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!

Slightly edited tensorflow official example doesn't run properly

I'm working on word2vec via distributed tensorflow. For a compatible reason, just slightly edit official word2vec to a Model kinda coding arch.
Code Snippet as follows:
def build():
self.global_step = tf.train.get_or_create_global_step()
with tf.variable_scope("weights", partitioner=partitioner):
self.embeddings = tf.get_variable(name="embeddings", shape=(self.vocab_size, self.embedding_size), initializer=tf.random_uniform_initializer(minval=-1.0, maxval=1.0))
self.nce_weights = tf.get_variable(name="nce_weights", shape=(self.vocab_size, self.embedding_size), initializer=tf.truncated_normal_initializer(stddev=1.0/math.sqrt(self.embedding_size)))
self.bias = tf.get_variable(name="bias", shape=(self.vocab_size), initializer=tf.zeros_initializer())
self.embeded = tf.nn.embedding_lookup(self.embeddings, inputs, partition_strategy='div')
print("lables: ", self.labels)
self.loss = tf.reduce_mean(
tf.nn.nce_loss(
weights = self.nce_weights,
biases = self.bias,
labels = self.labels,
inputs = self.embeded,
num_sampled = self.num_sampled,
num_classes = self.vocab_size,
partition_strategy="div"
)
)
self.optimizer = tf.train.GradientDescentOptimizer(self.learning_rate).minimize(self.loss, global_step=self.global_step)
# evaluate
normized = tf.sqrt(tf.reduce_sum(tf.square(self.embeddings), 1, keepdims=True))
normallized_embeddings = self.embeddings / normized
valid_data = np.r_[1:5]
self.valid_size = len(valid_data)
evaluate_examples = tf.constant(valid_data)
valid_embeddings = tf.nn.embedding_lookup(normallized_embeddings, evaluate_examples)
self.similarity = tf.matmul(valid_embeddings, normallized_embeddings, transpose_b=True)
The train method:
def train(args):
loss, _, global_step, embs = session.run([self.loss, self.optimizer, self.global_step, self.embeddings])
print(embs)
Training:
def main():
model = Word2vec(args)
model.build() # call the method above to build the graph
tf.global_variables_initializer()
with tf.Session() as sess:
while num_step < upperboud:
model.train(sess)
I print out the evaluation results during training, and found no changes all the time, but nce_weights are changing. And global_step and local_step are increasing. Not sure where is wrong, anyone can help point out ? Thanks

How to predict values with a trained Tensorflow model

I've trained my NN in Tensorflow and saved the model like this:
def neural_net(x):
layer_1 = tf.layers.dense(inputs=x, units=195, activation=tf.nn.sigmoid)
out_layer = tf.layers.dense(inputs=layer_1, units=6)
return out_layer
train_x = pd.read_csv("data_x.csv", sep=" ")
train_y = pd.read_csv("data_y.csv", sep=" ")
train_x = train_x / 6 - 0.5
train_size = 0.9
train_cnt = int(floor(train_x.shape[0] * train_size))
x_train = train_x.iloc[0:train_cnt].values
y_train = train_y.iloc[0:train_cnt].values
x_test = train_x.iloc[train_cnt:].values
y_test = train_y.iloc[train_cnt:].values
x = tf.placeholder("float", [None, 386])
y = tf.placeholder("float", [None, 6])
nn_output = neural_net(x)
cost = tf.reduce_mean(tf.losses.mean_squared_error(labels=y, predictions=nn_output))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)
training_epochs = 5000
display_step = 1000
batch_size = 30
keep_prob = tf.placeholder("float")
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(training_epochs):
total_batch = int(len(x_train) / batch_size)
x_batches = np.array_split(x_train, total_batch)
y_batches = np.array_split(y_train, total_batch)
for i in range(total_batch):
batch_x, batch_y = x_batches[i], y_batches[i]
_, c = sess.run([optimizer, cost],
feed_dict={
x: batch_x,
y: batch_y,
keep_prob: 0.8
})
saver.save(sess, 'trained_model', global_step=1000)
Now I want to use the trained model in a different file. Of course there are many many examples of restoring and saving the model, I went through lots of them. Still I couldn't make any of them work, there is always some kind of error. So this is my restore file, could you please help me to make it restore the saved model?
saver = tf.train.import_meta_graph('trained_model-1000.meta')
y_pred = []
with tf.Session() as sess:
saver.restore(sess, tf.train.latest_checkpoint('./'))
sess.run([y_pred], feed_dict={x: input_values})
E.g. this attempt gave me the error "The session graph is empty. Add operations to the graph before calling run()." So what operation should I add to the graph and how? I don't know what that operation should be in my model... I don't understand this whole concept of saving/restoring in Tensorflow. Or should I do the restoring completely differently? Thanks in advance!

Forgive me if I am wrong but tf.train.Saver() only saves the variable values not the graph itself. This means that if you want to load the model in a different file you need to rebuild the graph or somehow load the graph as well. Tensorflow documentation states:
The tf.train.Saver object not only saves variables to checkpoint files, it also restores variables. Note that when you restore variables from a file you do not have to initialize them beforehand.
Consider the following example:
One file that saves the model:
# Create some variables.
v1 = tf.get_variable("v1", shape=[3], initializer = tf.zeros_initializer)
v2 = tf.get_variable("v2", shape=[5], initializer = tf.zeros_initializer)
inc_v1 = v1.assign(v1+1)
dec_v2 = v2.assign(v2-1)
# Add an op to initialize the variables.
init_op = tf.global_variables_initializer()
# Add ops to save and restore all the variables.
saver = tf.train.Saver()
# Later, launch the model, initialize the variables, do some work, and save the
# variables to disk.
with tf.Session() as sess:
sess.run(init_op)
# Do some work with the model.
inc_v1.op.run()
dec_v2.op.run()
# Save the variables to disk.
save_path = saver.save(sess, "/tmp/model.ckpt")
print("Model saved in file: %s" % save_path)
The other file that loads the previously saved model:
tf.reset_default_graph()
# Create some variables.
v1 = tf.get_variable("v1", shape=[3])
v2 = tf.get_variable("v2", shape=[5])
# Add ops to save and restore all the variables.
saver = tf.train.Saver()
# Later, launch the model, use the saver to restore variables from disk, and
# do some work with the model.
with tf.Session() as sess:
# Restore variables from disk.
saver.restore(sess, "/tmp/model.ckpt")
print("Model restored.")
# Check the values of the variables
print("v1 : %s" % v1.eval())
print("v2 : %s" % v2.eval())

output = sess.run(nn_output, feed_dict={ x: batch_x, keep_prob: 0.8 })
Where nn_output is the name is the output variable of the last layer of you network. You can save you variable using:
saver = tf.train.Saver([nn_output])
saver.save(sess, 'my_test_model',global_step=1000) # save every 1000 steps
and therefore in your code:
out_layer = tf.layers.dense(inputs=layer_1, units=6)
should be :
out_layer = tf.layers.dense(inputs=layer_1, units=6, name='nn_output')
To restore:
with tf.Session() as sess:
saver = tf.train.import_meta_graph('my_test_model')
saver.restore(sess,tf.train.latest_checkpoint('./'))
Now you should have access to that node of the graph. If the name is not specified, it is difficult to recover that particular layer.

You can know use tf.saved_model.builder.SavedModelBuilder function.
The main lines for the saving:
builder = tf.saved_model.builder.SavedModelBuilder(graph_location)
builder.add_meta_graph_and_variables(sess, ["cnn_mnist"])
builder.save()
A code to save the model :
...
def main(_):
# Import data
mnist = input_data.read_data_sets(FLAGS.data_dir)
# Create the model
x = tf.placeholder(tf.float32, [None, 784])
# Define loss and optimizer
y_ = tf.placeholder(tf.int64, [None])
# Build the graph for the deep net
y_conv, keep_prob = deepnn(x) # an unknow model model
with tf.name_scope('loss'):
cross_entropy = tf.losses.sparse_softmax_cross_entropy(
labels=y_, logits=y_conv)
cross_entropy = tf.reduce_mean(cross_entropy)
with tf.name_scope('adam_optimizer'):
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
with tf.name_scope('accuracy'):
correct_prediction = tf.equal(tf.argmax(y_conv, 1), y_)
correct_prediction = tf.cast(correct_prediction, tf.float32)
accuracy = tf.reduce_mean(correct_prediction)
graph_location ="tmp/"
print('Saving graph to: %s' % graph_location)
**builder = tf.saved_model.builder.SavedModelBuilder(graph_location)**
train_writer = tf.summary.FileWriter(graph_location)
train_writer.add_graph(tf.get_default_graph())
saver = tf.train.Saver(max_to_keep=1)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
**builder.add_meta_graph_and_variables(sess, ["cnn_mnist"])**
for i in range(20000):
batch = mnist.train.next_batch(50)
if i % 100 == 0:
train_accuracy = accuracy.eval(feed_dict={
x: batch[0], y_: batch[1], keep_prob: 1.0})
print('step %d, training accuracy %g' % (i, train_accuracy))
train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
print('test accuracy %g' % accuracy.eval(feed_dict={
x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))
**builder.save()**
saver.save(sess, "tmp/my_checkpoint.ckpt", global_step =0)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir', type=str,
default='/tmp/tensorflow/mnist/input_data',
help='Directory for storing input data')
FLAGS, unparsed = parser.parse_known_args()
tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)
`
A code to restore the model :
import tensorflow as tf
# récupération des poids
export_dir = "tmp"
sess = tf.Session()
tf.saved_model.loader.load(sess,["cnn_mnist"], export_dir)
#trainable_var = tf.trainable_variables()
trainable_var = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
for var in trainable_var:
print(var.name)`

This question is old. But if someone else is struggling with doing predictions with a trained model (with TF 1.x) this code might help.
Pay attention that
Your network construction/defining code must be executed before the Saver() instance is created. Otherwise you get the error: ValueError: No variables to save. In the code below the LeNet(x) method constructs the network for input placeholder x.
You should not initialize the variables in the session. Because obviously you are loading them from the saved model.
# all the network construction code
# (e.g. defining the variables and layers)
# must be exectured before the creation of
# the Saver() object. Otherwise you get the
# error: ValueError: No variables to save.
logits = LeNet(x)
saver = tf.train.Saver()
index = random.randint(0, len(X_train))
image = X_train[index].squeeze()
label = y_train[index]
print("Label: ", label)
plt.figure(figsize=(1,1))
plt.imshow(image, cmap="gray")
plt.show()
with tf.Session() as sess:
saver.restore(sess, tf.train.latest_checkpoint('./checkpoints/'))
logits_output = sess.run(logits, feed_dict={x: image.reshape((1, 32, 32, 1))})
logits_output = logits_output.squeeze()
pred_output = np.exp(logits_output)/sum(np.exp(logits_output)) #softmax
print("Logits: ", logits_output)
print("Prediction output:", pred_output)
print("Predicted Label: ", np.argmax(pred_output))

tensorflow calculations without retraining

I want to train a model once in tensorflow, and then want to use the trained model for predicting some functions. Before we get into specifics, lets define a couple of functions ...
def runTF(func, inpDict):
sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)
result = sess.run(func, feed_dict = inpDict)
sess.close()
return result
and
def optTF(opt, res, others, inpDict, nSteps, printSteps=50):
os, re = [], []
sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)
for i in range(nSteps):
# First run the optimizer ...
sess.run(opt, feed_dict = inpDict)
# Save all the data you want to save
temp = sess.run( [res] + others, feed_dict = inpDict)
re.append(temp[0])
os.append(temp[1:])
if (i%printSteps) == 0:
print('{:5d}'.format(i))
sess.close()
return re, os
Here are a couple of steps for what I am doing ...
A. Generating some data
N = 500
features = 2
nSteps = 1000
X = np.array([np.random.random(N), np.random.random(N)])
data = [X.T, X[0].reshape(-1, 1)]
B. create a simple linear model
d = tf.placeholder(shape = np.shape(data[0]), dtype = tf.float32, name='d') # input layer
dOut = tf.placeholder(shape = np.shape(data[1]), dtype = tf.float32, name='dOut') # output layer
W = tf.Variable( np.random.randn(features, 1), dtype = tf.float32, name='W')
b = tf.Variable( np.zeros((1,1)), dtype = tf.float32, name='b')
result = tf.matmul(d, W)+b
cost = tf.reduce_mean((dOut - result)**2)
optimizer = tf.train.AdamOptimizer(learning_rate = 0.1).minimize(cost)
C. Lets run the optimiser
inpDict = {d: data[0], dOut:data[1]}
ot = optTF(optimizer, result, [W, cost], inpDict, 200, 50)
Here, I have checked the result, and see that it is what I want. So the optimiser is working fine. The model has been optimised. Now, I want to make a prediction with some other data. So I do ...
r = runTF(result, inpDict)
This new result is not what I would expect from the same trained model.
Now, as long as I stay within the same tf.Session(), we are ok. However, I do want to be able to do predictions even when I an done with a session. So my question is, how do I use a model once we have trained it in one session in a different session?
Note, the entire thing is something that I did in a different session?
Edited:
I edited the two functions to incorporate the saving ...
def runTF(func, inpDict, modelFile=None):
if modelFile is not None:
saver = tf.train.Saver()
sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)
if modelFile is not None:
ckpt = tf.train.get_checkpoint_state(modelFile)
print(modelFile, ckpt)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print('Session restored')
result = sess.run(func, feed_dict = inpDict)
sess.close()
return result
and
def optTF(opt, res, others, inpDict, nSteps, printSteps=50, modelFile='models/temp.ckpt'):
os, re = [], []
saver = tf.train.Saver()
sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)
for i in range(nSteps):
# First run the optimizer ...
sess.run(opt, feed_dict = inpDict)
# Save all the data you want to save
temp = sess.run( [res] + others, feed_dict = inpDict)
re.append(temp[0])
os.append(temp[1:])
if (i%printSteps) == 0:
print('{:5d}'.format(i))
path = saver.save(sess, modelFile)
print('Model saved in: {}'.format(path))
sess.close()
return re, os
And running the model as:
ot = optTF(optimizer, result, [cost]+weights+biases, inpDict, 200, 50)
r = runTF([result], inpDict, 'models/temp.ckpt')
Still nothing! I checked that:
The value of ckpt is None
The models folder has the following files:
checkpoint
temp.ckpt.index
temp.ckpt.data-00000-of-00001
temp.ckpt.meta

You need to save and restore the session you are creating and training. As in
init = tf.initialize_all_variables()
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(init)
if restore:
ckpt = tf.train.get_checkpoint_state(save_path)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
else:
# ... training code omitted ...
saver.save(sess, save_path)
checkout also https://www.tensorflow.org/programmers_guide/saved_model if you want to create a full model with estimator instead of just one session.

Tensorflow restored weights is not set

I'm trying to restore a model in Tensorflow which I've trained. The problem is that it does not seem like the weights are properly restored.
For the training I've got the weights and biases defined as:
W = {
'h1': tf.Variable(tf.random_normal([n_inputs, n_hidden_1]), name='wh1'),
'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2]), name='wh2'),
'o': tf.Variable(tf.random_normal([n_hidden_2, n_classes]), name='wo')
}
b = {
'b1': tf.Variable(tf.random_normal([n_hidden_1]), name='bh1'),
'b2': tf.Variable(tf.random_normal([n_hidden_2]), name='bh2'),
'o': tf.Variable(tf.random_normal([n_classes]), name='bo')
}
Then I do some training on my own custom 2D image dataset and save the model by calling the tf.saver
saver = tf.train.Saver()
saver.save(sess, 'tf.model')
Later I want to restore that model with the exact same weights, so I build the model as before (also with the random_normal initialization) and call the tf.saver.restore
saver = tf.train.import_meta_graph('tf.model.meta')
saver.restore(sess, tf.train.latest_checkpoint('./'))
Now, if i call:
temp = sess.run(W['h1'][0][0])
print temp
I get random values, and not the restored value of the weight.
I've drawn a blank on this one, can somebody point me in the right direction?
FYI, I've tried (without) luck to simply declare the tf.Variables, but I keep getting:
ValueError: initial_value must be specified.
even though Tensorflow themselves state that it should be possible to simply declare with no initial value (https://www.tensorflow.org/programmers_guide/variables part: Restoring Values)
Update 1
When I, as suggested, run
all_vars = tf.global_variables()
for v in all_vars:
print v.name
I get the following output:
wh1:0
wh2:0
wo:0
bh1:0
bh2:0
bo:0
wh1:0
wh2:0
wo:0
bh1:0
bh2:0
bo:0
beta1_power:0
beta2_power:0
wh1/Adam:0
wh1/Adam_1:0
wh2/Adam:0
wh2/Adam_1:0
wo/Adam:0
wo/Adam_1:0
bh1/Adam:0
bh1/Adam_1:0
bh2/Adam:0
bh2/Adam_1:0
bo/Adam:0
bo/Adam_1:0
Which shows that the variables indeed is read. However invoking
print sess.run("wh1:0")
Results in the error: Attempting to use uninitialized value wh1

So with the help of you guys, I ended up dividing the saving and restoring parts of my program into two files, to ensure that no unwanted variables were initialized.
Train and Save routines fnn.py
def build(self, topology):
"""
Builds the topology of the model
"""
# Sanity check
assert len(topology) == 4
n_inputs = topology[0]
n_hidden_1 = topology[1]
n_hidden_2 = topology[2]
n_classes = topology[3]
# Sanity check
assert self.img_h * self.img_w == n_inputs
# Instantiate TF Placeholders
self.x = tf.placeholder(tf.float32, [None, n_inputs], name='x')
self.y = tf.placeholder(tf.float32, [None, n_classes], name='y')
self.W = {
'h1': tf.Variable(tf.random_normal([n_inputs, n_hidden_1]), name='wh1'),
'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2]), name='wh2'),
'o': tf.Variable(tf.random_normal([n_hidden_2, n_classes]), name='wo')
}
self.b = {
'b1': tf.Variable(tf.random_normal([n_hidden_1]), name='bh1'),
'b2': tf.Variable(tf.random_normal([n_hidden_2]), name='bh2'),
'o': tf.Variable(tf.random_normal([n_classes]), name='bo')
}
# Create model
self.l1 = tf.nn.sigmoid(tf.add(tf.matmul(self.x, self.W['h1']), self.b['b1']))
self.l2 = tf.nn.sigmoid(tf.add(tf.matmul(self.l1, self.W['h2']), self.b['b2']))
logits = tf.add(tf.matmul(self.l2, self.W['o']), self.b['o'])
# Define predict operation
self.predict_op = tf.argmax(logits, 1)
probs = tf.nn.softmax(logits, name='probs')
# Define cost function
self.cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits, self.y))
# Adding these to collection so we can restore them again
tf.add_to_collection('inputs', self.x)
tf.add_to_collection('inputs', self.y)
tf.add_to_collection('outputs', logits)
tf.add_to_collection('outputs', probs)
tf.add_to_collection('outputs', self.predict_op)
def train(self, X, Y, n_epochs=10, learning_rate=0.001, logs_path=None):
"""
Trains the Model
"""
self.optimizer = tf.train.AdamOptimizer(learning_rate).minimize(self.cost)
costs = []
# Instantiate TF Saver
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
# start the threads used for reading files
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# Compute total number of batches
total_batch = int(self.get_num_examples() / self.batch_size)
# start training
for epoch in range(n_epochs):
for i in range(total_batch):
batch_xs, batch_ys = sess.run([X, Y])
# run the training step with feed of images
_, cost = sess.run([self.optimizer, self.cost], feed_dict={self.x: batch_xs,
self.y: batch_ys})
costs.append(cost)
print "step %d" % (epoch * total_batch + i)
#costs.append(cost)
print "Epoch %d" % epoch
saver.save(sess, self.model_file)
temp = sess.run(self.W['h1'][0][0])
print temp
if self.visu:
plt.plot(costs)
plt.show()
# finalize
coord.request_stop()
coord.join(threads)
Predict routine fnn_eval.py:
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
g = tf.get_default_graph()
# restore the model
self.saver = tf.train.import_meta_graph(self.model_file)
self.saver.restore(sess, tf.train.latest_checkpoint('./tfmodels/fnn/'))
wh1 = g.get_tensor_by_name("wh1:0")
print sess.run(wh1[0][0])
x, y = tf.get_collection('inputs')
logits, probs, predict_op = tf.get_collection('outputs')
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
predictions = []
print Y.eval()
for i in range(1):#range(self.get_num_examples()):
batch_xs = sess.run(X)
# Reshape batch_xs if only a single image is given
# (numpy is 4D: batch_size * heigth * width * channels)
batch_xs = np.reshape(batch_xs, (-1, self.img_w * self.img_h))
prediction, probabilities, logit = sess.run([predict_op, probs, logits], feed_dict={x: batch_xs})
predictions.append(prediction[0])
# finalize
coord.request_stop()
coord.join(threads)

I guess the problem might be caused by creating a new variable when you restore the model, not getting the already existed variable. I tried this code
saver = tf.train.import_meta_graph('./model.ckpt-10.meta')
w1 = None
for v in tf.global_variables():
print v.name
w1 = tf.get_variable('wh1', [])
init = tf.global_variables_initializer()
sess.run(init)
saver.restore(sess, './model.ckpt-10')
for v in tf.global_variables():
print v.name
and clearly you can see the output that it creates a new variable called wh1_1:0.
If you try this
w1 = None
for v in tf.global_variables():
print v.name
if v.name == 'wh1:0':
w1 = v
init = [tf.global_variables_initializer(), tf.local_variables_initializer()]
sess.run(init)
saver.restore(sess, './model.ckpt-10')
for v in tf.global_variables():
print v.name
temp = sess.run(w1)
print temp[0][0]
There will be no problem.
Tensorflow suggests that it is better to use tf.variable_scope() (link) like this
with tf.variable_scope("foo"):
v = tf.get_variable("v", [1])
with tf.variable_scope("foo", reuse=True):
v1 = tf.get_variable("v", [1])
assert v1 == v

I have meet the same problem when saving model to saved_model format. Anyone using the function add_meta_graph_and_variables to save the model for serving, be careful about this parameter "legacy_init_op: Legacy support for op or group of ops to execute after the restore op upon a load."

You want to pass in a var_list to the Saver.
In your case, the variable list would come from your W and b dictionaries: var_list = list(W.values())+list(b.values()). Then, to restore the model, pass in var_list to the Saver: saver = tf.train.Saver(var_list=var_list).
Next, you need to get your checkpoint state: model = tf.train.get_checkpoint_state(<your saved model directory>). After that you can restore the trained weights.
var_list = list(W.values())+list(b.values())
saver = tf.train.Saver(var_list=var_list)
model = tf.get_checkpoint_state('./model/')
with tf.Session() as sess:
saver.restore(sess,model.model_checkpoint_path)
#Now use the pretrained weights