We Keep Coding

Can't export to SavedModel in TensorFlow

I need to load a trained model in Tensorflow and save it to SavedModel format for use in Tensorflow Serving. I am following this tutorial. I am loading the model like this:
self.graph = tf.Graph()
def _graph_hourglass(self, inputs):
### All layers are here ###
...
def generate_model(self):
with tf.device(self.cpu):
with tf.name_scope('inputs'):
self.img = tf.compat.v1.placeholder(dtype=tf.float32,
shape=(None, 256, 256, 3), name='input_img')
self.output = self._graph_hourglass(self.img)
def model_init(self):
with self.graph.as_default():
self.generate_model()
def restore(self, load='path'):
with tf.name_scope('Session'):
with tf.device(self.cpu):
self.Session = tf.compat.v1.Session()
saver = tf.compat.v1.train.Saver()
saver.restore(self.Session, load)
def load_model(self, load = 'path'):
with self.graph.as_default():
self.restore(load)
...
self.model_init()
self.load_model(load='path/to/model')
Then I use SavedModelBuilder class from the tutorial:
export_path = 'new/path'
print('Exporting trained model to', export_path)
builder = tf.compat.v1.saved_model.builder.SavedModelBuilder(export_path)
tensor_info_x = tf.compat.v1.saved_model.utils.build_tensor_info(self.img)
tensor_info_y = tf.compat.v1.saved_model.utils.build_tensor_info(self.output)
prediction_signature = (
tf.compat.v1.saved_model.signature_def_utils.build_signature_def(
inputs={'image': tensor_info_x},
outputs={'joints': tensor_info_y},
method_name=tf.compat.v1.saved_model.signature_constants.PREDICT_METHOD_NAME))
builder.add_meta_graph_and_variables(
self.Session, [tf.compat.v1.saved_model.tag_constants.SERVING],
signature_def_map={
tf.compat.v1.saved_model.signature_constants
.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
prediction_signature,
},
strip_default_attrs=True)
builder.save()
print('Done exporting!')
Model is saved without errors, I see "Done exporting!" in console but model .pb file is empty! It is only 252 Bytes. What can be the problem here? Will appreciate any help.

How to save to saved_model.pb from tf.Session?

I am working with the ssd_mobilenet_v2_coco_2018_03_29 pretrained Tensorflow model. I want to change the input to fixed size, and save it under the saved_model.pb (I am using Neuron Compiler which require this format).
Here is how I change the input Tensor to fixed size:
graph = tf.Graph()
with graph.as_default():
fixed_image_tensor = tf.placeholder(tf.uint8, shape=(None, 300, 300, 3), name='image_tensor')
graph_def = tf.GraphDef()
with tf.io.gfile.GFile(frozen_pb_file, 'rb') as f:
serialized_graph = f.read()
graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(graph_def, name='', input_map={"image_tensor:0": fixed_image_tensor})
And now I save the modified graph to saved_model.pb format by using tf.saved_model.simple_save:
image_tensor = graph.get_tensor_by_name('image_tensor:0')
boxes_tensor = graph.get_tensor_by_name('detection_boxes:0')
scores_tensor = graph.get_tensor_by_name('detection_scores:0')
classes_tensor = graph.get_tensor_by_name('detection_classes:0')
num_detections_tensor = graph.get_tensor_by_name('num_detections:0')
sess = tf.Session(graph=graph)
tf.saved_model.simple_save(
session=sess,
export_dir='model/',
inputs={image_tensor.name: image_tensor},
outputs={
boxes_tensor.name: boxes_tensor,
scores_tensor.name: scores_tensor,
classes_tensor.name: classes_tensor,
num_detections_tensor.name: num_detections_tensor
}
)
The code create the following directory (variables is empty):
|-model/
|---variables/
|---saved_model.pb
The saved_model.pb is only 370 bytes, and must contains no actual information. I also try tf.saved_model.Builder like this and this, but still got the exact same result.
I can still use the sess for inference as usual with no problems. What did I do wrong? Are there any other approaches? I am using Tensorflow 1.15.0.

A bit rearranged code, TF1.13, got 67MBytes *.pb file. Reloaded generated saved_file, input has yours dimensions and all listed outputs:
import tensorflow as tf
frozen_pb_file = "./ssd_mobilenet_v2_coco_2018_03_29/frozen_inference_graph.pb"
graph = tf.Graph()
with graph.as_default():
fixed_image_tensor = tf.placeholder(tf.uint8, shape=(None, 300, 300, 3), name='image_tensor')
graph_def = tf.GraphDef()
with tf.io.gfile.GFile(frozen_pb_file, 'rb') as f:
serialized_graph = f.read()
graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(graph_def, name='', input_map={"image_tensor:0": fixed_image_tensor})
image_tensor = graph.get_tensor_by_name('image_tensor:0')
boxes_tensor = graph.get_tensor_by_name('detection_boxes:0')
scores_tensor = graph.get_tensor_by_name('detection_scores:0')
classes_tensor = graph.get_tensor_by_name('detection_classes:0')
num_detections_tensor = graph.get_tensor_by_name('num_detections:0')
sess = tf.Session(graph=graph)
file_writer = tf.summary.FileWriter(logdir='log', graph=graph)
tf.saved_model.simple_save(
session=sess,
export_dir='model/',
inputs={image_tensor.name: fixed_image_tensor},
outputs={
boxes_tensor.name: boxes_tensor,
scores_tensor.name: scores_tensor,
classes_tensor.name: classes_tensor,
num_detections_tensor.name: num_detections_tensor
}
)

Loading TF Records into Keras

I am trying to load a custom TFRecord file into my keras model. I attempted to follow this tutorial: https://medium.com/#moritzkrger/speeding-up-keras-with-tfrecord-datasets-5464f9836c36, but adapting for my use.
My goal is to have the functions work similar to ImageDataGenerator from Keras. I cannot use that function because I specific metadata from the images that the generator does not grab. I'm not including that metadata here because I just need the basic network to function first.
I also want to be able to apply this to a transfer learning application.
I keep getting this error: TypeError: Could not build a TypeSpec for None with type NoneType
I am using Tensorflow 2.2
def _parse_function(serialized):
features = \
{
'image': tf.io.FixedLenFeature([], tf.string),
'label': tf.io.FixedLenFeature([], tf.int64),
'shapex': tf.io.FixedLenFeature([], tf.int64),
'shapey': tf.io.FixedLenFeature([], tf.int64),
}
parsed_example = tf.io.parse_single_example(serialized=serialized,
features=features)
shapex = tf.cast(parsed_example['shapex'], tf.int32)
shapey = tf.cast(parsed_example['shapey'], tf.int32)
image_shape = tf.stack([shapex, shapey, 3])
image_raw = parsed_example['image']
# Decode the raw bytes so it becomes a tensor with type.
image = tf.io.decode_raw(image_raw, tf.uint8)
image = tf.reshape(image, image_shape)
# Get labels
label = tf.cast(parsed_example['label'], tf.float32)
return image, label
def imgs_inputs(type, perform_shuffle=False):
records_dir = '/path/to/tfrecord/'
record_paths = [os.path.join(records_dir,record_name) for record_name in os.listdir(records_dir)]
full_dataset = tf.data.TFRecordDataset(filenames=record_paths)
full_dataset = full_dataset.map(_parse_function, num_parallel_calls=16)
dataset_length = (len(list(full_dataset))) #Gets length of datase
iterator = tf.compat.v1.data.make_one_shot_iterator(databatch)
image, label = iterator.get_next()
#labels saved as values ex: [1,2,3], and are now converted to one hot encoded
label = to_categorical(label)
return image, label
image, label = imgs_inputs(type ='Train',perform_shuffle=True)
#Combine it with keras
# base_model = MobileNet(weights='imagenet', include_top=False, input_shape=(200,200,3), dropout=.3)
model_input = Input(shape=[200,200,3])
#Build your network
model_output = Flatten(input_shape=(200, 200, 3))(model_input)
model_output = Dense(19, activation='relu')(model_output)
#Create your model
train_model = Model(inputs=model_input, outputs=model_output)
#Compile your model
optimizer = Adam(learning_rate=.001)
train_model.compile(optimizer=optimizer,loss='mean_squared_error',metrics=['accuracy'],target_tensors=[label])
#Train the model
train_model.fit(epochs=10,steps_per_epoch=2)
image returns array of shape (100,200,200,3) which is a batch of 100 images
label returns array of shape(100,19) which is a batch of 100 labels (there are 19 labels)

The issue related to shapex and shapey but I don't know exactly why.
I set shapex = 200 and shapey=200. Then I rewrote the model to include the transfer learning.
base_model = MobileNet(weights='imagenet', include_top=False, input_shape=(200,200,3), dropout=.3)
x = base_model.output
types = Dense(19,activation='softmax')(x)
model = Model(inputs=base_model.input,outputs=types)
model.compile(
optimizer='adam',
loss = 'sparse_categorical_crossentropy',
metrics=['accuracy']
history = model.fit(get_batches(), steps_per_epoch=1000, epochs=10)
I found everything I needed on this Google Colab:
[https://colab.research.google.com/github/GoogleCloudPlatform/training-data-analyst/blob/master/courses/fast-and-lean-data-science/04_Keras_Flowers_transfer_learning_solution.ipynb#scrollTo=XLJNVGwHUDy1][1]
[1]: https://colab.research.google.com/github/GoogleCloudPlatform/training-data-analyst/blob/master/courses/fast-and-lean-data-science/04_Keras_Flowers_transfer_learning_solution.ipynb#scrollTo=XLJNVGwHUDy1

Tensorflow - can't initialize saved variables unless I recreate the "saver" object. Why?

I'm pretty sure I'm missing something about how tensorflow works because my solution doesn't make any sense.
I'm trying to train a neural network (from scratch, without using Estimators or other abstractions), save it, and load a simplified version of it for inference.
The following code trains but gives me the error: FailedPreconditionError (see above for traceback): Attempting to use uninitialized value hidden0/biases/Variable
[[Node: hidden0/biases/Variable/read = Identity[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"](hidden0/biases/Variable)]]. If I add the commented line - if I recreate the saver obect that I'm not going to use nor return - the code works just fine.
Why do I need to create a (useless) saver object in order to restore the saved weights?
import tensorflow as tf
import numpy as np
def add_fc_layer(input_tensor, input_dimensions, output_dimensions, layer_name, activation=None):
with tf.variable_scope(layer_name):
with tf.variable_scope('weights'):
weights = tf.Variable(tf.truncated_normal([input_dimensions, output_dimensions]))
with tf.variable_scope('biases'):
biases = tf.Variable(tf.zeros([output_dimensions]))
with tf.variable_scope('Wx_plus_b'):
preactivate = tf.matmul(input_tensor, weights) + biases
if activation is None:
return preactivate
with tf.variable_scope('activation'):
activations = activation(preactivate)
return activations
def make_network(model_phase):
if model_phase not in {"train", "test"}:
raise ValueError("invalid type")
hidden0_units = 25
hidden1_units = 15
hidden2_units = 10
input_size = 10
output_size = 4
with tf.variable_scope('InputVector'):
inputs = tf.placeholder(shape=[1, input_size], dtype=tf.float32)
hidden0_out = add_fc_layer(inputs, input_size, hidden0_units, "hidden0", activation=tf.nn.sigmoid)
hidden1_out = add_fc_layer(hidden0_out, hidden0_units, hidden1_units, "hidden1", activation=tf.nn.sigmoid)
hidden2_out = add_fc_layer(hidden1_out, hidden1_units, hidden2_units, "hidden2", activation=tf.nn.sigmoid)
out = add_fc_layer(hidden2_out, hidden2_units, output_size, "regression")
if model_phase == "test":
# UNCOMMENTIN THIS LINE MAKES THE SCRIPT WORK
# saver = tf.train.Saver(var_list=tf.trainable_variables())
return inputs, out
saver = tf.train.Saver(var_list=tf.trainable_variables())
with tf.variable_scope('training'):
with tf.variable_scope('groundTruth'):
ground_truth = tf.placeholder(shape=[1, output_size], dtype=tf.float32)
with tf.variable_scope('loss'):
loss = tf.reduce_sum(tf.square(ground_truth - out))
tf.summary.scalar('loss', loss)
with tf.variable_scope('optimizer'):
trainer = tf.train.AdamOptimizer(learning_rate=0.001)
with tf.variable_scope('gradient'):
updateModel = trainer.minimize(loss)
with tf.variable_scope('predict'):
predict = tf.random_shuffle(tf.boolean_mask(out, tf.equal(out, tf.reduce_max(out, axis=None))))[0]
writer = tf.summary.FileWriter('/tmp/test', tf.get_default_graph())
return inputs, out, ground_truth, updateModel, writer, saver
train_graph = tf.Graph()
with tf.Session(graph=train_graph) as sess:
tf.set_random_seed(42)
inputs, out, ground_truth, updateModel, writer, saver = make_network(model_phase='train')
init = tf.initialize_all_variables()
sess.run(init)
print('\nLearning...')
for _ in range(10):
sess.run([updateModel], feed_dict={inputs:np.arange(10)+np.random.random((1,10)), ground_truth:np.arange(4).reshape(1, 4)})
saver.save(sess,'./tensorflowModel.ckpt')
new_graph = tf.Graph()
with tf.Session(graph=new_graph) as sess:
inputs, out = make_network(model_phase='test')
saver = tf.train.import_meta_graph('./tensorflowModel.ckpt.meta')
saver.restore(sess, tf.train.latest_checkpoint('./'))
# evaluation
print('\nEvaluation...')
for _ in range(10):
_ = sess.run(out, feed_dict={inputs:np.arange(10).reshape(1,10)})

I don't know why creating an unused Saver makes the problem go away, but the code betrays a misunderstanding.
When you are restoring, you are creating the model graph twice. First, you call make_network() which creates the computation graph and variables. You then also call import_meta_graph which also creates a graph and variables. You should create a saver with simple saver = tf.train.Saver() instead of saver = tf.train.import_meta_graph('./tensorflowModel.ckpt.meta')

Outer dimension for outputs must be unknown, outer dimension of 'Const:0' is 12

I have created the classification model using retrain.py. Then I deployed the model in google ml engine. It gives the following error
Create Version failed. Model validation failed: Outer dimension for outputs must be unknown, the outer dimension of 'Const:0' is 12 For more information on how to export Tensorflow SavedModel, see https://www.tensorflow.org/api_docs/python/tf/saved_model.
StackOverflow suggests changing the output tensor dimension
def export_model(sess,keys, architecture, saved_model_dir):
print("reached export funct")
if architecture == 'inception_v3':
input_tensor = 'DecodeJpeg/contents:0'
elif architecture.startswith('mobilenet_'):
input_tensor = 'input:0'
else:
raise ValueError('Unknown architecture', architecture)
in_image = sess.graph.get_tensor_by_name(input_tensor)
inputs = {'image': tf.saved_model.utils.build_tensor_info(in_image)}
out_classes = sess.graph.get_tensor_by_name('final_result:0')
outputs = {'prediction':
tf.saved_model.utils.build_tensor_info(out_classes),
'classes':
tf.saved_model.utils.build_tensor_info(tf.convert_to_tensor(list(keys))),}
signature = tf.saved_model.signature_def_utils.build_signature_def(
inputs=inputs,
outputs=outputs,
method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME
)
print(
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY)
legacy_init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')
# Save out the SavedModel.
builder = tf.saved_model.builder.SavedModelBuilder(saved_model_dir)
builder.add_meta_graph_and_variables(
sess, [tf.saved_model.tag_constants.SERVING],
signature_def_map={
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
signature
},
legacy_init_op=legacy_init_op)
builder.save()