl_matrix2dq_t *in);
|
||||
|
||||
/**
|
||||
* @brief Subtract a quantized matrix from another, item-by-item: res=a-b
|
||||
*
|
||||
* @param a First matrix
|
||||
* @param b Second matrix
|
||||
* @param res Subtracted data. Can be equal to a or b to overwrite that.
|
||||
* @param shift Shift value. Only 0 or 1 makes sense here. <ToDo: check>
|
||||
*/
|
||||
void dl_matrixq_sub(const dl_matrix2dq_t *a, const dl_matrix2dq_t *b, dl_matrix2dq_t *res, int shift);
|
||||
|
||||
/**
|
||||
* @brief Multiply a pair of quantized matrices item-by-item: res=a*b
|
||||
*
|
||||
* @param a First multiplicand
|
||||
* @param b Second multiplicand
|
||||
* @param res Multiplicated data. Can be equal to a or b to overwrite that matrix.
|
||||
*/
|
||||
void dl_matrixq_mul(const dl_matrix2dq_t *a, const dl_matrix2dq_t *b, dl_matrix2dq_t *res);
|
||||
|
||||
/**
|
||||
* @brief Divide a pair of quantized matrices item-by-item: res=a/b
|
||||
*
|
||||
* @param a First matrix
|
||||
* @param b Second matrix
|
||||
* @param res Divided data. Can be equal to a or b to overwrite that.
|
||||
*/
|
||||
void dl_matrixq_div(const dl_matrix2dq_t *a, const dl_matrix2dq_t *b, dl_matrix2dq_t *out, int shift);
|
||||
|
||||
/**
|
||||
* @brief Check if two quantized matrices have the same shape, that is, the same amount of
|
||||
* rows and columns
|
||||
*
|
||||
* @param a First of the two matrices to compare
|
||||
* @param b Second of the two matrices to compare
|
||||
* @return true if the two matrices are shaped the same, false otherwise.
|
||||
*/
|
||||
int dl_matrixq_same_shape(const dl_matrix2dq_t *a, const dl_matrix2dq_t *b);
|
||||
|
||||
/**
|
||||
* @brief Concatenate the rows of two quantized matrices into a new matrix
|
||||
*
|
||||
* @param a First matrix
|
||||
* @param b Second matrix
|
||||
* @return A newly allocated quantized matrix with as values a|b
|
||||
*/
|
||||
dl_matrix2dq_t *dl_matrixq_concat(const dl_matrix2dq_t *a, const dl_matrix2dq_t *b);
|
||||
|
||||
/**
|
||||
* @brief Add a constant to every item of the quantized matrix
|
||||
*
|
||||
* @param subj Matrix to add the constant to
|
||||
* @param add The constant
|
||||
*/
|
||||
void dl_matrixq_add_const(dl_matrix2dq_t *subj, const fptp_t add, int shift);
|
||||
|
||||
/**
|
||||
* @brief Check the sanity of a quantized matrix
|
||||
*
|
||||
* Due to the nature of quantized matrices, depending on the calculations a quantized
|
||||
* matrix is the result of and the shift values chosen in those calculations, a quantized
|
||||
* matrix may have an exponent and mantissas that lead to a loss of precision, either because
|
||||
* most significant mantissa bits are unused, or because a fair amount of mantissas are
|
||||
* clipped. This function checks if this is the case and will report a message to stdout
|
||||
* if significant loss of precision is detected.
|
||||
*
|
||||
* @param m The quantized matrix to check
|
||||
* @param name A string to be displayed in the message if the sanity check fails
|
||||
* @return True if matrix is sane, false otherwise
|
||||
**/
|
||||
|
||||
int dl_matrixq_check_sanity(dl_matrix2dq_t *m, const char *name);
|
||||
|
||||
/**
|
||||
* @brief re-adjust the exponent of the matrix to fit the mantissa better
|
||||
*
|
||||
* This function will shift up all the data in the mantissas so there are no
|
||||
* most-significant bits that are unused in all mantissas. It will also adjust
|
||||
* the exponent to keep the actua values in the matrix the same.
|
||||
*
|
||||
* Some operations done on a matrix, especially operations that re-use the
|
||||
* result of earlier operations done in the same way, can lead to the loss of
|
||||
* data because the exponent of the quantized matrix is never re-adjusted. You
|
||||
* can do that implicitely by calling this function.
|
||||
*
|
||||
* @param m The matrix to re-adjust
|
||||
**/
|
||||
void dl_matrixq_readjust_exp(dl_matrix2dq_t *m);
|
||||
|
||||
|
||||
|
||||
/**
|
||||
* @brief Get the floating-point value of a specific item from the quantized matrix
|
||||
*
|
||||
* @param m Matrix to access
|
||||
* @param x Column address
|
||||
* @param y Row address
|
||||
* @return Value in that position
|
||||
*/
|
||||
fptp_t dl_matrixq_get(const dl_matrix2dq_t *m, const int x, const int y);
|
||||
|
||||
/**
|
||||
* @brief Set a specific item in the quantized matrix to the given
|
||||
* floating-point value
|
||||
*
|
||||
* @warning If the given value is more than the exponent in the quantized matrix
|
||||
* allows for, all mantissas in the matrix will be shifted down to make the value
|
||||
* 'fit'. If, however, the exponent is such that the value would result in a
|
||||
* quantized mantissa of 0, nothing is done.
|
||||
*
|
||||
* @param m Matrix to access
|
||||
* @param x Column address
|
||||
* @param y Row address
|
||||
* @param val Value to write to that position
|
||||
*/
|
||||
void dl_matrixq_set(dl_matrix2dq_t *m, const int x, const int y, fptp_t val);
|
||||
|
||||
#endif
|
@ -29,22 +29,41 @@ extern "C"
|
||||
#endif
|
||||
|
||||
#include "image_util.h"
|
||||
#include "dl_lib.h"
|
||||
#include "dl_lib_matrix3d.h"
|
||||
#include "mtmn.h"
|
||||
|
||||
typedef enum
|
||||
{
|
||||
FAST = 0,
|
||||
NORMAL = 1,
|
||||
} mtmn_resize_type;
|
||||
|
||||
typedef struct
|
||||
{
|
||||
float min_face; /// the minimum size of face can be detected
|
||||
float pyramid; /// the pyramid scale
|
||||
int pyramid_times; /// the pyramid resizing times
|
||||
threshold_config_t p_threshold; /// score, nms and candidate threshold of pnet
|
||||
threshold_config_t r_threshold; /// score, nms and candidate threshold of rnet
|
||||
threshold_config_t o_threshold; /// score, nms and candidate threshold of onet
|
||||
mtmn_resize_type type; /// image resize type. 'pyramid' will lose efficacy, when 'type'==FAST.
|
||||
} mtmn_config_t;
|
||||
|
||||
static inline mtmn_config_t mtmn_init_config()
|
||||
{
|
||||
mtmn_config_t mtmn_config;
|
||||
mtmn_config.type = FAST;
|
||||
mtmn_config.min_face = 80;
|
||||
mtmn_config.pyramid = 0.7;
|
||||
mtmn_config.pyramid = 0.707;
|
||||
mtmn_config.pyramid_times = 4;
|
||||
mtmn_config.p_threshold.score = 0.6;
|
||||
mtmn_config.p_threshold.nms = 0.7;
|
||||
mtmn_config.p_threshold.candidate_number = 100;
|
||||
mtmn_config.r_threshold.score = 0.6;
|
||||
mtmn_config.p_threshold.candidate_number = 20;
|
||||
mtmn_config.r_threshold.score = 0.7;
|
||||
mtmn_config.r_threshold.nms = 0.7;
|
||||
mtmn_config.r_threshold.candidate_number = 4;
|
||||
mtmn_config.o_threshold.score = 0.6;
|
||||
mtmn_config.o_threshold.nms = 0.4;
|
||||
mtmn_config.r_threshold.candidate_number = 10;
|
||||
mtmn_config.o_threshold.score = 0.7;
|
||||
mtmn_config.o_threshold.nms = 0.7;
|
||||
mtmn_config.o_threshold.candidate_number = 1;
|
||||
|
||||
return mtmn_config;
|
||||
|
@ -6,7 +6,7 @@ extern "C"
|
||||
#endif
|
||||
|
||||
#include "image_util.h"
|
||||
#include "dl_lib.h"
|
||||
#include "dl_lib_matrix3d.h"
|
||||
#include "frmn.h"
|
||||
|
||||
#define FACE_WIDTH 56
|
||||
@ -38,23 +38,22 @@ extern "C"
|
||||
|
||||
typedef struct
|
||||
{
|
||||
face_id_node *head; /*!< head pointer of the id list */
|
||||
face_id_node *tail; /*!< tail pointer of the id list */
|
||||
uint8_t count; /*!< number of enrolled ids */
|
||||
uint8_t confirm_times; /*!< images needed for one enrolling */
|
||||
face_id_node *head; /*!< head pointer of the id list */
|
||||
face_id_node *tail; /*!< tail pointer of the id list */
|
||||
uint8_t count; /*!< number of enrolled ids */
|
||||
uint8_t confirm_times; /*!< images needed for one enrolling */
|
||||
} face_id_name_list;
|
||||
|
||||
|
||||
typedef struct
|
||||
{
|
||||
uint8_t head; /*!< head index of the id list */
|
||||
uint8_t tail; /*!< tail index of the id list */
|
||||
uint8_t count; /*!< number of enrolled ids */
|
||||
uint8_t size; /*!< max len of id list */
|
||||
uint8_t confirm_times; /*!< images needed for one enrolling */
|
||||
dl_matrix3d_t **id_list; /*!< stores face id vectors */
|
||||
uint8_t head; /*!< head index of the id list */
|
||||
uint8_t tail; /*!< tail index of the id list */
|
||||
uint8_t count; /*!< number of enrolled ids */
|
||||
uint8_t size; /*!< max len of id list */
|
||||
uint8_t confirm_times; /*!< images needed for one enrolling */
|
||||
dl_matrix3d_t **id_list; /*!< stores face id vectors */
|
||||
} face_id_list;
|
||||
|
||||
|
||||
/**
|
||||
* @brief Initialize face id list
|
||||
*
|
||||
@ -86,6 +85,10 @@ extern "C"
|
||||
dl_matrix3du_t *src,
|
||||
dl_matrix3du_t *dest);
|
||||
|
||||
int8_t align_face2(fptp_t *landmark,
|
||||
dl_matrix3du_t *src,
|
||||
dl_matrix3du_t *dest);
|
||||
|
||||
dl_matrix3d_t *get_face_id(dl_matrix3du_t *aligned_face);
|
||||
|
||||
/**
|
||||
@ -104,11 +107,9 @@ extern "C"
|
||||
* @param id_list An ID list
|
||||
* @return int8_t Matched face id
|
||||
*/
|
||||
int8_t recognize_face(face_id_list *l,
|
||||
dl_matrix3du_t *algined_face);
|
||||
|
||||
face_id_node *recognize_face_with_name(face_id_name_list *l,
|
||||
dl_matrix3d_t *face_id);
|
||||
int8_t recognize_face(face_id_list *l, dl_matrix3du_t *algined_face);
|
||||
|
||||
face_id_node *recognize_face_with_name(face_id_name_list *l, dl_matrix3d_t *face_id);
|
||||
/**
|
||||
* @brief Produce face id according to the input aligned face, and save it to dest_id.
|
||||
*
|
||||
@ -119,12 +120,11 @@ extern "C"
|
||||
* @return 0 Enrollment finish
|
||||
* @return >=1 The left piece of aligned faces should be input
|
||||
*/
|
||||
int8_t enroll_face(face_id_list *l,
|
||||
dl_matrix3du_t *aligned_face);
|
||||
|
||||
int8_t enroll_face_with_name(face_id_name_list *l,
|
||||
dl_matrix3d_t *new_id,
|
||||
char *name);
|
||||
int8_t enroll_face(face_id_list *l, dl_matrix3du_t *aligned_face);
|
||||
|
||||
int8_t enroll_face_with_name(face_id_name_list *l,
|
||||
dl_matrix3d_t *new_id,
|
||||
char *name);
|
||||
|
||||
/**
|
||||
* @brief Alloc memory for aligned face.
|
||||
@ -133,7 +133,7 @@ extern "C"
|
||||
* @return uint8_t left count
|
||||
*/
|
||||
uint8_t delete_face(face_id_list *l);
|
||||
int8_t delete_face_with_name(face_id_name_list *l, char *name);
|
||||
int8_t delete_face_with_name(face_id_name_list *l, char *name);
|
||||
void delete_face_all_with_name(face_id_name_list *l);
|
||||
#if __cplusplus
|
||||
}
|
||||
|
@ -5,7 +5,8 @@ extern "C"
|
||||
{
|
||||
#endif
|
||||
|
||||
#include "dl_lib.h"
|
||||
#include "dl_lib_matrix3d.h"
|
||||
#include "dl_lib_matrix3dq.h"
|
||||
|
||||
/**
|
||||
* @brief
|
||||
|
@ -27,6 +27,7 @@ extern "C"
|
||||
{
|
||||
#endif
|
||||
#include <stdint.h>
|
||||
#include <math.h>
|
||||
#include "mtmn.h"
|
||||
|
||||
#define MAX_VALID_COUNT_PER_IMAGE (30)
|
||||
@ -57,6 +58,7 @@ extern "C"
|
||||
|
||||
typedef struct tag_box_list
|
||||
{
|
||||
fptp_t *score;
|
||||
box_t *box;
|
||||
landmark_t *landmark;
|
||||
int len;
|
||||
@ -142,12 +144,19 @@ extern "C"
|
||||
for (int i = 0; i < boxes->len; i++)
|
||||
{
|
||||
box_t *box = &(boxes->box[i]);
|
||||
float w, h;
|
||||
image_get_width_and_height(box, &w, &h);
|
||||
float l = DL_IMAGE_MAX(w, h);
|
||||
|
||||
box->box_p[0] = DL_IMAGE_MAX(0, box->box_p[0] + 0.5 * (w - l));
|
||||
box->box_p[1] = DL_IMAGE_MAX(0, box->box_p[1] + 0.5 * (h - l));
|
||||
int x1 = round(box->box_p[0]);
|
||||
int y1 = round(box->box_p[1]);
|
||||
int x2 = round(box->box_p[2]);
|
||||
int y2 = round(box->box_p[3]);
|
||||
|
||||
int w = x2 - x1 + 1;
|
||||
int h = y2 - y1 + 1;
|
||||
int l = DL_IMAGE_MAX(w, h);
|
||||
|
||||
box->box_p[0] = round(DL_IMAGE_MAX(0, x1) + 0.5 * (w - l));
|
||||
box->box_p[1] = round(DL_IMAGE_MAX(0, y1) + 0.5 * (h - l));
|
||||
|
||||
box->box_p[2] = box->box_p[0] + l - 1;
|
||||
if (box->box_p[2] > width)
|
||||
{
|
||||
@ -215,6 +224,25 @@ extern "C"
|
||||
*/
|
||||
void image_nms_process(image_list_t *image_list, fptp_t nms_threshold, int same_area);
|
||||
|
||||
/**
|
||||
* @brief
|
||||
*
|
||||
* @param dimage
|
||||
* @param dw
|
||||
* @param dh
|
||||
* @param dc
|
||||
* @param simage
|
||||
* @param sw
|
||||
* @param sc
|
||||
*/
|
||||
void image_zoom_in_twice(uint8_t *dimage,
|
||||
int dw,
|
||||
int dh,
|
||||
int dc,
|
||||
uint8_t *simage,
|
||||
int sw,
|
||||
int sc);
|
||||
|
||||
/**
|
||||
* @brief
|
||||
*
|
||||
|
@ -27,14 +27,7 @@
|
||||
extern "C"
|
||||
{
|
||||
#endif
|
||||
#include "dl_lib.h"
|
||||
|
||||
typedef enum
|
||||
{
|
||||
PNET = 0, /// P-Net
|
||||
RNET = 1, /// R-Net
|
||||
ONET = 2, /// O-Net
|
||||
} net_type_en;
|
||||
#include "dl_lib_matrix3d.h"
|
||||
|
||||
typedef struct
|
||||
{
|
||||
@ -45,22 +38,11 @@ extern "C"
|
||||
|
||||
typedef struct
|
||||
{
|
||||
net_type_en net_type; /// net type
|
||||
char *file_name; /// net name
|
||||
int w; /// net width
|
||||
int h; /// net height
|
||||
threshold_config_t threshold; /// threshold of net
|
||||
} net_config_t;
|
||||
|
||||
typedef struct
|
||||
{
|
||||
float min_face; /// the minimum size of face can be detected
|
||||
float pyramid; /// the pyramid scale
|
||||
threshold_config_t p_threshold; /// score, nms and candidate threshold of pnet
|
||||
threshold_config_t r_threshold; /// score, nms and candidate threshold of rnet
|
||||
threshold_config_t o_threshold; /// score, nms and candidate threshold of onet
|
||||
} mtmn_config_t;
|
||||
|
||||
typedef struct
|
||||
{
|
||||
dl_matrix3d_t *category;
|
||||
|
Reference in New Issue
Block a user