def get_nms_preds(b_clas, b_bb, idx, anchors):
a_ic = actn_to_bb(b_bb[idx], anchors, grid_sizes)
clas_pr, clas_ids = b_clas[idx].max(1)
clas_pr = clas_pr.sigmoid()
conf_scores = b_clas[idx].sigmoid().t().data
out1, out2, cc = [], [], []
for cl in range(0, len(conf_scores)-1):
c_mask = conf_scores[cl] > 0.1
if c_mask.sum() == 0: continue
scores = conf_scores[cl][c_mask]
l_mask = c_mask.unsqueeze(1).expand_as(a_ic)
boxes = a_ic[l_mask].view(-1, 4)
ids, count = nms(boxes.data, scores, 0.1, 50)
ids = ids[:count]
out1.append(scores[ids])
out2.append(boxes.data[ids])
cc.append([cl]*count)
if cc == []:
cc = [[0]]
cc = T(np.concatenate(cc))
if out1 == []:
out1 = [torch.Tensor()]
out1 = torch.cat(out1)
if out2 == []:
out2 = [torch.Tensor()]
out2 = torch.cat(out2)
bbox, clas, prs, thresh = out2, cc, out1, 0.
return predictions(to_np(bbox),
to_np(clas), to_np(prs) if prs is not None else None, thresh)
def predictions(bbox, clas=None, prs=None, thresh=0.3):
#bb = [bb_hw(o) for o in bbox.reshape(-1,4)]
bb = bbox
if prs is None: prs = [None]*len(bb)
if clas is None: clas = [None]*len(bb)
predictions = []
for i, (b, c, pr) in enumerate(zip(bb, clas, prs)):
if((b[2]>0) and (pr is None or pr > thresh)):
cat_str = 'pool' #'bg' if c[i]==len(id2cat) else id2cat[c[i]]
score = pr*100
bb_np = to_np(b).astype('float64')
predictions.append(pred2dict(bb_np, score, cat_str))
return predictions
def torch_gt(ax, ima, bbox, clas, prs=None, thresh=0.4):
return show_ground_truth(ax, ima, to_np((bbox*224).long()),to_np(clas), to_np(prs) if prs is not None else None, thresh)
model.load_state_dict(torch.load(dir))
if torch.cuda.is_available():
model.cuda()
model.eval()
test_sub = np.zeros((len(label_df['id']), 3), dtype=np.float)
batch = len(label_df['id']) // args.batch_size
for i, (test_data, event_labels) in enumerate(test_loader):
test_text, test_mask = to_var(test_data[0]), to_var(test_data[1])
test_text = test_text.long()
test_mask = test_mask.float()
test_outputs, domain_outputs = model(test_text, test_mask)
if i != batch:
test_sub[i * args.batch_size:(i + 1) *
args.batch_size, :] = to_np(test_outputs)
else:
test_sub[i *
args.batch_size:len(test_df['id']), :] = to_np(test_outputs)
score_t = 0
for i in range(len(label_df)):
score_t += np.dot(np.log(test_sub[i, :]),
label_df.loc[i, TARGET].T) / len(label_df)
score = 1 / (1 - score_t)
print(score)
test_pred = np.argmax(test_sub, axis=1)
test_true = [get_label(item[TARGET]) for index, item in label_df.iterrows()]
test_accuracy = metrics.accuracy_score(test_true, test_pred)
test_f1 = metrics.f1_score(test_true, test_pred, average='macro')
