add support to produce C code

Author: lucastsutsui

README.md

@@ -1,9 +1,9 @@
 
 # EmbML
 
-EmbML is a tool written in Python to automatically convert off-board-trained models into C++ source code files that can be compiled and executed in low-power microcontrollers. The main goal of EmbML is to produce classifier source codes that will run specifically in unresourceful hardware systems, using bare metal programming.
+EmbML is a tool written in Python to automatically convert off-board-trained models into C++ or C (C++ is the default option) source code files that can be compiled and executed in low-power microcontrollers. The main goal of EmbML is to produce classifier source codes that will run specifically in resource-constrained hardware systems, using bare metal programming.
 
-This tool takes as input a classification model that was trained in a desktop or server computer using WEKA or scikit-learn libraries. EmbML is responsible for converting the input model into a carefully crafted C++ code with support for embedded hardware, such as the avoidance of unnecessary use of main memory and implementation of fixed-point operations for non-integer numbers. 
+This tool takes as input a classification model that was trained in a desktop or server computer using WEKA or scikit-learn libraries. EmbML is responsible for converting the input model into a carefully crafted code in C++ or C with support for embedded hardware, such as the avoidance of unnecessary use of SRAM memory and implementation of fixed-point operations for non-integer numbers. 
 
 ## Input Models
 
@@ -35,7 +35,7 @@ You can install `embml` from [PyPi](https://pypi.org/project/embml/):
 pip install embml
 ```
 
-This tool is supported on Python 2.7 and Python 3.5 versions, and depends on the `javaobj` library.
+This tool is supported on Python 2.7 and Python 3.5 versions, and depends on the `javaobj` library (<https://pypi.org/project/javaobj-py3/>).
 
 ## How To Use
 
@@ -58,6 +58,10 @@ embml.wekaModel(inputModel, outputFile, opts)
 embml.wekaModel(inputDecisionTreeModel, outputFile, opts='-rules')
 embml.sklearnModel(inputDecisionTreeModel, outputFile, opts='-rules')
 
+# Examples of generating classifier codes in C programming language.
+embml.wekaModel(inputModel, outputFile, opts='-c')
+embml.sklearnModel(inputModel, outputFile, opts='-c')
+
 # Examples of generating classifier codes using fixed-point formats.
 embml.wekaModel(inputModel, outputFile, opts='-fxp 21 10') # Q21.10
 embml.sklearnModel(inputModel, outputFile, opts='-fxp 21 10') # Q21.10

embml/embml.py

@@ -25,6 +25,8 @@ def processOptions(opt):
  exit(1)
 
  opts['rules'] = ('-rules' in opt.split())
+ 
+ opts['C'] = ('-c' in opt.split())
 
  opts['sigApprox'] = ('-sigApprox' in opt.split())
  opts['pwl'] = ('-sigPwl' in opt.split())

embml/sklearnModels/embml_sklearn_LinearSVC.py

@@ -26,10 +26,15 @@ def write_output(classifier, opts):
  # Classify function
  funcs += utils.write_func_init("int", "classify") + \
  utils.write_dec("int", "indMax", initValue="0", tabs=1) + \
- utils.write_dec(decType, "scores[NUM_CLASSES]", tabs=1) + \
- utils.write_for("i = 0", "i < NUM_CLASSES", "i++", tabs=1) + \
+ utils.write_dec(decType, "scores[NUM_CLASSES]", tabs=1)
+
+ if opts['C']: 
+ funcs += utils.write_dec("int", "i", tabs=1) + \
+ utils.write_dec("int", "j", tabs=1)
+ 
+ funcs += utils.write_for("i = 0", "i < NUM_CLASSES", "i++", tabs=1, inC=opts['C']) + \
  utils.write_attribution("scores[i]", "intercept[i]", tabs=2) + \
- utils.write_for("j = 0", "j < INPUT_SIZE", "j++", tabs=2) + \
+ utils.write_for("j = 0", "j < INPUT_SIZE", "j++", tabs=2, inC=opts['C']) + \
  utils.write_attribution("scores[i]", \
  ("fxp_sum(scores[i], fxp_mul(coef[i][j], instance[j]))" \
  if opts['useFxp'] else \

embml/sklearnModels/embml_sklearn_LogisticRegression.py

@@ -26,10 +26,15 @@ def write_output(classifier, opts):
  # Classify function
  funcs += utils.write_func_init("int", "classify") + \
  utils.write_dec("int", "indMax", initValue="0", tabs=1) + \
- utils.write_dec(decType, "scores[NUM_CLASSES]", tabs=1) + \
- utils.write_for("i = 0", "i < NUM_CLASSES", "i++", tabs=1) + \
+ utils.write_dec(decType, "scores[NUM_CLASSES]", tabs=1)
+
+ if opts['C']: 
+ funcs += utils.write_dec("int", "i", tabs=1) + \
+ utils.write_dec("int", "j", tabs=1)
+ 
+ funcs += utils.write_for("i = 0", "i < NUM_CLASSES", "i++", tabs=1, inC=opts['C']) + \
  utils.write_attribution("scores[i]", "intercept[i]", tabs=2) + \
- utils.write_for("j = 0", "j < INPUT_SIZE", "j++", tabs=2) + \
+ utils.write_for("j = 0", "j < INPUT_SIZE", "j++", tabs=2, inC=opts['C']) + \
  utils.write_attribution("scores[i]", \
  ("fxp_sum(scores[i], fxp_mul(coef[i][j], instance[j]))" \
  if opts['useFxp'] else \

embml/sklearnModels/embml_sklearn_MLPClassifier.py

@@ -153,8 +153,13 @@ def write_output(classifier, opts):
  "const " + decType + " *coef, " + \
  "const " + decType + " *intercept, " + \
  "const int inputSize, const int outputSize") + \
- utils.write_dec("int", "i", initValue="0", tabs=1) + \
- utils.write_for("j = 0", "j < outputSize", "j++", tabs=1) + \
+ utils.write_dec("int", "i", initValue="0", tabs=1)
+ 
+ if opts['C']: 
+ funcs += utils.write_dec("int", "j", tabs=1) + \
+ utils.write_dec("int", "k", tabs=1)
+ 
+ funcs += utils.write_for("j = 0", "j < outputSize", "j++", tabs=1, inC=opts['C']) + \
  utils.write_dec(decType, "acc", initValue=(utils.toFxp(0.0, opts) \
  if opts['useFxp'] else \
  "0.0"), tabs=2) + \
@@ -177,14 +182,19 @@ def write_output(classifier, opts):
  # Classify function
  funcs += utils.write_func_init("int", "classify") + \
  utils.write_dec(decType, "*input", initValue="buffer1", tabs=1) + \
- utils.write_dec(decType, "*output", initValue="buffer2", tabs=1) + \
- utils.write_for("i = 0", "i < INPUT_SIZE", "i++", tabs=1) + \
+ utils.write_dec(decType, "*output", initValue="buffer2", tabs=1)
+ 
+ if opts['C']: 
+ funcs += utils.write_dec("int", "i", tabs=1) + \
+ utils.write_dec("int", "j", tabs=1)
+ 
+ funcs += utils.write_for("i = 0", "i < INPUT_SIZE", "i++", tabs=1, inC=opts['C']) + \
  utils.write_attribution("input[i]", "instance[i]", tabs=2) + \
  utils.write_end(tabs=1) + \
- utils.write_for("i = 0", "i < N_LAYERS - 1", "i++", tabs=1) + \
+ utils.write_for("i = 0", "i < N_LAYERS - 1", "i++", tabs=1, inC=opts['C']) + \
  utils.write_call("forward_pass(input, output, coefs[i], intercepts[i], sizes[i], sizes[i + 1])", tabs=2) + \
  utils.write_if("(i + 1) != (N_LAYERS - 1)", tabs=2) + \
- utils.write_for("j = 0", "j < sizes[i + 1]", "j++", tabs=3) + \
+ utils.write_for("j = 0", "j < sizes[i + 1]", "j++", tabs=3, inC=opts['C']) + \
  utils.write_attribution("output[j]", "activation_hidden(output[j])", tabs=4) + \
  utils.write_end(tabs=3) + \
  utils.write_dec(decType, "*tmp", initValue="input", tabs=3) + \
@@ -193,12 +203,12 @@ def write_output(classifier, opts):
  utils.write_end(tabs=2) + \
  utils.write_else(tabs=2) + \
  utils.write_dec(decType, "max_output", initValue="output[0]", tabs=3) + \
- utils.write_for("j = 1", "j < sizes[i + 1]", "j++", tabs=3) + \
+ utils.write_for("j = 1", "j < sizes[i + 1]", "j++", tabs=3, inC=opts['C']) + \
  utils.write_if("output[j] > max_output", tabs=4) + \
  utils.write_attribution("max_output", "output[j]", tabs=5) + \
  utils.write_end(tabs=4) + \
  utils.write_end(tabs=3) + \
- utils.write_for("j = 0", "j < sizes[i + 1]", "j++", tabs=3) + \
+ utils.write_for("j = 0", "j < sizes[i + 1]", "j++", tabs=3, inC=opts['C']) + \
  utils.write_attribution("output[j]", "activation_output(output[j], max_output)", tabs=4) + \
  utils.write_end(tabs=3) + \
  utils.write_end(tabs=2) + \
@@ -211,7 +221,7 @@ def write_output(classifier, opts):
  else:
  funcs += utils.write_dec("int", "indMax", "0", tabs=1) + \
  utils.write_for("i = 0", \
- "i < sizes[N_LAYERS - 1]", "i++", tabs=1) + \
+ "i < sizes[N_LAYERS - 1]", "i++", tabs=1, inC=opts['C']) + \
  utils.write_if("output[i] > output[indMax]", tabs=2) + \
  utils.write_attribution("indMax", "i", tabs=3) + \
  utils.write_end(tabs=2) + \

embml/sklearnModels/embml_sklearn_SVC_Kernel.py

@@ -64,8 +64,12 @@ def process(self, model, opts):
 def kernelFunction(kernel, decType, opts):
  funcCode = utils.write_dec(decType, "sum", (utils.toFxp(0, opts)\
  if opts['useFxp'] else \
- "0.0"), tabs=1) + \
- utils.write_for("i = 0", "i < INPUT_SIZE", "i++", tabs=1)
+ "0.0"), tabs=1)
+ 
+ if opts['C']: 
+ funcCode += utils.write_dec("int", "i", tabs=1)
+ 
+ funcCode += utils.write_for("i = 0", "i < INPUT_SIZE", "i++", tabs=1, inC=opts['C'])
 
  if kernel == 'linear':
  return funcCode + \
@@ -135,8 +139,14 @@ def write_output(classifier, opts):
 
  # Classify function
  funcs += utils.write_func_init("int", "classify") + \
- utils.write_dec(decType, "k_value[MODEL_L]", tabs=1) + \
- utils.write_for("i = 0", "i < MODEL_L", "i++", tabs=1) + \
+ utils.write_dec(decType, "k_value[MODEL_L]", tabs=1)
+ 
+ if opts['C']: 
+ funcs += utils.write_dec("int", "i", tabs=1) + \
+ utils.write_dec("int", "j", tabs=1) + \
+ utils.write_dec("int", "k", tabs=1)
+ 
+ funcs += utils.write_for("i = 0", "i < MODEL_L", "i++", tabs=1, inC=opts['C']) + \
  utils.write_attribution("k_value[i]",
  "k_function(support_vectors[i])",
  tabs=2) + \
@@ -146,13 +156,13 @@ def write_output(classifier, opts):
  "{0}", tabs=1) + \
  utils.write_dec(utils.chooseDataType(classifier.n_class),
  "p", "0", tabs=1) + \
- utils.write_for("i = 0", "i < NR_CLASS - 1", "i++", tabs=1) + \
- utils.write_for("j = i + 1", "j < NR_CLASS", "j++", tabs=2) + \
+ utils.write_for("i = 0", "i < NR_CLASS - 1", "i++", tabs=1, inC=opts['C']) + \
+ utils.write_for("j = i + 1", "j < NR_CLASS", "j++", tabs=2, inC=opts['C']) + \
  utils.write_dec(decType, "sum", (utils.toFxp(0, opts)\
  if opts['useFxp'] else \
  "0.0"), tabs=3) + \
  utils.write_for("k = end[j - 1][i]", \
- "k < end[j - 1][i + 1]", "k++", tabs=3) + \
+ "k < end[j - 1][i + 1]", "k++", tabs=3, inC=opts['C']) + \
  utils.write_attribution("sum",
  ("fxp_sum(sum, fxp_mul(dual_coef[j - 1][k], k_value[index_sv[j - 1][k]]))"\
  if opts['useFxp'] else\
@@ -161,7 +171,7 @@ def write_output(classifier, opts):
  tabs=4) + \
  utils.write_end(tabs=3) + \
  utils.write_for("k = end[i][j]", \
- "k < end[i][j + 1]", "k++", tabs=3) + \
+ "k < end[i][j + 1]", "k++", tabs=3, inC=opts['C']) + \
  utils.write_attribution("sum",
  ("fxp_sum(sum, fxp_mul(dual_coef[i][k], k_value[index_sv[i][k]]))"\
  if opts['useFxp'] else\
@@ -184,7 +194,7 @@ def write_output(classifier, opts):
  utils.write_end(tabs=2) + \
  utils.write_end(tabs=1) + \
  utils.write_dec("int", "indMax", "0", tabs=1) + \
- utils.write_for("i = 1", "i < NR_CLASS", "i++", tabs=1) + \
+ utils.write_for("i = 1", "i < NR_CLASS", "i++", tabs=1, inC=opts['C']) + \
  utils.write_if("vote[i] > vote[indMax]", tabs=2) + \
  utils.write_attribution("indMax", "i", tabs=3) + \
  utils.write_end(tabs=2) + \

embml/utils/utils.py

@@ -16,9 +16,13 @@ def write_end(tabs=0):
 def write_while(condition, tabs=0):
  return ('\t' * tabs) + 'while (' + condition + '){\n'
 
-def write_for(var, condition, inc, tabs=0):
- return ('\t' * tabs) + 'for (int ' + var + \
- '; ' + condition + '; ' + inc + '){\n'
+def write_for(var, condition, inc, tabs=0, inC=False):
+ if inC:
+ return ('\t' * tabs) + 'for (' + var + \
+ '; ' + condition + '; ' + inc + '){\n'
+ else:
+ return ('\t' * tabs) + 'for (int ' + var + \
+ '; ' + condition + '; ' + inc + '){\n'
 
 def write_if(condition, tabs=0):
  return ('\t' * tabs) + 'if (' + condition + '){\n'

embml/wekaModels/embml_weka_Logistic.py

@@ -72,12 +72,17 @@ def write_output(classifier, opts):
  funcs += utils.write_output_classes(classifier.classes) + '\n' + \
  utils.write_func_init("int", "classify") + \
  utils.write_dec(decType, "prob[NUM_CLASSES]", tabs=1) + \
- utils.write_dec(decType, "newInstance[NUM_PREDICTORS + 1]", tabs=1) + \
- utils.write_attribution("newInstance[0]", \
+ utils.write_dec(decType, "newInstance[NUM_PREDICTORS + 1]", tabs=1)
+ 
+ if opts['C']: 
+ funcs += utils.write_dec("int", "i", tabs=1) + \
+ utils.write_dec("int", "j", tabs=1)
+ 
+ funcs += utils.write_attribution("newInstance[0]", \
  (utils.toFxp(1.0, opts) \
  if opts['useFxp'] else \
  "1.0"), tabs=1) + \
- utils.write_for("i = 1", "i <= SELECTED_ATT_SIZE", "i++", tabs=1) + \
+ utils.write_for("i = 1", "i <= SELECTED_ATT_SIZE", "i++", tabs=1, inC=opts['C']) + \
  utils.write_if("m_SelectedAttributes[i] <= CLASS_INDEX", tabs=2) + \
  utils.write_attribution("newInstance[i]", \
  "instance[m_SelectedAttributes[i - 1]]", \
@@ -93,8 +98,8 @@ def write_output(classifier, opts):
  "v[NUM_CLASSES]", \
  initValue="{0}", \
  tabs=1) + \
- utils.write_for("i = 0", "i < NUM_CLASSES - 1", "i++", tabs=1) + \
- utils.write_for("j = 0", "j <= NUM_PREDICTORS", "j++", tabs=2) + \
+ utils.write_for("i = 0", "i < NUM_CLASSES - 1", "i++", tabs=1, inC=opts['C']) + \
+ utils.write_for("j = 0", "j <= NUM_PREDICTORS", "j++", tabs=2, inC=opts['C']) + \
  utils.write_attribution("v[i]", \
  ("fxp_sum(v[i], fxp_mul(m_Par[(i * (NUM_PREDICTORS + 1)) + j], newInstance[j]))" \
  if opts['useFxp'] else \
@@ -103,10 +108,10 @@ def write_output(classifier, opts):
  utils.write_end(tabs=2) + \
  utils.write_end(tabs=1) + \
  utils.write_attribution("v[NUM_CLASSES - 1]", "0", tabs=1) + \
- utils.write_for("i = 0", "i < NUM_CLASSES", "i++", tabs=1) + \
+ utils.write_for("i = 0", "i < NUM_CLASSES", "i++", tabs=1, inC=opts['C']) + \
  utils.write_dec(decType, "acc", \
  initValue=("0" if opts['useFxp'] else "0.0"), tabs=2) + \
- utils.write_for("j = 0", "j < NUM_CLASSES - 1", "j++", tabs=2) + \
+ utils.write_for("j = 0", "j < NUM_CLASSES - 1", "j++", tabs=2, inC=opts['C']) + \
  utils.write_attribution("acc", \
  ("fxp_sum(acc, fxp_exp(fxp_diff(v[j], v[i])))" \
  if opts['useFxp'] else \
@@ -121,7 +126,7 @@ def write_output(classifier, opts):
  "1.0 / (acc + expf(-v[i]))"), tabs=2) + \
  utils.write_end(tabs=1) + \
  utils.write_dec("int", "indexMax", initValue="0", tabs=1) + \
- utils.write_for("i = 1", "i < NUM_CLASSES", "i++", tabs=1) + \
+ utils.write_for("i = 1", "i < NUM_CLASSES", "i++", tabs=1, inC=opts['C']) + \
  utils.write_if("prob[i] > prob[indexMax]", tabs=2) + \
  utils.write_attribution("indexMax", "i", tabs=3) + \
  utils.write_end(tabs=2) + \

embml/wekaModels/embml_weka_MultilayerPerceptron.py

@@ -196,12 +196,17 @@ def write_output(classifier, opts):
  "const " + decType + " *coef, " + \
  "const " + decType + " *intercept, " + \
  "const int inputSize, const int outputSize") + \
- utils.write_dec("int", "i", initValue="0", tabs=1) + \
- utils.write_for("j = 0", "j < outputSize", "j++", tabs=1) + \
+ utils.write_dec("int", "i", initValue="0", tabs=1)
+ 
+ if opts['C']: 
+ funcs += utils.write_dec("int", "j", tabs=1) + \
+ utils.write_dec("int", "k", tabs=1)
+ 
+ funcs += utils.write_for("j = 0", "j < outputSize", "j++", tabs=1, inC=opts['C']) + \
  utils.write_dec(decType, "acc", initValue=(utils.toFxp(0.0, opts) \
  if opts['useFxp'] else \
  "0.0"), tabs=2) + \
- utils.write_for("k = 0", "k < inputSize", "k++", tabs=2) + \
+ utils.write_for("k = 0", "k < inputSize", "k++", tabs=2, inC=opts['C']) + \
  utils.write_attribution("acc", \
  "fxp_sum(acc, fxp_mul(coef[i++], input[k]))" \
  if opts['useFxp'] else \
@@ -219,8 +224,13 @@ def write_output(classifier, opts):
 
  # Classify function
  funcs += utils.write_output_classes(classifier.classes) + '\n' + \
- utils.write_func_init("int", "classify") + \
- utils.write_for("i = 0", "i < INPUT_SIZE", "i++", tabs=1) + \
+ utils.write_func_init("int", "classify")
+ 
+ if opts['C']: 
+ funcs += utils.write_dec("int", "i", tabs=1) + \
+ utils.write_dec("int", "j", tabs=1)
+ 
+ funcs += utils.write_for("i = 0", "i < INPUT_SIZE", "i++", tabs=1, inC=opts['C']) + \
  utils.write_if("m_attributeRanges[i] != " + ("0"\
  if opts['useFxp'] else \
  "0.0"), tabs=2) + \
@@ -240,12 +250,12 @@ def write_output(classifier, opts):
  utils.write_end(tabs=1) + \
  utils.write_dec(decType, "*input", initValue="buffer1", tabs=1) + \
  utils.write_dec(decType, "*output", initValue="buffer2", tabs=1) + \
- utils.write_for("i = 0", "i < INPUT_SIZE", "i++", tabs=1) + \
+ utils.write_for("i = 0", "i < INPUT_SIZE", "i++", tabs=1, inC=opts['C']) + \
  utils.write_attribution("input[i]", "instance[i]", tabs=2) + \
  utils.write_end(tabs=1) + \
- utils.write_for("i = 0", "i < N_LAYERS - 1", "i++", tabs=1) + \
+ utils.write_for("i = 0", "i < N_LAYERS - 1", "i++", tabs=1, inC=opts['C']) + \
  utils.write_call("forward_pass(input, output, coefs[i], intercepts[i], sizes[i], sizes[i + 1])", tabs=2) + \
- utils.write_for("j = 0", "j < sizes[i + 1]", "j++", tabs=2) + \
+ utils.write_for("j = 0", "j < sizes[i + 1]", "j++", tabs=2, inC=opts['C']) + \
  utils.write_attribution("output[j]", "activation_function(output[j])", tabs=3) + \
  utils.write_end(tabs=2) + \
  utils.write_if("(i + 1) != (N_LAYERS - 1)", tabs=2) + \
@@ -256,7 +266,7 @@ def write_output(classifier, opts):
  utils.write_end(tabs=1) + \
  utils.write_dec("int", "indMax", "0", tabs=1) + \
  utils.write_for("i = 0", \
- "i < sizes[N_LAYERS - 1]", "i++", tabs=1) + \
+ "i < sizes[N_LAYERS - 1]", "i++", tabs=1, inC=opts['C']) + \
  utils.write_if("output[i] > output[indMax]", tabs=2) + \
  utils.write_attribution("indMax", "i", tabs=3) + \
  utils.write_end(tabs=2) + \