def pipeline(subgraph_max_nodes):
dataset = get_dataset(data_args.dataset_dir, data_args.dataset_name)
input_dim = dataset.num_node_features
output_dim = dataset.num_classes
data = dataset[0]
node_indices = torch.where(data.test_mask * data.y != 0)[0]
gnnNets = GnnNets_NC(input_dim, output_dim, model_args)
checkpoint = torch.load(mcts_args.explain_model_path)
gnnNets.update_state_dict(checkpoint['net'])
gnnNets.to_device()
gnnNets.eval()
save_dir = os.path.join('./results', f"{mcts_args.dataset_name}"
f"_{model_args.model_name}"
f"_{reward_args.reward_method}")
if not os.path.isdir(save_dir):
os.mkdir(save_dir)
plotutils = PlotUtils(dataset_name=data_args.dataset_name)
fidelity_score_list = []
sparsity_score_list = []
for node_idx in tqdm(node_indices):
# find the paths and build the graph
result_path = os.path.join(save_dir, f"node_{node_idx}_score.pt")
# get data and prediction
logits, prob, _ = gnnNets(data.clone())
_, prediction = torch.max(prob, -1)
prediction = prediction[node_idx].item()
# build the graph for visualization
graph = to_networkx(data, to_undirected=True)
node_labels = {k: int(v) for k, v in enumerate(data.y)}
nx.set_node_attributes(graph, node_labels, 'label')
# searching using gnn score
mcts_state_map = MCTS(node_idx=node_idx, ori_graph=graph,
X=data.x, edge_index=data.edge_index,
num_hops=len(model_args.latent_dim),
n_rollout=mcts_args.rollout,
min_atoms=mcts_args.min_atoms,
c_puct=mcts_args.c_puct,
expand_atoms=mcts_args.expand_atoms)
value_func = GnnNets_NC2value_func(gnnNets,
node_idx=mcts_state_map.node_idx,
target_class=prediction)
score_func = reward_func(reward_args, value_func)
mcts_state_map.set_score_func(score_func)
# get searching result
if os.path.isfile(result_path):
gnn_results = torch.load(result_path)
else:
gnn_results = mcts_state_map.mcts(verbose=True)
torch.save(gnn_results, result_path)
tree_node_x = find_closest_node_result(gnn_results, subgraph_max_nodes)
# calculate the metrics
original_node_list = [i for i in tree_node_x.ori_graph.nodes]
masked_node_list = [i for i in tree_node_x.ori_graph.nodes
if i not in tree_node_x.coalition or i == mcts_state_map.node_idx]
original_score = gnn_score(original_node_list, tree_node_x.data,
value_func=value_func, subgraph_building_method='zero_filling')
masked_score = gnn_score(masked_node_list, tree_node_x.data,
value_func=value_func, subgraph_building_method='zero_filling')
sparsity_score = 1 - len(tree_node_x.coalition)/tree_node_x.ori_graph.number_of_nodes()
fidelity_score_list.append(original_score - masked_score)
sparsity_score_list.append(sparsity_score)
# visualization
subgraph_node_labels = nx.get_node_attributes(tree_node_x.ori_graph, name='label')
subgraph_node_labels = torch.tensor([v for k, v in subgraph_node_labels.items()])
plotutils.plot(tree_node_x.ori_graph, tree_node_x.coalition, y=subgraph_node_labels,
node_idx=mcts_state_map.node_idx,
figname=os.path.join(save_dir, f"node_{node_idx}.png"))
fidelity_scores = torch.tensor(fidelity_score_list)
sparsity_scores = torch.tensor(sparsity_score_list)
return fidelity_scores, sparsity_scores
from keras.preprocessing.image import load_img, img_to_array, ImageDataGenerator
from pandas import DataFrame as df
from joblib import Parallel, delayed
from util import paths_to_tensor
import numpy as np
import multiprocessing
if __name__ == '__main__':
"""
Realiza a extração de features das imagens utilizando a ResNet50 pretreinada (paralela) - data aug
"""
model = ResNet50(include_top=False, weights='imagenet')
print("Carregando os dados...")
train_files, train_targets = get_dataset("train", 1, True)
valid_files, valid_targets = get_dataset("val", 1, True)
data_augmentation = True
files_names = [["train_aug.csv", "train_target_aug.csv"],
["val_aug.csv", "val_target_aug.csv"]]
block_size = 500
n_examples = [len(train_files), len(valid_files)]
for i_file in [0]:
files = [train_files, valid_files][i_file]
target = [train_targets, valid_targets][i_file]
print("Incializando o pre data Augmentation...")
datagen_train = ImageDataGenerator(
def pipeline(max_nodes):
dataset = get_dataset(data_args.dataset_dir, data_args.dataset_name)
plotutils = PlotUtils(dataset_name=data_args.dataset_name)
input_dim = dataset.num_node_features
output_dim = dataset.num_classes
if data_args.dataset_name == 'mutag':
data_indices = list(range(len(dataset)))
else:
loader = get_dataloader(dataset,
batch_size=train_args.batch_size,
random_split_flag=data_args.random_split,
data_split_ratio=data_args.data_split_ratio,
seed=data_args.seed)
data_indices = loader['test'].dataset.indices
gnnNets = GnnNets(input_dim, output_dim, model_args)
checkpoint = torch.load(mcts_args.explain_model_path)
gnnNets.update_state_dict(checkpoint['net'])
gnnNets.to_device()
gnnNets.eval()
save_dir = os.path.join(
'./results', f"{mcts_args.dataset_name}_"
f"{model_args.model_name}_"
f"{reward_args.reward_method}")
if not os.path.isdir(save_dir):
os.mkdir(save_dir)
fidelity_score_list = []
sparsity_score_list = []
for i in tqdm(data_indices):
# get data and prediction
data = dataset[i]
_, probs, _ = gnnNets(Batch.from_data_list([data.clone()]))
prediction = probs.squeeze().argmax(-1).item()
original_score = probs.squeeze()[prediction]
# get the reward func
value_func = GnnNets_GC2value_func(gnnNets, target_class=prediction)
payoff_func = reward_func(reward_args, value_func)
# find the paths and build the graph
result_path = os.path.join(save_dir, f"example_{i}.pt")
# mcts for l_shapely
mcts_state_map = MCTS(data.x,
data.edge_index,
score_func=payoff_func,
n_rollout=mcts_args.rollout,
min_atoms=mcts_args.min_atoms,
c_puct=mcts_args.c_puct,
expand_atoms=mcts_args.expand_atoms)
if os.path.isfile(result_path):
results = torch.load(result_path)
else:
results = mcts_state_map.mcts(verbose=True)
torch.save(results, result_path)
# l sharply score
graph_node_x = find_closest_node_result(results, max_nodes=max_nodes)
masked_node_list = [
node for node in list(range(graph_node_x.data.x.shape[0]))
if node not in graph_node_x.coalition
]
fidelity_score = original_score - gnn_score(
masked_node_list,
data,
value_func,
subgraph_building_method='zero_filling')
sparsity_score = 1 - len(
graph_node_x.coalition) / graph_node_x.ori_graph.number_of_nodes()
fidelity_score_list.append(fidelity_score)
sparsity_score_list.append(sparsity_score)
# visualization
if hasattr(dataset, 'supplement'):
words = dataset.supplement['sentence_tokens'][str(i)]
plotutils.plot(graph_node_x.ori_graph,
graph_node_x.coalition,
words=words,
figname=os.path.join(save_dir, f"example_{i}.png"))
else:
plotutils.plot(graph_node_x.ori_graph,
graph_node_x.coalition,
x=graph_node_x.data.x,
figname=os.path.join(save_dir, f"example_{i}.png"))
fidelity_scores = torch.tensor(fidelity_score_list)
sparsity_scores = torch.tensor(sparsity_score_list)
return fidelity_scores, sparsity_scores
state = torch.load("%s/%d.pth" % (opt.checkpoint_dir, opt.epoch))
model.load_state_dict(state.get('weight', False))
opt = state.get('opt')
opt.epoch = temp_opt.epoch
if opt.model_name == 'unet_nested':
criterion = BCEDiceLoss()
elif opt.n_class > 1:
criterion = nn.CrossEntropyLoss()
else:
criterion = nn.BCEWithLogitsLoss()
optimizer = optim.AdamW(model.parameters())
dataloader = DataLoader(
load_dataset.get_dataset(opt),
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.n_cpu,
)
Tensor = torch.cuda.FloatTensor if cuda else torch.Tensor
log_list = []
for epoch in range(opt.epoch, opt.n_epochs):
epoch_loss = 0
num_batches = len(dataloader)
for i, imgs in enumerate(dataloader):
# Configure model input
data = Variable(imgs["input"].type(Tensor))
true_mask = Variable(imgs["gt"].type(Tensor))
def pipeline_NC(top_k):
dataset = get_dataset(data_args.dataset_dir, data_args.dataset_name)
input_dim = dataset.num_node_features
output_dim = dataset.num_classes
data = dataset[0]
node_indices = torch.where(data.test_mask * data.y != 0)[0].tolist()
gnnNets = GnnNets_NC(input_dim, output_dim, model_args)
checkpoint = torch.load(model_args.model_path)
gnnNets.update_state_dict(checkpoint['net'])
gnnNets.to_device()
gnnNets.eval()
save_dir = os.path.join(
'./results', f"{data_args.dataset_name}_"
f"{model_args.model_name}_"
f"pgexplainer")
if not os.path.isdir(save_dir):
os.mkdir(save_dir)
pgexplainer = PGExplainer(gnnNets)
if torch.cuda.is_available():
torch.cuda.synchronize()
tic = time.perf_counter()
pgexplainer.get_explanation_network(dataset, is_graph_classification=False)
if torch.cuda.is_available():
torch.cuda.synchronize()
toc = time.perf_counter()
training_duration = toc - tic
print(f"training time is {training_duration}s ")
duration = 0.0
data = dataset[0]
fidelity_score_list = []
sparsity_score_list = []
plotutils = PlotUtils(dataset_name=data_args.dataset_name)
for ori_node_idx in tqdm(node_indices):
tic = time.perf_counter()
if glob.glob(os.path.join(save_dir, f"node_{ori_node_idx}.pt")):
file = glob.glob(os.path.join(save_dir,
f"node_{ori_node_idx}.pt"))[0]
edge_mask, x, edge_index, y, subset = torch.load(file)
edge_mask = torch.from_numpy(edge_mask)
node_idx = int(torch.where(subset == ori_node_idx)[0])
pred_label = pgexplainer.get_node_prediction(
node_idx, x, edge_index)
else:
x, edge_index, y, subset, kwargs = \
pgexplainer.get_subgraph(node_idx=ori_node_idx, x=data.x, edge_index=data.edge_index, y=data.y)
node_idx = int(torch.where(subset == ori_node_idx)[0])
edge_mask = pgexplainer.explain_edge_mask(x, edge_index)
pred_label = pgexplainer.get_node_prediction(
node_idx, x, edge_index)
save_path = os.path.join(save_dir, f"node_{ori_node_idx}.pt")
edge_mask = edge_mask.cpu()
cache_list = [edge_mask.numpy(), x, edge_index, y, subset]
torch.save(cache_list, save_path)
duration += time.perf_counter() - tic
sub_data = Data(x=x, edge_index=edge_index, y=y)
graph = to_networkx(sub_data)
fidelity_score = top_k_fidelity(sub_data, edge_mask, top_k, gnnNets,
pred_label)
sparsity_score = top_k_sparsity(sub_data, edge_mask, top_k)
fidelity_score_list.append(fidelity_score)
sparsity_score_list.append(sparsity_score)
# visualization
plotutils.plot_soft_edge_mask(graph,
edge_mask,
top_k,
y=sub_data.y,
node_idx=node_idx,
un_directed=True,
figname=os.path.join(
save_dir,
f"example_{ori_node_idx}.png"))
fidelity_scores = torch.tensor(fidelity_score_list)
sparsity_scores = torch.tensor(sparsity_score_list)
return fidelity_scores, sparsity_scores
def pipeline_GC(top_k):
dataset = get_dataset(data_args.dataset_dir, data_args.dataset_name)
if data_args.dataset_name == 'mutag':
data_indices = list(range(len(dataset)))
pgexplainer_trainset = dataset
else:
loader = get_dataloader(dataset,
batch_size=train_args.batch_size,
random_split_flag=data_args.random_split,
data_split_ratio=data_args.data_split_ratio,
seed=data_args.seed)
data_indices = loader['test'].dataset.indices
pgexplainer_trainset = loader['train'].dataset
input_dim = dataset.num_node_features
output_dim = dataset.num_classes
gnnNets = GnnNets(input_dim, output_dim, model_args)
checkpoint = torch.load(model_args.model_path)
gnnNets.update_state_dict(checkpoint['net'])
gnnNets.to_device()
gnnNets.eval()
save_dir = os.path.join(
'./results', f"{data_args.dataset_name}_"
f"{model_args.model_name}_"
f"pgexplainer")
if not os.path.isdir(save_dir):
os.mkdir(save_dir)
pgexplainer = PGExplainer(gnnNets)
if torch.cuda.is_available():
torch.cuda.synchronize()
tic = time.perf_counter()
pgexplainer.get_explanation_network(pgexplainer_trainset)
if torch.cuda.is_available():
torch.cuda.synchronize()
toc = time.perf_counter()
training_duration = toc - tic
print(f"training time is {training_duration: .4}s ")
explain_duration = 0.0
plotutils = PlotUtils(dataset_name=data_args.dataset_name)
fidelity_score_list = []
sparsity_score_list = []
for data_idx in tqdm(data_indices):
data = dataset[data_idx]
if torch.cuda.is_available():
torch.cuda.synchronize()
tic = time.perf_counter()
prob = pgexplainer.eval_probs(data.x, data.edge_index)
pred_label = prob.argmax(-1).item()
if glob.glob(os.path.join(save_dir, f"example_{data_idx}.pt")):
file = glob.glob(os.path.join(save_dir,
f"example_{data_idx}.pt"))[0]
edge_mask = torch.from_numpy(torch.load(file))
else:
edge_mask = pgexplainer.explain_edge_mask(data.x, data.edge_index)
save_path = os.path.join(save_dir, f"example_{data_idx}.pt")
edge_mask = edge_mask.cpu()
torch.save(edge_mask.detach().numpy(), save_path)
if torch.cuda.is_available():
torch.cuda.synchronize()
toc = time.perf_counter()
explain_duration += (toc - tic)
graph = to_networkx(data)
fidelity_score = top_k_fidelity(data, edge_mask, top_k, gnnNets,
pred_label)
sparsity_score = top_k_sparsity(data, edge_mask, top_k)
fidelity_score_list.append(fidelity_score)
sparsity_score_list.append(sparsity_score)
# visualization
if hasattr(dataset, 'supplement'):
words = dataset.supplement['sentence_tokens'][str(data_idx)]
plotutils.plot_soft_edge_mask(graph,
edge_mask,
top_k,
x=data.x,
words=words,
un_directed=True,
figname=os.path.join(
save_dir,
f"example_{data_idx}.png"))
else:
plotutils.plot_soft_edge_mask(graph,
edge_mask,
top_k,
x=data.x,
un_directed=True,
figname=os.path.join(
save_dir,
f"example_{data_idx}.png"))
fidelity_scores = torch.tensor(fidelity_score_list)
sparsity_scores = torch.tensor(sparsity_score_list)
return fidelity_scores, sparsity_scores
from keras.applications.resnet50 import ResNet50, preprocess_input, decode_predictions
from keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D, BatchNormalization
from keras.layers import Dropout, Flatten, Dense
from keras.models import Sequential
from keras.callbacks import ModelCheckpoint
from keras.preprocessing.image import load_img, img_to_array, ImageDataGenerator
from PIL import ImageFile
from joblib import Parallel, delayed
from util import paths_to_tensor
import numpy as np
import multiprocessing
if __name__ == "__main__":
print("Carregando os dados...")
valid_files, valid_targets = get_dataset("val")
# pre-process the data for Keras
print("Convertendo para tensor...")
valid_tensors = paths_to_tensor(valid_files).astype('float32') / 255.0
model = Sequential()
model.add(
ResNet50(weights='imagenet',
include_top=False,
input_shape=(224, 224, 3)))
model.add(Flatten(name='flatten'))
model.add(Dense(83, activation='softmax'))
model.compile(optimizer='rmsprop',
loss='binary_crossentropy',
checkpoints = [
f for f in glob.glob(os.path.join(opt.checkpoint_dir, '*.pth'))
]
for checkpoint in checkpoints:
print("Starting Checkpoint", checkpoint)
state = torch.load(checkpoint)
saved_opt = state.get('opt')
model = load_model.load_model(saved_opt)
model.load_state_dict(state.get('weight', False))
if cuda:
model.cuda()
saved_opt.test = True
dataloader = DataLoader(
load_dataset.get_dataset(saved_opt),
batch_size=saved_opt.batch_size,
shuffle=False,
num_workers=saved_opt.n_cpu,
)
for i, imgs in enumerate(dataloader):
row = []
# Configure model input
data = imgs["input"].type(Tensor)
true_mask = imgs["gt"].type(Tensor)
model.eval()
with torch.no_grad():
predicted_mask = model(data)
if saved_opt.deep_supervision:
def train_GC():
# attention the multi-task here
print('start loading data====================')
dataset = get_dataset(data_args)
input_dim = dataset.num_node_features
output_dim = int(dataset.num_classes)
dataloader = get_dataloader(dataset, data_args, train_args)
print('start training model==================')
gnnNets = GnnNets(input_dim, output_dim, model_args)
gnnNets.to_device()
criterion = nn.CrossEntropyLoss()
optimizer = Adam(gnnNets.parameters(),
lr=train_args.learning_rate,
weight_decay=train_args.weight_decay)
avg_nodes = 0.0
avg_edge_index = 0.0
for i in range(len(dataset)):
avg_nodes += dataset[i].x.shape[0]
avg_edge_index += dataset[i].edge_index.shape[1]
avg_nodes /= len(dataset)
avg_edge_index /= len(dataset)
print(
f"graphs {len(dataset)}, avg_nodes{avg_nodes :.4f}, avg_edge_index_{avg_edge_index/2 :.4f}"
)
best_acc = 0.0
data_size = len(dataset)
print(f'The total num of dataset is {data_size}')
# save path for model
if not os.path.isdir('checkpoint'):
os.mkdir('checkpoint')
if not os.path.isdir(os.path.join('checkpoint', data_args.dataset_name)):
os.mkdir(os.path.join('checkpoint', f"{data_args.dataset_name}"))
ckpt_dir = f"./checkpoint/{data_args.dataset_name}/"
early_stop_count = 0
for epoch in range(train_args.max_epochs):
acc = []
loss_list = []
gnnNets.train()
for batch in dataloader['train']:
logits, probs, _ = gnnNets(batch)
loss = criterion(logits, batch.y)
# optimization
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_value_(gnnNets.parameters(),
clip_value=2.0)
optimizer.step()
## record
_, prediction = torch.max(logits, -1)
loss_list.append(loss.item())
acc.append(prediction.eq(batch.y).cpu().numpy())
# report train msg
print(f"Train Epoch:{epoch} |Loss: {np.average(loss_list):.3f} | "
f"Acc: {np.concatenate(acc, axis=0).mean():.3f}")
# report eval msg
eval_state = evaluate_GC(dataloader['eval'], gnnNets, criterion)
print(
f"Eval Epoch: {epoch} | Loss: {eval_state['loss']:.3f} | Acc: {eval_state['acc']:.3f}"
)
# only save the best model
is_best = (eval_state['acc'] > best_acc)
if eval_state['acc'] > best_acc:
early_stop_count = 0
else:
early_stop_count += 1
if early_stop_count > train_args.early_stopping:
break
if is_best:
best_acc = eval_state['acc']
early_stop_count = 0
if is_best or epoch % train_args.save_epoch == 0:
save_best(ckpt_dir, epoch, gnnNets, model_args.model_name,
eval_state['acc'], is_best)
print(f"The best validation accuracy is {best_acc}.")
# report test msg
checkpoint = torch.load(
os.path.join(ckpt_dir, f'{model_args.model_name}_best.pth'))
gnnNets.update_state_dict(checkpoint['net'])
test_state, _, _ = test_GC(dataloader['test'], gnnNets, criterion)
print(
f"Test: | Loss: {test_state['loss']:.3f} | Acc: {test_state['acc']:.3f}"
)
def train_NC():
print('start loading data====================')
import pdb
pdb.set_trace()
dataset = get_dataset(data_args)
input_dim = dataset.num_node_features
output_dim = int(dataset.num_classes)
# save path for model
if not os.path.isdir('checkpoint'):
os.mkdir('checkpoint')
if not os.path.isdir(
os.path.join('checkpoint', f"{data_args.dataset_name}")):
os.mkdir(os.path.join('checkpoint', f"{data_args.dataset_name}"))
ckpt_dir = f"./checkpoint/{data_args.dataset_name}/"
data = dataset[0]
gnnNets_NC = GnnNets_NC(input_dim, output_dim, model_args)
gnnNets_NC.to_device()
criterion = nn.CrossEntropyLoss()
optimizer = Adam(gnnNets_NC.parameters(),
lr=train_args.learning_rate,
weight_decay=train_args.weight_decay)
best_val_loss = float('inf')
best_acc = 0
val_loss_history = []
early_stop_count = 0
for epoch in range(1, train_args.max_epochs + 1):
gnnNets_NC.train()
logits, prob, _ = gnnNets_NC(data)
loss = criterion(logits[data.train_mask], data.y[data.train_mask])
optimizer.zero_grad()
loss.backward()
optimizer.step()
eval_info = evaluate_NC(data, gnnNets_NC, criterion)
eval_info['epoch'] = epoch
if eval_info['val_loss'] < best_val_loss:
best_val_loss = eval_info['val_loss']
val_acc = eval_info['val_acc']
val_loss_history.append(eval_info['val_loss'])
# only save the best model
is_best = (eval_info['val_acc'] > best_acc)
if eval_info['val_acc'] > best_acc:
early_stop_count = 0
else:
early_stop_count += 1
if early_stop_count > train_args.early_stopping:
break
if is_best:
best_acc = eval_info['val_acc']
if is_best or epoch % train_args.save_epoch == 0:
save_best(ckpt_dir, epoch, gnnNets_NC, model_args.model_name,
eval_info['val_acc'], is_best)
print(f'Epoch {epoch}, Train Loss: {eval_info["train_loss"]:.4f}, '
f'Train Accuracy: {eval_info["train_acc"]:.3f}, '
f'Val Loss: {eval_info["val_loss"]:.3f}, '
f'Val Accuracy: {eval_info["val_acc"]:.3f}')
# report test msg
checkpoint = torch.load(
os.path.join(ckpt_dir, f'{model_args.model_name}_best.pth'))
gnnNets_NC.update_state_dict(checkpoint['net'])
eval_info = evaluate_NC(data, gnnNets_NC, criterion)
print(
f'Test Loss: {eval_info["test_loss"]:.4f}, Test Accuracy: {eval_info["test_acc"]:.3f}'
)
from pandas import DataFrame as df
from joblib import Parallel, delayed
from util import paths_to_tensor
import numpy as np
import multiprocessing
if __name__ == '__main__':
"""
Realiza a extração de features das imagens utilizando a ResNet50 pretreinada (paralela)
"""
model = ResNet50(include_top=False, weights='imagenet')
print("Carregando os dados...")
train_files, train_targets = get_dataset("train")
valid_files, valid_targets = get_dataset("val")
files_names = ["dataset_train3.csv", "dataset_val3.csv"]
block_size = 500
# 0 - train, 1 - validação
for i_file in [0, 1]:
files = [train_files, valid_files][i_file]
count = int(np.ceil(len(files) / block_size))
for i in range(count):
print("{0} lotte: {1}/{2}".format(files_names[i_file], i, count))
begin = i * block_size