def main():
args = configs.arg_parse()
fix_seed(args.seed)
# Load the dataset
data = prepare_data(args.dataset, args.train_ratio, args.input_dim,
args.seed)
# Define and train the model
if args.dataset in ['Cora', 'PubMed']:
# Retrieve the model and training hyperparameters depending the data/model given as input
hyperparam = ''.join(['hparams_', args.dataset, '_', args.model])
param = ''.join(['params_', args.dataset, '_', args.model])
model = eval(args.model)(input_dim=data.num_features,
output_dim=data.num_classes,
**eval(hyperparam))
train_and_val(model, data, **eval(param))
_, test_acc = evaluate(data, model, data.test_mask)
print('Test accuracy is {:.4f}'.format(test_acc))
elif args.dataset in ['syn6', 'Mutagenicity']:
input_dims = data.x.shape[-1]
model = GcnEncoderGraph(input_dims,
args.hidden_dim,
args.output_dim,
data.num_classes,
args.num_gc_layers,
bn=args.bn,
dropout=args.dropout,
args=args)
train_gc(data, model, args)
_, test_acc = evaluate(data, model, data.test_mask)
print('Test accuracy is {:.4f}'.format(test_acc))
else:
# For pytorch geometric model
#model = GCNNet(args.input_dim, args.hidden_dim,
# data.num_classes, args.num_gc_layers, args=args)
input_dims = data.x.shape[-1]
model = GcnEncoderNode(data.num_features,
args.hidden_dim,
args.output_dim,
data.num_classes,
args.num_gc_layers,
bn=args.bn,
dropout=args.dropout,
args=args)
train_syn(data, model, args)
_, test_acc = evaluate(data, model, data.test_mask)
print('Test accuracy is {:.4f}'.format(test_acc))
# Save model
model_path = 'models/{}_model_{}.pth'.format(args.model, args.dataset)
if not os.path.exists(model_path) or args.save == True:
torch.save(model, model_path)
def main():
args = configs.arg_parse()
fix_seed(args.seed)
# Load the dataset
data = prepare_data(args.dataset, args.train_ratio,
args.input_dim, args.seed)
# Load the model
model_path = 'models/{}_model_{}.pth'.format(args.model, args.dataset)
model = torch.load(model_path)
# Evaluate the model
if args.dataset in ['Cora', 'PubMed']:
_, test_acc = evaluate(data, model, data.test_mask)
else:
test_acc = test(data, model, data.test_mask)
print('Test accuracy is {:.4f}'.format(test_acc))
# Explain it with GraphSVX
explainer = GraphSVX(data, model, args.gpu)
# Distinguish graph classfication from node classification
if args.dataset in ['Mutagenicity', 'syn6']:
explanations = explainer.explain_graphs(args.indexes,
args.hops,
args.num_samples,
args.info,
args.multiclass,
args.fullempty,
args.S,
'graph_classification',
args.feat,
args.coal,
args.g,
args.regu,
True)
else:
explanations = explainer.explain(args.indexes,
args.hops,
args.num_samples,
args.info,
args.multiclass,
args.fullempty,
args.S,
args.hv,
args.feat,
args.coal,
args.g,
args.regu,
True)
print('Sum explanations: ', [np.sum(explanation) for explanation in explanations])
print('Base value: ', explainer.base_values)
'reg_lambda': .1,
'subsample': .9,
'min_split_gain': .01,
'min_child_weight': 2,
'colsample_bytree': .9, # Subsample ratio of columns when constructing each tree.
'scale_pos_weight': 9, # because training data is unbalanced
'verbose': -1
}
features = list(pd.read_csv(DATA_FOLDER + '/v3/importances.csv', index_col=0).head(800).index)
train_features = [*features, "target"]
train = pd.read_pickle(DATA_FOLDER + '/v3/train.pkl')[train_features]
folds = prepare_folds(train)
models, result = train_folds(folds, config)
test = load_test(DATA_FOLDER + '/v3/test.pkl')[features]
test_target = evaluate(models, test)
print("AUC: %.4f, F1: %.4f" % (result['auc'], result['f1']))
importance = result['importances'].groupby(['feature']) \
.agg({'importance': 'mean'}) \
.sort_values(by="importance", ascending=False)
importance.to_csv(DATA_FOLDER + "/v3/importances.csv")
prepare_submission(test_target, "v3_AUC_%.4f_F1_%.4f" % (result['auc'], result['f1']))
gold_starts=gold_starts,
gold_ends=gold_ends,
cluster_ids=cluster_ids)
loader = DataLoader(
dataset,
batch_size=1,
shuffle=False
)
model.load_state_dict(torch.load(args['model_path']))
# print(model)
print(model.parameters)
evaluate(loader, model, device)
evaluate(loader, model, device)
# if args['mode'] == 'predict':
# all_sentences, tokens, batch_indices, mentions, gold_starts, gold_ends, clusters = process_ecb_plus(args['data_path'], args['mention_type'])
# tokenizer = AutoTokenizer.from_pretrained("SpanBERT/spanbert-base-cased", use_fast=True)
# encodings = tokenizer(all_sentences, return_offsets_mapping=True, is_split_into_words=True, truncation=True, padding=True)
# encoded_tokens = fix_tokens_with_offsets(tokens, encodings.offset_mapping, batch_indices)
# encoded_sentence_map = create_unmasked_sentence_map(encodings.offset_mapping, batch_indices)
# encoded_tokens, gold_starts, gold_ends, mentions = process_gold_mentions(encoded_tokens, gold_starts, gold_ends, mentions, encodings.attention_mask, batch_indices)
# cluster_ids = get_cluster_ids(mentions, clusters)
# dataset = ECBDataset(
# encodings=encodings,
# batch_indices=batch_indices,
# sentence_map=encoded_sentence_map,
