We Keep Coding

Display loss in a Tensorflow DQN without leaving tf.Session()

I have a DQN all set up and working, but I can't figure out how to display the loss without leaving the Tensorflow session.
I first thought it involved creating a new function or class, but I'm not sure where to put it in the code, and what specifically to put into the function or class.
observations = tf.placeholder(tf.float32, shape=[None, num_stops], name='observations')
actions = tf.placeholder(tf.int32,shape=[None], name='actions')
rewards = tf.placeholder(tf.float32,shape=[None], name='rewards')
# Model
Y = tf.layers.dense(observations, 200, activation=tf.nn.relu)
Ylogits = tf.layers.dense(Y, num_stops)
# sample an action from predicted probabilities
sample_op = tf.random.categorical(logits=Ylogits, num_samples=1)
# loss
cross_entropies = tf.losses.softmax_cross_entropy(onehot_labels=tf.one_hot(actions,num_stops), logits=Ylogits)
loss = tf.reduce_sum(rewards * cross_entropies)
# training operation
optimizer = tf.train.RMSPropOptimizer(learning_rate=0.001, decay=.99)
train_op = optimizer.minimize(loss)
I then run the network, which works without error.
with tf.Session() as sess:
'''etc. The network is run'''
sess.run(train_op, feed_dict={observations: observations_list,
actions: actions_list,
rewards: rewards_list})
I want to have loss from train_op displayed to the user.

try this
loss, _ = sess.run([loss, train_op], feed_dict={observations: observations_list,
actions: actions_list,
rewards: rewards_list})

How to plot different summary metrics on the same plot with Tensorboard?

I would like to be able plot the training loss per batch and the average validation loss for the validation set on the same plot in Tensorboard. I ran into this issue when my validation set was too large to fit into memory so required batching and the use of tf.metrics update ops.
This question could apply to any Tensorflow metrics you wanted to appear on the same graph in Tensorboard.
I am able to
plot these two graphs separately (see here)
plot the validation-loss-per-validation-batch on the same graph as the training-loss-per-training-batch (this was OK when the validation set could be a single batch and I could reuse the training summary op train_summ below)
In the example code below, my issue stems from the fact that my validation summary tf.summary.scalar with name=loss gets renamed to loss_1 and thus is moved to a separate graph in Tensorboard. From what I can work out Tensorboard takes "same name" and plots them on the same graph, regardless of what folder they are in. This is frustrating as train_summ (name=loss) is only ever written to the train folder and valid_summ (name=loss) is only ever written to the valid folder - but is still renamed to loss_1.
The example code:
# View graphs with (Linux): $ tensorboard --logdir=/tmp/my_tf_model
import tensorflow as tf
import numpy as np
import os
import tempfile
def train_data_gen():
yield np.random.normal(size=[3]), np.array([0.5, 0.5, 0.5])
def valid_data_gen():
yield np.random.normal(size=[3]), np.array([0.8, 0.8, 0.8])
batch_size = 25
n_training_batches = 4
n_valid_batches = 2
n_epochs = 5
summary_loc = os.path.join(tempfile.gettempdir(), 'my_tf_model')
print("Summaries written to" + summary_loc)
# Dummy data
train_data = tf.data.Dataset.from_generator(train_data_gen, (tf.float32, tf.float32)).repeat().batch(batch_size)
valid_data = tf.data.Dataset.from_generator(valid_data_gen, (tf.float32, tf.float32)).repeat().batch(batch_size)
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(handle,
train_data.output_types, train_data.output_shapes)
batch_x, batch_y = iterator.get_next()
train_iter = train_data.make_initializable_iterator()
valid_iter = valid_data.make_initializable_iterator()
# Some ops on the data
loss = tf.losses.mean_squared_error(batch_x, batch_y)
valid_loss, valid_loss_update = tf.metrics.mean(loss)
# Write to summaries
train_summ = tf.summary.scalar('loss', loss)
valid_summ = tf.summary.scalar('loss', valid_loss) # <- will be renamed to "loss_1"
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
train_handle, valid_handle = sess.run([train_iter.string_handle(), valid_iter.string_handle()])
sess.run([train_iter.initializer, valid_iter.initializer])
# Summary writers
writer_train = tf.summary.FileWriter(os.path.join(summary_loc, 'train'), sess.graph)
writer_valid = tf.summary.FileWriter(os.path.join(summary_loc, 'valid'), sess.graph)
global_step = 0 # implicit as no actual training
for i in range(n_epochs):
# "Training"
for j in range(n_training_batches):
global_step += 1
summ = sess.run(train_summ, feed_dict={handle: train_handle})
writer_train.add_summary(summary=summ, global_step=global_step)
# "Validation"
sess.run(tf.local_variables_initializer())
for j in range(n_valid_batches):
_, batch_summ = sess.run([valid_loss_update, train_summ], feed_dict={handle: valid_handle})
# The following will plot the batch loss for the validation set on the loss plot with the training data:
# writer_valid.add_summary(summary=batch_summ, global_step=global_step + j + 1)
summ = sess.run(valid_summ)
writer_valid.add_summary(summary=summ, global_step=global_step) # <- I want this on the training loss graph
What I have tried
Separate tf.summary.FileWriter objects (one for training, one for validation), as recommended by this issue and this question (think what I'm after is alluded to in the comment of that question)
The use of tf.summary.merge to merge all my training and validation/test metrics into overall summary ops; does useful book-keeping but doesn't plot what I want on the same graph
Use of the tf.summary.scalar family attribute (loss still gets renamed to loss_1)
(Complete hack solution) Use valid_loss, valid_loss_update = tf.metrics.mean(loss) on the training data and then run tf.local_variables_initializer() every training batch. This does give you the same summary op and thus puts things on the same graph but is surely not how you're meant to do this? It also doesn't generalise to other metrics.
Context
Tensorflow 1.9.0
Tensorboard 1.9.0
Python 3.5.2

The Tensorboard custom_scalar plugin is the way to solve this problem.
Here's the same example again with a custom_scalar to plot the two losses (per training batch + averaged over all validation batches) on the same plot:
# View graphs with (Linux): $ tensorboard --logdir=/tmp/my_tf_model
import os
import tempfile
import tensorflow as tf
import numpy as np
from tensorboard import summary as summary_lib
from tensorboard.plugins.custom_scalar import layout_pb2
def train_data_gen():
yield np.random.normal(size=[3]), np.array([0.5, 0.5, 0.5])
def valid_data_gen():
yield np.random.normal(size=[3]), np.array([0.8, 0.8, 0.8])
batch_size = 25
n_training_batches = 4
n_valid_batches = 2
n_epochs = 5
summary_loc = os.path.join(tempfile.gettempdir(), 'my_tf_model')
print("Summaries written to " + summary_loc)
# Dummy data
train_data = tf.data.Dataset.from_generator(
train_data_gen, (tf.float32, tf.float32)).repeat().batch(batch_size)
valid_data = tf.data.Dataset.from_generator(
valid_data_gen, (tf.float32, tf.float32)).repeat().batch(batch_size)
handle = tf.placeholder(tf.string, shape=[])
iterator = tf.data.Iterator.from_string_handle(handle, train_data.output_types,
train_data.output_shapes)
batch_x, batch_y = iterator.get_next()
train_iter = train_data.make_initializable_iterator()
valid_iter = valid_data.make_initializable_iterator()
# Some ops on the data
loss = tf.losses.mean_squared_error(batch_x, batch_y)
valid_loss, valid_loss_update = tf.metrics.mean(loss)
with tf.name_scope('loss'):
train_summ = summary_lib.scalar('training', loss)
valid_summ = summary_lib.scalar('valid', valid_loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
train_handle, valid_handle = sess.run([train_iter.string_handle(), valid_iter.string_handle()])
sess.run([train_iter.initializer, valid_iter.initializer])
writer_train = tf.summary.FileWriter(os.path.join(summary_loc, 'train'), sess.graph)
writer_valid = tf.summary.FileWriter(os.path.join(summary_loc, 'valid'), sess.graph)
layout_summary = summary_lib.custom_scalar_pb(
layout_pb2.Layout(category=[
layout_pb2.Category(
title='losses',
chart=[
layout_pb2.Chart(
title='losses',
multiline=layout_pb2.MultilineChartContent(tag=[
'loss/training', 'loss/valid'
]))
])
]))
writer_train.add_summary(layout_summary)
global_step = 0
for i in range(n_epochs):
for j in range(n_training_batches): # "Training"
global_step += 1
summ = sess.run(train_summ, feed_dict={handle: train_handle})
writer_train.add_summary(summary=summ, global_step=global_step)
sess.run(tf.local_variables_initializer())
for j in range(n_valid_batches): # "Validation"
_, batch_summ = sess.run([valid_loss_update, train_summ], feed_dict={handle: valid_handle})
summ = sess.run(valid_summ)
writer_valid.add_summary(summary=summ, global_step=global_step)
Here's the resulting output in Tensorboard.

Restoring a Tensorflow model that uses Iterators

I have a model that's trains my network using an Iterator; following the new Dataset API pipeline model that's now recommended by Google.
I read tfrecord files, feed data to the network, train nicely, and all is going well, I save my model in the end of the training so I can run Inference on it later. A simplified version of the code is as following:
""" Training and saving """
training_dataset = tf.contrib.data.TFRecordDataset(training_record)
training_dataset = training_dataset.map(ds._path_records_parser)
training_dataset = training_dataset.batch(BATCH_SIZE)
with tf.name_scope("iterators"):
training_iterator = Iterator.from_structure(training_dataset.output_types, training_dataset.output_shapes)
next_training_element = training_iterator.get_next()
training_init_op = training_iterator.make_initializer(training_dataset)
def train(num_epochs):
# compute for the number of epochs
for e in range(1, num_epochs+1):
session.run(training_init_op) #initializing iterator here
while True:
try:
images, labels = session.run(next_training_element)
session.run(optimizer, feed_dict={x: images, y_true: labels})
except tf.errors.OutOfRangeError:
saver_name = './saved_models/ucf-model'
print("Finished Training Epoch {}".format(e))
break
""" Restoring """
# restoring the saved model and its variables
session = tf.Session()
saver = tf.train.import_meta_graph(r'saved_models\ucf-model.meta')
saver.restore(session, tf.train.latest_checkpoint('.\saved_models'))
graph = tf.get_default_graph()
# restoring relevant tensors/ops
accuracy = graph.get_tensor_by_name("accuracy/Mean:0") #the tensor that when evaluated returns the mean accuracy of the batch
testing_iterator = graph.get_operation_by_name("iterators/Iterator") #my iterator used in testing.
next_testing_element = graph.get_operation_by_name("iterators/IteratorGetNext") #the GetNext operator for my iterator
# loading my testing set tfrecords
testing_dataset = tf.contrib.data.TFRecordDataset(testing_record_path)
testing_dataset = testing_dataset.map(ds._path_records_parser, num_threads=4, output_buffer_size=BATCH_SIZE*20)
testing_dataset = testing_dataset.batch(BATCH_SIZE)
testing_init_op = testing_iterator.make_initializer(testing_dataset) #to initialize the dataset
with tf.Session() as session:
session.run(testing_init_op)
while True:
try:
images, labels = session.run(next_testing_element)
accuracy = session.run(accuracy, feed_dict={x: test_images, y_true: test_labels}) #error here, x, y_true not defined
except tf.errors.OutOfRangeError:
break
My problem is mainly when I restore the model. How to feed testing data to the network?
When I restore my Iterator using testing_iterator = graph.get_operation_by_name("iterators/Iterator"), next_testing_element = graph.get_operation_by_name("iterators/IteratorGetNext"), I get the following error:
GetNext() failed because the iterator has not been initialized. Ensure that you have run the initializer operation for this iterator before getting the next element.
So I did try to initialize my dataset using: testing_init_op = testing_iterator.make_initializer(testing_dataset)). I got this error: AttributeError: 'Operation' object has no attribute 'make_initializer'
Another issue is, since an iterator is being used, there's no need to use placeholders in the training_model, as an iterator feed data directly to the graph. But this way, how to restore my feed_dict keys in the 3rd to last line, when I feed data to the "accuracy" op?
EDIT: if someone could suggest a way to add placeholders between the Iterator and the network input, then I could try running the graph by evaluating the "accuracy" tensor while feeding data to the placeholders and ignoring the iterator altogether.

When restoring a saved meta graph, you can restore the initialization operation with name and then use it again to initialize the input pipeline for inference.
That is, when creating the graph, you can do
dataset_init_op = iterator.make_initializer(dataset, name='dataset_init')
And then restore this operation by doing:
dataset_init_op = graph.get_operation_by_name('dataset_init')
Here is a self contained code snippet that compares results of a randomly initialized model before and after restoring.
Saving an Iterator
np.random.seed(42)
data = np.random.random([4, 4])
X = tf.placeholder(dtype=tf.float32, shape=[4, 4], name='X')
dataset = tf.data.Dataset.from_tensor_slices(X)
iterator = tf.data.Iterator.from_structure(dataset.output_types, dataset.output_shapes)
dataset_next_op = iterator.get_next()
# name the operation
dataset_init_op = iterator.make_initializer(dataset, name='dataset_init')
w = np.random.random([1, 4])
W = tf.Variable(w, name='W', dtype=tf.float32)
output = tf.multiply(W, dataset_next_op, name='output')
sess = tf.Session()
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
sess.run(dataset_init_op, feed_dict={X:data})
while True:
try:
print(sess.run(output))
except tf.errors.OutOfRangeError:
saver.save(sess, 'tmp/', global_step=1002)
break
And then you can restore the same model for inference as follows:
Restoring saved iterator
np.random.seed(42)
data = np.random.random([4, 4])
tf.reset_default_graph()
sess = tf.Session()
saver = tf.train.import_meta_graph('tmp/-1002.meta')
ckpt = tf.train.get_checkpoint_state(os.path.dirname('tmp/checkpoint'))
saver.restore(sess, ckpt.model_checkpoint_path)
graph = tf.get_default_graph()
# Restore the init operation
dataset_init_op = graph.get_operation_by_name('dataset_init')
X = graph.get_tensor_by_name('X:0')
output = graph.get_tensor_by_name('output:0')
sess.run(dataset_init_op, feed_dict={X:data})
while True:
try:
print(sess.run(output))
except tf.errors.OutOfRangeError:
break

I would suggest to use tf.contrib.data.make_saveable_from_iterator, which has been designed precisely for this purpose. It is much less verbose and does not require you to change existing code, in particular how you define your iterator.
Working example, when we save everything after step 5 has completed. Note how I don't even bother knowing what seed is used.
import tensorflow as tf
iterator = (
tf.data.Dataset.range(100)
.shuffle(10)
.make_one_shot_iterator())
batch = iterator.get_next(name='batch')
saveable_obj = tf.contrib.data.make_saveable_from_iterator(iterator)
tf.add_to_collection(tf.GraphKeys.SAVEABLE_OBJECTS, saveable_obj)
saver = tf.train.Saver()
with tf.Session() as sess:
tf.global_variables_initializer().run()
for step in range(10):
print('{}: {}'.format(step, sess.run(batch)))
if step == 5:
saver.save(sess, './foo', global_step=step)
# 0: 1
# 1: 6
# 2: 7
# 3: 3
# 4: 8
# 5: 10
# 6: 12
# 7: 14
# 8: 5
# 9: 17
Then later, if we resume from step 6, we get the same output.
import tensorflow as tf
saver = tf.train.import_meta_graph('./foo-5.meta')
with tf.Session() as sess:
saver.restore(sess, './foo-5')
for step in range(6, 10):
print('{}: {}'.format(step, sess.run('batch:0')))
# 6: 12
# 7: 14
# 8: 5
# 9: 17

I couldn't solve the problem related to initializing the iterator, but since I pre-process my dataset using map method, and I apply transformations defined by Python operations wrapped with py_func, which cannot be serialized for storing\restoring, I'll have to initialize my dataset when I want to restore it anyway.
So, the problem that remains is how to feed data to my graph when I restore it. I placed a tf.identity node between the iterator output and my network input. Upon restoring, I feed my data to the identity node. A better solution that I discovered later is using placeholder_with_default(), as described in this answer.

I would suggest having a look at CheckpointInputPipelineHook CheckpointInputPipelineHook, which implements saving iterator state for further training with tf.Estimator.

TensorFlow restore throwing "No Variable to save" error

I am working through some code to understand how to save and restore checkpoints in tensorflow. To do so, I implemented a simple neural netowork that works with MNIST digits and saved the .ckpt file like so:
from tensorflow.examples.tutorials.mnist import input_data
import numpy as np
learning_rate = 0.001
n_input = 784 # MNIST data input (img shape = 28*28)
n_classes = 10 # MNIST total classes 0-9
#import MNIST data
mnist = input_data.read_data_sets('.', one_hot = True)
#Features and Labels
features = tf.placeholder(tf.float32, [None, n_input])
labels = tf.placeholder(tf.float32, [None, n_classes])
#Weights and biases
weights = tf.Variable(tf.random_normal([n_input, n_classes]))
bias = tf.Variable(tf.random_normal([n_classes]))
#logits = xW + b
logits = tf.add(tf.matmul(features, weights), bias)
#Define loss and optimizer
cost = tf.reduce_mean(\
tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)\
.minimize(cost)
# Calculate accuracy
correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
import math
save_file = './train_model.ckpt'
batch_size = 128
n_epochs = 100
saver = tf.train.Saver()
# Launch the graph
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# Training cycle
for epoch in range(n_epochs):
total_batch = math.ceil(mnist.train.num_examples / batch_size)
# Loop over all batches
for i in range(total_batch):
batch_features, batch_labels = mnist.train.next_batch(batch_size)
sess.run(
optimizer,
feed_dict={features: batch_features, labels: batch_labels})
# Print status for every 10 epochs
if epoch % 10 == 0:
valid_accuracy = sess.run(
accuracy,
feed_dict={
features: mnist.validation.images,
labels: mnist.validation.labels})
print('Epoch {:<3} - Validation Accuracy: {}'.format(
epoch,
valid_accuracy))
# Save the model
saver.save(sess, save_file)
print('Trained Model Saved.')
This part works well, and I get the .ckpt file saved in the correct directory. The problem comes in when I try to restore the model in an attempt to work on it again. I use the following code to restore the model:
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, 'train_model.ckpt.meta')
print('model restored')
and end up with the error: ValueError: No variables to save
Not too sure, what the mistake here is. Any help is appreciated. Thanks in advance

A Graph is different to the Session. A graph is the set of operations joining tensors, each of which is a symbolic representation of a set of values. A Session assigns specific values to the Variable tensors, and allows you to run operations in that graph.
The chkpt file saves variable values - i.e. those saved in the weights and biases - but not the graph itself.
The solution is simple: re-run the graph construction (everything before the Session, then start your session and load values from the chkpt file.
Alternatively, you can check out this guide for exporting and importing MetaGraphs.

You should tell the Saver which Variables to restore, default Saver will get all the Variables from the default graph.
As in your case, you should add the constructing graph code before saver = tf.train.Saver()

duplicate a tensorflow graph

What is the best way of duplicating a TensorFlow graph and keep it uptodate?
Ideally I want to put the duplicated graph on another device (e.g. from GPU to CPU) and then time to time update the copy.

Short answer: You probably want checkpoint files (permalink).
Long answer:
Let's be clear about the setup here. I'll assume that you have two devices, A and B, and you are training on A and running inference on B.
Periodically, you'd like to update the parameters on the device running inference with new parameters found during training on the other.
The tutorial linked above is a good place to start. It shows you how tf.train.Saver objects work, and you shouldn't need anything more complicated here.
Here is an example:
import tensorflow as tf
def build_net(graph, device):
with graph.as_default():
with graph.device(device):
# Input placeholders
inputs = tf.placeholder(tf.float32, [None, 784])
labels = tf.placeholder(tf.float32, [None, 10])
# Initialization
w0 = tf.get_variable('w0', shape=[784,256], initializer=tf.contrib.layers.xavier_initializer())
w1 = tf.get_variable('w1', shape=[256,256], initializer=tf.contrib.layers.xavier_initializer())
w2 = tf.get_variable('w2', shape=[256,10], initializer=tf.contrib.layers.xavier_initializer())
b0 = tf.Variable(tf.zeros([256]))
b1 = tf.Variable(tf.zeros([256]))
b2 = tf.Variable(tf.zeros([10]))
# Inference network
h1 = tf.nn.relu(tf.matmul(inputs, w0)+b0)
h2 = tf.nn.relu(tf.matmul(h1,w1)+b1)
output = tf.nn.softmax(tf.matmul(h2,w2)+b2)
# Training network
cross_entropy = tf.reduce_mean(-tf.reduce_sum(labels * tf.log(output), reduction_indices=[1]))
optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)
# Your checkpoint function
saver = tf.train.Saver()
return tf.initialize_all_variables(), inputs, labels, output, optimizer, saver
The code for the training program:
def programA_main():
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
# Build training network on device A
graphA = tf.Graph()
init, inputs, labels, _, training_net, saver = build_net(graphA, '/cpu:0')
with tf.Session(graph=graphA) as sess:
sess.run(init)
for step in xrange(1,10000):
batch = mnist.train.next_batch(50)
sess.run(training_net, feed_dict={inputs: batch[0], labels: batch[1]})
if step%100==0:
saver.save(sess, '/tmp/graph.checkpoint')
print 'saved checkpoint'
...and code for an inference program:
def programB_main():
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
# Build inference network on device B
graphB = tf.Graph()
init, inputs, _, inference_net, _, saver = build_net(graphB, '/cpu:0')
with tf.Session(graph=graphB) as sess:
batch = mnist.test.next_batch(50)
saver.restore(sess, '/tmp/graph.checkpoint')
print 'loaded checkpoint'
out = sess.run(inference_net, feed_dict={inputs: batch[0]})
print out[0]
import time; time.sleep(2)
saver.restore(sess, '/tmp/graph.checkpoint')
print 'loaded checkpoint'
out = sess.run(inference_net, feed_dict={inputs: batch[0]})
print out[1]
If you fire up the training program and then the inference program, you'll see the inference program produces two different outputs (from the same input batch). This is a result of it picking up the parameters that the training program has checkpointed.
Now, this program obviously isn't your end point. We don't do any real synchronization, and you'll have to decide what "periodic" means with respect to checkpointing. But this should give you an idea of how to sync parameters from one network to another.
One final warning: this does not mean that the two networks are necessarily deterministic. There are known non-deterministic elements in TensorFlow (e.g., this), so be wary if you need exactly the same answer. But this is the hard truth about running on multiple devices.
Good luck!

I'll try to go with a pretty simplified answer, to see if the general approach is what OP is describing:
I'd implement it via the tf.train.Saver object.
Suppose you have your weights in a variable W1, W2, and b1
mysaver = tf.train.Saver(({'w1': W1, 'w2': W2, 'b1': b1}))
In the train loop you can add, every n iterations:
saver.save(session_var, 'model1', global_step=step)
And then in the loading instance, when needed, you run:
tf.train.Saver.restore(other_session_object, 'model1')
Hope this is similar to the solution you are asking.

Simply do the round trip tf.Graph > tf.GraphDef > tf.Graph:
import tensorflow as tf
def copy_graph(graph: tf.Graph) -> tf.Graph:
with tf.Graph().as_default() as copied_graph:
graph_def = graph.as_graph_def(add_shapes=True)
tf.graph_util.import_graph_def(graph_def)
return copied_graph