def test_graphattnet_train_dropout(self):
pickle_fname = 'abp_CV_fold_1_tlXlY_trn.pkl'
gcn_graph = GCNDataset.load_transkribus_pickle(pickle_fname)
gcn_graph_train = [gcn_graph[8], gcn_graph[18], gcn_graph[29]]
node_dim = gcn_graph[0].X.shape[1]
edge_dim = gcn_graph[0].E.shape[1] - 2.0
nb_class = gcn_graph[0].Y.shape[1]
gcn_model = GraphAttNet(node_dim,
nb_class,
num_layers=1,
learning_rate=0.01,
node_indim=-1,
nb_attention=3)
gcn_model.dropout_rate_node = 0.2
gcn_model.dropout_rate_attention = 0.2
gcn_model.create_model()
with tf.Session() as session:
session.run([gcn_model.init])
# Get the Test Prediction
g_acc, node_acc = gcn_model.test_lG(session, gcn_graph_train)
print('Mean Accuracy', g_acc, node_acc)
gcn_model.train_lG(session, gcn_graph)
g_acc, node_acc = gcn_model.test_lG(session, gcn_graph_train)
print('Mean Accuracy', g_acc, node_acc)
def test_predict_proba(self):
pickle_fname = 'abp_CV_fold_1_tlXlY_trn.pkl'
gcn_graph = GCNDataset.load_transkribus_pickle(pickle_fname)
gcn_graph_train = [gcn_graph[8], gcn_graph[18], gcn_graph[29]]
node_dim = gcn_graph[0].X.shape[1]
edge_dim = gcn_graph[0].E.shape[1] - 2.0
nb_class = gcn_graph[0].Y.shape[1]
gcn_model = GraphAttNet(node_dim,
nb_class,
num_layers=1,
learning_rate=0.01,
node_indim=-1,
nb_attention=3)
gcn_model.dropout_rate_node = 0.2
gcn_model.dropout_rate_attention = 0.2
gcn_model.create_model()
with tf.Session() as session:
session.run([gcn_model.init])
# Get the Test Prediction
gcn_model.train_lG(session, gcn_graph)
g_proba = gcn_model.prediction_prob(session, gcn_graph_train[1])
print(g_proba.shape)
print(type(g_proba))
print(gcn_graph_train[1].X.shape)
self.assertTrue(g_proba.shape == (gcn_graph_train[1].X.shape[0],
5))
def test_train_ensemble_NN_model(self):
#TODO Make a proper synthetic dataset for test Purpose
pickle_fname = 'abp_CV_fold_1_tlXlY_trn.pkl'
gcn_graph = GCNDataset.load_transkribus_pickle(pickle_fname)
gcn_graph_train = [gcn_graph[8], gcn_graph[18], gcn_graph[29]]
node_dim = gcn_graph[0].X.shape[1]
edge_dim = gcn_graph[0].E.shape[1] - 2.0
nb_class = gcn_graph[0].Y.shape[1]
gat_model = GraphAttNet(node_dim,
nb_class,
num_layers=1,
learning_rate=0.01,
node_indim=-1,
nb_attention=3)
gat_model.dropout_rate_node = 0.2
gat_model.dropout_rate_attention = 0.2
gat_model.create_model()
nb_layers = 3
lr = 0.001
nb_conv = 2
ecn_model = EdgeConvNet(
node_dim,
edge_dim,
nb_class,
num_layers=nb_layers,
learning_rate=lr,
mu=0.0,
node_indim=-1,
nconv_edge=nb_conv,
)
ecn_model.create_model()
#Check Graphs
#Are we recopying the models and graph definition implicitly ?
ensemble = EnsembleGraphNN([ecn_model, gat_model])
with tf.Session() as session:
session.run([ensemble.models[0].init])
for iter in range(500):
ensemble.train_lG(session, gcn_graph_train)
prediction = ensemble.predict_lG(session, gcn_graph_train)
print(prediction)
self.assertTrue(len(prediction) == len(gcn_graph_train))
print('Ensemble Prediction')
accs = ensemble.test_lG(session, gcn_graph_train)
print('Base Predictions')
for m in ensemble.models:
accs = m.test_lG(session, gcn_graph_train)
print(accs)
print(accs)
def test_dense_attn_layer(self):
gcn_graph = get_graph_test()
node_dim = gcn_graph.X.shape[1]
edge_dim = gcn_graph.E.shape[1] - 2.0
nb_class = gcn_graph.Y.shape[1]
gcn_model = GraphAttNet(node_dim,
nb_class,
num_layers=1,
learning_rate=0.01,
node_indim=8,
nb_attention=1)
gcn_model.create_model()
Wa = tf.eye(node_dim)
va = tf.ones([2, node_dim])
# elf.Ssparse, self.Tspars
alphas, nH = gcn_model.dense_graph_attention_layer(
gcn_model.node_input, Wa, va, gcn_model.nb_node,
gcn_model.dropout_p_attn, gcn_model.dropout_p_node)
alphas_shape = tf.shape(alphas)
init = tf.global_variables_initializer()
graph = gcn_graph
with tf.Session() as session:
session.run([init])
print('### Graph', graph.X.shape, graph.F.shape[0])
# print(graph.Sind)
# print(graph.Tind)
nb_node = graph.X.shape[0]
Aind = np.array(np.stack([graph.Sind[:, 0], graph.Tind[:, 1]],
axis=-1),
dtype='int64')
print("Adjacency Indices:", Aind.shape, Aind)
feed_batch = {
gcn_model.nb_node:
graph.X.shape[0],
gcn_model.nb_edge:
graph.F.shape[0],
gcn_model.node_input:
graph.X,
gcn_model.Ssparse:
np.array(graph.Sind, dtype='int64'),
gcn_model.Sshape:
np.array([graph.X.shape[0], graph.F.shape[0]], dtype='int64'),
gcn_model.Tsparse:
np.array(graph.Tind, dtype='int64'),
gcn_model.Aind:
Aind,
# self.F: graph.F,
gcn_model.y_input:
graph.Y,
# self.dropout_p_H: self.dropout_rate_H,
gcn_model.dropout_p_node:
0.0,
gcn_model.dropout_p_attn:
0.0,
}
[c_alphas, c_nH,
c_alphas_shape] = session.run([alphas, nH, alphas_shape],
feed_dict=feed_batch)
print('alphas', c_alphas, c_alphas_shape)
#sp_mat = sp.coo_matrix((c_alphas.data, (c_alphas.indices[:, 0], c_alphas.indices[:, 1])),
# shape=(nb_node, nb_node))
Att_dense = c_alphas
print(Att_dense)
#TODO Update True Value
self.assertTrue(c_alphas_shape[0] == 3)
self.assertTrue(c_alphas_shape[1] == 3)
def test_graphattnet_attnlayer_selfloop(self):
gcn_graph = get_graph_test()
node_dim = gcn_graph.X.shape[1]
edge_dim = gcn_graph.E.shape[1] - 2.0
nb_class = gcn_graph.Y.shape[1]
gcn_model = GraphAttNet(node_dim,
nb_class,
num_layers=1,
learning_rate=0.01,
node_indim=8,
nb_attention=1)
gcn_model.create_model()
Wa = tf.eye(node_dim)
va = tf.ones([2, node_dim])
# elf.Ssparse, self.Tspars
alphas, nH = gcn_model.simple_graph_attention_layer(
gcn_model.node_input,
Wa,
va,
gcn_model.Ssparse,
gcn_model.Tsparse,
gcn_model.Aind,
gcn_model.Sshape,
gcn_model.nb_edge,
gcn_model.dropout_p_attn,
gcn_model.dropout_p_node,
add_self_loop=True)
alphas_shape = tf.shape(alphas)
node_indices = tf.range(gcn_model.Sshape[0])
# Sparse Idendity
# Debug
id_indices = tf.stack([node_indices, node_indices], axis=1)
val = tf.squeeze(tf.matmul(gcn_model.node_input, va, transpose_b=True))
spI = tf.SparseTensor(
indices=id_indices,
values=val,
dense_shape=[gcn_model.Sshape[0], gcn_model.Sshape[0]])
init = tf.global_variables_initializer()
#AI=tf.sparse_add(alphas,spI)
graph = gcn_graph
with tf.Session() as session:
session.run([init])
print('### Graph', graph.X.shape, graph.F.shape[0])
# print(graph.Sind)
# print(graph.Tind)
nb_node = graph.X.shape[0]
Aind = np.array(np.stack([graph.Sind[:, 0], graph.Tind[:, 1]],
axis=-1),
dtype='int64')
print("Adjacency Indices:", Aind.shape, Aind)
feed_batch = {
gcn_model.nb_node:
graph.X.shape[0],
gcn_model.nb_edge:
graph.F.shape[0],
gcn_model.node_input:
graph.X,
gcn_model.Ssparse:
np.array(graph.Sind, dtype='int64'),
gcn_model.Sshape:
np.array([graph.X.shape[0], graph.F.shape[0]], dtype='int64'),
gcn_model.Tsparse:
np.array(graph.Tind, dtype='int64'),
gcn_model.Aind:
Aind,
# self.F: graph.F,
gcn_model.y_input:
graph.Y,
# self.dropout_p_H: self.dropout_rate_H,
gcn_model.dropout_p_node:
0.0,
gcn_model.dropout_p_attn:
0.0,
}
[c_alphas, c_nH, c_alphas_shape,
spI] = session.run([alphas, nH, alphas_shape, spI],
feed_dict=feed_batch)
print('alphas', c_alphas, c_alphas_shape)
print('spI', spI)
#print('AI',AI)
sp_mat = sp.coo_matrix(
(c_alphas.values,
(c_alphas.indices[:, 0], c_alphas.indices[:, 1])),
shape=(nb_node, nb_node))
Att_dense = sp_mat.todense()
print(Att_dense)
self.assertTrue(c_alphas_shape[0] == 3)
self.assertTrue(c_alphas_shape[1] == 3)
self.assertTrue(Att_dense[0, 2] == 0)