Spaces:
Runtime error
Runtime error
File size: 16,928 Bytes
4cea933 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 | import m2cgen as m2c
import numpy as np
def generate_c_code(eng, model_name: str, quantize: bool):
if not eng or model_name not in eng.models:
return {"error": f"Model {model_name} not found"}
model = eng.models[model_name]
# m2cgen handles Random Forest beautifully, but outputs FP32/double rules
if model_name == "rf":
try:
code = m2c.export_to_c(model)
if quantize:
code = "/* WARNING: M2CGen generated FP32 output. INT8 Quantization is not supported directly for Random Forest trees. */\n" + code
return {"code": code}
except Exception:
pass
# For LogReg, use m2cgen if FP32, otherwise manual generation for INT8
if not quantize and model_name == "logreg":
try:
code = m2c.export_to_c(model)
return {"code": code}
except Exception:
pass
# Manual C-code generation for all model types
try:
L = []
L.append("/* ====================================================== */")
L.append(f"/* TinyML C Export: {model_name} */")
q_text = "INT8 Quantized" if quantize else "FP32 Double"
L.append(f"/* Auto-generated for ARM Cortex-M / ESP32 ({q_text}) */")
L.append("/* ====================================================== */")
L.append("")
L.append("#include <math.h>")
L.append("#include <stdint.h>")
L.append("#include <string.h>")
L.append("")
if model_name == "svm" and hasattr(model, 'coef_'):
coefs = model.coef_
intercepts = model.intercept_
n_classes = len(model.classes_)
n_features = coefs.shape[1]
L.append(f"/* Linear SVM with {n_classes} classes, {n_features} features */")
L.append(f"#define N_FEATURES {n_features}")
L.append(f"#define N_CLASSES {n_classes}")
L.append(f"#define N_HYPERPLANES {coefs.shape[0]}")
L.append("")
if quantize:
scale_factor = 127.0 / max(np.max(np.abs(coefs)), np.max(np.abs(intercepts)), 1e-6)
L.append(f"/* Quantization Scale: {scale_factor:.4f} */")
L.append("static const int8_t SVM_COEF[N_HYPERPLANES][N_FEATURES] = {")
for row in coefs:
vals = ", ".join([str(int(round(v * scale_factor))) for v in row])
L.append(f" {{{vals}}},")
L.append("};")
L.append("")
vals = ", ".join([str(int(round(v * scale_factor))) for v in intercepts])
L.append(f"static const int8_t SVM_INTERCEPT[N_HYPERPLANES] = {{{vals}}};")
L.append("")
L.append("int predict(int8_t *features) {")
L.append(" int32_t scores[N_CLASSES] = {0};")
L.append(" int h = 0;")
L.append(" for (int i = 0; i < N_CLASSES; i++) {")
L.append(" for (int j = i + 1; j < N_CLASSES; j++) {")
L.append(" int32_t decision = SVM_INTERCEPT[h];")
L.append(" for (int f = 0; f < N_FEATURES; f++) {")
L.append(" decision += (int32_t)SVM_COEF[h][f] * features[f];")
L.append(" }")
L.append(" if (decision > 0) scores[i] += 1;")
L.append(" else scores[j] += 1;")
L.append(" h++;")
L.append(" }")
L.append(" }")
L.append(" int best = 0;")
L.append(" for (int c = 1; c < N_CLASSES; c++) {")
L.append(" if (scores[c] > scores[best]) best = c;")
L.append(" }")
L.append(" return best;")
L.append("}")
else:
L.append("static const double SVM_COEF[N_HYPERPLANES][N_FEATURES] = {")
for row in coefs:
vals = ", ".join([f"{v:.6f}" for v in row])
L.append(f" {{{vals}}},")
L.append("};")
L.append("")
vals = ", ".join([f"{v:.6f}" for v in intercepts])
L.append(f"static const double SVM_INTERCEPT[N_HYPERPLANES] = {{{vals}}};")
L.append("")
L.append("int predict(double *features) {")
L.append(" double scores[N_CLASSES] = {0};")
L.append(" int h = 0;")
L.append(" for (int i = 0; i < N_CLASSES; i++) {")
L.append(" for (int j = i + 1; j < N_CLASSES; j++) {")
L.append(" double decision = SVM_INTERCEPT[h];")
L.append(" for (int f = 0; f < N_FEATURES; f++) {")
L.append(" decision += SVM_COEF[h][f] * features[f];")
L.append(" }")
L.append(" if (decision > 0) scores[i] += 1.0;")
L.append(" else scores[j] += 1.0;")
L.append(" h++;")
L.append(" }")
L.append(" }")
L.append(" int best = 0;")
L.append(" for (int c = 1; c < N_CLASSES; c++) {")
L.append(" if (scores[c] > scores[best]) best = c;")
L.append(" }")
L.append(" return best;")
L.append("}")
elif model_name == "logreg" and hasattr(model, 'coef_'):
coefs = model.coef_
intercepts = model.intercept_
n_classes = coefs.shape[0] if len(model.classes_) > 2 else 2
n_features = coefs.shape[1]
L.append(f"/* Logistic Regression with {n_classes} classes, {n_features} features */")
L.append(f"#define N_FEATURES {n_features}")
L.append(f"#define N_CLASSES {coefs.shape[0]}")
L.append("")
if quantize:
scale_factor = 127.0 / max(np.max(np.abs(coefs)), np.max(np.abs(intercepts)), 1e-6)
L.append(f"/* Quantization Scale: {scale_factor:.4f} */")
L.append("static const int8_t LOGREG_COEF[N_CLASSES][N_FEATURES] = {")
for row in coefs:
vals = ", ".join([str(int(round(v * scale_factor))) for v in row])
L.append(f" {{{vals}}},")
L.append("};")
L.append("")
vals = ", ".join([str(int(round(v * scale_factor))) for v in intercepts])
L.append(f"static const int8_t LOGREG_INTERCEPT[N_CLASSES] = {{{vals}}};")
L.append("")
L.append("int predict(int8_t *features) {")
L.append(" int32_t scores[N_CLASSES];")
L.append(" for (int c = 0; c < N_CLASSES; c++) {")
L.append(f" scores[c] = LOGREG_INTERCEPT[c] * {int(scale_factor)};")
L.append(" for (int f = 0; f < N_FEATURES; f++) {")
L.append(" scores[c] += (int32_t)LOGREG_COEF[c][f] * features[f];")
L.append(" }")
L.append(" }")
L.append(" int best = 0;")
L.append(" for (int c = 1; c < N_CLASSES; c++) {")
L.append(" if (scores[c] > scores[best]) best = c;")
L.append(" }")
L.append(" return best;")
L.append("}")
else:
L.append("static const double LOGREG_COEF[N_CLASSES][N_FEATURES] = {")
for row in coefs:
vals = ", ".join([f"{v:.6f}" for v in row])
L.append(f" {{{vals}}},")
L.append("};")
L.append("")
vals = ", ".join([f"{v:.6f}" for v in intercepts])
L.append(f"static const double LOGREG_INTERCEPT[N_CLASSES] = {{{vals}}};")
L.append("")
L.append("int predict(double *features) {")
L.append(" double scores[N_CLASSES];")
L.append(" for (int c = 0; c < N_CLASSES; c++) {")
L.append(" scores[c] = LOGREG_INTERCEPT[c];")
L.append(" for (int f = 0; f < N_FEATURES; f++) {")
L.append(" scores[c] += LOGREG_COEF[c][f] * features[f];")
L.append(" }")
L.append(" }")
L.append(" int best = 0;")
L.append(" for (int c = 1; c < N_CLASSES; c++) {")
L.append(" if (scores[c] > scores[best]) best = c;")
L.append(" }")
L.append(" return best;")
L.append("}")
elif model_name == "small_nn" and hasattr(model, 'coefs_'):
layers = model.coefs_
biases = model.intercepts_
arch = " -> ".join([str(l.shape[0]) for l in layers] + [str(layers[-1].shape[1])])
L.append(f"/* MLP Neural Network: {len(layers)} layers */")
L.append(f"/* Architecture: {arch} */")
L.append("")
for idx, (W, b) in enumerate(zip(layers, biases)):
n_in, n_out = W.shape
L.append(f"#define L{idx}_IN {n_in}")
L.append(f"#define L{idx}_OUT {n_out}")
L.append(f"static const double W{idx}[{n_in}][{n_out}] = {{")
for row in W:
vals = ", ".join([f"{v:.6f}" for v in row])
L.append(f" {{{vals}}},")
L.append("};")
bvals = ", ".join([f"{v:.6f}" for v in b])
L.append(f"static const double B{idx}[{n_out}] = {{{bvals}}};")
L.append("")
L.append("static inline double relu(double x) { return x > 0 ? x : 0; }")
L.append("")
if quantize:
# Calculate global max for int8 scaling
max_val = max([np.max(np.abs(w)) for w in layers] + [np.max(np.abs(b)) for b in biases] + [1e-6])
scale_factor = 127.0 / max_val
L.append(f"/* INT8 Quantization Scale Factor: {scale_factor:.4f} */")
for idx, (W, b) in enumerate(zip(layers, biases)):
n_in, n_out = W.shape
L.append(f"static const int8_t W{idx}[{n_in}][{n_out}] = {{")
for row in W:
vals = ", ".join([str(int(round(v * scale_factor))) for v in row])
L.append(f" {{{vals}}},")
L.append("};")
bvals = ", ".join([str(int(round(v * scale_factor))) for v in b])
L.append(f"static const int8_t B{idx}[{n_out}] = {{{bvals}}};")
L.append("")
L.append("static inline int32_t relu_int(int32_t x) { return x > 0 ? x : 0; }")
L.append("")
L.append("int predict(int8_t *input) {")
for idx in range(len(layers)):
n_in = layers[idx].shape[0]
n_out = layers[idx].shape[1]
is_last = (idx == len(layers) - 1)
src = "input" if idx == 0 else f"a{idx-1}"
L.append(f" int32_t a{idx}[{n_out}];")
L.append(f" for (int j = 0; j < {n_out}; j++) {{")
L.append(f" a{idx}[j] = B{idx}[j] * {int(scale_factor)}; /* scale bias */")
L.append(f" for (int i = 0; i < {n_in}; i++) {{")
L.append(f" a{idx}[j] += (int32_t){src}[i] * W{idx}[i][j];")
L.append(f" }}")
if not is_last:
L.append(f" a{idx}[j] = relu_int(a{idx}[j]) / {int(scale_factor)}; /* rescale */")
L.append(f" }}")
last_idx = len(layers) - 1
last_out = layers[-1].shape[1]
L.append(f" int best = 0;")
L.append(f" for (int c = 1; c < {last_out}; c++) {{")
L.append(f" if (a{last_idx}[c] > a{last_idx}[best]) best = c;")
L.append(f" }}")
L.append(f" return best;")
L.append("}")
else:
for idx, (W, b) in enumerate(zip(layers, biases)):
n_in, n_out = W.shape
L.append(f"#define L{idx}_IN {n_in}")
L.append(f"#define L{idx}_OUT {n_out}")
L.append(f"static const double W{idx}[{n_in}][{n_out}] = {{")
for row in W:
vals = ", ".join([f"{v:.6f}" for v in row])
L.append(f" {{{vals}}},")
L.append("};")
bvals = ", ".join([f"{v:.6f}" for v in b])
L.append(f"static const double B{idx}[{n_out}] = {{{bvals}}};")
L.append("")
L.append("static inline double relu(double x) { return x > 0 ? x : 0; }")
L.append("")
L.append("int predict(double *input) {")
for idx in range(len(layers)):
n_in = layers[idx].shape[0]
n_out = layers[idx].shape[1]
is_last = (idx == len(layers) - 1)
src = "input" if idx == 0 else f"a{idx-1}"
L.append(f" double a{idx}[{n_out}];")
L.append(f" for (int j = 0; j < {n_out}; j++) {{")
L.append(f" a{idx}[j] = B{idx}[j];")
L.append(f" for (int i = 0; i < {n_in}; i++) {{")
L.append(f" a{idx}[j] += {src}[i] * W{idx}[i][j];")
L.append(f" }}")
if not is_last:
L.append(f" a{idx}[j] = relu(a{idx}[j]);")
L.append(f" }}")
last_idx = len(layers) - 1
last_out = layers[-1].shape[1]
L.append(f" int best = 0;")
L.append(f" for (int c = 1; c < {last_out}; c++) {{")
L.append(f" if (a{last_idx}[c] > a{last_idx}[best]) best = c;")
L.append(f" }}")
L.append(f" return best;")
L.append("}")
elif model_name == "knn" and hasattr(model, '_fit_X'):
n_samples = min(model._fit_X.shape[0], 100)
n_feats = model._fit_X.shape[1]
L.append(f"/* KNN Lookup Table: {n_samples} reference samples */")
L.append(f"#define N_NEIGHBORS {model.n_neighbors}")
L.append(f"#define N_SAMPLES {n_samples}")
L.append(f"#define N_FEATURES {n_feats}")
L.append("")
L.append("static const double REF[N_SAMPLES][N_FEATURES] = {")
for row in model._fit_X[:n_samples]:
vals = ", ".join([f"{v:.4f}" for v in row])
L.append(f" {{{vals}}},")
L.append("};")
L.append("")
labels_str = ", ".join([str(int(l)) for l in model._y[:n_samples]])
L.append(f"static const int LABELS[N_SAMPLES] = {{{labels_str}}};")
L.append("")
L.append("int predict(double *features) {")
L.append(" double dists[N_SAMPLES];")
L.append(" for (int i = 0; i < N_SAMPLES; i++) {")
L.append(" dists[i] = 0.0;")
L.append(" for (int f = 0; f < N_FEATURES; f++) {")
L.append(" double d = features[f] - REF[i][f];")
L.append(" dists[i] += d * d;")
L.append(" }")
L.append(" }")
L.append(" int votes[10] = {0};")
L.append(" for (int k = 0; k < N_NEIGHBORS; k++) {")
L.append(" int mi = 0;")
L.append(" for (int i = 1; i < N_SAMPLES; i++) {")
L.append(" if (dists[i] < dists[mi]) mi = i;")
L.append(" }")
L.append(" votes[LABELS[mi]]++;")
L.append(" dists[mi] = 1e18;")
L.append(" }")
L.append(" int best = 0;")
L.append(" for (int i = 1; i < 10; i++) {")
L.append(" if (votes[i] > votes[best]) best = i;")
L.append(" }")
L.append(" return best;")
L.append("}")
else:
return {"error": f"Model {model_name} cannot be exported to C."}
return {"code": "\n".join(L)}
except Exception as e:
return {"error": f"Export failed: {str(e)}"}
|