train_ds.read()
dev_ds.read()
test_ds = None
if args.test is not None:
test_ds = DataSet(file=args.test, reader=jlr, formatter=formatter)
test_ds.read()
train_feats, dev_feats, test_feats = f.load(train_ds, dev_ds, test_ds)
f.save_vocab(mname)
input_shape = train_feats[0].shape[1]
model = SimpleMLP(input_shape,100,3)
if gpu():
model.cuda()
if model_exists(mname) and os.getenv("TRAIN").lower() not in ["y","1","t","yes"]:
model.load_state_dict(torch.load("models/{0}.model".format(mname)))
else:
train(model, train_feats, 500, 1e-2, 90,dev_feats,early_stopping=EarlyStopping(mname))
torch.save(model.state_dict(), "models/{0}.model".format(mname))
print_evaluation(model, dev_feats, FEVERLabelSchema())
if args.test is not None:
print_evaluation(model, test_feats, FEVERLabelSchema())
waseem_te_composite = CompositeDataset(name="waseem_composite_test")
for dataset in datasets_te:
dataset.read()
waseem_te_composite.add(dataset)
davidson_te = DataSet(os.path.join("data","davidson.te.csv"),reader=csvreader,formatter=df,name="davidson_test")
davidson_te.read()
features = Features(get_feature_functions(mname))
train_fs, dev_fs, test_fs_primary, test_fs_aux = features.load(waseem_tr_composite, waseem_de_composite, waseem_te_composite, davidson_te)
print("Number of features in primary: {0}".format(train_fs[0].shape[1]))
model = MLP(train_fs[0].shape[1],get_model_shape(),3)
if gpu():
model.cuda()
if model_exists(mname) and os.getenv("TRAIN").lower() not in ["y","1","t","yes"]:
model.load_state_dict(torch.load("models/{0}.model".format(mname)))
else:
train(model, train_fs, 50, 1e-3, 45, dev=dev_fs,
lr_schedule=lambda a, b: exp_lr_scheduler(a, b, 0.5, 5))
torch.save(model.state_dict(), "models/{0}.model".format(mname))
create_log_dir(mname)
print_evaluation(model,test_fs_primary, WaseemLabelSchema(),log="logs/{0}/primary.jsonl".format(mname))
print_evaluation(model,test_fs_aux, WaseemLabelSchema(),log="logs/{0}/aux.jsonl".format(mname))
}
if __name__ == "__main__":
bodies = Bodies("data/fnc-1/train_bodies.csv",
"data/fnc-1/competition_test_bodies.csv")
f = Features([FNCTermFrequencyFeatureFunction(bodies)])
csvr = CSVReader()
formatter = FNCFormatter(FNCLabelSchema())
train_ds = DataSet(file="data/fnc-1/train_stances.csv",
reader=csvr,
formatter=formatter)
test_ds = DataSet(file="data/fnc-1/competition_test_stances.csv",
reader=csvr,
formatter=formatter)
train_ds.read()
test_ds.read()
train_feats, _, test_feats = f.load(train_ds, None, test_ds)
input_shape = train_feats[0].shape[1]
model = SimpleMLP(input_shape, 100, 4)
if gpu():
model.cuda()
train(model, train_feats, 500, 1e-2, 90, test_feats, clip=5)
print("Number of features: {0}".format(train_fs[0].shape[1]))
model = MLP(train_fs[0].shape[1], get_model_shape(), 3)
if gpu():
model.cuda()
if model_exists(mname) and os.getenv("TRAIN").lower() not in [
"y", "1", "t", "yes"
]:
model.load_state_dict(torch.load("models/{0}.model".format(mname)))
else:
train(model,
train_fs,
50,
1e-3,
60,
dev=dev_fs,
early_stopping=EarlyStopping(mname),
lr_schedule=lambda a, b: exp_lr_scheduler(a, b, 0.5, 5))
torch.save(model.state_dict(), "models/{0}.model".format(mname))
create_log_dir(mname)
print_evaluation(model,
dev_fs,
DavidsonLabelSchema(),
log="logs/{0}/dev.jsonl".format(mname))
print_evaluation(model,
test_fs,
DavidsonLabelSchema(),
log="logs/{0}/test.jsonl".format(mname))
reader=jlr,
formatter=formatter)
train_ds.read()
dev_ds.read()
test_ds.read()
train_feats, dev_feats, test_feats = f.load(train_ds, dev_ds, test_ds)
input_shape = train_feats[0].shape[1]
model = SimpleMLP(input_shape, 100, 3)
if gpu():
model.cuda()
if model_exists(mname):
model.load_state_dict(torch.load("models/{0}.model".format(mname)))
final_model = model
else:
final_model = train(model,
train_feats,
500,
1e-2,
90,
dev_feats,
early_stopping=EarlyStopping())
torch.save(model.state_dict(), "models/{0}.model".format(mname))
print_evaluation(final_model, dev_feats, FEVERLabelSchema())
print_evaluation(final_model, test_feats, FEVERLabelSchema())