def write_weight_Conv2D_cpp(weight, bias, file_name, weight_array_name, bias_array_name, type_name, isFixed=False, fractal_width=0, array_type=None):
# Weight generated by Keras has 4D array with (height, width, input_depth, output_depth)
# Transpose weight axis from (height, width, input_depth, output_depth) to (output_depth, input_depth ,height, width)
with open(file_name, 'w') as f:
# reshape weight array
weight = weight.transpose(3, 2, 0, 1)
if isFixed is True:
weight = float2fixed.float2fixed_array(array_type, fractal_width, weight)
bias = float2fixed.float2fixed_array(array_type, fractal_width, bias)
# headers
todaytime = str(datetime.datetime.today())
f.write("/*\n")
f.write(" * author : shtsno24\n")
f.write(" * Date : " + todaytime + "\n")
if isFixed is True:
f.write(str(" * array_type : {}\n").format(weight.dtype))
f.write(str(" * fractal_width : {} bit\n").format(fractal_width))
f.write(str(" * bit_width : {} bit\n").format(str(8 * np.dtype(array_type).itemsize)))
f.write(" *\n */\n")
# include <cstdint> and <vector>
f.write("#pragma once\n")
f.write("#include <cstdint>\n")
f.write("#include <vector>\n\n")
f.write("using namespace std;\n\n")
# define data_width
if isFixed is True:
f.write(str("#define data_width_{} {}\n").format(str(weight_array_name[:-2]), str(8 * np.dtype(array_type).itemsize)))
f.write(str("#define fractal_width_{} {}\n\n").format(str(weight_array_name[:-2]), str(fractal_width)))
# weights
f.write(str("const vector< uint16_t> shape_{}_w = ").format(str(weight_array_name[:-2])))
f.write("{%d, %d, %d, %d};\n" % (weight.shape[0], weight.shape[1], weight.shape[2], weight.shape[3]))
f.write("const vector< vector< vector< vector< " + type_name + "> > > > " + weight_array_name + " =\n")
write_array_ND(weight, f)
f.write(";")
f.write("\n\n")
# bias
f.write(str("const uint16_t shape_{}_b = {};\n").format(str(weight_array_name[:-2]), bias.shape[0]))
f.write("const vector< " + type_name + "> " + bias_array_name + " = ")
# f.write(" = " % bias.shape)
write_array_ND(bias, f)
f.write(";")
f.write("\n")
import numpy as np
from keras_mnist_DCAE import float2fixed
import keras_weight_generator
import os
if os.path.isdir("test_data") is False:
os.mkdir("test_data")
tsv_file = input("file name : ")
tsv_array = np.loadtxt(tsv_file + ".tsv", dtype=np.float32, delimiter='\t')
tsv_array = tsv_array.reshape((1,) + tsv_array.shape)
print(tsv_array.shape, tsv_array.dtype)
tsv_array_fix16 = float2fixed.float2fixed_array(np.int16, 16-3, tsv_array)
print(tsv_array_fix16.shape, tsv_array_fix16.dtype, tsv_array_fix16)
with open("test_data/test_data.h", "w") as f:
f.write("#include <stdint.h>\n\n")
f.write(str("int16_t test_input_fix16[{}][{}][{}] = ").format(tsv_array_fix16.shape[0], tsv_array_fix16.shape[1], tsv_array_fix16.shape[2]))
keras_weight_generator.write_array_ND(tsv_array_fix16, f)
f.write(";\n\n")
f.write(str("float test_input_float32[{}][{}][{}] = ").format(tsv_array.shape[0], tsv_array.shape[1], tsv_array.shape[2]))
keras_weight_generator.write_array_ND(tsv_array, f)
f.write(";\n\n")
def write_weight_SeparableConv2D_c(weight_depthwise, weight_pointwise, bias, file_name, weight_array_name, bias_array_name, type_name, isFixed=False, fractal_width=0, array_type=None):
# Weight generated by Keras has 4D array with (height, width, input_depth, output_depth)
# Transpose weight axis from (height, width, input_depth, output_depth) to (output_depth, input_depth ,height, width)
with open(file_name, 'w') as f:
# reshape weight array
weight_depthwise = weight_depthwise.transpose(3, 2, 0, 1)
weight_pointwise = weight_pointwise.transpose(3, 2, 0, 1)
bias_pointwise = np.copy(bias)
if isFixed is True:
weight_depthwise = float2fixed.float2fixed_array(array_type, fractal_width, weight_depthwise)
weight_pointwise = float2fixed.float2fixed_array(array_type, fractal_width, weight_pointwise)
bias_pointwise = float2fixed.float2fixed_array(array_type, fractal_width, bias_pointwise)
# bias_depthwise = np.zeros((bias.shape), dtype=bias_pointwise.dtype)
bias_depthwise = np.zeros((weight_pointwise.shape[1],), dtype=bias_pointwise.dtype)
# headers
todaytime = str(datetime.datetime.today())
f.write("/*\n")
f.write(" * author : shtsno24\n")
f.write(" * Date : " + todaytime + "\n")
if isFixed is True:
f.write(str(" * array_type : {}\n").format(weight_depthwise.dtype))
f.write(str(" * fractal_width : {} bit\n").format(fractal_width))
f.write(str(" * bit_width : {} bit\n").format(str(8 * np.dtype(array_type).itemsize)))
f.write(" *\n */\n")
# include <stdint.h>
f.write(str("#pragma once\n"))
f.write(str("#include <stdint.h>\n\n"))
# define data_width
if isFixed is True:
f.write(str("#define data_width_{} {}\n").format(str(weight_array_name[:-2]), str(8 * np.dtype(array_type).itemsize)))
f.write(str("#define fractal_width_{} {}\n\n").format(str(weight_array_name[:-2]), str(fractal_width)))
# weights(depthwise)
f.write(str("const uint16_t shape_{}_w_d[] = ").format(str(weight_array_name[:-2])))
f.write("{%d, %d, %d, %d};\n" % (weight_depthwise.shape[0], weight_depthwise.shape[1], weight_depthwise.shape[2], weight_depthwise.shape[3]))
f.write("const " + type_name + " " + weight_array_name + "_d")
f.write(str("[{}][{}][{}][{}] =\n").format(weight_depthwise.shape[0], weight_depthwise.shape[1], weight_depthwise.shape[2], weight_depthwise.shape[3]))
write_array_ND(weight_depthwise, f)
f.write(";")
f.write("\n\n")
# weights(pointwise)
f.write(str("const uint16_t shape_{}_w_p[] = ").format(str(weight_array_name[:-2])))
f.write("{%d, %d, %d, %d};\n" % (weight_pointwise.shape[0], weight_pointwise.shape[1], weight_pointwise.shape[2], weight_pointwise.shape[3]))
f.write("const " + type_name + " " + weight_array_name + "_p")
# f.write(str("[{}][{}][{}][{}] =\n").format(weight_pointwise.shape[0], weight_pointwise.shape[1], weight_pointwise.shape[2], weight_pointwise.shape[3]))
f.write(str("[{}] =\n").format(weight_pointwise.shape[0] * weight_pointwise.shape[1] * weight_pointwise.shape[2] * weight_pointwise.shape[3]))
write_array_ND(weight_pointwise.reshape(-1), f)
f.write(";")
f.write("\n\n")
# bias(depthwise)
f.write(str("const uint16_t shape_{}_b_d = {};\n").format(str(weight_array_name[:-2]), bias_depthwise.shape[0]))
f.write("const " + type_name + " " + bias_array_name + "_d")
f.write("[%d] = " % bias_depthwise.shape)
write_array_ND(bias_depthwise, f)
f.write(";")
f.write("\n")
# bias(pointwise)
f.write(str("const uint16_t shape_{}_b_p = {};\n").format(str(weight_array_name[:-2]), bias.shape[0]))
f.write("const " + type_name + " " + bias_array_name + "_p")
f.write("[%d] = " % bias_pointwise.shape)
write_array_ND(bias_pointwise, f)
f.write(";")
f.write("\n")