def get_tensorflow_object_last(self, dst_units):
"""
Parameters
----------
dst_units: int
Number of units that the recurrent cell will have
"""
try:
c_ = getattr(tf.keras.layers, self.type)
except Exception:
print_failure(
"The layer of layer_type '" + self.type +
"' is not a valid tf.keras layer. Please check the "
"documentation to write the correct way to define this layer. ")
self.parameters['units'] = dst_units # can we assume that it will always be units??
try:
layer = c_(**self.parameters)
except Exception:
parameters_string = ''
for k, v in self.parameters.items():
parameters_string += k + ': ' + v + '\n'
print_failure(
"One of the parameters passed to the layer of layer_type '" + self.type + "' is incorrect. \n " +
"You have defined the following parameters: \n" + parameters_string)
return layer
def calculate(self, product_input1, product_input2):
"""
Parameters
----------
product_input1: tensor
Input 1
product_input2: tensor
Input 2
"""
try:
if self.type_product == 'dot_product':
result = tf.tensordot(product_input1, product_input2, axes=[[1], [1]])
# the correct values are in the diagonal (IMPROVE THIS)
# This does the dot product row by row (so independently for each adjacency)
result = tf.linalg.tensor_diag_part(result)
result = tf.expand_dims(result, axis=-1)
elif self.type_product == 'element_wise':
result = tf.math.multiply(product_input1, product_input2)
elif self.type_product == 'mat_mult':
result = tf.tensordot(product_input1, product_input2, axes=[[2], [1]])
result = tf.squeeze(result, axis=2)
result = tf.cast(result, tf.float32)
return result
except:
print_failure(
'The product operation between ' + product_input1 + ' and ' + product_input2 +
' failed. Check that the dimensions are compatible.')
def get_tensorflow_object(self, dst_dim):
"""
Parameters
----------
dst_dim: int
Dimension of the destination nodes. Thus, number of units of the RNN model
"""
self.parameters['units'] = dst_dim
try:
c_ = getattr(tf.keras.layers, self.type + 'Cell')
except Exception:
print_failure(
"Error when trying to define a RNN of layer_type '" + self.type +
"' since this layer_type does not exist. Check the valid RNN cells that Keras allow to define.")
try:
layer = c_(**self.parameters)
except Exception:
print_failure(
"Error when creating the RNN of layer_type '" + self.type +
"' since invalid parameters were passed. Check the documentation to check which "
"parameters are acceptable or check the spelling of the parameters' names.")
return layer
def find_total_input_dim(self, dimensions, calculations):
"""
Parameters
----------
dimensions: dict
Dictionary with the dimensions of each tensor (indexed by name)
calculations: dict
Dictionary with the current calculations throughout the execution of the GNN model
"""
if self.input is not None:
input_nn = self.input
input_dim = 0
dimension = None
for i in input_nn:
if '_initial_state' in i:
i = i.split('_initial_state')[0]
if i in dimensions:
dimension = dimensions[i]
elif i + '_out_dim' in calculations:
dimension = calculations[i + '_out_dim'] # take the dimension from here or from self.dimensions
else:
print_failure("Keyword " + i + " used in the model definition was not recognized")
input_dim += dimension
return input_dim
def stream_read_json(self, f):
"""
Parameters
----------
f:
Input data
"""
# check that it is a valid array of objects
pos1 = f.read(1)
if pos1 != '[':
print_failure(
"Error because the dataset files must be an array of json objects, and not single json objects")
start_pos = 1
while True:
try:
obj = json.load(f)
yield obj
return
except json.JSONDecodeError as e:
f.seek(start_pos)
json_str = f.read(e.pos)
obj = json.loads(json_str)
start_pos += e.pos + 1
a = f.read(1) # this 1 is the coma or the final symbol
if a == self.end_symbol or a == ']':
yield obj
return
yield obj
def obtain_total_input_dim_message(self, dimensions, calculations, dst_name, src):
"""
Parameters
----------
dimensions: dict
Dictionary with the dimensions of each tensor (indexed by name)
calculations: dict
Dictionary with the current calculations throughout the execution of the GNN model
dst_name: str
Name of the destination entity
src: Source_mp object
Object that includes the information about the source entity of the mp
"""
# Find out the dimension of the model
input_nn = self.input
input_dim = 0
for i in input_nn:
if i == 'source':
input_dim += int(dimensions.get(src.name))
elif i == 'destination':
input_dim += int(dimensions.get(dst_name))
elif i in dimensions:
input_dim += int(dimensions[i])
elif i + '_dim' in calculations:
input_dim += dimensions
else:
print_failure("Keyword " + i + " used in the message passing was not recognized.")
return input_dim
def calculate(self, src_states, adj_src, dst_states, adj_dst):
"""
Parameters
----------
src_states: tensor
Input 1
adj_src: tensor
Adj src -> dest
dst_states: tensor
Input 2
adj_dst: tensor
Adj dst -> src
"""
# obtain the extended input (by extending it to the number of adjacencies between them)
try:
extended_src = tf.gather(src_states, adj_src)
except Exception:
print_failure('Extending the adjacency list ' + str(self.adj_list) +
' was not possible. Check that the indexes of the source of the adjacency '
'list match the input given.')
try:
extended_dst = tf.gather(dst_states, adj_dst)
except Exception:
print_failure('Extending the adjacency list ' + str(self.adj_list) +
' was not possible. Check that the indexes of the destination of '
'the adjacency list match the input given.')
return extended_src, extended_dst
def __prepocess_parameters(self):
for k, v in self.parameters.items():
if v == 'None':
self.parameters[k] = None
elif v == 'True':
self.parameters[k] = True
elif v == 'False':
self.parameters[k] = False
elif 'regularizer' in k:
try:
self.parameters[k] = tf.keras.regularizers.l2(float(self.parameters.get(k)))
except Exception:
print_failure("The " + k + " parameter '" + str(self.parameters.get(k)) +
"' in layer of layer_type " + self.type +
" is invalid. Please make sure it is a numerical value.")
elif 'activation' in k: # already ensures that it was not None
try:
self.parameters['activation'] = getattr(tf.nn, v)
except Exception:
print_failure("The activation '" + v +
"' is not a valid function from the tf.nn library. Please check the documentation "
"and the spelling of the function.")
def __init__(self, op):
"""
Parameters
----------
op: dict
Dictionary with the data defining this general operation
"""
self.type = op.get('type')
self.output_name = op.get('output_name', None)
self.output_label = op.get('output_label', None)
if self.output_label is not None:
# There may be more than one output_label
self.output_label = [output.split('$')[-1] for output in
self.output_label] # delete the $ from the output label
# parse the input of the operation
self.input = []
self.source_dataset = False
self.destination_dataset = False
if 'input' in op:
for input_item in op.get('input'):
if '$source' == input_item or '$destination' == input_item:
print_failure(
'The keywords source and destination are reserved keywords. Thus, they cannot name feature '
'from the dataset. Check that you really meant to use $, indicating that its a feature '
'from the dataset')
else:
self.input.append(input_item.split('$')[-1]) # delete the $ from the inputs (if any)
def create_aggregations(self, attrs):
"""
Parameters
----------
attrs: dict
Dictionary with the required attributes for the aggregation (defining the set of operations)
"""
aggregations = []
single_embedding = None
multiple_embedding = None
for attr in attrs:
attr_type = attr.get('type')
if attr_type == 'interleave':
aggregations.append(InterleaveAggr(attr))
multiple_embedding = True
elif attr_type == 'neural_network':
aggregations.append(
FeedForwardOperation(attr, model_role='aggregation'))
single_embedding = True
elif attr_type == 'concat':
aggregations.append(ConcatAggr(attr))
multiple_embedding = True
elif attr_type == 'sum':
aggregations.append(SumAggr(attr))
single_embedding = True
elif attr_type == 'mean':
aggregations.append(MeanAggr(attr))
single_embedding = True
elif attr_type == 'min':
aggregations.append(MinAggr(attr))
single_embedding = True
elif attr_type == 'max':
aggregations.append(MaxAggr(attr))
single_embedding = True
elif attr_type == 'std':
aggregations.append(StdAggr(attr))
single_embedding = True
elif attr_type == 'attention':
aggregations.append(AttentionAggr(attr))
single_embedding = True
elif attr_type == 'edge_attention':
aggregations.append(EdgeAttentionAggr(attr))
single_embedding = True
elif attr_type == 'convolution':
aggregations.append(ConvAggr(attr))
single_embedding = True
else: # this is for the ordered aggregation
multiple_embedding = True
if single_embedding and multiple_embedding:
print_failure(
"You cannot combine aggregations which return a sequence of tensors, "
"and aggregations that return a single embedding")
elif single_embedding:
return aggregations, 0
else:
return aggregations, 1
def generate_from_array(self,
data_samples,
entity_names,
feature_names,
output_names,
adj_names,
interleave_names,
additional_input,
training,
shuffle=False):
"""
Parameters
----------
data_samples: [array]
Array of samples to be processed
entity_names: [array]
Name of the entities to be found in the dataset
feature_names: [array]
Name of the features to be found in the dataset
output_names: [str]
Names of the output data to be found in the dataset
interleave_names: [array]
First parameter is the name of the interleave, and the second the destination entity
additional_input: [array]
Name of other vectors that need to be retrieved because they appear in other parts of the model definition
training: bool
Indicates if we are training, and thus a label is required.
shuffle: bool
Shuffle parameter of the dataset
"""
data_samples = [json.loads(x) for x in data_samples]
self.entity_names = [x for x in entity_names]
self.feature_names = [x for x in feature_names]
self.output_names = output_names
self.adj_names = adj_names
self.interleave_names = [[i[0], i[1]] for i in interleave_names]
self.additional_input = [x for x in additional_input]
self.training = training
for sample in data_samples:
try:
processed_sample = self.__process_sample(sample)
yield processed_sample
except StopIteration:
pass
except KeyboardInterrupt:
sys.exit()
except Exception as inf:
print_failure("\n There was an unexpected error: \n" + str(
inf) + "\n Please make sure that all the names used in the sample passed ")
def calculate(self, inputs):
"""
Parameters
----------
inputs: tensor
"""
try:
result = tf.concat(inputs, axis=self.axis)
return result
except Exception:
print_failure(
'The concat operation failed. Check that the dimensions are compatible.')
sys.exit(1)
def __read_yaml(self, path, file_name=''):
"""
Parameters
----------
path: str
Path of the json file with the model description
file_name: str
Name of the file we aim to read
"""
if os.path.isfile(path):
with open(path, 'r') as stream:
try:
return yaml.safe_load(stream)
except yaml.YAMLError as exc:
print_failure("There was the following error in the " + file_name + " file.\n" + str(exc))
else:
print_failure("The " + file_name + " file was not found in: " + path)
def get_tensorflow_object(self):
try:
c_ = getattr(tf.keras.layers, self.type)
except Exception:
print_failure(
"The layer of layer_type '" + self.type +
"' is not a valid tf.keras layer. Please check the documentation to "
"write the correct way to define this layer. ")
try:
layer = c_(**self.parameters)
except Exception:
parameters_string = ''
for k, v in self.parameters.items():
parameters_string += k + ': ' + v + '\n'
print_failure(
"One of the parameters passed to the layer of layer_type '" + self.type + "' is incorrect. \n " +
"You have defined the following parameters: \n" + parameters_string)
return layer
def construct_tf_model(self, input_dim, dst_dim=None, is_readout=False, dst_name=None):
"""
Parameters
----------
input_dim: int
Dimension of the input of the model
dst_dim: int
Dimension of the destination hs if any
is_readout: bool
Is a model used for the readout?
dst_name: str
Name of the destination entity
"""
model = tf.keras.models.Sequential()
model.add(tf.keras.Input(shape=input_dim))
layer_counter = 1
n = len(self.layers)
for j in range(n):
current_layer = self.layers[j]
try:
# if it's the last layer and we have defined an output dimension
if j == (n - 1) and dst_dim is not None:
layer_model = current_layer.get_tensorflow_object_last(dst_dim)
else:
layer_model = current_layer.get_tensorflow_object()
model.add(layer_model)
except:
if dst_dim is None:
if is_readout:
print_failure('The layer ' + str(layer_counter) +
' of the readout is not correctly defined. Check keras documentation '
'to make sure all the parameters are correct.')
else:
print_failure('The layer ' + str(layer_counter) +
' of the message creation neural network in the message passing to ' +
str(dst_name) + ' is not correctly defined. Check keras documentation to '
'make sure all the parameters are correct.')
else:
print_failure('The layer ' + str(
layer_counter) + ' of the update neural network in message passing to ' + str(dst_name) +
' is not correctly defined. Check keras documentation to make sure all the '
'parameters are correct.')
layer_counter += 1
output_shape = model.output_shape[-1]
return [model, output_shape]
def __process_sample(self, sample, file=None):
"""
Parameters
----------
sample: dict
Input sample which is a serialized version (in JSON) of a networkx graph.
file: str
Path to these file (which is useful for error-checking purposes)
"""
# load the model
G = json_graph.node_link_graph(sample)
# Only directed graphs are supported. Error checking message if the graph is undirected
if not G.is_directed():
print_failure("IGNNITION received as input an undirected graph, even though it only supports "
"(at the moment) directed graphs. Please consider reformating your code accordingly. "
"You can double the edges between two nodes (e.g., edge 1-2 can be transformed into 1->2 "
"and 2->1) to simulate the same behaviour.")
if G.is_multigraph():
print_failure("IGNNITION received as input a multigraph, while these are not yet supported. This means, "
"that for every pair of nodes, only one edge with the same source and destination can exist "
"(e.g., you cannot have two edges 1->2 and 1->2. Notice that 1->2 and 2->1 does not incur "
"in this problem.")
entity_counter = {}
mapping = {}
data = {}
for name in self.entity_names:
entity_counter[name] = 0
list_nodes = list(G.nodes())
for node_name in list_nodes:
attributes = G.nodes[node_name]
if 'entity' not in attributes:
print_failure(
"Error in the dataset file located in '" + file + ". The node named'" + str(node_name)
+ "' was not assigned an entity.")
entity_name = attributes['entity']
new_node_name = entity_name + '_{}'
num_node = entity_counter[entity_name]
entity_counter[entity_name] += 1
mapping[node_name] = new_node_name.format(num_node)
# save the number of nodes of each entity
for name in self.entity_names:
data['num_' + name] = entity_counter[name]
# rename the name of the nodes to a mapping that also indicates its entity layer_type
D_G = nx.relabel_nodes(G, mapping)
# discard if the graph is empty
if not D_G.edges():
print_info("\nA sample was discarded because the graph is empty (has no edges).")
raise StopIteration
# load the features (all the features are set to be lists. So we always return a list of lists)
for f in self.feature_names:
try:
features_dict = nx.get_node_attributes(D_G, f)
feature_vals = np.array(list(features_dict.values()))
entity_names = set([name.split('_')[0] for name in features_dict.keys()]) #indicates the (unique)
# names of the entities that have that feature
if len(entity_names) > 1:
entities_string = functools.reduce(lambda x,y: str(x) + ',' + str(y), entity_names )
print_failure("The feature " + f + " was defined in several entities(" + entities_string +
"). The feature names should be unique for each layer_type of node.")
# it should always be a 2d array
if len(np.shape(feature_vals)) == 1:
feature_vals = np.expand_dims(feature_vals, axis=-1)
if feature_vals.size == 0:
message = "The feature " + f + " was used in the model_description.yaml file " \
"but was not defined in the dataset."
if file is not None:
message = "Error in the dataset file located in '" + file + ".\n" + message
raise Exception(message)
else:
data[f] = feature_vals
except:
message = "The feature " + f + " was used in the model_description.yaml file " \
"but was not defined in the dataset."
if file is not None:
message = "Error in the dataset file located in '" + file + ".\n" + message
raise Exception(message)
# take other inputs if needed (check that they might be global features)
for a in self.additional_input:
# 1) try to see if this name has been defined as a node attribute
node_dict = nx.get_node_attributes(D_G, a)
node_attr = np.array(list(node_dict.values()))
entity_names = set([name.split('_')[0] for name in node_dict.keys()]) # indicates the (unique) names
# of the entities that have that feature
if len(entity_names) > 1:
entities_string = functools.reduce(lambda x, y: str(x) + ',' + str(y), entity_names)
print_failure(
"The feature " + a + " was defined in several entities(" + entities_string +
"). The feature names should be unique for each layer_type of node.")
# it should always be a 2d array
if len(np.shape(node_attr)) == 1:
node_attr = np.expand_dims(node_attr, axis=-1)
# 2) try to see if this name has been defined as an edge feature
edge_dict = nx.get_edge_attributes(D_G, a)
edge_attr = np.array(list(edge_dict.values()))
entity_names = set([(pair[0].split('_')[0], pair[1].split('_')[0]) for pair in edge_dict.keys()]) #
# indicates the (unique) names of the entities that have that feature
# obtain the entities, with a small token indicating if it is source or destination
# Problem: When we transform an undirected graph to directed, we double all the edges. Hence,
# we still need to differentiate between source and destination entities?? Solution: Allow only directed??
# for now, check that the name is unique for every src-dst. Problem: One node connected to another but
# the reverse to other nodes??
if len(entity_names) > 2:
#print(entity_names)
entities_string = functools.reduce(lambda x, y: str(x) + ',' + str(y), entity_names)
print_failure(
"The edge feature " + a + " was defined in connecting two different source-destination entities(" +
entities_string + "). Make sure that an edge feature is unique for a given pair of entities "
"(types of nodes).")
# it should always be a 2d array
if len(np.shape(edge_attr)) == 1:
edge_attr = np.expand_dims(edge_attr, axis=-1)
# 3) try to see if this name has been defined as a graph feature
graph_attr = [D_G.graph[a]] if a in D_G.graph else []
# Check that this name has not been defined both as node features and as edge_features
if node_attr.size != 0 and edge_attr.size != 0 and len(graph_attr) != 0:
print_failure("The feature " + a + "was defined both as node feature, edge feature and graph feature. "
"Please use unique names in this case.")
elif node_attr.size != 0 and edge_attr.size != 0:
print_failure("The feature " + a + "was defined both as node feature and as edge feature. Please use "
"unique names in this case.")
elif node_attr.size != 0 and len(graph_attr) != 0:
print_failure("The feature " + a + "was defined both as node feature and as graph feature. Please use "
"unique names in this case.")
# Return the correct value
if node_attr.size != 0:
data[a] = node_attr
elif edge_attr.size != 0:
data[a] = edge_attr
elif a in D_G.graph:
data[a] = graph_attr
else:
message = 'The data named "' + a + '" was used in the model_description.yaml file ' \
'but was not defined in the dataset.'
if file is not None:
message = "Error in the dataset file located in '" + file + ".\n" + message
raise Exception(message)
if self.training:
# collect the output (if there is more than one, concatenate them on axis=1
# limitation: all the outputs must be of the same layer_type (same number of elements)
final_output = []
for output in self.output_names:
try:
aux = list(nx.get_node_attributes(D_G, output).values())
if not aux: # When having global/graph-level output
aux = D_G.graph[output]
aux = aux if isinstance(aux, list) else [aux]
except Exception:
print_failure(
f"Error when trying to get output with name: {output}. "
"Check the data which corresponds to the output_label in the readout block."
)
# if it is a 1d array, transform it into a 2d array
if len(np.array(aux).shape) == 1:
aux = np.expand_dims(aux, -1)
# try to concatenate them together. If error, it means that the two labels are incompatible
final_output.extend(aux)
data['__ignnition_{}_len'.format(output)] = len(aux)
# find the adjacencies
edges_list = list(D_G.edges())
processed_neighbours = {}
# create the adjacency lists that we are required to pass
for adj_name_item in self.adj_names:
src_entity = adj_name_item.split('_to_')[0]
dst_entity = adj_name_item.split('_to_')[1]
data['src_' + src_entity + '_to_' + dst_entity] = []
data['dst_' + src_entity + '_to_' + dst_entity] = []
data['seq_' + src_entity + '_to_' + dst_entity] = []
for e in edges_list:
src_node, dst_node = e
src_num = int(src_node.split('_')[-1])
dst_num = int(dst_node.split('_')[-1])
src_entity = D_G.nodes[src_node]['entity']
dst_entity = D_G.nodes[dst_node]['entity']
if dst_node not in processed_neighbours:
processed_neighbours[dst_node] = 0
if src_entity + '_to_' + dst_entity in self.adj_names:
data['src_' + src_entity + '_to_' + dst_entity].append(src_num)
data['dst_' + src_entity + '_to_' + dst_entity].append(dst_num)
data['seq_' + src_entity + '_to_' + dst_entity].append(processed_neighbours[dst_node])
processed_neighbours[dst_node] += 1 # this is useful to check which sequence number to use
# check that the dataset contains all the adjacencies needed
if not self.warnings_shown:
for adj_name_item in self.adj_names:
if data['src_' + adj_name_item] == []:
src_entity = adj_name_item.split('_to_')[0]
dst_entity = adj_name_item.split('_to_')[1]
print_info(
"WARNING: The GNN definition uses edges between " + src_entity + " and " + dst_entity +
" but these were not found in the input graph. The MP defined between these two entities "
"will be ignored.\nIn case the graph ought to contain such edges, one reason for this error "
"is a mistake in defining the graph as directional, when the edges have been defined as "
"undirected. Please check the documentation.")
self.warnings_shown = True
# this collects the sequence for the interleave aggregation (if any)
for i in self.interleave_names:
name, dst_entity = i
interleave_definition = list(D_G.graph[name].values()) # this must be a graph variable
involved_entities = {}
total_sequence = []
total_size, n_total, counter = 0, 0, 0
for src_entity in interleave_definition:
total_size += 1
if src_entity not in involved_entities:
involved_entities[src_entity] = counter # each entity a different value (identifier)
seq = data['seq_' + src_entity + '_to_' + dst_entity]
n_total += max(seq) + 1 # superior limit of the size of any destination
counter += 1
# obtain all the original definition in a numeric format
total_sequence.append(involved_entities[src_entity])
# we exceed the maximum length for sake to make it multiple. Then we will cut it
repetitions = math.ceil(float(n_total) / total_size)
result = np.array((total_sequence * repetitions)[:n_total])
for entity in involved_entities:
id = involved_entities[entity]
data['indices_' + entity + '_to_' + dst_entity] = np.where(result == id)[0].tolist()
if self.training:
return data, final_output
else:
return data
def __validate_model_description(self, data):
"""
Parameters
----------
data: dict
Dictionary with the initial data
"""
entities = data['entities']
stages = data['message_passing']['stages']
src_names, dst_names, called_nn_names, input_names = [], [], [], []
output_names = ['source', 'destination']
# check the hidden state creation
for entity_item in entities:
state_ops = entity_item['initial_state']
for op in state_ops:
if 'input' in op:
input_names += op['input']
if 'output' in op:
output_names += op['output']
# check the message passing
for stage in stages:
stage_mp = stage.get('stage_message_passings')
for mp in stage_mp: # for every message-passing
dst_names.append(mp.get('destination_entity'))
sources = mp.get('source_entities')
# check the message functions
for src in sources:
src_names.append(src.get('name'))
messages = src.get('message', None)
if messages is not None:
for op in messages: # for every operation
if op.get('type') == 'neural_network':
called_nn_names.append(op.get('nn_name'))
input_names += op.get('input')
if 'output_name' in op:
output_names.append(op.get('output_name'))
# check the aggregation functions
aggregations = mp.get('aggregation')
for aggr in aggregations:
if aggr.get('type') == 'neural_network':
input_names += aggr.get('input')
if 'output_name' in aggr:
output_names.append(aggr.get('output_name'))
readout_op = data.get('readout')
called_nn_names += [op.get('nn_name') for op in readout_op if op.get('type') == 'neural_network']
if 'output_label' not in readout_op[-1]:
print_failure('The last operation of the readout MUST contain the definition of the output_label')
else:
input_names += readout_op[-1]['output_label']
# now check the entities
entity_names = [a.get('name') for a in data.get('entities')]
nn_names = [n.get('nn_name') for n in data.get('neural_networks')]
# check if the name of two NN defined match
if len(nn_names) != len(set(nn_names)):
print_failure("The names of two NN are repeated. Please ensure that each NN has a unique name.")
# check the source entities
for a in src_names:
if a not in entity_names:
print_failure(
'The source entity "' + a + '" was used in a message passing. However, there is no such entity. \n '
'Please check the spelling or define a new entity.')
# check the destination entities
for d in dst_names:
if d not in entity_names:
print_failure(
'The destination entity "' + d + '" was used in a message passing. However, there is no such '
'entity. \n Please check the spelling or define a new entity.')
# check the nn_names
for name in called_nn_names:
if name not in nn_names:
print_failure(
'The name "' + name + '" is used as a reference to a neural network (nn_name), even though the '
'neural network was not defined. \n Please make sure the name is correctly '
'spelled or define a neural network named ' + name)
# ensure that all the inputs (that are not output of another operation) start with a $
for i in input_names:
if i not in output_names and i[0] != '$':
print_failure('The input name ' + i + ' references data from the dataset but does not start with $')
for i in output_names:
if i[0] == '$':
print_failure(
'The keyword ' + i + ' starts with $ even though it does not represent data from the dataset.')
