def __init__(self, encoderParam, decoderParam, verbose=1):
'''
A-Bot Model
Uses an encoder network for input sequences (questions, answers and
history) and a decoder network for generating a response (answer).
'''
super(Answerer, self).__init__()
self.encType = encoderParam['type']
self.decType = decoderParam['type']
# Encoder
if verbose:
print('Encoder: ' + self.encType)
print('Decoder: ' + self.decType)
if 'hre' in self.encType:
self.encoder = hre_enc.Encoder(**encoderParam)
else:
raise Exception('Unknown encoder {}'.format(self.encType))
# Decoder
if 'gen' == self.decType:
self.decoder = gen_dec.Decoder(**decoderParam)
else:
raise Exception('Unkown decoder {}'.format(self.decType))
# Share word embedding parameters between encoder and decoder
self.decoder.wordEmbed = self.encoder.wordEmbed
# Initialize weights
utils.initializeWeights(self.encoder)
utils.initializeWeights(self.decoder)
self.reset()
def __init__(self,
encoderParam,
decoderParam,
imgFeatureSize=0,
verbose=1,
multiGPU=False):
'''
Q-Bot Model
Uses an encoder network for input sequences (questions, answers and
history) and a decoder network for generating a response (question).
'''
super(Questioner, self).__init__()
self.encType = encoderParam['type']
self.decType = decoderParam['type']
self.dropout = encoderParam['dropout']
self.rnnHiddenSize = encoderParam['rnnHiddenSize']
self.imgFeatureSize = imgFeatureSize
encoderParam = encoderParam.copy()
encoderParam['isAnswerer'] = False
# Encoder
if verbose:
print('Encoder: ' + self.encType)
print('Decoder: ' + self.decType)
if 'hre' in self.encType:
self.encoder = hre_enc.Encoder(**encoderParam)
if multiGPU:
self.encoder = nn.DataParallel(self.encoder)
else:
raise Exception('Unknown encoder {}'.format(self.encType))
# Decoder
if 'gen' == self.decType:
self.decoder = gen_dec.Decoder(**decoderParam)
if multiGPU:
self.decoder = nn.DataParallel(self.decoder)
else:
raise Exception('Unkown decoder {}'.format(self.decType))
# Share word embedding parameters between encoder and decoder
if multiGPU:
self.decoder.module.wordEmbed = self.encoder.module.wordEmbed
else:
self.decoder.wordEmbed = self.encoder.wordEmbed
# Setup feature regressor
if self.imgFeatureSize:
self.featureNet = nn.Linear(self.rnnHiddenSize,
self.imgFeatureSize)
if multiGPU:
self.featureNet = nn.DataParallel(self.featureNet)
self.featureNetInputDropout = nn.Dropout(0.5)
# Initialize weights
utils.initializeWeights(self.encoder)
utils.initializeWeights(self.decoder)
self.reset()
def __init__(self,
encoderParam,
decoderParam,
imgFeatureSize=0,
verbose=1):
"""
Q-Bot Model
Uses an encoder network for input sequences (questions, answers and
history) and a decoder network for generating a response (question).
"""
super(Questioner, self).__init__()
self.encType = encoderParam["type"]
self.decType = decoderParam["type"]
self.dropout = encoderParam["dropout"]
self.rnnHiddenSize = encoderParam["rnnHiddenSize"]
self.imgFeatureSize = imgFeatureSize
encoderParam = encoderParam.copy()
encoderParam["isAnswerer"] = False
# Encoder
if verbose:
print("Encoder: " + self.encType)
print("Decoder: " + self.decType)
if "hre" in self.encType:
self.encoder = hre_enc.Encoder(**encoderParam)
else:
raise Exception("Unknown encoder {}".format(self.encType))
# Decoder
if "gen" == self.decType:
self.decoder = gen_dec.Decoder(**decoderParam)
else:
raise Exception("Unkown decoder {}".format(self.decType))
# Share word embedding parameters between encoder and decoder
self.decoder.wordEmbed = self.encoder.wordEmbed
# Setup feature regressor
if self.imgFeatureSize:
self.featureNet = nn.Linear(self.rnnHiddenSize,
self.imgFeatureSize)
self.featureNetInputDropout = nn.Dropout(0.5)
# Initialize weights
utils.initializeWeights(self.encoder)
utils.initializeWeights(self.decoder)
self.reset()
def __init__(self, config, imgFeatureSize=0, verbose=1):
"""
QAA-Bot Model
Uses an encoder network for input sequences (questions, answers and
history), a decoder network for generating a response (question),
and answerer networks for answering its own questions.
"""
super(QAABot, self).__init__()
encoderParam = config["encoder"]
decoderParam = config["decoder"]
self.encType = encoderParam["type"]
self.decType = decoderParam["type"]
self.dropout = encoderParam["dropout"]
self.rnnHiddenSize = encoderParam["rnn_hidden_size"]
self.imgFeatureSize = imgFeatureSize
encoderParam = encoderParam.copy()
# Encoder
if verbose:
print("Encoder: " + self.encType)
print("Decoder: " + self.decType)
if "hre" in self.encType:
self.encoder = hre_enc.create_hre(config["encoder"])
else:
raise Exception("Unknown encoder {}".format(self.encType))
# Decoder
if "gen" == self.decType:
self.decoder = gen_dec.create_gen(config["decoder"])
else:
raise Exception("Unkown decoder {}".format(self.decType))
# Answerers
self.answerer = NotImplemented
self.answerer2 = NotImplemented
# Share word embedding parameters between encoder and decoder
self.decoder.wordEmbed = self.encoder.wordEmbed
# Create discriminator
self.featureNet = nn.Linear(self.rnnHiddenSize, 1)
self.featureNetInputDropout = nn.Dropout(0.5)
# Initialize weights
utils.initializeWeights(self.encoder)
utils.initializeWeights(self.decoder)
self.reset()
