def main(self):
self.textData = TextData(self.args)
self.args['vocabularySize'] = self.textData.getVocabularySize()
print(self.textData.getVocabularySize())
self.model = Model(self.args)
self.train()
def set_up_things(self, args=None):
self.args = {}
self.args['rootDir'] = os.getcwd() # Use the current working directory
self.args['corpus'] = 'cornell'
self.args['maxLength'] = 10
self.args['hiddenSize'] = 256
self.args['numLayers'] = 2
self.args['embeddingSize'] = 32
self.args['softmaxSamples'] = 0
self.args['numEpochs'] = 50
self.args['saveEvery'] = 5000
self.args['batchSize'] = 10
self.args['learningRate'] = 0.001
self.args['reset'] = False
test_yes = True
self.args['interactive'] = test_yes
self.args['test'] = test_yes
self.loadModelParams(
) # Update the self.modelDir and self.globStep, for now, not used when loading Model (but need to be called before _getSummaryName)
self.textData = TextData(self.args)
self.model = Model(self.args, self.textData)
self.writer = tf.train.SummaryWriter(self.modelDir)
if '12' in tf.__version__: # HACK: Solve new tf Saver V2 format
self.saver = tf.train.Saver(max_to_keep=200,
write_version=1) # Arbitrary limit ?
else:
self.saver = tf.train.Saver(max_to_keep=200)
self.sess = tf.Session()
print('Initialize variables...')
self.sess.run(tf.initialize_all_variables())
self.managePreviousModel(self.sess)
def __init__(self, encoder_hidden_units, input_embedding_size, bath_size):
self.textData = TextData("train.tsv", "data", 100, "test_", 800)
self.vocab_size = self.textData.getVocabularySize()
self.input_embedding_size = input_embedding_size
self.encoder_hidden_units = encoder_hidden_units
self.batch_size = bath_size
self.buildNetwork()
def main():
print('Welcome to DeepQA v0.1 !')
print()
args = parseArgs()
textData = TextData(args)
pass
def main(self):
args['datasetsize'] = 'big'
if args['model_arch'] in ['lstmgrid']:
args['batchSize'] = 64
elif args['model_arch'] in ['lstmibgan']:
args['classify_type'] = 'single'
args['batchSize'] = 64
elif args['model_arch'] in ['lstmibgan_law']:
args['classify_type'] = 'single'
args['batchSize'] = 64
args['task'] = 'law'
elif args['model_arch'] in ['lstmibgan_toi']:
args['classify_type'] = 'single'
args['batchSize'] = 64
args['task'] = 'toi'
self.textData = TextData('cail')
self.start_token = self.textData.word2index['START_TOKEN']
self.end_token = self.textData.word2index['END_TOKEN']
args['vocabularySize'] = self.textData.getVocabularySize()
if args['model_arch'] in ['lstmibgan_law']:
args['chargenum'] = self.textData.getLawNum()
elif args['model_arch'] in ['lstmibgan_toi']:
args['chargenum'] = 11
else:
args['chargenum'] = self.textData.getChargeNum()
print(self.textData.getVocabularySize())
if args['model_arch'] == 'lstm':
print('Using LSTM model.')
self.model = LSTM_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmatt':
print('Using LSTM attention model.')
self.model = LSTM_att_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'transformer':
print('Using Transformer model.')
self.model = TransformerModel(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmib':
print('Using LSTM information bottleneck model.')
self.model = LSTM_IB_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'].startswith('lstmibgan'):
print('Using LSTM information bottleneck GAN model. Task: ' +
args['task'])
LM = torch.load(args['rootDir'] + '/LM' + args['datasetsize'] +
'.pkl',
map_location=args['device'])
for param in LM.parameters():
param.requires_grad = False
LSTM_IB_GAN.train(self.textData, LM)
elif args['model_arch'] == 'lstmibcp':
print('Using LSTM information bottleneck model. -- complete words')
self.model = LSTM_IB_CP_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmcapib':
print('Using LSTM capsule information bottleneck model.')
self.model = LSTM_capsule_IB_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmiterib':
print('Using LSTM iteratively information bottleneck model.')
self.model = LSTM_iterIB_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmcap':
print('Using LSTM capsule model.')
self.model = LSTM_capsule_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmgrid':
print('Using LSTM grid model.')
self.model = LSTM_grid_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmgmib':
print('Using LSTM Gaussian Mixture IB model.')
self.model = LSTM_GMIB_Model(self.textData.word2index,
self.textData.index2word)
self.model = self.model.to(args['device'])
self.train()
def main(self):
args['batchSize'] = 32
self.textData = TextData('cail')
self.start_token = self.textData.word2index['START_TOKEN']
self.end_token = self.textData.word2index['END_TOKEN']
args['vocabularySize'] = self.textData.getVocabularySize()
args['chargenum'] = self.textData.getChargeNum()
print(self.textData.getVocabularySize())
if args['model_arch'] == 'lstm':
print('Using LSTM model.')
self.model = LSTM_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmatt':
print('Using LSTM attention model.')
self.model = LSTM_att_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'transformer':
print('Using Transformer model.')
self.model = TransformerModel(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmib':
print('Using LSTM information bottleneck model.')
self.model = LSTM_IB_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmibgan':
print('Using LSTM information bottleneck GAN model.')
LSTM_IB_GAN.train(self.textData)
elif args['model_arch'] == 'lstmibcp':
print('Using LSTM information bottleneck model. -- complete words')
self.model = LSTM_IB_CP_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmcapib':
print('Using LSTM capsule information bottleneck model.')
self.model = LSTM_capsule_IB_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmiterib':
print('Using LSTM iteratively information bottleneck model.')
self.model = LSTM_iterIB_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmcap':
print('Using LSTM capsule model.')
self.model = LSTM_capsule_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmgrid':
print('Using LSTM grid model.')
self.model = LSTM_grid_Model(self.textData.word2index,
self.textData.index2word)
self.train()
elif args['model_arch'] == 'lstmgmib':
print('Using LSTM Gaussian Mixture IB model.')
self.model = nn.parallel.DataParallel(
LSTM_GMIB_Model(self.textData.word2index,
self.textData.index2word))
self.train()
args['device'] = "cuda:0" if torch.cuda.is_available() else "cpu"
self.model.to(args['device'])
def main(self, args=None):
"""
Launch the training and/or the interactive mode
"""
print('Welcome to DeepQA v0.1 !')
print()
print('TensorFlow detected: v{}'.format(tf.__version__))
# General initialisation
self.args = self.parseArgs(args)
if not self.args.rootDir:
self.args.rootDir = os.getcwd() # Use the current working directory
#tf.logging.set_verbosity(tf.logging.INFO) # DEBUG, INFO, WARN (default), ERROR, or FATAL
self.loadModelParams() # Update the self.modelDir and self.globStep, for now, not used when loading Model (but need to be called before _getSummaryName)
self.textData = TextData(self.args)
# TODO: Add a mode where we can force the input of the decoder // Try to visualize the predictions for
# each word of the vocabulary / decoder input
# TODO: For now, the model are trained for a specific dataset (because of the maxLength which define the
# vocabulary). Add a compatibility mode which allow to launch a model trained on a different vocabulary (
# remap the word2id/id2word variables).
if self.args.createDataset:
print('Dataset created! Thanks for using this program')
return # No need to go further
# Prepare the model
with tf.device(self.getDevice()):
self.model = Model(self.args, self.textData)
# Saver/summaries
self.writer = tf.summary.FileWriter(self._getSummaryName())
self.saver = tf.train.Saver(max_to_keep=200)
# TODO: Fixed seed (WARNING: If dataset shuffling, make sure to do that after saving the
# dataset, otherwise, all which cames after the shuffling won't be replicable when
# reloading the dataset). How to restore the seed after loading ??
# Also fix seed for random.shuffle (does it works globally for all files ?)
# Running session
self.sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True, # Allows backup device for non GPU-available operations (when forcing GPU)
log_device_placement=False) # Too verbose ?
) # TODO: Replace all sess by self.sess (not necessary a good idea) ?
if self.args.debug:
self.sess = tf_debug.LocalCLIDebugWrapperSession(self.sess)
self.sess.add_tensor_filter("has_inf_or_nan", tf_debug.has_inf_or_nan)
print('Initialize variables...')
self.sess.run(tf.global_variables_initializer())
# Reload the model eventually (if it exist.), on testing mode, the models are not loaded here (but in predictTestset)
if self.args.test != Chatbot.TestMode.ALL:
self.managePreviousModel(self.sess)
# Initialize embeddings with pre-trained word2vec vectors
if self.args.initEmbeddings:
self.loadEmbedding(self.sess)
if self.args.test:
if self.args.test == Chatbot.TestMode.INTERACTIVE:
self.mainTestInteractive(self.sess)
elif self.args.test == Chatbot.TestMode.ALL:
print('Start predicting...')
self.predictTestset(self.sess)
print('All predictions done')
elif self.args.test == Chatbot.TestMode.DAEMON:
print('Daemon mode, running in background...')
else:
raise RuntimeError('Unknown test mode: {}'.format(self.args.test)) # Should never happen
else:
self.mainTrain(self.sess)
if self.args.test != Chatbot.TestMode.DAEMON:
self.sess.close()
print("The End! Thanks for using this program")