def buildVae(self):
self.vae_mean, self.vae_SE_ln = oneLayerBidirectionalLstmEncoder(self.placeHolder_samples, \
self.placeHolder_seqLens, \
self.hiddenDim_vae_encoder, \
self.hiddenDim_vae_z)
self.Z = vaeSampler(self.vae_mean, self.vae_SE_ln, self.placeHolder_noises)
self.reconstruction_vae = oneLayerBidirectionalLstmDecoder(self.Z, self.placeHolder_seqLens, self.hiddenDim_vae_decoder, self.reconstructionTargetDim)
def buildVae(self):
self.vae_mean, self.vae_SE_ln = oneLayerBidirectionalLstmEncoder(self.placeHolder_samples, \
self.placeHolder_seqLens, \
self.hiddenDim_vae_encoder, \
self.hiddenDim_vae_z)
self.Z = vaeSampler(self.vae_mean, self.vae_SE_ln, self.placeHolder_noises)
self.reconstruction_vae = oneLayerBidirectionalLstmDecoder(self.Z, self.placeHolder_seqLens, self.hiddenDim_vae_decoder, self.reconstructionTargetDim)
_, self.reconstruction_splited = tf.split(self.reconstruction_vae, [1, self.segLen_max - 1], 1)
_, self.Z_split = tf.split(self.Z, [1, self.segLen_max - 1], 1)
def buildNet(self):
#BaseNetwork.buildNet(self)
self.placeHolder_samples = tf.placeholder(
"float", [None, self.segLen_max, self.sampleDim])
self.placeHolder_targets_reconstruction = tf.placeholder(
"float", [None, self.segLen_max, self.reconstructionTargetDim])
self.placeHolder_noises = tf.placeholder(
"float", [None, self.segLen_max, self.hiddenDim_vae_z])
self.placeHolder_seqLens = tf.placeholder(tf.int64, [None])
self.placeHoder_learningRate = tf.placeholder("float", [])
self.mask = tf.sequence_mask(self.placeHolder_seqLens,
self.segLen_max,
dtype=tf.float32)
self.placeHolder_labels = tf.placeholder(tf.int64, [None])
#self.vectorizedLabels = tf.one_hot(self.placeHolder_labels, self.classNumAll)
self.vectorizedLabels = tf.reshape(
tf.one_hot(self.placeHolder_labels, self.classNumAll),
[-1, self.classNumAll])
self.vae_mean, self.vae_SE_ln = oneLayerBidirectionalLstmEncoder(self.placeHolder_samples, \
self.placeHolder_seqLens, \
self.hiddenDim_vae_encoder, \
self.hiddenDim_vae_z)
self.Z = vaeSampler(self.vae_mean, self.vae_SE_ln,
self.placeHolder_noises)
self.reconstruction_vae = oneLayerBidirectionalLstmDecoder(
self.Z, self.placeHolder_seqLens, self.hiddenDim_vae_decoder,
self.reconstructionTargetDim)
self.vae_mean_squared = tf.square(self.vae_mean,
name="vae_encoder_mean_squared")
self.vae_SE = tf.exp(self.vae_SE_ln, name="vae_encoder_SE")
self.loss_vae, self.KLD, _ = vaeLoss_forLstm1(
self.reconstruction_vae, self.placeHolder_targets_reconstruction,
self.vae_mean_squared, self.vae_SE, self.vae_SE_ln)
#_, self.loss_vae, _ = vaeLoss_forLstm(self.reconstruction_vae, self.placeHolder_targets_reconstruction, self.vae_mean_squared, self.vae_SE, self.vae_SE_ln)
if self.mConfig.finetune is False:
self.maskedLoss_vae = tf.multiply(self.mask, self.loss_vae)
self.meanLoss = tf.divide(tf.reduce_sum(self.maskedLoss_vae),
tf.reduce_sum(self.mask))
else:
self.prediction = blstmSequencePredictionNetwork(self.reconstruction_vae, self.placeHolder_seqLens, self.segLen_max, \
self.classNumAll, self.mConfig.blstmLayerNum, self.mConfig.blstmLayerNeuronNum)
self.correctPrediction = tf.equal(
tf.argmax(self.prediction, 1),
tf.argmax(self.vectorizedLabels, 1))
self.correctPredcitNum = tf.reduce_sum(
tf.cast(self.correctPrediction, tf.int64))
self.loss_classification = tf.nn.softmax_cross_entropy_with_logits(
logits=self.prediction, labels=self.vectorizedLabels)
self.meanLoss = tf.reduce_mean(self.loss_classification)
self.optimizer = tf.train.AdamOptimizer(
learning_rate=self.placeHoder_learningRate)
self.params = tf.trainable_variables()
self.variables_to_restore = None
self.gradients = tf.gradients(self.meanLoss, self.params)
self.clippedGradients, self.orgGradientNorm = tf.clip_by_global_norm(
self.gradients, self.maxGradient)
self.train_op = self.optimizer.apply_gradients(
zip(self.clippedGradients, self.params))
self.init = tf.global_variables_initializer()
self.gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)
self.sess = tf.Session(config=tf.ConfigProto(
gpu_options=self.gpu_options))
self.sess.run(self.init)
self.restore()