def twoLayerFcEncoder_norm(bottom, hiddenDims_1, hiddenDims_z):
linear1 = linearLayer("vae_encoder_1", bottom, hiddenDims_1, True)
linear2 = linearLayer("vae_encoder_mu", linear1, hiddenDims_z, False)
return linear2
def myModel2(samples, seqenceLengths, sequenceLength_max, outputNum):
time_dim = 1
batch_dim = 0
n_classes = 20
samples_bw = tf.reverse_sequence(samples, seqenceLengths, time_dim,
batch_dim, 'reverse_bottom')
output_fw, output_bw, _, _ = bidirectionalLstmNetwork(
'BidirectionalLstm', samples, samples_bw, 1, [1024], seqenceLengths)
output_bw_reversed = tf.reverse_sequence(output_bw, seqenceLengths,
time_dim, batch_dim,
'reverse_blstm_out')
biLstmOut1 = tf.concat([output_fw, output_bw_reversed], 2)
biLstmOut = tf.reshape(biLstmOut1, [-1, 2 * 1024])
fc1 = linearLayer('FC_1', biLstmOut, 1024, True)
fc1_dropout = tf.nn.dropout(fc1, 0.5)
prediction = linearLayer('Linear_1', fc1_dropout, outputNum, False)
return prediction
def oneLayerBidirectionalLstmDecoder(bottom, sequenceLengths, hiddenDim_1,
hiddenDim_target):
time_dim = 1
batch_dim = 0
batchSize = bottom.shape[0].value
maxSeqLen = bottom.shape[1].value
samples_bw = tf.reverse_sequence(bottom, sequenceLengths, time_dim,
batch_dim, 'vae_decoder_reverse_bottom')
output_fw, output_bw, _, _ = bidirectionalLstmNetwork(
'vae_decoder_BidirectionalLstm', bottom, samples_bw, 1, [hiddenDim_1],
sequenceLengths)
output_bw_reversed = tf.reverse_sequence(output_bw, sequenceLengths,
time_dim, batch_dim,
'vae_decoder_reverse_blstm_out')
biLstmOut1 = tf.concat([output_fw, output_bw_reversed], 2)
biLstmOut = tf.reshape(biLstmOut1, [-1, 2 * hiddenDim_1])
linear_decoder_1 = linearLayer('vae_decoder_FC_1', biLstmOut, hiddenDim_1,
True)
reconstruction = tf.reshape(linearLayer("vae_decoder_reconstruction",
linear_decoder_1, hiddenDim_target,
False),
shape=[-1, maxSeqLen, hiddenDim_target])
return reconstruction
def blstmSequencePredictionNetwork(samples,
seqenceLengths,
sequenceLength_max,
outputNum,
layerNum=3,
neuronNum=[512]):
time_dim = 1
batch_dim = 0
samples_bw = tf.reverse_sequence(samples, seqenceLengths, time_dim,
batch_dim, 'reverse_bottom')
output_fw, output_bw, _, _ = bidirectionalLstmNetwork(
'BidirectionalLstm', samples, samples_bw, layerNum, neuronNum,
seqenceLengths)
output_bw_reversed = tf.reverse_sequence(output_bw, seqenceLengths,
time_dim, batch_dim,
'reverse_blstm_out')
biLstmOut1 = tf.concat([output_fw, output_bw_reversed], 2)
batch_size = tf.cast(tf.shape(biLstmOut1)[0], dtype=tf.int64)
# Start indices for each sample
index = tf.range(0, batch_size) * sequenceLength_max + (seqenceLengths - 1)
# Indexing
outputs = tf.gather(tf.reshape(biLstmOut1, [-1, 2 * neuronNum[-1]]), index)
fc1 = linearLayer('FC_1', outputs, neuronNum[-1], True)
fc1_dropout = tf.nn.dropout(fc1, 0.5)
prediction = linearLayer('Linear_1', fc1_dropout, outputNum, False)
return prediction
def blstmFramePredictionNetwork(samples,
seqenceLengths,
sequenceLength_max,
outputNum,
layerNum=3,
neuronNum=[512]):
time_dim = 1
batch_dim = 0
samples_bw = tf.reverse_sequence(samples, seqenceLengths, time_dim,
batch_dim, 'reverse_bottom')
output_fw, output_bw, _, _ = bidirectionalLstmNetwork(
'BidirectionalLstm', samples, samples_bw, layerNum, neuronNum,
seqenceLengths)
output_bw_reversed = tf.reverse_sequence(output_bw, seqenceLengths,
time_dim, batch_dim,
'reverse_blstm_out')
biLstmOut1 = tf.concat([output_fw, output_bw_reversed], 2)
biLstmOut = tf.reshape(biLstmOut1, [-1, 2 * neuronNum[-1]])
fc1 = linearLayer('FC_1', biLstmOut, neuronNum[-1], True)
fc1_dropout = tf.nn.dropout(fc1, 0.5)
prediction = tf.reshape(
linearLayer('Linear_1', fc1_dropout, outputNum, False),
[-1, sequenceLength_max, outputNum])
return prediction
def oneLayerBidirectionalLstmEncoder(bottom, sequenceLengths, hiddenDims_1,
hiddenDims_z):
time_dim = 1
batch_dim = 0
batchSize = bottom.shape[0].value
maxSequLen = bottom.shape[1].value
samples_bw = tf.reverse_sequence(bottom, sequenceLengths, time_dim,
batch_dim, 'vae_encoder_reverse_bottom')
output_fw, output_bw, _, _ = bidirectionalLstmNetwork(
'vae_encoder_BidirectionalLstm', bottom, samples_bw, 1, [hiddenDims_1],
sequenceLengths)
output_bw_reversed = tf.reverse_sequence(output_bw, sequenceLengths,
time_dim, batch_dim,
'vae_encoder_reverse_blstm_out')
biLstmOut1 = tf.concat([output_fw, output_bw_reversed], 2)
biLstmOut = tf.reshape(biLstmOut1, [-1, 2 * hiddenDims_1])
linear1 = linearLayer('vae_encoder_FC_1', biLstmOut, hiddenDims_1, True)
mean = tf.reshape(linearLayer("vae_encoder_mu", linear1, hiddenDims_z,
False),
shape=[-1, maxSequLen, hiddenDims_z])
SE_ln = tf.reshape(linearLayer("vae_encoder_SE_log", linear1, hiddenDims_z,
False),
shape=[-1, maxSequLen, hiddenDims_z])
return mean, SE_ln
def twoLayerFcDecoder(bottom, hiddenDim_1, hiddenDim_target):
linear_decoder_1 = linearLayer("vae_decoder_1", bottom, hiddenDim_1, True)
reconstruction = linearLayer("vae_decoder_reconstruction",
linear_decoder_1, hiddenDim_target, False)
return reconstruction
def twoLayerFcEncoder(bottom, hiddenDims_1, hiddenDims_z):
linear1 = linearLayer("vae_encoder_1", bottom, hiddenDims_1, True)
mean = linearLayer("vae_encoder_mu", linear1, hiddenDims_z, False)
SE_ln = linearLayer("vae_encoder_SE_log", linear1, hiddenDims_z, False)
return mean, SE_ln