def test_inspect_adjacency(self):
# This "test" is best viewed in PyCharm debugger, put a breakpoint at "print(df)" and look at the
# resulting adacency-weight matrix. You will see that it makes sense.
t2g = Text2GraphTransformer(min_df=1, window_size=3, rm_stopwords=True)
X = [
"Time is an illusion. Lunchtime doubly so.",
"The ships hung in the sky in much the same way that bricks don't.",
"If there's anything more important than my ego around, I want it caught and shot now.",
"Would it save you a lot of time if I just gave up and went mad now?"
]
res = t2g.fit_transform(X, y=[1, 0, 1], test_idx=[2])
# expand adjacency matrix
A = np.zeros((t2g.n_vocabs_ + 4, t2g.n_vocabs_ + 4))
for (e1, e2), weight in zip(res.edge_index.T, res.edge_attr):
A[e1, e2] = weight
inv_map = {v: k for k, v in t2g.vocabulary.items()}
inv_map[t2g.n_vocabs_] = "Document 1"
inv_map[t2g.n_vocabs_ + 1] = "Document 2"
inv_map[t2g.n_vocabs_ + 2] = "Document 3"
inv_map[t2g.n_vocabs_ + 3] = "Document 4"
names = [inv_map[i] for i in range(t2g.n_vocabs_ + 4)]
df = pd.DataFrame(data=A, columns=names, index=names)
print(df)
val_idx = np.random.choice(len(x), int(train_val_split * len(x)), replace=False)
train_idx = np.array([x for x in range(len(x)) if x not in val_idx])
x_test = test['Text'].tolist()
y_test = test[labels].tolist()
test_idx = np.arange(len(x), len(x) + len(x_test))
# Combine training & test data set
x = x + x_test
y = y + y_test
y = LabelEncoder().fit_transform(y)
print("Data loaded!")
t2g = Text2GraphTransformer(n_jobs=8, min_df=5, save_path=None, verbose=1, max_df=max_df, window_size=window_size)
# t2g = Text2GraphTransformer(n_jobs=8, min_df=1, save_path=save_path, verbose=1, max_df=1.0)
ls = os.listdir("textgcn/graphs")
# if not ls:
if True:
g = t2g.fit_transform(x, y, test_idx=test_idx, val_idx=val_idx)
print("Graph built!")
else:
g = t2g.load_graph(os.path.join(save_path, ls[0]))
print(f"Graph loaded from {os.path.join(save_path, ls[0])}!")
print(f"n_classes={len(np.unique(g.y))}")
# gcn = JumpingKnowledgeNetwork(g.x.shape[1], len(np.unique(y)), n_hidden_gcn=100, dropout=dropout, activation=th.nn.SELU)
# gcn = EGCN(g.x.shape[1], len(np.unique(y)), n_hidden_gcn=100, embedding_dim=2000, dropout=dropout)
gcn = model(g.x.shape[1], len(np.unique(y)), n_hidden_gcn=100, dropout=dropout)
y += y_test
y = LabelEncoder().fit_transform(y)
print("Data loaded!")
################################################ Text to Graph ################################################
kf = KFold(n_splits=k_split, shuffle=True)
frameIterator = 0
timestamp = datetime.now().strftime("%d_%b_%y_%H_%M_%S")
csv_name = "DBPEDIA_Flat_HypOpt_" + lable_category + "_" + timestamp + ".csv"
for mdf in dfs:
t2g = Text2GraphTransformer(n_jobs=1,
min_df=50,
save_path=None,
verbose=1,
max_df=mdf)
g = t2g.fit_transform(x, y, test_idx=test_idx, val_idx=val_idx)
print("Graph built!")
# Mask for doc nodes
mask = th.logical_or(g.train_mask, g.test_mask)
indizes = th.nonzero(mask)
for dropout in dos:
for lr in lrs:
for model in models:
model_name = "GCN" if model == GCN else "EGCN"
try:
y_top1 = LabelEncoder().fit_transform(y_top1)
y_top2 = LabelEncoder().fit_transform(y_top2)
del x_val
del x_test
del y_top1_val
del y_top2_val
del y_top1_test
del y_top2_test
print("Data loaded!")
t2g = Text2GraphTransformer(n_jobs=1,
min_df=100,
save_path=save_path,
verbose=1,
max_df=max_df,
max_length=MAX_LENGTH,
window_size=window_size)
g1 = t2g.fit_transform(x,
y_top1,
test_idx=test_idx,
val_idx=val_idx,
hierarchy_feats=None)
print("Graph built!")
gcn1 = model(g1.x.shape[1],
len(np.unique(y_top1)),
n_hidden_gcn=n_hidden,
x_test = test['Text'].tolist()
y_test = test['Cat2'].tolist()
y_test_top = test['Cat1'].tolist()
test_idx = np.arange(len(x), len(x) + len(x_test))
y = y + y_test
y_top = y_top + y_test_top
x = x + x_test
y = LabelEncoder().fit_transform(y)
y_top = LabelEncoder().fit_transform(y_top)
print("Data loaded!")
t2g = Text2GraphTransformer(n_jobs=8, min_df=5, save_path=save_path, verbose=1, max_df=0.6)
g1 = t2g.fit_transform(x, y_top, test_idx=test_idx, val_idx=val_idx, hierarchy_feats=None)
g1 = g1.to(device)
gcn1 = gcn1.to(device)
with th.no_grad():
pred_test = np.argmax(gcn1(g1)[g1.test_mask].cpu().detach().numpy(), axis=1)
gcn1 = gcn1.to(cpu)
g2 = t2g.fit_transform(x, y, test_idx=test_idx, val_idx=val_idx, hierarchy_feats=None)
g2 = g2.to(device)
predictions = np.zeros(len(g2.y))
predictions[:] = -1