Coverage for mlprodict/cli/convert_validate.py: 100%

1""" 

2@file 

3@brief Command line about validation of prediction runtime. 

4""" 

5import os 

6import pickle 

7from logging import getLogger 

8import warnings 

9from pandas import read_csv 

10from ..onnx_conv import to_onnx 

11from ..onnxrt import OnnxInference 

12from ..onnx_tools.optim import onnx_optimisations 

13from ..onnxrt.validate.validate_difference import measure_relative_difference 

14from ..onnx_conv import guess_schema_from_data, guess_schema_from_model 

15 

16 

17def convert_validate(pkl, data=None, schema=None, 

18 method="predict", name='Y', 

19 target_opset=None, 

20 outonnx="model.onnx", 

21 runtime='python', metric="l1med", 

22 use_double=None, noshape=False, 

23 optim='onnx', rewrite_ops=True, 

24 options=None, fLOG=print, verbose=1, 

25 register=True): 

26 """ 

27 Converts a model stored in *pkl* file and measure the differences 

28 between the model and the ONNX predictions. 

29 

30 :param pkl: pickle file 

31 :param data: data file, loaded with pandas, 

32 converted to a single array, the data is used to guess 

33 the schema if *schema* not specified 

34 :param schema: initial type of the model 

35 :param method: method to call 

36 :param name: output name 

37 :param target_opset: target opset 

38 :param outonnx: produced ONNX model 

39 :param runtime: runtime to use to compute predictions, 

40 'python', 'python_compiled', 

41 'onnxruntime1' or 'onnxruntime2' 

42 :param metric: the metric 'l1med' is given by function 

43 :func:`measure_relative_difference 

44 <mlprodict.onnxrt.validate.validate_difference.measure_relative_difference>` 

45 :param noshape: run the conversion with no shape information 

46 :param use_double: use double for the runtime if possible, 

47 two possible options, ``"float64"`` or ``'switch'``, 

48 the first option produces an ONNX file with doubles, 

49 the second option loads an ONNX file (float or double) 

50 and replaces matrices in ONNX with the matrices coming from 

51 the model, this second way is just for testing purposes 

52 :param optim: applies optimisations on the first ONNX graph, 

53 use 'onnx' to reduce the number of node Identity and 

54 redundant subgraphs 

55 :param rewrite_ops: rewrites some converters from :epkg:`sklearn-onnx` 

56 :param options: additional options for conversion, 

57 dictionary as a string 

58 :param verbose: verbose level 

59 :param register: registers additional converters implemented by this package 

60 :param fLOG: logging function 

61 :return: a dictionary with the results 

62 

63 .. cmdref:: 

64 :title: Converts and compares an ONNX file 

65 :cmd: -m mlprodict convert_validate --help 

66 :lid: l-cmd-convert_validate 

67 

68 The command converts and validates a :epkg:`scikit-learn` model. 

69 An example to check the prediction of a logistic regression. 

70 

71 :: 

72 

73 import os 

74 import pickle 

75 import pandas 

76 from sklearn.datasets import load_iris 

77 from sklearn.model_selection import train_test_split 

78 from sklearn.linear_model import LogisticRegression 

79 from mlprodict.__main__ import main 

80 from mlprodict.cli import convert_validate 

81 

82 iris = load_iris() 

83 X, y = iris.data, iris.target 

84 X_train, X_test, y_train, _ = train_test_split(X, y, random_state=11) 

85 clr = LogisticRegression() 

86 clr.fit(X_train, y_train) 

87 

88 pandas.DataFrame(X_test).to_csv("data.csv", index=False) 

89 with open("model.pkl", "wb") as f: 

90 pickle.dump(clr, f) 

91 

92 And the command line to check the predictions 

93 using a command line. 

94 

95 :: 

96 

97 convert_validate --pkl model.pkl --data data.csv 

98 --method predict,predict_proba 

99 --name output_label,output_probability 

100 --verbose 1 

101 """ 

102 from skl2onnx.common.data_types import FloatTensorType, DoubleTensorType # delayed 

103 if fLOG is None: 

104 verbose = 0 # pragma: no cover 

105 if use_double not in (None, 'float64', 'switch'): 

106 raise ValueError( # pragma: no cover 

107 "use_double must be either None, 'float64' or 'switch'") 

108 if optim == '': 

109 optim = None # pragma: no cover 

110 if target_opset == '': 

111 target_opset = None # pragma: no cover 

112 if verbose == 0: 

113 logger = getLogger('skl2onnx') 

114 logger.disabled = True 

115 if not os.path.exists(pkl): 

116 raise FileNotFoundError( # pragma: no cover 

117 "Unable to find model '{}'.".format(pkl)) 

118 if os.path.exists(outonnx): 

119 warnings.warn("File '{}' will be overwritten.".format(outonnx)) 

120 if verbose > 0: 

121 fLOG("[convert_validate] load model '{}'".format(pkl)) 

122 with open(pkl, "rb") as f: 

123 model = pickle.load(f) 

124 

125 if use_double == 'float64': 

126 tensor_type = DoubleTensorType 

127 else: 

128 tensor_type = FloatTensorType 

129 if options in (None, ''): 

130 options = None 

131 else: 

132 from ..onnxrt.validate.validate_scenarios import ( 

133 interpret_options_from_string) 

134 options = interpret_options_from_string(options) 

135 if verbose > 0: 

136 fLOG("[convert_validate] options={}".format(repr(options))) 

137 

138 if register: 

139 from ..onnx_conv import ( 

140 register_converters, register_rewritten_operators) 

141 register_converters() 

142 register_rewritten_operators() 

143 

144 # data and schema 

145 if data is None or not os.path.exists(data): 

146 if schema is None: 

147 schema = guess_schema_from_model(model, tensor_type) 

148 if verbose > 0: 

149 fLOG("[convert_validate] model schema={}".format(schema)) 

150 df = None 

151 else: 

152 if verbose > 0: 

153 fLOG("[convert_validate] load data '{}'".format(data)) 

154 df = read_csv(data) 

155 if verbose > 0: 

156 fLOG("[convert_validate] convert data into matrix") 

157 if schema is None: 

158 schema = guess_schema_from_data(df, tensor_type) 

159 if schema is None: 

160 schema = [ # pragma: no cover 

161 ('X', tensor_type([None, df.shape[1]]))] 

162 if len(schema) == 1: 

163 df = df.values # pylint: disable=E1101 

164 if verbose > 0: 

165 fLOG("[convert_validate] data schema={}".format(schema)) 

166 

167 if noshape: 

168 if verbose > 0: 

169 fLOG( # pragma: no cover 

170 "[convert_validate] convert the model with no shape information") 

171 schema = [(name, col.__class__([None, None])) for name, col in schema] 

172 onx = to_onnx( 

173 model, initial_types=schema, rewrite_ops=rewrite_ops, 

174 target_opset=target_opset, options=options) 

175 else: 

176 if verbose > 0: 

177 fLOG("[convert_validate] convert the model with shapes") 

178 onx = to_onnx( 

179 model, initial_types=schema, target_opset=target_opset, 

180 rewrite_ops=rewrite_ops, options=options) 

181 

182 if optim is not None: 

183 if verbose > 0: 

184 fLOG("[convert_validate] run optimisations '{}'".format(optim)) 

185 onx = onnx_optimisations(onx, optim=optim) 

186 if verbose > 0: 

187 fLOG("[convert_validate] saves to '{}'".format(outonnx)) 

188 memory = onx.SerializeToString() 

189 with open(outonnx, 'wb') as f: 

190 f.write(memory) 

191 

192 if verbose > 0: 

193 fLOG("[convert_validate] creates OnnxInference session") 

194 sess = OnnxInference( 

195 onx, runtime=runtime, runtime_options=dict( 

196 log_severity_level=3)) 

197 if use_double == "switch": 

198 if verbose > 0: 

199 fLOG("[convert_validate] switch to double") 

200 sess.switch_initializers_dtype(model) 

201 

202 if verbose > 0: 

203 fLOG("[convert_validate] compute prediction from model") 

204 

205 if ',' in method: 

206 methods = method.split(',') 

207 else: 

208 methods = [method] 

209 if ',' in name: 

210 names = name.split(',') 

211 else: 

212 names = [name] 

213 

214 if len(names) != len(methods): 

215 raise ValueError( 

216 "Number of methods and outputs do not match: {}, {}".format( 

217 names, methods)) 

218 

219 if metric != 'l1med': 

220 raise ValueError( # pragma: no cover 

221 "Unknown metric '{}'".format(metric)) 

222 

223 if df is None: 

224 # no test on data 

225 return dict(onnx=memory) 

226 

227 if verbose > 0: 

228 fLOG("[convert_validate] compute predictions from ONNX with name '{}'" 

229 "".format(name)) 

230 

231 ort_preds = sess.run( 

232 {'X': df}, verbose=max(verbose - 1, 0), fLOG=fLOG) 

233 

234 metrics = [] 

235 out_skl_preds = [] 

236 out_ort_preds = [] 

237 for method_, name_ in zip(methods, names): 

238 if verbose > 0: 

239 fLOG("[convert_validate] compute predictions with method '{}'".format( 

240 method_)) 

241 meth = getattr(model, method_) 

242 skl_pred = meth(df) 

243 out_skl_preds.append(df) 

244 

245 if name_ not in ort_preds: 

246 raise KeyError( 

247 "Unable to find output name '{}' in {}".format( 

248 name_, list(sorted(ort_preds)))) 

249 

250 ort_pred = ort_preds[name_] 

251 out_ort_preds.append(ort_pred) 

252 diff = measure_relative_difference(skl_pred, ort_pred) 

253 if verbose > 0: 

254 fLOG("[convert_validate] {}={}".format(metric, diff)) 

255 metrics.append(diff) 

256 

257 return dict(skl_pred=out_skl_preds, ort_pred=out_ort_preds, 

258 metrics=metrics, onnx=memory)