init

demo/lenet.c

@@ -0,0 +1,136 @@
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "catcoon.h"
+
+#define INPUT_W 28
+#define INPUT_H 28
+
+const char *parameters_files[]={
+	"conv1_w.bin",
+	"conv1_b.bin",
+	"conv2_w.bin",
+	"conv2_b.bin",
+	"fc1_w.bin",
+	"fc1_b.bin",
+	"fc2_w.bin",
+	"fc2_b.bin",
+};
+const char *parameters_path = "/home/evi/catcoon-pytorch-model/lenet/dump";
+
+int main(int argc, const char *argv[])
+{
+	cc_int32 i, j;
+	cc_float32 v;
+	char filepath[256];
+	/* Image */
+	utim_image_t *img, *img_read;
+	/* Layers */
+	cc_tensor_t *input,*l1, *l1_pool, *l2, *l2_pool, *l2_flat, *l3, *l4;
+	/* Parameters */
+	cc_tensor_t *conv1_w, *conv1_b,
+		*conv2_w, *conv2_b, *fc1_w, *fc1_b, *fc2_w, *fc2_b;
+	cc_int32 shape_conv1_w[] = {5, 5, 1, 32, 0};
+	cc_int32 shape_conv1_b[] = {32, 0};
+	cc_int32 shape_conv2_w[] = {5, 5, 32, 64, 0};
+	cc_int32 shape_conv2_b[] = {64, 0};
+	cc_int32 shape_flat[]    = {-1, 1, 1, 0};
+	cc_int32 shape_fc1_w[]   = {1, 1, 3136, 128, 0};
+	cc_int32 shape_fc1_b[]   = {128, 0};
+	cc_int32 shape_fc2_w[]   = {1, 1, 128, 10, 0};
+	cc_int32 shape_fc2_b[]   = {10, 0};
+
+	if (argc < 2) {
+		fprintf(stderr, "usage: lenet [filename]\n");
+		exit(255);
+	}
+
+	img_read = utim_read(argv[1]);
+	utim_img2gray(img_read);
+	img = utim_resize(img_read, INPUT_H, INPUT_W, 0);
+	input = cc_image2tensor(img, "input");
+	utim_free_image(img);
+	utim_free_image(img_read);
+	input = cc_cast_tensor(input, CC_FLOAT32, "input");
+	v = 255.;
+	input = cc_tensor_by_scalar(input, '/', &v, "input");
+
+	i = 0;
+	sprintf(filepath, "%s/%s",
+		parameters_path, parameters_files[i++]);
+	conv1_w = cc_load_bin(filepath,
+		shape_conv1_w, CC_FLOAT32, "conv1_w");
+
+	sprintf(filepath, "%s/%s",
+		parameters_path, parameters_files[i++]);
+	conv1_b = cc_load_bin(filepath,
+		shape_conv1_b, CC_FLOAT32, "conv1_b");
+
+	sprintf(filepath, "%s/%s",
+		parameters_path, parameters_files[i++]);
+	conv2_w = cc_load_bin(filepath,
+		shape_conv2_w, CC_FLOAT32, "conv2_w");
+
+	sprintf(filepath, "%s/%s",
+		parameters_path, parameters_files[i++]);
+	conv2_b = cc_load_bin(filepath,
+		shape_conv2_b, CC_FLOAT32, "conv2_b");
+
+	sprintf(filepath, "%s/%s",
+		parameters_path, parameters_files[i++]);
+	fc1_w = cc_load_bin(filepath,
+		shape_fc1_w, CC_FLOAT32, "fc1_w");
+
+	sprintf(filepath, "%s/%s",
+		parameters_path, parameters_files[i++]);
+	fc1_b = cc_load_bin(filepath,
+		shape_fc1_b, CC_FLOAT32, "fc1_b");
+
+	sprintf(filepath, "%s/%s",
+		parameters_path, parameters_files[i++]);
+	fc2_w = cc_load_bin(filepath,
+		shape_fc2_w, CC_FLOAT32, "fc2_w");
+
+	sprintf(filepath, "%s/%s",
+		parameters_path, parameters_files[i++]);
+	fc2_b = cc_load_bin(filepath,
+		shape_fc2_b, CC_FLOAT32, "fc2_b");
+
+	l1 = cc_conv2d(input, conv1_w, conv1_b, 1, 2, 0, "l1_conv");
+
+	l1 = cc_relu(l1, NULL);
+
+	l1_pool = cc_max_pool2d(l1, 2, "l1_pool");
+
+	l2 = cc_conv2d(l1_pool, conv2_w, conv2_b, 1, 2, 0, "l2_conv");
+
+	l2 = cc_relu(l2, NULL);
+
+	l2_pool = cc_max_pool2d(l2, 2, "l2_pool");
+
+	/* l2_flat = cc_fmap2d_flat(l2_pool, "l2_flat"); */
+	l2_flat = cc_tensor_reshape(l2_pool, shape_flat);
+
+	/* l3 = cc_conv2d(l2_flat, fc1_w, fc1_b, 1, 0, 0, "fc1"); */
+	l3 = cc_fully_connected(l2_flat, fc1_w, fc1_b, "fc1");
+
+	l3 = cc_relu(l3, NULL);
+
+	/* l4 = cc_conv2d(l3, fc2_w, fc2_b, 1, 0, 0, "fc2"); */
+	l4 = cc_fully_connected(l3, fc2_w, fc2_b, "fc2");
+
+	l4 = cc_softmax(l4, NULL);
+
+	v = 0.;
+	for (i = 0, j = 0; i < 10; ++i) {
+		if (*((cc_float32*)l4->data + i) > v) {
+			v = *((cc_float32*)l4->data + i);
+			j = i;
+		}
+		printf("[%d]: %f\n", i, *((cc_float32*)l4->data + i));
+	}
+	printf("Result of \"%s\": [%d]\n", argv[1], j);
+	cc_tsrmgr_list();
+	cc_clear();
+	return 0;
+}

demo/simple.c

@@ -0,0 +1,11 @@
+#include "catcoon.h"
+
+int main(int argc, const char *argv[])
+{
+	cc_tensor_t *tensor;
+	cc_int32 shape[] = {3, 3, 3, 0};
+	tensor = cc_create_tensor(shape, CC_FLOAT32, "tensor0");
+	cc_print_tensor_property(tensor);
+	cc_free_tensor(tensor);
+	return 0;
+}

makefile

@@ -0,0 +1,89 @@
+cc = gcc
+
+# Debug flag, cc_assert
+DFLAG += -DENABLE_CC_ASSERT
+# AddressSanitizer for gcc/clang
+DFLAG += # -g -fsanitize=address -fno-omit-frame-pointer
+
+CFLAG += # -std=c89
+CFLAG += -Wall # -Wpedantic
+
+OFLAG += -Os
+
+# Enable OpenMP
+OFLAG += -DENABLE_OPENMP -fopenmp
+
+CFLAG += $(DFLAG) $(WFLAG) $(OFLAG)
+
+# Enable automatic tensor manager
+CFLAG += -DAUTO_TSRMGR
+
+# 3rd party source/library configurations
+# stb for jpg, png, tga format images,
+# will only support bmp image if disabled
+3RDSRC_CF += stb
+
+ifneq ($(findstring -std=c89, $(CFLAG)),)
+	CFLAG += -DCONFIG_STD_C89
+	ifneq ($(findstring $(cc), gcc clang),)
+		WFLAG += -Wno-long-long
+	endif
+endif
+
+LINK += -lm
+INC  += -I ./src/
+ALL_O = catcoon.o cc_tensor.o cc_dtype.o cc_image.o cc_array.o cc_tsrmgr.o \
+	cc_basic.o cc_fmap2d.o cc_pad2d.o cc_conv2d.o cc_pool2d.o cc_cpufn.o \
+	cc_fullycon.o global_fn_cfg.o util_rbt.o util_list.o util_log.o \
+	util_image.o
+
+APP_NAMES = simple lenet
+
+ifeq ($(OS),Windows_NT)
+	RM = del
+	APPS =  $(foreach v, $(APP_NAMES), $(v).exe)
+else
+	RM = rm
+	APPS = $(APP_NAMES)
+endif
+
+all: $(APPS)
+
+%.o: ./src/%.c
+	$(cc) -c -o $@ $< $(CFLAG) $(INC)
+# APPS For Linux
+%: ./demo/%.c $(ALL_O)
+	$(cc) -o $@ $< $(ALL_O) $(CFLAG) $(INC) $(LINK)
+# APPS For Windows
+%.exe: ./demo/%.c $(ALL_O)
+	$(cc) -o $@ $< $(ALL_O) $(CFLAG) $(INC) $(LINK)
+
+global_fn_cfg.o : $(patsubst %, ./src/%, global_fn_cfg.h global_fn_cfg.c)
+
+catcoon.o     : $(patsubst %, ./src/%, catcoon.h catcoon.c)
+cc_array.o    : $(patsubst %, ./src/%, cc_array.h cc_array.c)
+cc_basic.o    : $(patsubst %, ./src/%, cc_basic.h cc_basic.c)
+cc_conv2d.o   : $(patsubst %, ./src/%, cc_conv2d.h cc_conv2d.c)
+cc_cpufn.o    : $(patsubst %, ./src/%, cc_cpufn.h cc_cpufn.c)
+cc_dtype.o    : $(patsubst %, ./src/%, cc_dtype.h cc_dtype.c)
+cc_fmap2d.o   : $(patsubst %, ./src/%, cc_fmap2d.h cc_fmap2d.c)
+cc_fullycon.o : $(patsubst %, ./src/%, cc_fullycon.h cc_fullycon.c)
+cc_image.o    : $(patsubst %, ./src/%, cc_image.h cc_image.c)
+cc_pad2d.o    : $(patsubst %, ./src/%, cc_pad2d.h cc_pad2d.c)
+cc_pool2d.o   : $(patsubst %, ./src/%, cc_pool2d.h cc_pool2d.c)
+cc_tsrmgr.o   : $(patsubst %, ./src/%, cc_tsrmgr.h cc_tsrmgr.c)
+cc_tensor.o   : $(patsubst %, ./src/%, cc_tensor.h cc_tensor.c)
+
+util_log.o   : $(patsubst %, ./src/%, util_log.h util_log.c)
+util_rbt.o   : $(patsubst %, ./src/%, util_rbt.h util_rbt.c)
+util_list.o  :  ./src/util_list.h ./src/util_list.c
+	$(cc) -c -o $@ ./src/util_list.c $(CFLAG) -DENABLE_FOPS
+UTIM_COND = 
+ifneq ($(findstring stb, $(3RDSRC_CF)),)
+	UTIM_COND += -DUSE3RD_STB_IMAGE -I ./src/3rd_party/stb/
+endif
+util_image.o : ./src/util_image.h ./src/util_image.c
+	$(cc) -c -o $@ ./src/util_image.c $(CFLAG) $(UTIM_COND)
+
+clean:
+	$(RM) *.o && $(RM) $(APPS)