def imdb_experient():
model_class = gnn_models.SagCWL2GCN
dsm = eval_ds.IMDB_8()
model = model_class(
act="sigmoid", local_act="sigmoid", squeeze_output=True,
conv_layer_dims=[dsm.dim_wl2_features(), 40, 40, 40, 1],
att_conv_layer_dims=[dsm.dim_wl2_features(), 1],
conv_stack_tf="keep_input",
bias=True)
opt = keras.optimizers.Adam(0.0007)
model.compile(
optimizer=opt,
loss="binary_crossentropy",
metrics=["accuracy"])
train, val = dsm.get_train_fold(0, output_type=model_class.input_type)
test = dsm.get_test_fold(0, output_type=model_class.input_type)
evaluate.train(
model, train, test, verbose=2, epochs=1000,
label=f"{dsm.name}_{model.name}")
print(model.evaluate(test))
def wl2_power_experiment():
# model_class = gnn_models.AvgGIN
model_class = gnn_models.AvgCWL2GCN
model_class = gnn_models.with_fc(model_class)
dsm = synthetic.threesix_dataset(stored=True)(
wl2_neighborhood=1) # ok(3, 2, 1)
if model_class.input_type == "dense":
in_dim = dsm.dim_dense_features()
elif model_class.input_type == "wl1":
in_dim = dsm.dim_wl1_features()
else:
in_dim = dsm.dim_wl2_features()
if in_dim == 0:
in_dim = 1
opt = keras.optimizers.Adam(0.1)
model = model_class(
act="sigmoid", squeeze_output=True,
layer_dims=[in_dim, 4, 1],
fc_layer_dims=[1, 10, 1],
neighborhood_mask=1, # ok(3, 2), nok(-1, 1)
bias=False, no_local_hash=True)
model.compile(
optimizer=opt,
loss="binary_crossentropy",
metrics=["accuracy"])
ds = dsm.get_all(
output_type=model_class.input_type,
shuffle=True)
evaluate.train(
model, ds, verbose=1,
epochs=200, patience=200,
label=f"{dsm.name}_{model.name}")
print(
list(dsm.get_all(output_type="dense"))[0][1].numpy(),
model.predict(dsm.get_all(output_type=model_class.input_type)))
def synthetic_experiment2():
model_class = gnn_models.AvgCWL2GCN
dsm = synthetic.balanced_triangle_classification_dataset(stored=True)(
with_holdout=False,
wl2_neighborhood=1,
wl2_batch_size=dict(batch_graph_count=228))
if model_class.input_type == "dense":
in_dim = dsm.dim_dense_features()
else:
in_dim = dsm.dim_wl2_features()
if in_dim == 0:
in_dim = 1
opt = keras.optimizers.Adam(0.0005)
model = model_class(
act="sigmoid", local_act="relu",
squeeze_output=True,
layer_dims=[in_dim, 32, 32, 32, 1],
att_conv_layer_dims=[in_dim, 32, 32, 32, 1],
bias=True, no_local_hash=True)
model.compile(
optimizer=opt,
loss="binary_crossentropy",
metrics=["accuracy"])
i = 5
ds = dsm.get_train_fold(
i, output_type=model_class.input_type)
ds_test = dsm.get_test_fold(
i, output_type=model_class.input_type)
evaluate.train(
model, ds, ds_test, verbose=2,
epochs=5000, patience=2000,
label=f"{dsm.name}_{model.name}")
def proteins_experient():
model_class = gnn_models.AvgWL2GCN
dsm = eval_ds.Proteins_6()
model = model_class(
act="sigmoid", squeeze_output=True,
conv_layer_dims=[dsm.dim_wl2_features(), 64, 64, 64, 1],
vert_only_pool=False,
bias=True)
opt = keras.optimizers.Adam(0.0001)
model.compile(
optimizer=opt,
loss="binary_crossentropy",
metrics=["accuracy"])
ds, ds_val = dsm.get_train_fold(
1, output_type=model_class.input_type)
evaluate.train(
model, ds, ds_val, verbose=1,
label=f"{dsm.name}_{model.name}")
def kernel_experiment():
model_class = kernel_models.WL_sp
model = model_class(C=0.001)
dsm = synthetic.balanced_triangle_classification_dataset(stored=True)(
with_holdout=False)
for i in range(10):
ds = dsm.get_train_fold(
i, output_type=model_class.input_type)
ds_test = dsm.get_test_fold(
i, output_type=model_class.input_type)
#ds = dsm.get_all(output_type=model_class.input_type)
print(i)
print(evaluate.train(model, ds, ds_test, label=f"{dsm.name}_{model.name}").history)
def dd_experient():
model_class = gnn_models.AvgWL2GCN
dsm = eval_ds.DD_2()
model = model_class(
act="sigmoid", squeeze_output=True,
layer_dims=[dsm.dim_wl2_features(), 64, 64, 64, 1],
bias=True)
opt = keras.optimizers.Adam(0.00001)
model.compile(
optimizer=opt,
loss="binary_crossentropy",
metrics=["accuracy"])
ds_raw = dsm.get_all(
output_type=model_class.input_type)
ds = ds_raw
evaluate.train(
model, ds, verbose=1,
label=f"{dsm.name}_{model.name}")
print(model.evaluate(ds))