Related
Problem
As the title suggests I have been trying to create a pipeline for training an Autoencoder model using TFX. The problem I'm having is fitting the tf.Dataset returned by the DataAccessor.tf_dataset_factory object to the Autoencoder.
Below I summarise the steps I've taken through this project, and have some Questions at the bottom if you wish to skip the background information.
Intro
TFX Pipeline
The TFX components I have used so far have been:
CsvExampleGenerator (the dataset has 82 columns, all numeric, and the sample csv has 739 rows)
StatisticsGenerator / SchemaGenerator, the schema has been edited as is now loaded in using an Importer
Transform
Trainer (this is the component I am currently having problems with)
Model
The model that I am attempting to train is based off of the example laid out here https://www.tensorflow.org/tutorials/generative/autoencoder. However, my model is being trained on tabular data, searching for anomalous results, as opposed to image data.
As I have tried a couple of solutions I have tried using both the Keras.layers and Keras.model format for defining the model and I outline both below:
Subclassing Keras.Model
class Autoencoder(keras.models.Model):
def __init__(self, features):
super(Autoencoder, self).__init__()
self.encoder = tf.keras.Sequential([
keras.layers.Dense(82, activation = 'relu'),
keras.layers.Dense(32, activation = 'relu'),
keras.layers.Dense(16, activation = 'relu'),
keras.layers.Dense(8, activation = 'relu')
])
self.decoder = tf.keras.Sequential([
keras.layers.Dense(16, activation = 'relu'),
keras.layers.Dense(32, activation = 'relu'),
keras.layers.Dense(len(features), activation = 'sigmoid')
])
def call(self, x):
inputs = [keras.layers.Input(shape = (1,), name = f) for f in features]
dense = keras.layers.concatenate(inputs)
encoded = self.encoder(dense)
decoded = self.decoder(encoded)
return decoded
Subclassing Keras.Layers
def _build_keras_model(features: List[str]) -> tf.keras.Model:
inputs = [keras.layers.Input(shape = (1,), name = f) for f in features]
dense = keras.layers.concatenate(inputs)
dense = keras.layers.Dense(32, activation = 'relu')(dense)
dense = keras.layers.Dense(16, activation = 'relu')(dense)
dense = keras.layers.Dense(8, activation = 'relu')(dense)
dense = keras.layers.Dense(16, activation = 'relu')(dense)
dense = keras.layers.Dense(32, activation = 'relu')(dense)
outputs = keras.layers.Dense(len(features), activation = 'sigmoid')(dense)
model = keras.Model(inputs = inputs, outputs = outputs)
model.compile(
optimizer = 'adam',
loss = 'mae'
)
return model
TFX Trainer Component
For creating the Trainer Component I have been mainly following the implementation details laid out here: https://www.tensorflow.org/tfx/guide/trainer
As well as following the default penguins example: https://www.tensorflow.org/tfx/tutorials/tfx/penguin_simple#write_model_training_code
run_fn defintion
def run_fn(fn_args: tfx.components.FnArgs) -> None:
tft_output = tft.TFTransformOutput(fn_args.transform_output)
train_dataset = _input_fn(
file_pattern = fn_args.train_files,
data_accessor = fn_args.data_accessor,
tf_transform_output = tft_output,
batch_size = fn_args.train_steps
)
eval_dataset = _input_fn(
file_pattern = fn_args.eval_files,
data_accessor = fn_args.data_accessor,
tf_transform_output = tft_output,
batch_size = fn_args.custom_config['eval_batch_size']
)
# model = Autoencoder(
# features = fn_args.custom_config['features']
# )
model = _build_keras_model(features = fn_args.custom_config['features'])
model.compile(optimizer = 'adam', loss = 'mse')
model.fit(
train_dataset,
steps_per_epoch = fn_args.train_steps,
validation_data = eval_dataset,
validation_steps = fn_args.eval_steps
)
...
_input_fn definition
def _apply_preprocessing(raw_features, tft_layer):
transformed_features = tft_layer(raw_features)
return transformed_features
def _input_fn(
file_pattern,
data_accessor: tfx.components.DataAccessor,
tf_transform_output: tft.TFTransformOutput,
batch_size: int) -> tf.data.Dataset:
"""
Generates features and label for tuning/training.
Args:
file_pattern: List of paths or patterns of input tfrecord files.
data_accessor: DataAccessor for converting input to RecordBatch.
tf_transform_output: A TFTransformOutput.
batch_size: representing the number of consecutive elements of returned
dataset to combine in a single batch
Returns:
A dataset that contains features where features is a
dictionary of Tensors.
"""
dataset = data_accessor.tf_dataset_factory(
file_pattern,
tfxio.TensorFlowDatasetOptions(batch_size = batch_size),
tf_transform_output.transformed_metadata.schema
)
transform_layer = tf_transform_output.transform_features_layer()
def apply_transform(raw_features):
return _apply_preprocessing(raw_features, transform_layer)
return dataset.map(apply_transform).repeat()
This differs from the _input_fn example given above as I was following the example in the next tfx tutorial found here: https://www.tensorflow.org/tfx/tutorials/tfx/penguin_tft#run_fn
Also for reference, there is no Target within the example data so there is no label_key to be passed to the tfxio.TensorFlowDatasetOptions object.
Error
When trying to run the Trainer component using a TFX InteractiveContext object I receive the following error.
ValueError: No gradients provided for any variable: ['dense_460/kernel:0', 'dense_460/bias:0', 'dense_461/kernel:0', 'dense_461/bias:0', 'dense_462/kernel:0', 'dense_462/bias:0', 'dense_463/kernel:0', 'dense_463/bias:0', 'dense_464/kernel:0', 'dense_464/bias:0', 'dense_465/kernel:0', 'dense_465/bias:0'].
From my own attempts to solve this I believe the problem lies in the way that an Autoencoder is trained. From the Autoencoder example linked here https://www.tensorflow.org/tutorials/generative/autoencoder the data is fitted like so:
autoencoder.fit(x_train, x_train,
epochs=10,
shuffle=True,
validation_data=(x_test, x_test))
therefore it stands to reason that the tf.Dataset should also mimic this behaviour and when testing with plain Tensor objects I have been able to recreate the error above and then solve it when adding the target to be the same as the training data in the .fit() function.
Things I've Tried So Far
Duplicating Train Dataset
model.fit(
train_dataset,
train_dataset,
steps_per_epoch = fn_args.train_steps,
validation_data = eval_dataset,
validation_steps = fn_args.eval_steps
)
Raises error due to Keras not accepting a 'y' value when a dataset is passed.
ValueError: `y` argument is not supported when using dataset as input.
Returning a dataset that is a tuple with itself
def _input_fn(...
dataset = data_accessor.tf_dataset_factory(
file_pattern,
tfxio.TensorFlowDatasetOptions(batch_size = batch_size),
tf_transform_output.transformed_metadata.schema
)
transform_layer = tf_transform_output.transform_features_layer()
def apply_transform(raw_features):
return _apply_preprocessing(raw_features, transform_layer)
dataset = dataset.map(apply_transform)
return dataset.map(lambda x: (x, x))
This raises an error where the keys from the features dictionary don't match the output of the model.
ValueError: Found unexpected keys that do not correspond to any Model output: dict_keys(['feature_string', ...]). Expected: ['dense_477']
At this point I switched to using the keras.model Autoencoder subclass and tried to add output keys to the Model using an output which I tried to create dynamically in the same way as the inputs.
def call(self, x):
inputs = [keras.layers.Input(shape = (1,), name = f) for f in x]
dense = keras.layers.concatenate(inputs)
encoded = self.encoder(dense)
decoded = self.decoder(encoded)
outputs = {}
for feature_name in x:
outputs[feature_name] = keras.layers.Dense(1, activation = 'sigmoid')(decoded)
return outputs
This raises the following error:
TypeError: Cannot convert a symbolic Keras input/output to a numpy array. This error may indicate that you're trying to pass a symbolic value to a NumPy call, which is not supported. Or, you may be trying to pass Keras symbolic inputs/outputs to a TF API that does not register dispatching, preventing Keras from automatically converting the API call to a lambda layer in the Functional Model.
I've been looking into solving this issue but am no longer sure if the data is being passed correctly and am beginning to think I'm getting side-tracked from the actual problem.
Questions
Has anyone managed to get an Autoencoder working when connected via TFX examples?
Did you alter the tf.Dataset or handled the examples in a different way to the _input_fn demonstrated?
So I managed to find an answer to this and wanted to leave what I found here in case anyone else stumbles onto a similar problem.
It turns out my feelings around the error were correct and the solution did indeed lie in how the tf.Dataset object was presented.
This can be demonstrated when I ran some code which simulated the incoming data using randomly generated tensors.
tensors = [tf.random.uniform(shape = (1, 82)) for i in range(739)]
# This gives us a list of 739 tensors which hold 1 value for 82 'features' simulating the dataset I had
dataset = tf.data.Dataset.from_tensor_slices(tensors)
dataset = dataset.map(lambda x : (x, x))
# This returns a dataset which marks the training set and target as the same
# which is what the Autoecnoder model is looking for
model.fit(dataset ...)
Following this I proceeded to do the same thing with the dataset returned by the _input_fn. Given that the tfx DataAccessor object returns a features_dict however I needed to combine the tensors in that dict together to create a single tensor.
This is how my _input_fn looks now:
def create_target_values(features_dict: Dict[str, tf.Tensor]) -> tuple:
value_tensor = tf.concat(list(features_dict.values()), axis = 1)
return (features_dict, value_tensor)
def _input_fn(
file_pattern,
data_accessor: tfx.components.DataAccessor,
tf_transform_output: tft.TFTransformOutput,
batch_size: int) -> tf.data.Dataset:
"""
Generates features and label for tuning/training.
Args:
file_pattern: List of paths or patterns of input tfrecord files.
data_accessor: DataAccessor for converting input to RecordBatch.
tf_transform_output: A TFTransformOutput.
batch_size: representing the number of consecutive elements of returned
dataset to combine in a single batch
Returns:
A dataset that contains (features, target_tensor) tuple where features is a
dictionary of Tensors, and target_tensor is a single Tensor that is a concatenated tensor of all the
feature values.
"""
dataset = data_accessor.tf_dataset_factory(
file_pattern,
tfxio.TensorFlowDatasetOptions(batch_size = batch_size),
tf_transform_output.transformed_metadata.schema
)
dataset = dataset.map(lambda x: create_target_values(features_dict = x))
return dataset.repeat()
I am trying to build a NN model for my Policy Gradient (deep reinforcement learning) agent by using Keras Functional API. What I intend to do is to mask invalid actions by reducing their probability distribution to zero in the logit layer:
def __build_policy_network(self):
inputs = keras.layers.Input(shape=(self.input_dim,))
advantages = keras.layers.Input(shape=(1,))
valid_actions = keras.layers.Input(shape=(3,))
dense_1 = keras.layers.Dense(units=self.fc1_size, activation="relu", kernel_initializer="he_uniform")(inputs)
dense_2 = keras.layers.Dense(units=self.fc2_size, activation="relu", kernel_initializer="he_uniform")(dense_1)
probs_logits = keras.layers.Dense(units=self.nb_actions, activation='softmax')(dense_2)
masked_probs = keras.layers.Multiply()([probs_logits, valid_actions])
probs = keras.layers.Lambda(lambda x: x / keras.backend.sum(x, axis=1))(masked_probs)
def custom_loss(y_true, y_pred):
out = keras.backend.clip(y_pred, 1e-8, 1 - 1e-8)
log_lik = y_true * keras.backend.log(out)
return keras.backend.sum(-log_lik * advantages)
policy = keras.models.Model([inputs, advantages], [probs])
policy.compile(optimizer=keras.optimizers.Adam(lr=self.alpha), loss=custom_loss)
predict = keras.models.Model([inputs, valid_actions], [probs])
return policy, predict
However, I run into the infamous error ValueError: Graph disconnected: cannot obtain value for tensor Tensor("input_3:0", shape=(None, 3), dtype=float32) at layer "multiply". When I comment out either of the advantages or valid_actions input layers (and of course, removing their corresponding lines) I can successfully run the code. I should mention that valid_actions input layer is only passed to mask invalid probabilities and is not required for loss calculation.
I really appreciate it if someone can help me with this.
Thanks in advance for your time
Your loss involves also advantages so you need to pass it inside the loss. You can do it with .add_loss.
policy model needs also valid_actions as input to produce probs.
predict model seems ok and can be used at inference time.
Here the full example with .add_loss.
inputs = keras.layers.Input(shape=(30,))
advantages = keras.layers.Input(shape=(1,))
valid_actions = keras.layers.Input(shape=(3,))
true = keras.layers.Input(shape=(3,))
dense_1 = keras.layers.Dense(units=64, activation="relu", kernel_initializer="he_uniform")(inputs)
dense_2 = keras.layers.Dense(units=32, activation="relu", kernel_initializer="he_uniform")(dense_1)
probs_logits = keras.layers.Dense(units=3, activation='softmax')(dense_2)
masked_probs = keras.layers.Multiply()([probs_logits, valid_actions])
probs = keras.layers.Lambda(lambda x: x / keras.backend.sum(x, axis=1))(masked_probs)
def custom_loss(y_true, y_pred, advantages):
out = keras.backend.clip(y_pred, 1e-8, 1 - 1e-8)
log_lik = y_true * keras.backend.log(out)
return keras.backend.sum(-log_lik * advantages)
policy = keras.models.Model([inputs, advantages, valid_actions, true], [probs])
policy.add_loss( custom_loss(true, probs, advantages) )
policy.compile(optimizer=keras.optimizers.Adam(lr=0.001), loss=None)
predict = keras.models.Model([inputs, valid_actions], [probs])
Thanks a lot to MarcoCerliani for his time and follow-ups, I succeeded to find a working solution to my problem finally. The mistake was that I modified the probs output layer which is required for loss calculation with the valid_actions input layer which is indeed only required for the predict model. As stated by this answer:
Keras cannot just ignore an input layer as the output depends on it.
all I need to do is pass probs_logits (the unmodified output layer by the valid_actions layer) to the policy model for loss calculation and pass the probs output layer (manipulated by the valid actions layer) to predict model:
def __build_policy_network(self):
// previous lines of code left unchanged
policy = keras.models.Model([inputs, advantages], [probs_logits])
policy.compile(optimizer=keras.optimizers.Adam(lr=self.alpha), loss=custom_loss)
predict = keras.models.Model([inputs, valid_actions], [probs])
return policy, predict
I am trying to construct a basic "vanilla gradient" saliency heatmap (gradient-based feature attribution) for MNIST using keras. I know there are libraries such as this one to compute saliency heatmaps, but I would like to construct this from scratch since the vanilla gradient approach seems conceptually straightforward to implement. I have trained the following digit classifier in Keras using functional model definition:
input = layers.Input(shape=(28,28,1), name='input')
conv2d_1 = layers.Conv2D(32, kernel_size=(3, 3), activation='relu')(input)
maxpooling2d_1 = layers.MaxPooling2D(pool_size=(2, 2), name='maxpooling2d_1')(conv2d_1)
conv2d_2 = layers.Conv2D(64, kernel_size=(3, 3), activation='relu')(maxpooling2d_1)
maxpooling2d_2 = layers.MaxPooling2D(pool_size=(2, 2))(conv2d_2)
flatten = layers.Flatten(name='flatten')(maxpooling2d_2)
dropout = layers.Dropout(0.5, name='dropout')(flatten)
dense = layers.Dense(num_classes, activation='softmax', name='dense')(dropout)
model = keras.models.Model(inputs=input, outputs=dense)
Now, I want to compute the saliency map for a single MNIST image. Since the final layer has a softmax activation and the denominator is a normalization term (so that the output nodes add up to 1), I believe that I need to either take the pre-softmax output or change the activation of the trained model linear for computing saliency maps. I will do the latter.
model.layers[-1].activation = tf.keras.activations.linear # swap activation to linear
input = loaded_model.layers[0].input
output = loaded_model.layers[-1].output
input_image = x_test[0] # shape is (28, 28, 1)
pred = np.argmax(loaded_model.predict(np.expand_dims(input_image, axis=0))) # predicted class
However, I am not sure what to do beyond this. I know I can use the following K.gradients(output, input) to compute gradients. That being said, I believe I should compute the gradient of the predicted class with respect to the input image, versus computing the gradient of the entire output. How would I do this? Also, I'm not sure how to evaluate the saliency heatmap for a specific image/prediction. I imagine I will have to use sess = tf.keras.backend.get_session() and sess.run(), but not sure exactly. I would greatly appreciate any help with completing the saliency heatmap code. Thanks!
If you add the activation as a single layer after the last dense layer with:
keras.layers.Activation('softmax')
you can do:
linear_model = keras.Model(input=model, output=model.layers[-2].output)
To then compute the gradients like:
def get_saliency_map(model, image, class_idx):
with tf.GradientTape() as tape:
tape.watch(image)
predictions = model(image)
loss = predictions[:, class_idx]
# Get the gradients of the loss w.r.t to the input image.
gradient = tape.gradient(loss, image)
# take maximum across channels
gradient = tf.reduce_max(gradient, axis=-1)
# convert to numpy
gradient = gradient.numpy()
# normaliz between 0 and 1
min_val, max_val = np.min(gradient), np.max(gradient)
smap = (gradient - min_val) / (max_val - min_val + keras.backend.epsilon())
return smap
So I am trying to write a simple softmax classifier in TensorFlow.
Here is the code:
# Neural network parameters
n_hidden_units = 500
n_classes = 10
# training set placeholders
input_X = tf.placeholder(dtype='float32',shape=(None,X_train.shape[1], X_train.shape[2]),name="input_X")
input_y = tf.placeholder(dtype='int32', shape=(None,), name="input_y")
# hidden layer
dim = X_train.shape[1]*X_train.shape[2] # dimension of each traning data point
flatten_X = tf.reshape(input_X, shape=(-1, dim))
weights_hidden_layer = tf.Variable(initial_value=np.zeros((dim,n_hidden_units)), dtype ='float32')
bias_hidden_layer = tf.Variable(initial_value=np.zeros((1,n_hidden_units)), dtype ='float32')
hidden_layer_output = tf.nn.relu(tf.matmul(flatten_X, weights_hidden_layer) + bias_hidden_layer)
# output layer
weights_output_layer = tf.Variable(initial_value=np.zeros((n_hidden_units,n_classes)), dtype ='float32')
bias_output_layer = tf.Variable(initial_value=np.zeros((1,n_classes)), dtype ='float32')
output_logits = tf.matmul(hidden_layer_output, weights_output_layer) + bias_output_layer
predicted_y = tf.nn.softmax(output_logits)
# loss
one_hot_labels = tf.one_hot(input_y, depth=n_classes, axis = -1)
loss = tf.losses.softmax_cross_entropy(one_hot_labels, output_logits)
# optimizer
optimizer = tf.train.MomentumOptimizer(0.01, 0.5).minimize(
loss, var_list=[weights_hidden_layer, bias_hidden_layer, weights_output_layer, bias_output_layer])
This compiles, and I have checked the shape of all the tensor and it coincides with what I expect.
However, I tried to run the optimizer using the following code:
# running the optimizer
s = tf.InteractiveSession()
s.run(tf.global_variables_initializer())
for i in range(5):
s.run(optimizer, {input_X: X_train, input_y: y_train})
loss_i = s.run(loss, {input_X: X_train, input_y: y_train})
print("loss at iter %i:%.4f" % (i, loss_i))
And the loss kept being the same in all iterations!
I must have messed up something, but I fail to see what.
Any ideas? I also appreciate if somebody leaves comments regarding code style and/or tensorflow tips.
You have made a mistake. You are initializing your weights using np.zeros. Use np.random.normal. You can choose mean for this Gaussian Distribution by using number of inputs going to a particular neuron. You can read more about it here.
The reason that you want to initialize with Gaussian Distribution is because you want to break symmetry. If all the weights are initialized by zero, then you can use backpropogation to see that all the weights will evolved same.
One could visualize the weight histogram using TensorBoard to make it easier. I executed your code for this. A few more lines are needed to set up Tensorboard logging but the histogram summary of weights can be easily added.
Initialized to zeros
weights_hidden_layer = tf.Variable(initial_value=np.zeros((784,n_hidden_units)), dtype ='float32')
tf.summary.histogram("weights_hidden_layer",weights_hidden_layer)
Xavier initialization
initializer = tf.contrib.layers.xavier_initializer()
weights_hidden_layer = tf.Variable(initializer(shape=(784,n_hidden_units)), dtype ='float32')
tf.summary.histogram("weights_hidden_layer",weights_hidden_layer)
I want to feed a CNN with the tensor "images". I want this tensor to contain images from the training set ( which have FIXED size ) when the placeholder is_training is True, otherwise I want it to contain images from the test set ( which are of NOT FIXED size ).
This is needed because in training I take a random fixed crop from the training images, while in test I want to perform a dense evaluation and feed the entire images inside the network ( it is fully convolutional so it will accept them)
The current NOT WORKING way is to create two different iterators, and try to select the training/test input with tf.cond at the session.run(images,{is_training:True/False}).
The problem is that BOTH the iterators are evaluated. The training and test dataset are also of different size so I cannot iterate both of them until the end. Is there a way to make this work? Or to rewrite this in a smarter way?
I've seen some questions/answers about this but they always used tf.assign which takes a numpy array and assigns it to a tensor. In this case I cannot use tf.assign because I already have a tensor coming from the iterators.
The current code that I have is this one. It simply checks the shape of the tensor "images":
train_filenames, train_labels = list_images(args.train_dir)
val_filenames, val_labels = list_images(args.val_dir)
graph = tf.Graph()
with graph.as_default():
# Preprocessing (for both training and validation):
def _parse_function(filename, label):
image_string = tf.read_file(filename)
image_decoded = tf.image.decode_jpeg(image_string, channels=3)
image = tf.cast(image_decoded, tf.float32)
return image, label
# Preprocessing (for training)
def training_preprocess(image, label):
# Random flip and crop
image = tf.image.random_flip_left_right(image)
image = tf.random_crop(image, [args.crop,args.crop, 3])
return image, label
# Preprocessing (for validation)
def val_preprocess(image, label):
flipped_image = tf.image.flip_left_right(image)
batch = tf.stack([image,flipped_image],axis=0)
return batch, label
# Training dataset
train_filenames = tf.constant(train_filenames)
train_labels = tf.constant(train_labels)
train_dataset = tf.contrib.data.Dataset.from_tensor_slices((train_filenames, train_labels))
train_dataset = train_dataset.map(_parse_function,num_threads=args.num_workers, output_buffer_size=args.batch_size)
train_dataset = train_dataset.map(training_preprocess,num_threads=args.num_workers, output_buffer_size=args.batch_size)
train_dataset = train_dataset.shuffle(buffer_size=10000)
batched_train_dataset = train_dataset.batch(args.batch_size)
# Validation dataset
val_filenames = tf.constant(val_filenames)
val_labels = tf.constant(val_labels)
val_dataset = tf.contrib.data.Dataset.from_tensor_slices((val_filenames, val_labels))
val_dataset = val_dataset.map(_parse_function,num_threads=1, output_buffer_size=1)
val_dataset = val_dataset.map(val_preprocess,num_threads=1, output_buffer_size=1)
train_iterator = tf.contrib.data.Iterator.from_structure(batched_train_dataset.output_types,batched_train_dataset.output_shapes)
val_iterator = tf.contrib.data.Iterator.from_structure(val_dataset.output_types,val_dataset.output_shapes)
train_images, train_labels = train_iterator.get_next()
val_images, val_labels = val_iterator.get_next()
train_init_op = train_iterator.make_initializer(batched_train_dataset)
val_init_op = val_iterator.make_initializer(val_dataset)
# Indicates whether we are in training or in test mode
is_training = tf.placeholder(tf.bool)
def f_true():
with tf.control_dependencies([tf.identity(train_images)]):
return tf.identity(train_images)
def f_false():
return val_images
images = tf.cond(is_training,f_true,f_false)
num_images = images.shape
with tf.Session(graph=graph) as sess:
sess.run(train_init_op)
#sess.run(val_init_op)
img = sess.run(images,{is_training:True})
print(img.shape)
The problem is that when I want to use only the training iterator, I comment the line to initialize the val_init_op but there is the following error:
FailedPreconditionError (see above for traceback): 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.
[[Node: IteratorGetNext_1 = IteratorGetNext[output_shapes=[[2,?,?,3], []], output_types=[DT_FLOAT, DT_INT32], _device="/job:localhost/replica:0/task:0/cpu:0"](Iterator_1)]]
If I do not comment that line everything works as expected, when is_training is true I get training images and when is_training is False I get validation images. The issue is that both the iterators need to be initialized and when I evaluate one of them, the other is incremented too. Since as I said they are of different size this causes an issue.
I hope there is a way to solve it! Thanks in advance
The trick is to call iterator.get_next() inside the f_true() and f_false() functions:
def f_true():
train_images, _ = train_iterator.get_next()
return train_images
def f_false():
val_images, _ = val_iterator.get_next()
return val_images
images = tf.cond(is_training, f_true, f_false)
The same advice applies to any TensorFlow op that has a side effect, like assigning to a variable: if you want that side effect to happen conditionally, the op must be created inside the appropriate branch function passed to tf.cond().