I'm studying CBHG module for extracting representations from sequence in Tacotron.
CBHG is consisted of (1-D convolution bank - highway network - bidirectional GRU).
Inputs of 1-d conv is made by 'lookup_table' for embedding a~z.
After 1-D conv caculate, there is 'tf.concat' for concated the results.
And there is projection with this result.
I think it's associate with word2vec embedding like 'CBOW' but It's very hard to me.
What is the effect of projection layer after concated result by n-grams convolution? Extracting meaningful 'n' in n-grams conv banks?
Please help me.
with tf.variable_scope('conv_bank'):
# Convolution bank: concatenate on the last axis
# to stack channels from all convolutions
conv_fn = lambda k: \
conv1d(inputs, k, bank_channel_size,
tf.nn.relu, is_training, 'conv1d_%d' % k)
conv_outputs = tf.concat(
[conv_fn(k) for k in range(1, bank_size+1)], axis=-1,
)
# Maxpooling:
maxpool_output = tf.layers.max_pooling1d(
conv_outputs,
pool_size=maxpool_width,
strides=1,
padding='same')
# Two projection layers:
proj_out = maxpool_output
for idx, proj_size in enumerate(proj_sizes):
activation_fn = None if idx == len(proj_sizes) - 1 else tf.nn.relu
proj_out = conv1d(
proj_out, proj_width, proj_size, activation_fn,
is_training, 'proj_{}'.format(idx + 1))
Related
I have some neural network with following code snippets, note that batch_size == 1 and input_dim == output_dim:
net_in = tf.Variable(tf.zeros(shape = [batch_size, input_dim]), dtype=tf.float32)
input_placeholder = tf.compat.v1.placeholder(shape = [batch_size, input_dim], dtype=tf.float32)
assign_input = net_in.assign(input_placeholder)
# Some matmuls, activations, dropouts, normalizations...
net_out = tf.tanh(output_before_activation)
def loss_fn(output, input):
#input.shape = output.shape = (batch_size, input_dim)
output = tf.reshape(output, [input_dim,]) # shape them into 1d vectors
input = tf.reshape(input, [input_dim,])
return my_fn_that_only_takes_in_vectors(output, input)
# Create session, preprocess data ...
for epoch in epoch_num:
for batch in range(total_example_num // batch_size):
sess.run(assign_input, feed_dict = {input_placeholder : some_appropriate_numpy_array})
sess.run(optimizer.minimize(loss_fn(net_out, net_in)))
Currently the neural network above works fine, but it is very slow because it updates gradient every sample (batch size = 1). I would like to set batch size > 1, but my_fn_that_only_takes_in_vectors cannot accommodate matrices whose first dimension is not 1. Due to the nature of my custom loss, flattening the batch input into a vector of length (batch_size * input_dim) seems to not work.
How would I write my new custom loss_fn now that the input and output are N x input_dim where N > 1? In Keras this would not have been an issue because keras somehow takes the average of the gradients of each example in the batch. For my TensorFlow function, should I take each row as a vector individually, pass them to my_fn_that_only_takes_in_vectors, then take the average of the results?
You can use a function that computes the loss on the whole batch, and works independently on the batch size. Basically the operations are applied to the whole first dimension of the input (the first dimension represents the element number in the batch). Here is an example, I hope this helps to see how the operations are carried out:
def my_loss(y_true, y_pred):
dx2 = tf.math.squared_difference(y_true[:, 0], y_true[:, 2]) # shape (BatchSize, )
dy2 = tf.math.squared_difference(y_true[:, 1], y_true[:, 3]) # shape: (BatchSize, )
denominator = dx2 + dy2 # shape: (BatchSize, )
dst_vec = tf.math.squared_difference(y_true, y_pred) # shape: (Batch, n_labels)
numerator = tf.reduce_sum(dst_vec, axis=-1) # shape: (BatchSize,)
loss_vector = tf.cast(numerator / denominator, dtype="float32") # shape: (BatchSize,) this is a vector containing the loss of each element of the batch
loss = tf.reduce_sum(loss_vector ) #if you want to sum the losses
return loss
I am not sure whether you need to return the sum or the avg of the losses for the batch.
If you sum, make sure to use a validation dataset with same batch size, otherwise the loss is not comparable.
I want to prune over the highest weight values in a tf layer. I'm thinking about using tf.nn.top_k but I'm not exactly sure how I would go about doing this.
Documentation: https://www.tensorflow.org/model_optimization/api_docs/python/tfmot/sparsity/keras/prune_low_magnitude
Code:
pruning_params = {
'pruning_schedule': PolynomialDecay(initial_sparsity=0.2,
final_sparsity=0.8, begin_step=1000, end_step=2000),
'block_size': (2, 3),
'block_pooling_type': 'MAX'
}
model = keras.Sequential([
layers.Dense(10, activation='relu', input_shape=(100,)),
prune_low_magnitude(layers.Dense(2, activation='tanh'), **pruning_params)
])
Assuming that w is the weight matrix of the layer you want to prune, and k is the percentage of weights that should be pruned, this should do the trick for you:
# Convert k from percentage to integer representing the number of weights
k = tf.cast(tf.round(tf.size(w, out_type=tf.float32) * tf.constant(k)), dtype=tf.int32)
# Reshape flatten the weight matrix
w_reshaped = tf.reshape(w, [-1])
# Select the indices of the largest k weights
_, indices = tf.nn.top_k(w_reshaped, k, sorted=True, name=None)
# Set the elements matching the indices to 0
mask = tf.scatter_nd_update(tf.Variable(tf.ones_like(w_reshaped, dtype=tf.float32), name="mask", trainable=False), tf.reshape(indices, [-1, 1]), tf.zeros([k], tf.float32))
# Update the weight matrix w
w.assign(tf.reshape(w_reshaped * mask, tf.shape(w)))
This is based on this Github repo. Please note that in that project, I am pruning the smallest k weights.
I want to understand better those shape of the Tensorflow´s BasicLSTMCell Kernel and Bias.
#tf_export("nn.rnn_cell.BasicLSTMCell")
class BasicLSTMCell(LayerRNNCell):
input_depth = inputs_shape[1].value
h_depth = self._num_units
self._kernel = self.add_variable(
_WEIGHTS_VARIABLE_NAME,
shape=[input_depth + h_depth, 4 * self._num_units])
self._bias = self.add_variable(
_BIAS_VARIABLE_NAME,
shape=[4 * self._num_units],
initializer=init_ops.zeros_initializer(dtype=self.dtype))
Why does the kernel have the shape=[input_depth + h_depth, 4 * self._num_units]) and the bias the shape = [4 * self._num_units] ? Maybe the factor 4 come from the forget gate, block input, input gate and output gate? And what´s the reason for the summation of input_depth and h_depth?
More information about my LSTM Network:
num_input = 12, timesteps = 820, num_hidden = 64, num_classes = 2.
With tf.trainables_variables() i get the following information:
Variable name: Variable:0 Shape: (64, 2) Parameters: 128
Variable name: Variable_1:0 Shape: (2,) Parameters: 2
Variable name: rnn/basic_lstm_cell/kernel:0 Shape: (76, 256) Parameters: 19456
Variable name: rnn/basic_lstm_cell/bias:0 Shape: (256,) Parameters: 256
The following Code defines my LSTM Network.
def RNN(x, weights, biases):
x = tf.unstack(x, timesteps, 1)
lstm_cell = rnn.BasicLSTMCell(num_hidden)
outputs, states = rnn.static_rnn(lstm_cell, x, dtype=tf.float32)
return tf.matmul(outputs[-1], weights['out']) + biases['out']
First, about summing input_depth and h_depth: RNNs generally follow equations like h_t = W*h_t-1 + V*x_t to compute the state h at time t. That is, we apply a matrix multiplication to the last state and the current input and add the two. This is actually equivalent to concatenating h_t-1 and x_t (let's just call this c), "stacking" the two matrices W and V (let's just call this S) and computing S*c.
Now we only have one matrix multiplication instead of two; I believe this can be parallelized more effectively so this is done for performance reasons. Since h_t-1 has size h_depth and x has size input_depth we need to add the two dimensionalities for the concatenated vector c.
Second, you are right about the factor 4 coming from the gates. This is essentially the same as above: Instead of carrying out four separate matrix multiplications for the input and each of the gates, we carry out one multiplication that results in a big vector that is the input and all four gate values concatenated. Then we can just split this vector into four parts. In the LSTM cell source code this happens in lines 627-633.
I am looking methods to embed variable length sequences with float values to fixed size vectors. The input formats as following:
[f1,f2,f3,f4]->[f1,f2,f3,f4]->[f1,f2,f3,f4]-> ... -> [f1,f2,f3,f4]
[f1,f2,f3,f4]->[f1,f2,f3,f4]->[f1,f2,f3,f4]->[f1,f2,f3,f4]-> ... -> [f1,f2,f3,f4]
...
[f1,f2,f3,f4]-> ... -> ->[f1,f2,f3,f4]
Each line is a variable length sequnece, with max length 60. Each unit in one sequece is a tuple of 4 float values. I have already paded zeros to fill all sequences to the same length.
The following architecture seems solve my problem if I use the output as the same as input, I need the thought vector in the center as the embedding for the sequences.
In tensorflow, I have found tow candidate methods tf.contrib.legacy_seq2seq.basic_rnn_seq2seq and tf.contrib.legacy_seq2seq.embedding_rnn_seq2seq.
However, these tow methos seems to be used to solve NLP problem, and the input must be discrete value for words.
So, is there another functions to solve my problems?
All you need is only an RNN, not the seq2seq model, since seq2seq goes with an additional decoder which is unecessary in your case.
An example code:
import numpy as np
import tensorflow as tf
from tensorflow.contrib import rnn
input_size = 4
max_length = 60
hidden_size=64
output_size = 4
x = tf.placeholder(tf.float32, shape=[None, max_length, input_size], name='x')
seqlen = tf.placeholder(tf.int64, shape=[None], name='seqlen')
lstm_cell = rnn.BasicLSTMCell(hidden_size, forget_bias=1.0)
outputs, states = tf.nn.dynamic_rnn(cell=lstm_cell, inputs=x, sequence_length=seqlen, dtype=tf.float32)
encoded_states = states[-1]
W = tf.get_variable(
name='W',
shape=[hidden_size, output_size],
dtype=tf.float32,
initializer=tf.random_normal_initializer())
b = tf.get_variable(
name='b',
shape=[output_size],
dtype=tf.float32,
initializer=tf.random_normal_initializer())
z = tf.matmul(encoded_states, W) + b
results = tf.sigmoid(z)
###########################
## cost computing and training components goes here
# e.g.
# targets = tf.placeholder(tf.float32, shape=[None, input_size], name='targets')
# cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=targets, logits=z))
# optimizer = tf.train.AdamOptimizer(learning_rate=0.1).minimize(cost)
###############################
init = tf.global_variables_initializer()
batch_size = 4
data_in = np.zeros((batch_size, max_length, input_size), dtype='float32')
data_in[0, :4, :] = np.random.rand(4, input_size)
data_in[1, :6, :] = np.random.rand(6, input_size)
data_in[2, :20, :] = np.random.rand(20, input_size)
data_in[3, :, :] = np.random.rand(60, input_size)
data_len = np.asarray([4, 6, 20, 60], dtype='int64')
with tf.Session() as sess:
sess.run(init)
#########################
# training process goes here
#########################
res = sess.run(results,
feed_dict={
x: data_in,
seqlen: data_len})
print(res)
To encode sequence to a fixed length vector you typically use recurrent neural networks (RNNs) or convolutional neural networks (CNNs).
If you use a recurrent neural network you can use the output at the last time step (last element in your sequence). This corresponds to the thought vector in your question. Have a look at tf.dynamic_rnn. dynamic_rnn requires you to specify to type of RNN cell you want to use. tf.contrib.rnn.LSTMCell and tf.contrib.rnn.GRUCell are most common.
If you want to use CNNs you need to use 1 dimensional convolutions. To build CNNs you need tf.layers.conv1d and tf.layers.max_pooling1d
I have found a solution to my problem, using the following architecture,
,
The LSTMs layer below encode the series x1,x2,...,xn. The last output, the green one, is duplicated to the same count as the input for the decoding LSTM layers above. The tensorflow code is as following
series_input = tf.placeholder(tf.float32, [None, conf.max_series, conf.series_feature_num])
print("Encode input Shape", series_input.get_shape())
# encoding layer
encode_cell = tf.contrib.rnn.MultiRNNCell(
[tf.contrib.rnn.BasicLSTMCell(conf.rnn_hidden_num, reuse=False) for _ in range(conf.rnn_layer_num)]
)
encode_output, _ = tf.nn.dynamic_rnn(encode_cell, series_input, dtype=tf.float32, scope='encode')
print("Encode output Shape", encode_output.get_shape())
# last output
encode_output = tf.transpose(encode_output, [1, 0, 2])
last = tf.gather(encode_output, int(encode_output.get_shape()[0]) - 1)
# duplite the last output of the encoding layer
decoder_input = tf.stack([last for _ in range(conf.max_series)], axis=1)
print("Decoder input shape", decoder_input.get_shape())
# decoding layer
decode_cell = tf.contrib.rnn.MultiRNNCell(
[tf.contrib.rnn.BasicLSTMCell(conf.series_feature_num, reuse=False) for _ in range(conf.rnn_layer_num)]
)
decode_output, _ = tf.nn.dynamic_rnn(decode_cell, decoder_input, dtype=tf.float32, scope='decode')
print("Decode output", decode_output.get_shape())
# Loss Function
loss = tf.losses.mean_squared_error(labels=series_input, predictions=decode_output)
print("Loss", loss)
what I have is the following, which I believe is a network with one hidden LSTM layer:
# Parameters
learning rate = 0.001
training_iters = 100000
batch_size = 128
display_step = 10
# Network Parameters
n_input = 13
n_steps = 10
n_hidden = 512
n_classes = 13
# tf Graph input
x = tf.placeholder("float", [None, n_steps, n_input])
y = tf.placeholder("float", [None, n_classes])
# Define weights
weights = {
'out' : tf.Variable(tf.random_normal([n_hidden, n_classes]))
}
biases = {
'out' : tf.Variable(tf.random_normal([n_classes]))
}
However, I am trying to build an LSTM network using TensorFlow to predict power consumption. I have been looking around to find a good example, but I could not find any model with 2 hidden LSTM layers. Here's the model that I would like to build:
1 input layer,
1 output layer,
2 hidden LSTM layers(with 512 neurons in each),
time step(sequence length): 10
Could anyone guide me to build this using TensorFlow? ( from defining weights, building input shape, training, predicting, use of optimizer or cost function, etc), any help would be much appreciated.
Thank you so much in advance!
Here is how I do it in a translation model with GRU cells. You can just replace the GRU with an LSTM. It is really easy just use tf.nn.rnn_cell.MultiRNNCell with a list of the multiple cells it should wrap. In the code bellow I am manually unrolling it but you can pass it to tf.nn.dynamic_rnn or tf.nn.rnn as well.
y = input_tensor
with tf.variable_scope('encoder') as scope:
rnn_cell = rnn.MultiRNNCell([rnn.GRUCell(1024) for _ in range(3)])
state = tf.zeros((BATCH_SIZE, rnn_cell.state_size))
output = [None] * TIME_STEPS
for t in reversed(range(TIME_STEPS)):
y_t = tf.reshape(y[:, t, :], (BATCH_SIZE, -1))
output[t], state = rnn_cell(y_t, state)
scope.reuse_variables()
y = tf.pack(output, 1)
First you need some placeholders to put your training data (one batch)
x_input = tf.placeholder(tf.float32, [batch_size, truncated_series_length, 1])
y_output = tf.placeholder(tf.float32, [batch_size, truncated_series_length, 1])
A LSTM need a state, which consists of two components, the hidden state and the cell state, very good guide here: https://arxiv.org/pdf/1506.00019.pdf. For every layer in the LSTM you have one cell state and one hidden state.
The problem is that Tensorflow stores this in a LSTMStateTuple which you can not send into placeholder. So you need to store it in a Tensor, and then unpack it into a tuple:
state_placeholder = tf.placeholder(tf.float32, [num_layers, 2, batch_size, state_size])
l = tf.unpack(state_placeholder, axis=0)
rnn_tuple_state = tuple(
[tf.nn.rnn_cell.LSTMStateTuple(l[idx][0], l[idx][1])
for idx in range(num_layers)]
)
Then you can use the built-in Tensorflow API to create the stacked LSTM layer.
cell = tf.nn.rnn_cell.LSTMCell(state_size, state_is_tuple=True)
cell = tf.nn.rnn_cell.MultiRNNCell([cell]*num_layers, state_is_tuple=True)
outputs, state = tf.nn.dynamic_rnn(cell, x_input, initial_state=rnn_tuple_state)
From here you continue with the outputs to calculate logits and then a loss with respect to the y_inputs.
Then you run each batch with the sess.run-command, with truncated backpropagation (good explanation here http://r2rt.com/styles-of-truncated-backpropagation.html)
init_state = np.zeros((num_layers, 2, batch_size, state_size))
...current_state... = sess.run([...state...], feed_dict={x_input:batch_in, state_placeholder:current_state ...})
current_state = np.array(current_state)
You will have to convert the state to a numpy array before feeding it again.
Perhaps it is better to use a librarly like Tflearn or Keras instead?