Develop Reference

is tensor2tensor/avg_checkpoints script wrong

I am training two sub-models in parallel, and want to average them after both of them are finished.
The model is implemented out of tensor2tensor (but still use tensorflow). Part of the definition is like this:
def build_model(self):
layer_sizes = [n,n,n]
kernel_sizes = [n,n,n]
tf.reset_default_graph()
self.graph = tf.Graph()
with self.graph.as_default():
self.num_classes = num_classes
# placeholder for parameter
self.learning_rate = tf.placeholder(tf.float32, name="learning_rate")
self.dropout_keep_prob = tf.placeholder(tf.float32, name="dropout_keep_prob")
self.phase = tf.placeholder(tf.bool, name="phase")
# Placeholders for regular data
self.input_x = tf.placeholder(tf.float32, [None, None, input_feature_dim], name="input_x")
self.input_y = tf.placeholder(tf.float32, [None, num_classes], name="input_y")
h = self.input_x
......[remaining codes]
I save it following:
def save_model(sess, output):
saver = tf.train.Saver()
save_path = saver.save(sess, os.path.join(output, 'model'))
When I load it, I use:
def load_model(self, sess, input_dir, logger):
if logger is not None:
logger.info("Start loading graph ...")
saver = tf.train.import_meta_graph(os.path.join(input_dir, 'model.meta'))
saver.restore(sess, os.path.join(input_dir, 'model'))
self.graph = sess.graph
self.input_x = self.graph.get_tensor_by_name("input_x:0")
self.input_y = self.graph.get_tensor_by_name("input_y:0")
self.num_classes = self.input_y.shape[1]
self.learning_rate = self.graph.get_tensor_by_name("learning_rate:0")
self.dropout_keep_prob = self.graph.get_tensor_by_name("dropout_keep_prob:0")
self.phase = self.graph.get_tensor_by_name("phase:0")
self.loss = self.graph.get_tensor_by_name("loss:0")
self.optimizer = self.graph.get_operation_by_name("optimizer")
self.accuracy = self.graph.get_tensor_by_name("accuracy/accuracy:0")
I use the avg_checkpoint to average two sub-models:
python utils/avg_checkpoints.py
--checkpoints path/to/checkpoint1,path/to/checkpoint2
--num_last_checkpoints 2
--output_path where/to/save/the/output
But I find problems when I check the avg_checkpoints code further:
for checkpoint in checkpoints:
reader = tf.train.load_checkpoint(checkpoint)
for name in var_values:
tensor = reader.get_tensor(name)
var_dtypes[name] = tensor.dtype
var_values[name] += tensor
tf.logging.info("Read from checkpoint %s", checkpoint)
for name in var_values: # Average.
var_values[name] /= len(checkpoints)
with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
tf_vars = [
tf.get_variable(v, shape=var_values[v].shape, dtype=var_dtypes[v])
for v in var_values
]
placeholders = [tf.placeholder(v.dtype, shape=v.shape) for v in tf_vars]
assign_ops = [tf.assign(v, p) for (v, p) in zip(tf_vars, placeholders)]
global_step = tf.Variable(
0, name="global_step", trainable=False, dtype=tf.int64)
saver = tf.train.Saver(tf.all_variables())
# Build a model consisting only of variables, set them to the average values.
with tf.Session() as sess:
for p, assign_op, (name, value) in zip(placeholders, assign_ops,
six.iteritems(var_values)):
sess.run(assign_op, {p: value})
# Use the built saver to save the averaged checkpoint.
saver.save(sess, FLAGS.output_path, global_step=global_step)
It only save the variables, not all tensors. E.G. when I load it using the above load_model function, it cannot have "input_x:0" tensor. Is this script wrong, or I should modify it based on my use?
I am using TF r1.13. Thanks!

TensorFlow tf.train.Saver() not working on tf.contrib.layers.fully_connected()

So I wrote this generalised TensorFlow code and want to save and restore models. But apparently the error is that there is no variables to save. I did everything as given in this official example. Ignore the __init__ method except the last line, since it only takes relevant parameters to train the model with, also there is no Syntax Errors. The error it produces is given below the code.
class Neural_Network(object):
def __init__(self, numberOfLayers, nodes, activations, learningRate,
optimiser = 'GradientDescent', regularizer = None,
dropout = 0.5, initializer = tf.contrib.layers.xavier_initializer()):
self.numberOfLayers = numberOfLayers
self.nodes = nodes
self.activations = activations
self.learningRate = learningRate
self.regularizer = regularizer
self.dropout = dropout
self.initializer = initializer
if(optimiser == 'GradientDescent'):
self.optimiser = tf.train.GradientDescentOptimizer(self.learningRate)
elif(optimiser == 'AdamOptimiser'):
self.optimiser = tf.train.AdamOptimizer(self.learningRate)
self.saver = tf.train.Saver()
def create_Neural_Net(self, numberOfFeatures):
self.numberOfFeatures = numberOfFeatures
self.X = tf.placeholder(dtype = tf.float32, shape = (None, self.numberOfFeatures), name = 'Input_Dataset')
#self.output = None
for i in range(0, self.numberOfLayers):
if(i == 0):
layer = tf.contrib.layers.fully_connected(self.X, self.nodes[i],
activation_fn = self.activations[i],
weights_initializer = self.initializer,
biases_initializer = self.initializer)
elif(i == self.numberOfLayers-1):
self.output = tf.contrib.layers.fully_connected(layer, self.nodes[i],
activation_fn = self.activations[i],
weights_initializer = self.initializer,
biases_initializer = self.initializer)
else:
layer = tf.contrib.layers.fully_connected(layer, self.nodes[i],
activation_fn = self.activations[i],
weights_initializer = self.initializer,
biases_initializer = self.initializer)
def train_Neural_Net(self, dataset, labels, epochs):
entropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits = self.output, labels = labels, name = 'cross_entropy')
loss = tf.reduce_mean(entropy, name = 'loss')
hypothesis = tf.nn.softmax(self.output)
correct_preds = tf.equal(tf.argmax(hypothesis, 1), tf.argmax(labels, 1))
accuracy = tf.reduce_sum(tf.cast(correct_preds, tf.float32))
train_op = self.optimiser.minimize(loss)
self.loss=[]
self.accuracy = []
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(0, epochs):
_, l, acc = sess.run([train_op, loss, accuracy], feed_dict = {self.X:dataset})
print('Loss in epoch ' + str(i) + ' is: ' + str(l))
self.loss.append(l)
self.accuracy.append(acc)
self.saver.save(sess, './try.ckpt')
return self.loss, self.accuracy
And ran this code as:
nn = Neural_Network(2, [20,3], [tf.nn.relu, tf.nn.relu], 0.001, optimiser = 'AdamOptimiser')
nn.create_Neural_Net(4)
nn.train_Neural_Net(dataset, labels, 1000)
The error it gives is:
ValueError: No variables to save
So what is wrong in this code? And how can I fix it?

With Tensorflow, 2 class classification using Neural Network

I'm trying 2 class classification of images by neural network using Tensorflow.
I want to extract 1000 pixels randomly.
However, I am in trouble with error:
"logits = inference(images_placeholder, keep_prob)
File "train5.py", line 83, in inference
list = random.sample(x_image(IMAGE_PIXELS),SAMPLE_PIXELS)
TypeError: 'Tensor' object is not callable"
Please tell me what should I do.
I will attach a code below.
import sys
sys.path.append('/usr/local/opt/opencv3/lib/python3.5.4/site-packages')
import cv2
import numpy as np
import tensorflow as tf
import tensorflow.python.platform
import tensorboard as tb
import os
import math
import time
import random
start_time = time.time()
# TensorBoard output information directory
log_dir = '/tmp/data1' #tensorboard --logdir=/tmp/data1
#directory delete and reconstruction
if tf.gfile.Exists(log_dir):
tf.gfile.DeleteRecursively(log_dir)
tf.gfile.MakeDirs(log_dir)
# Reserve memory
config = tf.ConfigProto(
gpu_options=tf.GPUOptions(allow_growth=True))
sess = sess = tf.Session(config=config)
NUM_CLASSES = 2
IMAGE_SIZE_x = 1024
IMAGE_SIZE_y = 768
IMAGE_CHANNELS = 1
IMAGE_PIXELS = IMAGE_SIZE_x*IMAGE_SIZE_y*IMAGE_CHANNELS
SAMPLE_PIXELS = 1000
flags = tf.app.flags
FLAGS = flags.FLAGS
flags.DEFINE_string('train', 'train.txt', 'File name of train data')
flags.DEFINE_string('test', 'test.txt', 'File name of train data')
flags.DEFINE_string('image_dir', 'trdata', 'Directory of images')
flags.DEFINE_string('train_dir', '/tmp/data', 'Directory to put the training data.')
flags.DEFINE_integer('max_steps', 20000, 'Number of steps to run trainer.')
flags.DEFINE_integer('batch_size', 10, 'Batch size'
'Must divide evenly into the dataset sizes.')
flags.DEFINE_float('learning_rate', 1e-5, 'Initial learning rate.')
def inference(images_placeholder, keep_prob):
""" Function to create predictive model
argument:
images_placeholder: image placeholder
keep_prob: dropout rate placeholder
Return:
y_conv:
"""
# Initialie with normal distribution with weight of 0.1
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial)
# Initialized with normal distribution with bias of 0.1
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
# Reshape input
x_image = images_placeholder
# ramdom sumpling pixels
list = random.sample(x_image(IMAGE_PIXELS),SAMPLE_PIXELS)
x_list = [samples[i] for i in list]
# Input
with tf.name_scope('fc1') as scope:
W_fc1 = weight_variable([x_list,10])
b_fc1 = bias_variable([10])
h_fc1 = tf.nn.relu(tf.matmul(x_image,W_fc1) + b_fc1)
# Affine1
with tf.name_scope('fc2') as scope:
W_fc2 = weight_variable([10,10])
b_fc2 = bias_variable([10])
h_fc2 = tf.nn.relu(tf.matmul(h_fc1,W_fc2) + b_fc2)
# Affine2
with tf.name_scope('fc3') as scope:
W_fc3 = weight_variable([10,10])
b_fc3 = bias_variable([10])
h_fc3 = tf.nn.relu(tf.matmul(h_fc2,W_fc3) + b_fc3)
# Affine3
with tf.name_scope('fc4') as scope:
W_fc4 = weight_variable([10,10])
b_fc4 = bias_variable([10])
h_fc4 = tf.nn.relu(tf.matmul(h_fc3,W_fc4) + b_fc4)
# Affine4
with tf.name_scope('fc5') as scope:
W_fc5 = weight_variable([10,2])
b_fc5 = bias_variable([2])
# softmax regression
with tf.name_scope('softmax') as scope:
y_out=tf.nn.softmax(tf.matmul(h_fc4, W_fc5) + b_fc5)
# return
return y_out
def loss(logits, labels):
""" loss function
引数:
logits: logit tensor, float - [batch_size, NUM_CLASSES]
labels: labrl tensor, int32 - [batch_size, NUM_CLASSES]
返り値:
cross_entropy:tensor, float
"""
# cross entropy
cross_entropy = -tf.reduce_sum(labels*tf.log(tf.clip_by_value(logits,1e-10,1.0)))
# TensorBoard
tf.summary.scalar("cross_entropy", cross_entropy)
return cross_entropy
def training(loss, learning_rate):
#Adam
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
return train_step
def accuracy(logits, labels):
correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
tf.summary.scalar("accuracy", accuracy)
return accuracy
if __name__ == '__main__':
f = open(FLAGS.train, 'r')
# array data
train_image = []
train_label = []
for line in f:
# Separate space and remove newlines
line = line.rstrip()
l = line.split()
# Load data and resize
img = cv2.imread(FLAGS.image_dir + '/' + l[0])
img = cv2.resize(img, (IMAGE_SIZE_x, IMAGE_SIZE_y))
#transrate grayscale
img_gry = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# transrate one row and 0-1 float
train_image.append(img_gry.flatten().astype(np.float32)/255.0)
# Prepare with label 1-of-k method
tmp = np.zeros(NUM_CLASSES)
tmp[int(l[1])] = 1
train_label.append(tmp)
# transrate numpy
train_image = np.asarray(train_image)
train_label = np.asarray(train_label)
f.close()
f = open(FLAGS.test, 'r')
test_image = []
test_label = []
for line in f:
line = line.rstrip()
l = line.split()
img = cv2.imread(FLAGS.image_dir + '/' + l[0])
img = cv2.resize(img, (IMAGE_SIZE_x, IMAGE_SIZE_y))
#transrate grayscale
img_gry = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# transrate one row and 0-1 float
test_image.append(img_gry.flatten().astype(np.float32)/255.0)
tmp = np.zeros(NUM_CLASSES)
tmp[int(l[1])] = 1
test_label.append(tmp)
test_image = np.asarray(test_image)
test_label = np.asarray(test_label)
f.close()
with tf.Graph().as_default():
# Put the image Tensor
images_placeholder = tf.placeholder("float", shape=(None, IMAGE_PIXELS))
# Put the label Tensor
labels_placeholder = tf.placeholder("float", shape=(None, NUM_CLASSES))
# Put dropout rate Tensor
keep_prob = tf.placeholder("float")
# Load inference() and make model
logits = inference(images_placeholder, keep_prob)
# Load loss() and calculate loss
loss_value = loss(logits, labels_placeholder)
# Load training() and train
train_op = training(loss_value, FLAGS.learning_rate)
# calculate accuracy
acc = accuracy(logits, labels_placeholder)
# save
saver = tf.train.Saver()
# Make Session
sess = tf.Session()
# Initialize variable
sess.run(tf.global_variables_initializer())
# TensorBoard
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(FLAGS.train_dir, sess.graph)
# Start training
for step in range(FLAGS.max_steps):
for i in range(int(len(train_image)/FLAGS.batch_size)):
batch = FLAGS.batch_size*i
sess.run(train_op, feed_dict={
images_placeholder: train_image[batch:batch+FLAGS.batch_size],
labels_placeholder: train_label[batch:batch+FLAGS.batch_size],
keep_prob: 0.5})
# Accuracy calculation for every steps
train_accuracy = sess.run(acc, feed_dict={
images_placeholder: train_image,
labels_placeholder: train_label,
keep_prob: 1.0})
print("step %d, training accuracy %g" %(step, train_accuracy))
# Added value to be displayed in Tensorflow every 1step
summary_str = sess.run(summary_op, feed_dict={
images_placeholder: train_image,
labels_placeholder: train_label,
keep_prob: 1.0})
summary_writer.add_summary(summary_str, step)
# Display accuracy on test data after training
print(" test accuracy %g"%sess.run(acc, feed_dict={
images_placeholder: test_image,
labels_placeholder: test_label,
keep_prob: 1.0}))
duration = time.time() - start_time
print('%.3f sec' %duration)
# Save model
save_path = saver.save(sess, os.getcwd() + "\\model.ckpt")

The error is this:
images_placeholder = tf.placeholder("float", shape=(None, IMAGE_PIXELS))
...
x_image = images_placeholder
list = random.sample(x_image(IMAGE_PIXELS),SAMPLE_PIXELS)
x_image, just like images_placeholder is a variable node, so x_image(...) doesn't make sense and obviously leads to an error "TypeError: 'Tensor' object is not callable".
I assume you're trying to sample SAMPLE_PIXELS from each image in a batch. Note that random.sample won't work here, because x_image is a symbolic variable, its value is only known during the session. You have to use tf.boolean_mask with a random mask in order to select random pixels from the image.

tensorflow static_rnn error: input must be a sequence

I'm trying to feed my own 3D data to a LSTM. The data have: height = 365, width = 310, time = unknown / inconsistent, consist of 0 and 1, each block of data that produce an output are separated to a single file.
import tensorflow as tf
import os
from tensorflow.contrib import rnn
filename = "C:/Kuliah/EmotionRecognition/Train1/D2N2Sur.txt"
hm_epochs = 10
n_classes = 12
n_chunk = 443
n_hidden = 500
data = tf.placeholder(tf.bool, name='data')
cat = tf.placeholder("float", [None, n_classes])
weights = {
'out': tf.Variable(tf.random_normal([n_hidden, n_classes]))
}
biases = {
'out': tf.Variable(tf.random_normal([n_classes]))
}
def RNN(x, weights, biases):
lstm_cell = rnn.BasicLSTMCell(n_hidden, forget_bias=1.0)
outputs, states = rnn.static_rnn(lstm_cell, x, dtype=tf.float32)
return tf.matmul(outputs[-1], weights['out']) + biases['out']
pred = RNN(data, weights, biases)
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=cat))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)
correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(cat, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
saver = tf.train.Saver()
temp = [[]]
d3 = [[]]
counter = 0
with tf.Session() as sess:
#load
#saver.restore(sess, "C:/Kuliah/EmotionRecognition/model.ckpt")
sess.run(tf.global_variables_initializer())
with open(filename) as inf:
for line in inf:
bla = list(line)
bla.pop(len(bla) - 1)
for index, item in enumerate(bla):
if (item == '0'):
bla[index] = False
else:
bla[index] = True
temp.append(bla)
counter += 1
if counter%365==0: #height 365
temp.pop(0)
d3.append(temp)
temp = [[]]
temp.pop(0)
d3.append(temp)
batch_data = d3.reshape()
sess.run(optimizer, feed_dict={data: d3, cat: 11})
acc = sess.run(accuracy, feed_dict={data: d3, cat: 11})
loss = sess.run(loss, feed_dict={data: d3, cat: 11})
print(acc)
print(loss)
#save
saver.save(sess, "C:/Kuliah/EmotionRecognition/model.ckpt")
this code throw me an error:
Traceback (most recent call last):
File "C:/Kuliah/EmotionRecognition/Main", line 31, in <module>
pred = RNN(data, weights, biases)
File "C:/Kuliah/EmotionRecognition/Main", line 28, in RNN
outputs, states = rnn.static_rnn(lstm_cell, x, dtype=tf.float32)
File "C:\Users\Anonymous\AppData\Roaming\Python\Python35\site-packages\tensorflow\python\ops\rnn.py", line 1119, in static_rnn
raise TypeError("inputs must be a sequence")
TypeError: inputs must be a sequence

When you call pred = RNN(data, weights, biases), the data argument should be a sequence of length the length of your RNN. But in your case, it's a data = tf.placeholder(tf.bool, name='data').
You could try pred = RNN([data], weights, biases).
See the string doc of the method:
inputs: A length T list of inputs, each a Tensor of shape
[batch_size, input_size], or a nested tuple of such elements.
If the length of your RNN is unknow, you should consider use tf.nn.dynamic_rnn.

add new output for pre-trained model

I am confused to add new class for a pre-trained model, what i have done until now restore the pre-trained checkpoint and create a matrix of size m * C+1 and a vector of length C+1, then initialize the first C rows/elements of these from the existing weights and freeze the previous layer by training just the FC layer in the Optimizer.minimize(). However when i run the code, i got this error :
Traceback (most recent call last):
File "/home/tensorflow/tensorflow/models/image/mnist/new_dataset/Nets.py", line 482, in <module>
new_op_w = optimizer_new.minimize(loss, var_list = resize_var_w)
File "/home/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/training/optimizer.py", line 279, in minimize
grad_loss=grad_loss)
File "/home/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/training/optimizer.py", line 337, in compute_gradients
processors = [_get_processor(v) for v in var_list]
File "/home/tensorflow/local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 502, in __iter__
raise TypeError("'Tensor' object is not iterable.")
TypeError: 'Tensor' object is not iterable.
and that's the code :
with tf.Session(graph=graph) as sess:
if os.path.isfile(ckpt):
aver.restore(sess, 'path_to_checkpoint.ckpt')
w_b_new = {
'weight_4': tf.Variable(tf.random_normal([num_hidden, 1], stddev=0.1), name = 'weight_4'),
'bias_4' : tf.Variable(tf.constant(1.0, shape=[1]), name = 'bias_4'),}
change_1 = tf.unstack(w_b_not['weight_4'])
change_2 = tf.unstack(w_b_not['bias_4'])
change_3 = tf.unstack(w_b_new['weight_4'])
change_4 = tf.unstack(w_b_new['bias_4'])
changestep1 = []
for i in range(len(change_1)):
changestep1.append(tf.unstack(change_1[i]))
changestep3 = []
for i in range(len(change_3)):
changestep3.append(tf.unstack(change_3[i]))
for j in range(len(changestep3[i])):
changestep1[i].append(changestep3[i][j])
changestep1[i] = tf.stack(changestep1[i])
final1 = tf.stack(changestep1)
resize_var_w = tf.assign(w_b_not['weight_4'], final1, validate_shape=False)
final2 = tf.concat([w_b_not['bias_4'] , w_b_new['bias_4']], axis=0)
resize_var = tf.assign(w_b_not['bias_4'], final2, validate_shape=False)
optimizer_new = tf.train.GradientDescentOptimizer(0.01)
new_op_w = optimizer_new.minimize(loss, var_list = resize_var_w)
new_op_b = optimizer_new.minimize(loss, var_list = resize_var)
for step in range(num_steps,num_steps + num_train_steps):
offset = (step * batch_size) % (train_labels.shape[0] - batch_size)
batch_data = train_dataset[offset:(offset + batch_size), :, :, :]
batch_labels = train_labels[offset:(offset + batch_size), :]
feed_dict = {tf_train_dataset : batch_data, tf_train_labels : batch_labels , keep_prob:0.5}
_,_, l, predictions = sess.run([new_op_w,new_op_b, loss, train_prediction ], feed_dict=feed_dict)
if (step % 50 == 0):
print('%d\t%f\t%.1f%%\t%.1f%%' % (step, l, accuracy(predictions, batch_labels), accuracy(valid_prediction.eval(), valid_labels)))
print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval() , test_labels))
save_path_w_b = saver.save(sess, "path_checkpoint.ckpt")
print("Model saved in file: %s" % save_path_w_b)

According to the TensorFlow docs on GradientDescentOptimizer's minimize method, "var_list" must be a list of Variable objects. According to your code, resize_var_w is a single tensor.
EDIT
To be more specific:
If you give the optimizer var_list, as the name implies, this must be a list of variables. During backprop the optimizer will loop var_list and only update the variables in the list, as opposed to all trainable variables in the graph. A single variable is not iterable.
If you only want to update a single Tensor, you can simply try:
resize_var_w = [tf.assign(w_b_not['weight_4'], final1, validate_shape=False)]
I did not test, but should work.