def evaluate(args, exe, test_prog, model, test_dataset, featurizer):
"""
Define the evaluate function
In the dataset, a proportion of labels are blank. So we use a `valid` tensor
to help eliminate these blank labels in both training and evaluation phase.
"""
data_gen = test_dataset.iter_batch(batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=True,
collate_fn=featurizer.collate_fn)
total_pred = []
total_label = []
total_valid = []
for batch_id, feed_dict in enumerate(data_gen):
pred, = exe.run(test_prog,
feed=feed_dict,
fetch_list=[model.pred],
return_numpy=False)
total_pred.append(np.array(pred))
total_label.append(feed_dict['finetune_label'])
total_valid.append(feed_dict['valid'])
total_pred = np.concatenate(total_pred, 0)
total_label = np.concatenate(total_label, 0)
total_valid = np.concatenate(total_valid, 0)
return calc_rocauc_score(total_label, total_pred, total_valid)
def evaluate(args, exe, test_prog, agent, test_data_list, epoch_id):
"""Evaluate the model on test dataset."""
collate_fn = MoleculeCollateFunc(agent.graph_wrapper,
task_type='cls',
num_cls_tasks=args.num_tasks,
with_graph_label=True,
with_pos_neg_mask=False)
data_loader = Dataloader(test_data_list,
batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=False,
collate_fn=collate_fn)
total_data, eval_data = len(test_data_list), 0
total_pred, total_label, total_valid = [], [], []
for batch_id, feed_dict in enumerate(data_loader):
pred, = exe.run(test_prog,
feed=feed_dict,
fetch_list=[agent.pred],
return_numpy=False)
total_pred.append(np.array(pred))
total_label.append(feed_dict['label'])
total_valid.append(feed_dict['valid'])
total_pred = np.concatenate(total_pred, 0)
total_label = np.concatenate(total_label, 0)
total_valid = np.concatenate(total_valid, 0)
return calc_rocauc_score(total_label, total_pred, total_valid)
def evaluate(args, exe, test_prog, model, test_dataset, featurizer):
"""tbd"""
data_gen = test_dataset.iter_batch(batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=True,
collate_fn=featurizer.collate_fn)
total_pred = []
total_label = []
total_valid = []
for batch_id, feed_dict in enumerate(data_gen):
pred, = exe.run(test_prog,
feed=feed_dict,
fetch_list=[model.pred],
return_numpy=False)
total_pred.append(np.array(pred))
total_label.append(feed_dict['finetune_label'])
total_valid.append(feed_dict['valid'])
total_pred = np.concatenate(total_pred, 0)
total_label = np.concatenate(total_label, 0)
total_valid = np.concatenate(total_valid, 0)
return calc_rocauc_score(total_label, total_pred, total_valid)