Blame view
scripts/training.py
17.3 KB
f2d3bd141
![]() |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 |
#!/usr/bin/env python # -*- coding: utf-8 -*- # Authors: Parcollet Titouan # Imports import editdistance import h5py import datasets.timit from datasets.timit import Timit from datasets.utils import construct_conv_stream, phone_to_phoneme_dict import complexnn from complexnn import * import h5py as H import keras from keras.callbacks import Callback, ModelCheckpoint, LearningRateScheduler from keras.initializers import Orthogonal from keras.layers import Layer, Dropout, AveragePooling1D, AveragePooling2D, AveragePooling3D, add, Add, concatenate, Concatenate, Input, Flatten, Dense, Convolution2D, BatchNormalization, Activation, Reshape, ConvLSTM2D, Conv2D, Lambda from keras.models import Model, load_model, save_model from keras.optimizers import SGD, Adam, RMSprop from keras.regularizers import l2 from keras.utils.np_utils import to_categorical import keras.backend as K import keras.models as KM from keras.utils.training_utils import multi_gpu_model import logging as L import numpy as np import os, pdb, socket, sys, time import theano as T from keras.backend.tensorflow_backend import set_session from models_timit import getTimitResnetModel2D,ctc_lambda_func import tensorflow as tf import itertools import random # # Generator wrapper for timit # def timitGenerator(stream): while True: for data in stream.get_epoch_iterator(): yield data # # Custom metrics # class EditDistance(Callback): def __init__(self, func, dataset, quaternion, save_prefix): self.func = func if(dataset in ['train','test','dev']): self.dataset_type = dataset self.save_prefix = save_prefix self.dataset = Timit(str(dataset)) self.full_phonemes_dict = self.dataset.get_phoneme_dict() self.ind_phonemes_dict = self.dataset.get_phoneme_ind_dict() self.rng = np.random.RandomState(123) self.data_stream = construct_conv_stream(self.dataset, self.rng, 20, 1000,quaternion) else: raise ValueError("Unknown dataset for edit distance "+dataset) def labels_to_text(self,labels): ret = [] for c in labels: if c == len(self.full_phonemes_dict) - 2: ret.append("") else: c_ = self.full_phonemes_dict[c + 1] ret.append(phone_to_phoneme_dict.get(c_, c_)) ret = [k for k, g in itertools.groupby(ret)] return list(filter(lambda c: c != "", ret)) def decode_batch(self, out, mask): ret = [] for j in range(out.shape[0]): out_best = list(np.argmax(out[j], 1))[:int(mask[j])] out_best = [k for k, g in itertools.groupby(out_best)] # map from 61-d to 39-d out_str = self.labels_to_text(out_best) ret.append(out_str) return ret def on_epoch_end(self, epoch, logs={}): mean_norm_ed = 0. num = 0 for data in self.data_stream.get_epoch_iterator(): x, y = data y_pred = self.func([x[0]])[0] decoded_y_pred = self.decode_batch(y_pred, x[1]) decoded_gt = [] for i in range(x[2].shape[0]): decoded_gt.append(self.labels_to_text(x[2][i][:int(x[3][i])])) num += len(decoded_y_pred) for i, (_pred, _gt) in enumerate(zip(decoded_y_pred, decoded_gt)): edit_dist = editdistance.eval(_pred, _gt) mean_norm_ed += float(edit_dist) / float(len(_gt)) mean_norm_ed = mean_norm_ed / num # Dump To File Logs at every epoch for clusters sbatch f=open(str(self.save_prefix)+"_"+str(self.dataset_type)+"_PER.txt",'ab') mean = np.array([mean_norm_ed]) np.savetxt(f,mean) f.close() L.getLogger("train").info("PER on "+str(self.dataset_type)+" : "+str(mean_norm_ed)+" at epoch "+str(epoch)) # # Callbacks: # class TrainLoss(Callback): def __init__(self, savedir): self.savedir = savedir def on_epoch_end(self, epoch, logs={}): f=open(str(self.savedir)+"_train_loss.txt",'ab') f2=open(str(self.savedir)+"_dev_loss.txt",'ab') value = float(logs['loss']) np.savetxt(f,np.array([value])) f.close() value = float(logs['val_loss']) np.savetxt(f2,np.array([value])) f2.close() # # Print a newline after each epoch, because Keras doesn't. Grumble. # class PrintNewlineAfterEpochCallback(Callback): def on_epoch_end(self, epoch, logs={}): sys.stdout.write(" ") # # Save checkpoints. # class SaveLastModel(Callback): def __init__(self, workdir, save_prefix, model_mono,period=10): self.workdir = workdir self.model_mono = model_mono self.chkptsdir = os.path.join(self.workdir, "chkpts") self.save_prefix = save_prefix if not os.path.isdir(self.chkptsdir): os.mkdir(self.chkptsdir) self.period_of_epochs = period self.linkFilename = os.path.join(self.chkptsdir, str(save_prefix)+"ModelChkpt.hdf5") self.linkFilename_weight = os.path.join(self.chkptsdir, str(save_prefix)+"ModelChkpt_weight.hdf5") def on_epoch_end(self, epoch, logs={}): if (epoch + 1) % self.period_of_epochs == 0: # Filenames baseHDF5Filename = str(self.save_prefix)+"ModelChkpt{:06d}.hdf5".format(epoch+1) baseHDF5Filename_weight = str(self.save_prefix)+"ModelChkpt{:06d}_weight.hdf5".format(epoch+1) baseYAMLFilename = str(self.save_prefix)+"ModelChkpt{:06d}.yaml".format(epoch+1) hdf5Filename = os.path.join(self.chkptsdir, baseHDF5Filename) hdf5Filename_weight = os.path.join(self.chkptsdir, baseHDF5Filename_weight) yamlFilename = os.path.join(self.chkptsdir, baseYAMLFilename) # YAML yamlModel = self.model_mono.to_yaml() with open(yamlFilename, "w") as yamlFile: yamlFile.write(yamlModel) # HDF5 KM.save_model(self.model_mono, hdf5Filename) self.model_mono.save_weights(hdf5Filename_weight) with H.File(hdf5Filename, "r+") as f: f.require_dataset("initialEpoch", (), "uint64", True)[...] = int(epoch+1) f.flush() with H.File(hdf5Filename_weight, "r+") as f: f.require_dataset("initialEpoch", (), "uint64", True)[...] = int(epoch+1) f.flush() # Symlink to new HDF5 file, then atomically rename and replace. os.symlink(baseHDF5Filename_weight, self.linkFilename_weight+".rename") os.rename (self.linkFilename_weight+".rename", self.linkFilename_weight) # Symlink to new HDF5 file, then atomically rename and replace. os.symlink(baseHDF5Filename, self.linkFilename+".rename") os.rename (self.linkFilename+".rename", self.linkFilename) # Print L.getLogger("train").info("Saved checkpoint to {:s} at epoch {:5d}".format(hdf5Filename, epoch+1)) # # Summarize environment variable. # def summarizeEnvvar(var): if var in os.environ: return var+"="+os.environ.get(var) else: return var+" unset" # # TRAINING PROCESS # def train(d): # # # Log important data about how we were invoked. # L.getLogger("entry").info("INVOCATION: "+" ".join(sys.argv)) L.getLogger("entry").info("HOSTNAME: "+socket.gethostname()) L.getLogger("entry").info("PWD: "+os.getcwd()) L.getLogger("entry").info("CUDA DEVICE: "+str(d.device)) os.environ["CUDA_VISIBLE_DEVICES"]=str(d.device) # # Setup GPUs # config = tf.ConfigProto() # # Don't pre-allocate memory; allocate as-needed # config.gpu_options.allow_growth = True # # Only allow a total of half the GPU memory to be allocated # config.gpu_options.per_process_gpu_memory_fraction = d.memory # # Create a session with the above options specified. # K.tensorflow_backend.set_session(tf.Session(config=config)) summary = " " summary += "Environment: " summary += summarizeEnvvar("THEANO_FLAGS")+" " summary += " " summary += "Software Versions: " summary += "Theano: "+T.__version__+" " summary += "Keras: "+keras.__version__+" " summary += " " summary += "Arguments: " summary += "Path to Datasets: "+str(d.datadir)+" " summary += "Number of GPUs: "+str(d.datadir)+" " summary += "Path to Workspace: "+str(d.workdir)+" " summary += "Model: "+str(d.model)+" " summary += "Number of Epochs: "+str(d.num_epochs)+" " summary += "Number of Start Filters: "+str(d.start_filter)+" " summary += "Number of Layers: "+str(d.num_layers)+" " summary += "Optimizer: "+str(d.optimizer)+" " summary += "Learning Rate: "+str(d.lr)+" " summary += "Learning Rate Decay: "+str(d.decay)+" " summary += "Clipping Norm: "+str(d.clipnorm)+" " summary += "Clipping Value: "+str(d.clipval)+" " summary += "Dropout Probability: "+str(d.dropout)+" " if d.optimizer in ["adam"]: summary += "Beta 1: "+str(d.beta1)+" " summary += "Beta 2: "+str(d.beta2)+" " else: summary += "Momentum: "+str(d.momentum)+" " summary += "Save Prefix: "+str(d.save_prefix)+" " L.getLogger("entry").info(summary[:-1]) # # Load dataset # L.getLogger("entry").info("Loading dataset {:s} ...".format(d.dataset)) np.random.seed(d.seed % 2**32) # # Create training data generator # dataset = Timit('train') rng=np.random.RandomState(123) if d.model =="quaternion": data_stream_train = construct_conv_stream(dataset, rng, 200, 1000, quaternion=True) else: data_stream_train = construct_conv_stream(dataset, rng, 200, 1000, quaternion=False) # # Create dev data generator # dataset = Timit('dev') rng=np.random.RandomState(123) if d.model =="quaternion": data_stream_dev = construct_conv_stream(dataset, rng, 200, 10000, quaternion=True) else: data_stream_dev = construct_conv_stream(dataset, rng, 200, 1000, quaternion=False) L.getLogger("entry").info("Training set length: "+str(Timit('train').num_examples)) L.getLogger("entry").info("Validation set length: "+str(Timit('dev').num_examples)) L.getLogger("entry").info("Test set length: "+str(Timit('test').num_examples)) L.getLogger("entry").info("Loaded dataset {:s}.".format(d.dataset)) # # Optimizers # if d.optimizer in ["sgd", "nag"]: opt = SGD (lr = d.lr, momentum = d.momentum, decay = d.decay, nesterov = (d.optimizer=="nag"), clipnorm = d.clipnorm) elif d.optimizer == "rmsprop": opt = RMSProp(lr = d.lr, decay = d.decay, clipnorm = d.clipnorm) elif d.optimizer == "adam": opt = Adam (lr = d.lr, beta_1 = d.beta1, beta_2 = d.beta2, decay = d.decay, clipnorm = d.clipnorm) else: raise ValueError("Unknown optimizer "+d.optimizer) # # Initial Entry or Resume ? # initialEpoch = 0 chkptFilename = os.path.join(d.workdir, "chkpts", str(d.save_prefix)+"ModelChkpt.hdf5") chkptFilename_weight = os.path.join(d.workdir, "chkpts", str(d.save_prefix)+"ModelChkpt_weight.hdf5") isResuming = os.path.isfile(chkptFilename) isResuming_weight = os.path.isfile(chkptFilename_weight) if isResuming or isResuming_weight: # Reload Model and Optimizer if d.dataset == "timit": L.getLogger("entry").info("Re-Creating the model from scratch.") model_mono,test_func = getTimitResnetModel2D(d) model_mono.load_weights(chkptFilename_weight) with H.File(chkptFilename_weight, "r") as f: initialEpoch = int(f["initialEpoch"][...]) L.getLogger("entry").info("Training will restart at epoch {:5d}.".format(initialEpoch+1)) L.getLogger("entry").info("Compilation Started.") else: L.getLogger("entry").info("Reloading a model from "+chkptFilename+" ...") np.random.seed(d.seed % 2**32) model = KM.load_model(chkptFilename, custom_objects={ "QuaternionConv2D": QuaternionConv2D, "QuaternionConv1D": QuaternionConv1D, "GetIFirst": GetIFirst, "GetJFirst": GetJFirst, "GetKFirst": GetKFirst, "GetRFirst": GetRFirst, }) L.getLogger("entry").info("... reloading complete.") with H.File(chkptFilename, "r") as f: initialEpoch = int(f["initialEpoch"][...]) L.getLogger("entry").info("Training will restart at epoch {:5d}.".format(initialEpoch+1)) L.getLogger("entry").info("Compilation Started.") else: model_mono,test_func = getTimitModel2D(d) L.getLogger("entry").info("Compilation Started.") # # Multi GPU: Can only save the model_mono because of keras bug # if d.gpus >1: model = multi_gpu_model(model_mono, gpus=d.gpus) else: model = model_mono # # Compile with CTC koss function # model.compile(opt, loss={'ctc': lambda y_true, y_pred: y_pred}) # # Precompile several backend functions # if d.summary: model.summary() L.getLogger("entry").info("# of Parameters: {:10d}".format(model.count_params())) L.getLogger("entry").info("Compiling Train Function...") t =- time.time() model._make_train_function() t += time.time() L.getLogger("entry").info(" {:10.3f}s".format(t)) L.getLogger("entry").info("Compiling Predict Function...") t =- time.time() model._make_predict_function() t += time.time() L.getLogger("entry").info(" {:10.3f}s".format(t)) L.getLogger("entry").info("Compiling Test Function...") t =- time.time() model._make_test_function() t += time.time() L.getLogger("entry").info(" {:10.3f}s".format(t)) L.getLogger("entry").info("Compilation Ended.") # # Create Callbacks # newLineCb = PrintNewlineAfterEpochCallback() saveLastCb = SaveLastModel(d.workdir, d.save_prefix, model_mono, period=10) callbacks = [] # # End of line for better looking # callbacks += [newLineCb] if d.model=="quaternion": quaternion = True else: quaternion = False |
745363964
![]() |
444 445 446 |
if not os.path.exists(d.workdir+"/LOGS"): os.makedirs(d.workdir+"/LOGS") |
f2d3bd141
![]() |
447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 |
savedir = d.workdir+"/LOGS/"+d.save_prefix # # Save the Train loss # trainLoss = TrainLoss(savedir) # # Compute accuracies and save # editDistValCb = EditDistance(test_func,'dev',quaternion, savedir) editDistTestCb = EditDistance(test_func,'test',quaternion, savedir) callbacks += [trainLoss] callbacks += [editDistValCb] callbacks += [editDistTestCb] callbacks += [newLineCb] # # Save the model # callbacks += [saveLastCb] # # Enter training loop. # L .getLogger("entry").info("**********************************************") if isResuming: L.getLogger("entry").info("*** Reentering Training Loop @ Epoch {:5d} ***".format(initialEpoch+1)) else: L.getLogger("entry").info("*** Entering Training Loop @ First Epoch ***") L .getLogger("entry").info("**********************************************") # # TRAIN # ######## # Make sure to give the right number of mini_batch size # needed to complete ONE epoch (according to your data generator) ######## epochs_train = 1144 epochs_dev = 121 model.fit_generator(generator = timitGenerator(data_stream_train), steps_per_epoch = epochs_train, epochs = d.num_epochs, verbose = 1, validation_data = timitGenerator(data_stream_dev), validation_steps = epochs_dev, callbacks = callbacks, initial_epoch = initialEpoch) |