egs/wsj/s5/steps/libs/nnet3/report/log_parse.py

   1
   2
   3 # Copyright 2016    Vijayaditya Peddinti
   4 #                   Vimal Manohar
   5 # Apache 2.0.
   6
   7 from __future__ import division
   8 import datetime
   9 import re
  10 import sys
  11
  12 import libs.common as common_lib
  13
  14
  15 def parse_progress_logs_for_nonlinearity_stats(exp_dir):
  16     """ Parse progress logs for mean and std stats for non-linearities.
  17
  18     e.g. for a line that is parsed from progress.*.log:
  19     exp/nnet3/lstm_self_repair_ld5_sp/log/progress.9.log:component name=Lstm3_i
  20     type=SigmoidComponent, dim=1280, self-repair-scale=1e-05, count=1.96e+05,
  21     value-avg=[percentiles(0,1,2,5 10,20,50,80,90
  22     95,98,99,100)=(0.05,0.09,0.11,0.15 0.19,0.27,0.50,0.72,0.83
  23     0.88,0.92,0.94,0.99), mean=0.502, stddev=0.23],
  24     deriv-avg=[percentiles(0,1,2,5 10,20,50,80,90
  25     95,98,99,100)=(0.009,0.04,0.05,0.06 0.08,0.10,0.14,0.17,0.18
  26     0.19,0.20,0.20,0.21), mean=0.134, stddev=0.0397]
  27     """
  28
  29     progress_log_files = "%s/log/progress.*.log" % (exp_dir)
  30     stats_per_component_per_iter = {}
  31
  32     progress_log_lines = common_lib.run_kaldi_command(
  33         'grep -e "value-avg.*deriv-avg" {0}'.format(progress_log_files))[0]
  34
  35     parse_regex = re.compile(
  36         ".*progress.([0-9]+).log:component name=(.+) "
  37         "type=(.*)Component,.*"
  38         "value-avg=\[.*mean=([0-9\.\-e]+), stddev=([0-9\.e\-]+)\].*"
  39         "deriv-avg=\[.*mean=([0-9\.\-e]+), stddev=([0-9\.e\-]+)\]")
  40
  41     for line in progress_log_lines.split("\n"):
  42         mat_obj = parse_regex.search(line)
  43         if mat_obj is None:
  44             continue
  45         # groups = ('9', 'Lstm3_i', 'Sigmoid', '0.502', '0.23',
  46         # '0.134', '0.0397')
  47         groups = mat_obj.groups()
  48         iteration = int(groups[0])
  49         component_name = groups[1]
  50         component_type = groups[2]
  51         value_mean = float(groups[3])
  52         value_stddev = float(groups[4])
  53         deriv_mean = float(groups[5])
  54         deriv_stddev = float(groups[6])
  55         try:
  56             stats_per_component_per_iter[component_name][
  57                 'stats'][iteration] = [value_mean, value_stddev,
  58                                        deriv_mean, deriv_stddev]
  59         except KeyError:
  60             stats_per_component_per_iter[component_name] = {}
  61             stats_per_component_per_iter[component_name][
  62                 'type'] = component_type
  63             stats_per_component_per_iter[component_name]['stats'] = {}
  64             stats_per_component_per_iter[component_name][
  65                 'stats'][iteration] = [value_mean, value_stddev,
  66                                        deriv_mean, deriv_stddev]
  67
  68     return stats_per_component_per_iter
  69
  70
  71 def parse_difference_string(string):
  72     dict = {}
  73     for parts in string.split():
  74         sub_parts = parts.split(":")
  75         dict[sub_parts[0]] = float(sub_parts[1])
  76     return dict
  77
  78
  79 class MalformedClippedProportionLineException(Exception):
  80     def __init__(self, line):
  81         Exception.__init__(self,
  82                            "Malformed line encountered while trying to "
  83                            "extract clipped-proportions.\n{0}".format(line))
  84
  85
  86 def parse_progress_logs_for_clipped_proportion(exp_dir):
  87     """ Parse progress logs for clipped proportion stats.
  88
  89     e.g. for a line that is parsed from progress.*.log:
  90     exp/chain/cwrnn_trial2_ld5_sp/log/progress.245.log:component
  91     name=BLstm1_forward_c type=ClipGradientComponent, dim=512,
  92     norm-based-clipping=true, clipping-threshold=30,
  93     clipped-proportion=0.000565527,
  94     self-repair-clipped-proportion-threshold=0.01, self-repair-target=0,
  95     self-repair-scale=1
  96     """
  97
  98     progress_log_files = "%s/log/progress.*.log" % (exp_dir)
  99     component_names = set([])
 100     progress_log_lines = common_lib.run_kaldi_command(
 101         'grep -e "{0}" {1}'.format(
 102             "clipped-proportion", progress_log_files))[0]
 103     parse_regex = re.compile(".*progress\.([0-9]+)\.log:component "
 104                              "name=(.*) type=.* "
 105                              "clipped-proportion=([0-9\.e\-]+)")
 106
 107     cp_per_component_per_iter = {}
 108
 109     max_iteration = 0
 110     component_names = set([])
 111     for line in progress_log_lines.split("\n"):
 112         mat_obj = parse_regex.search(line)
 113         if mat_obj is None:
 114             if line.strip() == "":
 115                 continue
 116             raise MalformedClippedProportionLineException(line)
 117         groups = mat_obj.groups()
 118         iteration = int(groups[0])
 119         max_iteration = max(max_iteration, iteration)
 120         name = groups[1]
 121         clipped_proportion = float(groups[2])
 122         if clipped_proportion > 1:
 123             raise MalformedClippedProportionLineException(line)
 124         if iteration not in cp_per_component_per_iter:
 125             cp_per_component_per_iter[iteration] = {}
 126         cp_per_component_per_iter[iteration][name] = clipped_proportion
 127         component_names.add(name)
 128     component_names = list(component_names)
 129     component_names.sort()
 130
 131     # re arranging the data into an array
 132     # and into an cp_per_iter_per_component
 133     cp_per_iter_per_component = {}
 134     for component_name in component_names:
 135         cp_per_iter_per_component[component_name] = []
 136     data = []
 137     data.append(["iteration"]+component_names)
 138     for iter in range(max_iteration+1):
 139         if iter not in cp_per_component_per_iter:
 140             continue
 141         comp_dict = cp_per_component_per_iter[iter]
 142         row = [iter]
 143         for component in component_names:
 144             try:
 145                 row.append(comp_dict[component])
 146                 cp_per_iter_per_component[component].append(
 147                     [iter, comp_dict[component]])
 148             except KeyError:
 149                 # if clipped proportion is not available for a particular
 150                 # component it is set to None
 151                 # this usually happens during layer-wise discriminative
 152                 # training
 153                 row.append(None)
 154         data.append(row)
 155
 156     return {'table': data,
 157             'cp_per_component_per_iter': cp_per_component_per_iter,
 158             'cp_per_iter_per_component': cp_per_iter_per_component}
 159
 160
 161 def parse_progress_logs_for_param_diffP(exp_dir, pattern, logger=None):
 162     """ Parse progress logs for per-component parameter differences.
 163
 164     e.g. for a line that is parsed from progress.*.log:
 165     exp/chain/cwrnn_trial2_ld5_sp/log/progress.245.log:LOG
 166     (nnet3-show-progress:main():nnet3-show-progress.cc:144) Relative parameter
 167     differences per layer are [ Cwrnn1_T3_W_r:0.0171537
 168     Cwrnn1_T3_W_x:1.33338e-07 Cwrnn1_T2_W_r:0.048075 Cwrnn1_T2_W_x:1.34088e-07
 169     Cwrnn1_T1_W_r:0.0157277 Cwrnn1_T1_W_x:0.0212704 Final_affine:0.0321521
 170     Cwrnn2_T3_W_r:0.0212082 Cwrnn2_T3_W_x:1.33691e-07 Cwrnn2_T2_W_r:0.0212978
 171     Cwrnn2_T2_W_x:1.33401e-07 Cwrnn2_T1_W_r:0.014976 Cwrnn2_T1_W_x:0.0233588
 172     Cwrnn3_T3_W_r:0.0237165 Cwrnn3_T3_W_x:1.33184e-07 Cwrnn3_T2_W_r:0.0239754
 173     Cwrnn3_T2_W_x:1.3296e-07 Cwrnn3_T1_W_r:0.0194809 Cwrnn3_T1_W_x:0.0271934 ]
 174     """
 175
 176     if pattern not in set(["Relative parameter differences",
 177                            "Parameter differences"]):
 178         raise Exception("Unknown value for pattern : {0}".format(pattern))
 179
 180     progress_log_files = "%s/log/progress.*.log" % (exp_dir)
 181     progress_per_iter = {}
 182     component_names = set([])
 183     progress_log_lines = common_lib.run_kaldi_command(
 184         'grep -e "{0}" {1}'.format(pattern, progress_log_files))[0]
 185     parse_regex = re.compile(".*progress\.([0-9]+)\.log:"
 186                              "LOG.*{0}.*\[(.*)\]".format(pattern))
 187     for line in progress_log_lines.split("\n"):
 188         mat_obj = parse_regex.search(line)
 189         if mat_obj is None:
 190             continue
 191         groups = mat_obj.groups()
 192         iteration = groups[0]
 193         differences = parse_difference_string(groups[1])
 194         component_names = component_names.union(differences.keys())
 195         progress_per_iter[int(iteration)] = differences
 196
 197     component_names = list(component_names)
 198     component_names.sort()
 199     # rearranging the parameter differences available per iter
 200     # into parameter differences per component
 201     progress_per_component = {}
 202     for cn in component_names:
 203         progress_per_component[cn] = {}
 204
 205     max_iter = max(progress_per_iter.keys())
 206     total_missing_iterations = 0
 207     gave_user_warning = False
 208     for iter in range(max_iter + 1):
 209         try:
 210             component_dict = progress_per_iter[iter]
 211         except KeyError:
 212             continue
 213
 214         for component_name in component_names:
 215             try:
 216                 progress_per_component[component_name][iter] = component_dict[
 217                     component_name]
 218             except KeyError:
 219                 total_missing_iterations += 1
 220                 # the component was not found this iteration, may be because of
 221                 # layerwise discriminative training
 222                 pass
 223         if (total_missing_iterations/len(component_names) > 20
 224                 and not gave_user_warning and logger is not None):
 225             logger.warning("There are more than {0} missing iterations per "
 226                            "component. Something might be wrong.".format(
 227                                 total_missing_iterations/len(component_names)))
 228             gave_user_warning = True
 229
 230     return {'progress_per_component': progress_per_component,
 231             'component_names': component_names,
 232             'max_iter': max_iter}
 233
 234
 235 def parse_train_logs(exp_dir):
 236     train_log_files = "%s/log/train.*.log" % (exp_dir)
 237     train_log_lines = common_lib.run_kaldi_command(
 238         'grep -e Accounting {0}'.format(train_log_files))[0]
 239     parse_regex = re.compile(".*train\.([0-9]+)\.([0-9]+)\.log:# "
 240                              "Accounting: time=([0-9]+) thread.*")
 241
 242     train_times = {}
 243     for line in train_log_lines.split('\n'):
 244         mat_obj = parse_regex.search(line)
 245         if mat_obj is not None:
 246             groups = mat_obj.groups()
 247             try:
 248                 train_times[int(groups[0])][int(groups[1])] = float(groups[2])
 249             except KeyError:
 250                 train_times[int(groups[0])] = {}
 251                 train_times[int(groups[0])][int(groups[1])] = float(groups[2])
 252     iters = train_times.keys()
 253     for iter in iters:
 254         values = train_times[iter].values()
 255         train_times[iter] = max(values)
 256     return train_times
 257
 258
 259 def parse_prob_logs(exp_dir, key='accuracy'):
 260     train_prob_files = "%s/log/compute_prob_train.*.log" % (exp_dir)
 261     valid_prob_files = "%s/log/compute_prob_valid.*.log" % (exp_dir)
 262     train_prob_strings = common_lib.run_kaldi_command(
 263         'grep -e {0} {1}'.format(key, train_prob_files), wait=True)[0]
 264     valid_prob_strings = common_lib.run_kaldi_command(
 265         'grep -e {0} {1}'.format(key, valid_prob_files))[0]
 266
 267     # LOG
 268     # (nnet3-chain-compute-prob:PrintTotalStats():nnet-chain-diagnostics.cc:149)
 269     # Overall log-probability for 'output' is -0.399395 + -0.013437 = -0.412832
 270     # per frame, over 20000 fra
 271
 272     # LOG
 273     # (nnet3-chain-compute-prob:PrintTotalStats():nnet-chain-diagnostics.cc:144)
 274     # Overall log-probability for 'output' is -0.307255 per frame, over 20000
 275     # frames.
 276
 277     parse_regex = re.compile(
 278         ".*compute_prob_.*\.([0-9]+).log:LOG "
 279         ".nnet3.*compute-prob:PrintTotalStats..:"
 280         "nnet.*diagnostics.cc:[0-9]+. Overall ([a-zA-Z\-]+) for "
 281         "'output'.*is ([0-9.\-e]+) .*per frame")
 282
 283     train_loss = {}
 284     valid_loss = {}
 285
 286     for line in train_prob_strings.split('\n'):
 287         mat_obj = parse_regex.search(line)
 288         if mat_obj is not None:
 289             groups = mat_obj.groups()
 290             if groups[1] == key:
 291                 train_loss[int(groups[0])] = groups[2]
 292     for line in valid_prob_strings.split('\n'):
 293         mat_obj = parse_regex.search(line)
 294         if mat_obj is not None:
 295             groups = mat_obj.groups()
 296             if groups[1] == key:
 297                 valid_loss[int(groups[0])] = groups[2]
 298     iters = list(set(valid_loss.keys()).intersection(train_loss.keys()))
 299     iters.sort()
 300     return map(lambda x: (int(x), float(train_loss[x]),
 301                           float(valid_loss[x])), iters)
 302
 303
 304 def generate_accuracy_report(exp_dir, key="accuracy"):
 305     times = parse_train_logs(exp_dir)
 306     data = parse_prob_logs(exp_dir, key)
 307     report = []
 308     report.append("%Iter\tduration\ttrain_loss\tvalid_loss\tdifference")
 309     for x in data:
 310         try:
 311             report.append("%d\t%s\t%g\t%g\t%g" % (x[0], str(times[x[0]]),
 312                                                   x[1], x[2], x[2]-x[1]))
 313         except KeyError:
 314             continue
 315
 316     total_time = 0
 317     for iter in times.keys():
 318         total_time += times[iter]
 319     report.append("Total training time is {0}\n".format(
 320                     str(datetime.timedelta(seconds=total_time))))
 321     return ["\n".join(report), times, data]