np.set_printoptions(formatter={'float_kind': "{:.4f}".format})

with torch.no_grad():

net.eval()

# load weights

load_weight(model_file, net)

final_res = {'version': 'VERSION 1.3', 'results': {},

'external_data': {'used': True, 'details': 'Features from I3D Network'}}

num_correct = 0.

num_total = 0.

for _data, _label, _, vid_name, vid_num_seg in test_loader:

batch_size = _data.shape[0]

vid = vid_name[0]

_data = _data.cuda()

_label = _label.cuda()

# FORWARD PASS

cas, (action_flow, action_rgb) = net(_data, is_training=False)

combined_cas = misc_utils.instance_selection_function(torch.softmax(cas.detach(), -1), action_flow.permute(0, 2, 1).detach(), action_rgb.permute(0, 2, 1))

_, topk_indices = torch.topk(combined_cas, config.num_segments // 8, dim=1)

# class prediction

cas_top = torch.gather(cas, 1, topk_indices)

cas_top = torch.mean(cas_top, dim=1)

score_supp = F.softmax(cas_top, dim=1)

label_np = _label.cpu().numpy()

score_np = score_supp[0, :].cpu().numpy()

score_np[np.where(score_np < config.class_thresh)] = 0

score_np[np.where(score_np >= config.class_thresh)] = 1

if np.all(score_np == 0):

arg = np.argmax(score_supp[0, :].cpu().data.numpy())

score_np[arg] = 1

correct_pred = np.sum(label_np == score_np, axis=1)

num_correct += np.sum((correct_pred == config.num_classes).astype(np.float32))

num_total += correct_pred.shape[0]

pred = np.where(score_np > config.class_thresh)[0]

# action prediction

if len(pred) != 0:

cas_pred = combined_cas[0].cpu().numpy()[:, pred]

cas_pred = np.reshape(cas_pred, (config.num_segments, -1, 1))

# cas_pred = misc_utils.upgrade_resolution(cas_pred, config.scale)

proposal_dict = {}

for t in range(len(config.act_thresh)):

cas_temp = cas_pred.copy()

zero_location = np.where(cas_temp[:, :, 0] < config.act_thresh[t])

cas_temp[zero_location] = 0

cas_input = cas_pred.copy()

seg_list = []

for c in range(len(pred)):

pos = np.where(cas_temp[:, c, 0] > 0)

seg_list.append(pos)

proposals = misc_utils.get_proposal_oic(seg_list, cas_pred.copy(), score_supp[0, :].cpu().data.numpy(),

pred, config.scale, vid_num_seg[0].cpu().item(), config.fps,

config.num_segments, config.gamma)

for j in range(len(proposals)):

if not proposals[j]:

continue

class_id = proposals[j][0][0]

if class_id not in proposal_dict.keys():

proposal_dict[class_id] = []

proposal_dict[class_id] += proposals[j]

final_proposals = []

for class_id in proposal_dict.keys():

final_proposals.append(misc_utils.basnet_nms(proposal_dict[class_id], config.nms_thresh,

config.soft_nms, config.nms_alpha))

final_res['results'][vid_name[0]] = misc_utils.result2json(final_proposals)

test_acc = num_correct / num_total

json_path = os.path.join(config.model_path, 'temp_result.json')

with open(json_path, 'w') as f:

json.dump(final_res, f)

f.close()

mean_ap, _ = evaluate(config.gt_path, json_path, None, tiou_thresholds=np.linspace(0.1, 0.9, 9), plot=False,

subset='test', verbose=config.verbose)

print(_)