def evaluation(self, batch_data, model_outputs, *args, **kwargs):
from imix.models.vqa_models.mcan_mix import list2dict
from imix.engine.organizer import is_by_iter
if is_by_iter():
batch_data = list2dict(batch_data)
return [{'batch_score': model_outputs['batch_score']}], [{'batch_size': model_outputs['batch_size']}]
def evaluation(self, batch_data, model_outputs, *args, **kwargs):
from imix.models.vqa_models.mcan_mix import list2dict
from imix.engine.organizer import is_by_iter
if is_by_iter():
batch_data = list2dict(batch_data)
eval_result = self.evaluate(batch_data, model_outputs)
return eval_result
def evaluation(self, batch_data, model_outputs, *args, **kwargs):
from imix.models.vqa_models.mcan_mix import list2dict
from imix.engine.organizer import is_by_iter
if is_by_iter():
batch_data = list2dict(batch_data)
labels = list(batch_data['answers_scores'].split(1))
q_ids, scores = batch_data['question_id'].split(1), model_outputs['scores'].to('cpu').split(1)
predictions = list({'question_id': q_id, 'scores': score} for q_id, score in zip(q_ids, scores))
# predictions, labels = self.data_pre_process(predictions, labels, *args,
# **kwargs)
predictions = self.data_pre_process(predictions, *args, **kwargs)
return predictions, labels
def evaluation(self, batch_data, model_outputs, *args, **kwargs):
from imix.models.vqa_models.mcan_mix import list2dict
from imix.engine.organizer import is_by_iter
if is_by_iter():
batch_data = list2dict(batch_data)
predictions = []
ques_id = batch_data['ques_id']
score, label = model_outputs['scores'].max(1)
for qid, l in zip(ques_id, label.cpu().numpy()):
if torch.is_tensor(qid):
predictions.append({qid.item(): l})
else:
predictions.append({qid: l})
return predictions, score
def preprocess_data(self, batched_inputs: Dict) -> Dict:
"""padding data and turn data into cuda format.
Args:
batched_inputs: a Dict containing input data
Returns:
Dict: a dict containing input data
"""
from imix.engine.organizer import is_by_iter
if is_by_iter():
batched_inputs = list2dict(batched_inputs)
img_feat = batched_inputs['features']
input_ids = batched_inputs['input_ids']
input_mask = batched_inputs['input_mask']
b, c, h, w = img_feat.shape
feat = img_feat.view(b, c, -1)
padded_feat = torch.zeros((b, c, 1024), dtype=torch.float)
padded_feat[:, :, :h * w] = feat
feat = padded_feat.unsqueeze(-1)
# feat = feat.squeeze(0)
feat = feat.cuda()
input_ids = input_ids.cuda()
input_mask = input_mask.cuda()
batched_inputs['feature'] = feat
batched_inputs['input_ids'] = input_ids
batched_inputs['input_mask'] = ~input_mask
if self.training:
answers_scores = batched_inputs['answers_scores']
answers_scores = answers_scores.cuda()
batched_inputs['answers_scores'] = answers_scores
return batched_inputs
def submit(self, batch_data, model_outputs, *args, **kwargs):
# scores, labels = model_outputs['scores'].max(1)
# q_ids = batch_data['question_id'].detach().numpy()
# labels = labels.cpu().detach().numpy()
# q2a = batch_data['quesid2ans']
# predictions = list({
# 'question_id': int(qid),
# 'answer': q2a[l][0]
# } for qid, l in zip(q_ids, labels))
from imix.models.vqa_models.mcan_mix import list2dict
from imix.engine.organizer import is_by_iter
if is_by_iter():
batch_data = list2dict(batch_data)
q_ids, scores = batch_data['question_id'].split(1), model_outputs['scores'].to('cpu').split(1)
predictions = list({'question_id': q_id, 'scores': score} for q_id, score in zip(q_ids, scores))
predictions = self.data_pre_process(predictions, *args, **kwargs)
# by yinyin q_ids should be question str;q2a should be the answer str
q2a = batch_data['quesid2ans']
predictions = list({
'questionid': str(qid),
'prediction': str(m[0][l])
} for qid, l, m in zip(q_ids, predictions, q2a))
return predictions
