I'm loading a pretrained network into Tensorflow using the methods below that are within a Network class (hence the calls to self.xyz). First, define_network() is called, then I do initialization of other variables and optimizers, then load_model() is called.
However, despite using tf.variable_scope(self.name) the variables from the graph are loaded into the generic space of variables. This is problematic as I have two instances of this class that each load the same network and I want to separate the out into different scopes.
How can I load the variables into a specific scope?
P.S. Feel free to correct me on any errors in my code!
def load_model(self):
with tf.variable_scope(self.name) as scope:
self.saver.restore(self.sess, self.model_path)
print("Loaded model from {}".format(self.model_path))
def define_model(self):
with tf.variable_scope(self.name) as scope:
self.saver = tf.train.import_meta_graph(self.model_path + '.meta')
print("Loaded model from {}".format(self.model_path + '.meta'))
graph = tf.get_default_graph()
self.inputs = []
inp_names = ['i_hand1:0', 'i_hand2:0', 'i_flop1:0', 'i_flop2:0', 'i_flop3:0',
'i_turn:0', 'i_river:0', 'i_other:0', 'i_allowed_mod:0', 'keras_learning_phase:0']
for inp in inp_names:
self.inputs.append(tf.get_default_graph().get_tensor_by_name(inp))
self.outputs = tf.get_default_graph().get_tensor_by_name("Tanh:0")
self.add_output_conversions()
all_vars = tf.trainable_variables()
for var in all_vars:
self.var[var.name] = var
I think your problem can be solved by adding an argument into
self.saver = tf.train.import_meta_graph(self.model_path + '.meta', 'import_scope'=self.name)
Here's the reference
Related
I wanted to share my findings on how to export a tf model for serving directly from session without creating model checkpoint. my use case requires minimum time to create a pb file, therefore I wanted to get a model.pb file directly from session without creating model checkpoint.
most examples online (and documentation refers to the common case of creating a model checkpoint and loading it in order to create a tf-serving (pb) file. of course this use case is good in case export performance time is not an issue.
import tensorflow as tf
from tensorflow.python.framework import importer
output_path = '/export_directory' # be sure to create it before export
input_ops = ['name/s_of_model_input/s']
output_ops = ['name/s_of_model_output/s']
session = tf.compat.v1.Session()
def get_ops_dict(ops, graph, name='op_'):
out_dict = dict()
for i, op in enumerate(ops):
out_dict[name + str(i)] = tf.compat.v1.saved_model.build_tensor_info(graph.get_tensor_by_name(op + ':0'))
return out_dict
def add_meta_graph(pbtxt_tmp_path, graph_def):
with tf.Graph().as_default() as graph:
importer.import_graph_def(graph_def, name="")
os.unlink(pbtxt_tmp_path)
# used to rename model input/outputs
inputs_dict = get_ops_dict(input_ops, graph, name='input_')
outputs_dict = get_ops_dict(output_ops, graph, name='output_')
prediction_signature = (
tf.compat.v1.saved_model.signature_def_utils.build_signature_def(
inputs=inputs_dict,
outputs=outputs_dict,
method_name=tf.saved_model.PREDICT_METHOD_NAME))
legacy_init_op = tf.group(tf.compat.v1.tables_initializer(), name='legacy_init_op')
builder = tf.compat.v1.saved_model.builder.SavedModelBuilder(output_path+'/export')
builder.add_meta_graph_and_variables(
session,
tags=[tf.saved_model.SERVING],
signature_def_map={
tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY: prediction_signature},
legacy_init_op=legacy_init_op)
builder.save()
return prediction_signature
def export_model(session, output_path, output_ops):
graph_def = session.graph_def
tf.io.write_graph(graph_or_graph_def=graph_def, logdir=output_path,
name='model.pbtxt', as_text=False)
frozen_graph_def = tf.compat.v1.graph_util.convert_variables_to_constants(
session, graph_def, output_ops)
prediction_signature = add_meta_graph(output_path+'/model.pbtxt', frozen_graph_def)
Hi have a Python script where I instantiate two objects of a neural network class.
Each object defines its own session and provide methods for saving the graph.
import tensorflow as tf
import os, shutil
class TestNetwork:
def __init__(self, id):
self.id = id
tf.reset_default_graph()
self.s = tf.placeholder(tf.float32, [None, 2], name='s')
w_initializer, b_initializer = tf.random_normal_initializer(0., 1.0), tf.constant_initializer(0.1)
self.k = tf.layers.dense(self.s, 2, kernel_initializer=w_initializer,
bias_initializer=b_initializer, name= 'k')
'''Defines self.session and initialize the variables'''
session_conf = tf.ConfigProto(
allow_soft_placement = True,
log_device_placement = False)
self.session = tf.Session(config = session_conf)
self.session.run(tf.global_variables_initializer())
def save_model(self, output_dir):
'''Save the network graph and weights to disk'''
if os.path.exists(output_dir):
# if provided output_dir already exists, remove it
shutil.rmtree(output_dir)
builder = tf.saved_model.builder.SavedModelBuilder(output_dir)
builder.add_meta_graph_and_variables(
self.session,
[tf.saved_model.tag_constants.SERVING],
clear_devices=True)
# create a new directory output_dir and store the saved model in it
builder.save()
t1 = TestNetwork(1)
t2 = TestNetwork(2)
t1.save_model("t1_model")
t2.save_model("t2_model")
The error I get is
TypeError: Cannot interpret feed_dict key as Tensor: The name
'save/Const:0' refers to a Tensor which does not exist. The operation,
'save/Const', does not exist in the graph.
I read something saying that this error is due to tf.train.Saver.
Thus I added the following line at the end of the __init__ method:
self.saver = tf.train.Saver(tf.global_variables(), max_to_keep = 5)
However I still get the error.
tf.reset_default_graph will clear the default graph stack and resets the global default graph.
NOTE: The default graph is a property of the current thread. This
function applies only to the current thread. Calling this function
while a tf.Session or tf.InteractiveSession is active will result in
undefined behavior. Using any previously created tf.Operation or
tf.Tensor objects after calling this function will result in undefined
behavior.
You should specify Graph separately, and define all of these in the corresponding graph scope.
def __init__(self, id):
self.id = id
self.graph = tf.Graph()
with self.graph.as_default():
self.s = tf.placeholder(tf.float32, [None, 2], name='s')
w_initializer, b_initializer = tf.random_normal_initializer(0., 1.0), tf.constant_initializer(0.1)
self.k = tf.layers.dense(self.s, 2, kernel_initializer=w_initializer,
bias_initializer=b_initializer, name= 'k')
init = tf.global_variables_initializer()
'''Defines self.session and initialize the variables'''
session_conf = tf.ConfigProto(
allow_soft_placement = True,
log_device_placement = False)
self.session = tf.Session(config = session_conf,graph=self.graph)
self.session.run(init)
tf.train.Saver is another way to save model variables.
Edit
If you get empty "variable", you should save model in graph:
def save_model(self, output_dir):
'''Save the network graph and weights to disk'''
if os.path.exists(output_dir):
# if provided output_dir already exists, remove it
shutil.rmtree(output_dir)
with self.graph.as_default():
builder = tf.saved_model.builder.SavedModelBuilder(output_dir)
builder.add_meta_graph_and_variables(
self.session,
[tf.saved_model.tag_constants.SERVING],
clear_devices=True)
# create a new directory output_dir and store the saved model in it
builder.save()
I have very simple model which consists of one tf.Variable() and here is who code:
import tensorflow as tf
save_path="model1/model1.ckpt"
num_input = 2
n_nodes_hl1 = 2
with tf.variable_scope("model1"):
hidden_1_layer = {
'weights' : tf.Variable(tf.random_normal([num_input, n_nodes_hl1]), name='Weight1')
}
def train_model():
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
save_model(sess)
def save_model(sess):
saver = tf.train.Saver(tf.global_variables(), save_path)
saver.save(sess, save_path)
def load_model(sess):
saver = tf.train.Saver(tf.global_variables(), save_path)
saver.restore(sess, save_path)
def run_model():
print("model1 running...")
with tf.Session() as sess:
load_model(sess)
x = sess.run(hidden_1_layer)
print(x)
#train_model()
The second model is completely the same, but with changed names "model1" to "model2". Both models are trained, saved and work perfect separately. So now I want to test them using following script:
import model1 as m1
import model2 as m2
m1.run_model()
m2.run_model()
And here I got an error message:
NotFoundError (see above for traceback): Key model2/Weight2 not found in checkpoint
So it looks like running imports causes adding all variables to common graph (even though they are in separate variable scopes) and then it cannot find variable from model2 saved in checkpoint in model1.
Can anyone solve my problem?
Is it possible in Tensorflow to run a few different models in one script?
EDIT - PROBLEM SOLVED
The solution is very easy. What you have to do is to create separate graphs for each model like. It means that all tensors you declare or calculate must be within that graph. You also must put it as an argument in Session, like: tf.Session(graph=self.graph)
Whole example below:
import tensorflow as tf
save_path="model1/model1.ckpt"
class model1:
num_input = 2
n_nodes_hl1 = 2
def init(self):
self.graph = tf.Graph()
with self.graph.as_default():
with tf.variable_scope("model1"):
self.hidden_1_layer = {
'weights' : tf.Variable(tf.random_normal([self.num_input, self.n_nodes_hl1]), name='Weight1')
}
def train_model(self):
init = tf.global_variables_initializer()
with tf.Session(graph = self.graph) as sess:
sess.run(init)
self.save_model(sess)
def save_model(self, sess):
saver = tf.train.Saver(tf.global_variables(), save_path)
saver.save(sess, save_path)
def load_model(self, sess):
saver = tf.train.Saver(tf.global_variables(), save_path)
saver.restore(sess, save_path)
def run_model(self):
print("model1 running...")
with tf.Session(graph = self.graph) as sess:
self.load_model(sess)
x = sess.run(self.hidden_1_layer)
print(x)
Oh! the common "I want to use several models" question! just make sure that you reset the graph after each model:
tf.reset_default_graph()
Your code would look like:
import tensorflow as tf
import model1 as m1
m1.run_model()
tf.reset_default_graph()
import model2 as m2
m2.run_model()
Why? The moment you create a variable in tensorflow using tf.Variable, that variable is added to the default graph. If you import both models one after the other, you just created all the variables in the default graph! This is by far the easiest solution. Consider the default graph as a blackboard: you can draw your fancy ML model, but you need to wipe it clean before reuse!
NOTE: If you are wondering, the alternative is to create separate graphs for each of the models, but it is much more worrysome and I only recommend it for times when you must have both models at the same time.
EXTRA: Encapsulating your model in a Tensorflow class
A fancier way to do it while avoiding several graphs (seriously, it is horrible!) is to encapsulate the whole model in a class. Thus, your code would look like this:
import tensorflow as tf
class model():
self.num_input = 2
self.n_nodes_hl1 = 2
def init(self, new_save_path)
self.save_path=new_save_path
tf.reset_default_graph()
with tf.variable_scope("model1"):
self.hidden_1_layer = {
'weights' : tf.Variable(tf.random_normal([self.num_input,
self.n_nodes_hl1]), name='Weight1')
}
self.saver = tf.train.Saver(tf.global_variables(), self.save_path)
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
def save_model(self):
self.saver.save(self.sess, self.save_path)
def load_model(self):
self.saver.restore(self.sess, self.save_path)
def run_model(self):
print("model1 running...")
load_model()
x = sess.run(self.hidden_1_layer)
print(x)
#train_model(self)
This way you could simply do:
import model
m1 = model('model1/model1.ckpt') # These two lines could be put into one
m1.run_model() # m1 = model('model1/model1.ckpt').run_model()
m2 = model('model2/model2.ckpt')
m2.run_model()
You still want it in a for loop?
import model
model_file_list = ['model1/model1.ckpt', 'model2/model2.ckpt']
for model_file in model_list:
m = model(model_file ).run_model()
# Run tests, print stuff, save stuff here!
I'm trying to fit multiple small Keras models in parallel on a single GPU. Because of reasons i need to get them out of a list and train them one step at a time. Since I was not lucky with the standard multiprocessing module i use pathos.
What I tried to do is something like this:
from pathos.multiprocessing import ProcessPool as Pool
import tensorflow as tf
import keras.backend as K
def multiprocess_step(self, model):
K.set_session(sess)
with sess.graph.as_default():
model = step(model, sess)
return model
def step(model, sess):
K.set_session(sess)
with sess.graph.as_default():
model.fit(x=data['X_train'], y=data['y_train'],
batch_size=batch_size
validation_data=(data['X_test'], data['y_test']),
verbose=verbose,
shuffle=True,
initial_epoch=self.step_num - 1)
return model
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = "0"
sess = tf.Session(config=config)
K.set_session(sess)
with sess.graph.as_default():
pool = Pool(8).map
model_list = pool(multiprocess_step, model_list)
but whatever I try I keep getting an error claiming that the models dont seem to be on the same graph...
ValueError: Tensor("training/RMSprop/Variable:0", shape=(25, 352), dtype=float32_ref) must be from the same graph as Tensor("RMSprop/rho/read:0", shape=(), dtype=float32).
The exception originates in the model.fit() row so I must have done something wrong with the assignment of the session graph even though I tried to set that in every possible location?
Does anyone have experience with something similar?
The following was suggested on the Keras issue tracker. I'm not sure about the relative merits of the approach compared to using multiprocessing.
in_1 = Input()
lstm_1 = LSTM(...)(in_1)
out_1 = Dense(...)(lstm_1)
in_2 = Input()
lstm_2 = LSTM(...)(in_2)
out_2 = Dense(...)(lstm_2)
model_1 = Model(input=in_1, output=out_1)
model_2 = Model(input=in_2, output=out_2)
model = Model(input = [in_1, in_2], output = [out_1, out_2])
model.compile(...)
model.fit(...)
model_1.predict(...)
model_2.predict(...)
Considering the backend is set to tensorflow for the keras. you can use code and do parallel processing for multiple model invocation/ multiple model loading.
def model1(dir_model):
model = os.path.join(dir_model, 'model.json')
dir_weights = os.path.join(dir_model, 'model.h5')
graph1 = Graph()
with graph1.as_default():
session1 = Session(graph=graph1, config=config)
with session1.as_default():
with open(model, 'r') as data:
model_json = data.read()
model_1 = model_from_json(model_json)
model_1.load_weights(dir_weights)
return model_1,gap_weights,session1,graph1
def model_2(dir_model):
model = os.path.join(dir_model, 'model.json')
dir_weights = os.path.join(dir_model, 'model.h5')
graph2 = Graph()
with graph2.as_default():
session2 = Session(graph=graph2, config=config)
with session2.as_default():
with open(model, 'r') as data:
model_json = data.read()
model_2 = model_from_json(model_json)
model_2.load_weights(dir_weights)
return model_2,session2,graph2
and for invocation of the specific model do the following experiments.
for model 1 predict do the following
K.set_session(session2)
with graph2.as_default():
img_pred[img_name] =
patch_dict[np.argmax(np.squeeze(model_2.predict(img_invoke)))
and for the model 2 it follows same as
K.set_session(session2)
with graph2.as_default():
img_pred[img_name] =
patch_dict[np.argmax(np.squeeze(model_2.predict(img_invoke)))]
This article illustrates how to add Runtime statistics to Tensorboard:
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, _ = sess.run([merged, train_step],
feed_dict=feed_dict(True),
options=run_options,
run_metadata=run_metadata)
train_writer.add_run_metadata(run_metadata, 'step%d' % i)
train_writer.add_summary(summary, i)
print('Adding run metadata for', i)
which creates the following details in Tensorboard:
This is fairly straightforward on a single machine. How could one do this in a distributed environment using Estimators?
I use the following hook, based on ProfilerHook, to have the estimator output the run metadata into the model directory and inspect it later with Tensorboard.
import tensorflow as tf
from tensorflow.python.training.session_run_hook import SessionRunHook, SessionRunArgs
from tensorflow.python.training import training_util
from tensorflow.python.training.basic_session_run_hooks import SecondOrStepTimer
class MetadataHook(SessionRunHook):
def __init__ (self,
save_steps=None,
save_secs=None,
output_dir=""):
self._output_tag = "step-{}"
self._output_dir = output_dir
self._timer = SecondOrStepTimer(
every_secs=save_secs, every_steps=save_steps)
def begin(self):
self._next_step = None
self._global_step_tensor = training_util.get_global_step()
self._writer = tf.summary.FileWriter (self._output_dir, tf.get_default_graph())
if self._global_step_tensor is None:
raise RuntimeError("Global step should be created to use ProfilerHook.")
def before_run(self, run_context):
self._request_summary = (
self._next_step is None or
self._timer.should_trigger_for_step(self._next_step)
)
requests = {"global_step": self._global_step_tensor}
opts = (tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
if self._request_summary else None)
return SessionRunArgs(requests, options=opts)
def after_run(self, run_context, run_values):
stale_global_step = run_values.results["global_step"]
global_step = stale_global_step + 1
if self._request_summary:
global_step = run_context.session.run(self._global_step_tensor)
self._writer.add_run_metadata(
run_values.run_metadata, self._output_tag.format(global_step))
self._writer.flush()
self._next_step = global_step + 1
def end(self, session):
self._writer.close()
To use it, one creates the estimator instance (my_estimator) as usual, whether it is pre-made one or a custom estimator. The desired operation is called passing an instance of the class above as a hook. For example:
hook = MetadataHook(save_steps=1, output_dir=<model dir>)
my_estimator.train( train_input_fn, hooks=[hook] )
The run metadata will be placed in the model dir and can be inspected by TensorBoard.
You may use tf.train.ProfilerHook. However the catch is that it was released at 1.14.
Example usage:
estimator = tf.estimator.LinearClassifier(...)
hooks = [tf.train.ProfilerHook(output_dir=model_dir, save_secs=600, show_memory=False)]
estimator.train(input_fn=train_input_fn, hooks=hooks)
Executing the hook will generate files timeline-xx.json in output_dir.
Then open chrome://tracing/ in chrome browser and load the file. You will get a time usage timeline like below.