This is my Graph tensor.
featureA_dim = 768
featureB_dim = 61
featureC_dim = 1
def create_graph_tensor(node_df,edge_df_tmp,genomic_edge_df_tmp):
graph_tensor = tfgnn.GraphTensor.from_pieces(
node_sets = {"repeats":
tfgnn.NodeSet.from_fields(sizes = [-1],
features ={
'featureA':np.array(featureA_data, dtype='float32').reshape(-1,featureA_dim),
'featureB':np.array(featureB_data, dtype='float32').reshape(-1,featureB_dim),
'featureC':np.array(featureC_data, dtype='float32').reshape(-1,featureC_dim)
})},
edge_sets = {"order":
tfgnn.EdgeSet.from_fields(sizes = [len(genomic_edge_df_tmp)],
features = {},
adjacency = tfgnn.Adjacency.from_indices(
source = ("repeats", np.array(genomic_edge_df_tmp['previous_node_id'], dtype='int32').reshape(-1,1)),
target = ("repeats", np.array(genomic_edge_df_tmp['following_node_id'], dtype='int32').reshape(-1,1)) )),
"possibility": tfgnn.EdgeSet.from_fields(sizes = [len(edge_df_tmp)],
features = {'label': np.array(edge_df_tmp['label'], dtype='int32').reshape(-1,1)},
adjacency = tfgnn.Adjacency.from_indices(
source = ("repeats", np.array(edge_df_tmp['previous_node_id'], dtype='int32')),
target = ("repeats", np.array(edge_df_tmp['following_node_id'], dtype='int32'))))})
return graph_tensor
For the above graph tensor, I have processed the edge to separate the label.
def edge_merge_encode(graph):
edge_features = graph.edge_sets['possibility'].get_features_dict()
label = edge_features.pop('label')
new_graph = graph.replace_features(edge_sets={'possibility':edge_features})
return new_graph, label
def create_dataset(graph,function):
dataset = tf.data.Dataset.from_tensors(graph)
return dataset.map(function)
#INPUT:DO AWAY LABELS(Edge Datasets)
full_edge_dataset = create_dataset(full_tensor,edge_merge_encode)
train_edge_dataset = create_dataset(train_tensor,edge_merge_encode)
The train_edge_dataset has GraphTensorSpec as follows.
<_MapDataset element_spec=(GraphTensorSpec({'context': ContextSpec({'features': {}, 'sizes': TensorSpec(shape=(1,), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, None), 'node_sets': {'repeats': NodeSetSpec({'features': {'featureA': TensorSpec(shape=(942, 768), dtype=tf.float32, name=None), 'featureB': TensorSpec(shape=(942, 61), dtype=tf.float32, name=None), 'featureC': TensorSpec(shape=(942, 1), dtype=tf.float32, name=None)}, 'sizes': TensorSpec(shape=(1,), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, None)}, 'edge_sets': {'order': EdgeSetSpec({'features': {}, 'adjacency': AdjacencySpec({'#index.0': TensorSpec(shape=(939, 1), dtype=tf.int32, name=None), '#index.1': TensorSpec(shape=(939, 1), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, {'#index.0': 'repeats', '#index.1': 'repeats'}), 'sizes': TensorSpec(shape=(1,), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, None), 'possibility': EdgeSetSpec({'features': {}, 'adjacency': AdjacencySpec({'#index.0': TensorSpec(shape=(282,), dtype=tf.int32, name=None), '#index.1': TensorSpec(shape=(282,), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, {'#index.0': 'repeats', '#index.1': 'repeats'}), 'sizes': TensorSpec(shape=(1,), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, None)}}, TensorShape([]), tf.int32, tf.int64, None), TensorSpec(shape=(282, 1), dtype=tf.int32, name=None))>
I have trained train_edge_dataset with below graph-updating layer architecture.
`
def set_initial_node_state(node_set, node_set_name):
features = [tf.keras.layers.Dense(64,activation="relu")(node_set['featureA']),
tf.keras.layers.Dense(32,activation="relu")(node_set['featureB']),
tf.keras.layers.Dense(2,activation="relu")(node_set['featureC']) ]
return tf.keras.layers.Concatenate()(features)
def set_initial_edge_state(edge_set, edge_set_name):
if edge_set_name == "order":
features = tfgnn.keras.layers.MakeEmptyFeature()(edge_set)
return features
elif edge_set_name == "possibility":
features = tfgnn.keras.layers.MakeEmptyFeature()(edge_set)
return features
graph_spec,_ = train_edge_dataset.element_spec
input_graph = tf.keras.layers.Input(type_spec=graph_spec)
graph = tfgnn.keras.layers.MapFeatures(
node_sets_fn=set_initial_node_state,
edge_sets_fn=set_initial_edge_state
)(input_graph)
#MODEL:SPCIFYING LAYERS
graph_updates = 3
for i in range(graph_updates):
graph = tfgnn.keras.layers.GraphUpdate(
edge_sets = {
'possibility': tfgnn.keras.layers.EdgeSetUpdate(
next_state = tfgnn.keras.layers.NextStateFromConcat( tf.keras.layers.Dense(8))),
},
node_sets = {
'repeats': tfgnn.keras.layers.NodeSetUpdate(
{
"order": tfgnn.keras.layers.SimpleConv(
tf.keras.layers.Dense(32, "relu"),
"mean",
receiver_tag=tfgnn.TARGET),
"possibility": tfgnn.keras.layers.SimpleConv(
tf.keras.layers.Dense(32, "relu"),
"mean",
receiver_tag=tfgnn.TARGET)
},
tfgnn.keras.layers.NextStateFromConcat(tf.keras.layers.Dense(64)))})(graph)
#OUTPUT:FINAL OUTPUT LAYER
logits = tf.keras.layers.Dense(8,activation='relu')(graph.edge_sets['possibility'][tfgnn.HIDDEN_STATE])
logits = tf.keras.layers.Dense(1,activation='sigmoid')(logits)
The model was compiled as follows.
edge_model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=0.00001),
loss=['binary_crossentropy'],
metrics=['Accuracy'])
es = tf.keras.callbacks.EarlyStopping(
monitor='val_loss',
mode='min',
verbose=1,
patience=1000,
restore_best_weights=True)
history = edge_model.fit(train_edge_dataset,
epochs=10,
validation_data=val_edge_dataset,
callbacks=[es],
verbose=1)
Training with train_edge_dataset went well, but training with full_edge_dataset is problematic.
The GraphTensorSpec for full_edge_dataset is as follows.
<_MapDataset element_spec=(GraphTensorSpec({'context': ContextSpec({'features': {}, 'sizes': TensorSpec(shape=(1,), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, None), 'node_sets': {'repeats': NodeSetSpec({'features': {'featureA': TensorSpec(shape=(942, 768), dtype=tf.float32, name=None), 'featureB': TensorSpec(shape=(942, 61), dtype=tf.float32, name=None), 'featureC': TensorSpec(shape=(942, 1), dtype=tf.float32, name=None)}, 'sizes': TensorSpec(shape=(1,), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, None)}, 'edge_sets': {'order': EdgeSetSpec({'features': {}, 'adjacency': AdjacencySpec({'#index.0': TensorSpec(shape=(939, 1), dtype=tf.int32, name=None), '#index.1': TensorSpec(shape=(939, 1), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, {'#index.0': 'repeats', '#index.1': 'repeats'}), 'sizes': TensorSpec(shape=(1,), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, None), 'possibility': EdgeSetSpec({'features': {}, 'adjacency': AdjacencySpec({'#index.0': TensorSpec(shape=(376,), dtype=tf.int32, name=None), '#index.1': TensorSpec(shape=(376,), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, {'#index.0': 'repeats', '#index.1': 'repeats'}), 'sizes': TensorSpec(shape=(1,), dtype=tf.int32, name=None)}, TensorShape([]), tf.int32, tf.int64, None)}}, TensorShape([]), tf.int32, tf.int64, None), TensorSpec(shape=(376, 1), dtype=tf.int32, name=None))>
I have checked the label and logit shape, and I see no problem. The logits for full_edge_dataset could not be retrieved because the full_edge_dataset could not go through the layers. (the task that I want is link prediction case)
Shape of label for full_edge_dataset tensor during model fitting: (376, 1)
Shape of label for train_edge_dataset tensor during model fitting: (282, 1)
Shape of logits for train_edge_dataset after final layer: (282, 1)
The exact error message is as follows.
Traceback (most recent call last):
File "gnn_final.py", line 282, in <module>edge_model.fit(full_edge_dataset,
File "~/.local/lib/python3.8/site-packages/keras/src/utils/traceback_utils.py", line 70, in error_handlerraise e.with_traceback(filtered_tb) from None
File "~/.local/lib/python3.8/site-packages/tensorflow/python/eager/execute.py", line 53, in quick_executetensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,tensorflow.python.framework.errors_impl.InvalidArgumentError: Graph execution error:
Detected at node 'IteratorGetNext' defined at (most recent call last):
File "gnn_final.py", line 282, in <module>edge_model.fit(full_edge_dataset,
File "~/.local/lib/python3.8/site-packages/keras/src/utils/traceback_utils.py", line 65, in error_handler
return fn(*args, **kwargs)
File "~/.local/lib/python3.8/site-packages/keras/src/engine/training.py", line 1742, in fit
tmp_logs = self.train_function(iterator)
File "~/.local/lib/python3.8/site-packages/keras/src/engine/training.py", line 1338, in train_function
return step_function(self, iterator)
File "~/.local/lib/python3.8/site-packages/keras/src/engine/training.py", line 1321, in step_function
data = next(iterator)
Node: 'IteratorGetNext'Detected at node 'IteratorGetNext' defined at (most recent call last):
File "gnn_final.py", line 282, in <module>edge_model.fit(full_edge_dataset,File "~/.local/lib/python3.8/site-packages/keras/src/utils/traceback_utils.py", line 65, in error_handler
return fn(*args, **kwargs)
File "~/.local/lib/python3.8/site-packages/keras/src/engine/training.py", line 1742, in fit
tmp_logs = self.train_function(iterator)
File "~/.local/lib/python3.8/site-packages/keras/src/engine/training.py", line 1338, in train_function
return step_function(self, iterator)File "~/.local/lib/python3.8/site-packages/keras/src/engine/training.py", line 1321, in step_function
data = next(iterator)
Node: 'IteratorGetNext'
2 root error(s) found.
(0) INVALID_ARGUMENT: Data type mismatch at component 4: expected variant but got int32.
[[{{node IteratorGetNext}}]]
[[IteratorGetNext/_4]]
(1) INVALID_ARGUMENT: Data type mismatch at component 4: expected variant but got int32.[[{{node IteratorGetNext}}]]
0 successful operations.
0 derived errors ignored. [Op:__inference_train_function_12592]