def evaluate(vocab, decoder, eval_range, prediction_txt_path, reference):
# 载入测试数据集
eval_loader = get_eval_loader(eval_range, feature_h5_path)
result = {}
for i, (videos, video_ids) in enumerate(eval_loader):
# 构造mini batch的Variable
videos = Variable(videos)
if use_cuda:
videos = videos.cuda()
outputs, attens = decoder.sample(videos)
outputs = outputs.data.squeeze(2)
for (tokens, vid) in zip(outputs, video_ids):
s = decoder.decode_tokens(tokens)
result[vid] = s
prediction_txt = open(prediction_txt_path, 'w')
for vid, s in result.items():
prediction_txt.write('%d\t%s\n' % (vid, s)) # 注意，MSVD数据集的视频文件名从1开始
prediction_txt.close()
# 开始根据生成的结果计算各种指标
metrics = measure(prediction_txt_path, reference)
return metrics
def evaluate(vocab, banet, eval_range, prediction_txt_path, reference):
# 载入测试数据集
eval_loader = get_eval_loader(eval_range, feature_h5_path)
result = {}
for i, (videos, video_ids) in enumerate(eval_loader):
# 构造mini batch的Variable
videos = Variable(videos)
if use_cuda:
videos = videos.cuda()
outputs, _ = banet(videos, None)
for (tokens, vid) in zip(outputs, video_ids):
s = banet.decoder.decode_tokens(tokens.data)
result[vid] = s
prediction_txt = open(prediction_txt_path, 'w')
for vid, s in result.items():
prediction_txt.write('%d\t%s\n' % (vid, s)) # 注意，MSVD数据集的视频文件名从1开始
prediction_txt.close()
# 开始根据生成的结果计算各种指标
metrics = measure(prediction_txt_path, reference)
return metrics
def test_trees(model_path):
""" use the trained model to generate parse trees for text """
# load model and options
checkpoint = torch.load(model_path, map_location='cpu')
opt = checkpoint['opt']
# load vocabulary used by the model
vocab = pickle.load(open(os.path.join(opt.data_path, 'vocab.pkl'), 'rb'))
opt.vocab_size = len(vocab)
# construct model
model = VGNSL(opt)
# load model state
model.load_state_dict(checkpoint['model'])
print('Loading dataset')
data_loader = get_eval_loader(
opt.data_path, 'test', vocab, opt.batch_size, opt.workers,
load_img=False, img_dim=opt.img_dim
)
cap_embs = None
logged = False
trees = list()
for i, (images, captions, lengths, ids) in enumerate(data_loader):
# make sure val logger is used
model.logger = print
lengths = torch.Tensor(lengths).long()
if torch.cuda.is_available():
lengths = lengths.cuda()
# compute the embeddings
model_output = model.forward_emb(images, captions, lengths, volatile=True)
img_emb, cap_span_features, left_span_features, right_span_features, word_embs, tree_indices, all_probs, \
span_bounds = model_output[:8]
candidate_trees = list()
for j in range(len(ids)):
candidate_trees.append(generate_tree(captions, tree_indices, j, vocab))
appended_trees = ['' for _ in range(len(ids))]
for j in range(len(ids)):
appended_trees[ids[j] - min(ids)] = clean_tree(candidate_trees[j])
trees.extend(appended_trees)
cap_emb = torch.cat([cap_span_features[l-2][i].reshape(1, -1) for i, l in enumerate(lengths)], dim=0)
del images, captions, img_emb, cap_emb
ground_truth = [line.strip() for line in open(
os.path.join(opt.data_path, 'test_ground-truth.txt'))]
return trees, ground_truth
def evaluate(model, file_path, labels_name,num_sample):
print 'loading test data...'
hashcode = np.zeros((num_sample,nbits),dtype = np.float32)
label_array = Array()
hashcode_array = Array()
rem = num_sample%test_batch_size
labels = sio.loadmat(labels_name)['labels']
eval_loader = get_eval_loader(file_path,batch_size=test_batch_size)
label_array.setmatrcs(labels)
batch_num = len(eval_loader)
time0 = time.time()
for i, data in enumerate(eval_loader):
data = {key: value.cuda() for key, value in data.items()}
my_H,_,_ = model.forward(data["visual_word"])
my_H = torch.mean(my_H,1)
BinaryCode = torch.sign(my_H)
if i == batch_num-1:
hashcode[i*test_batch_size:,:] = BinaryCode[:rem,:].data.cpu().numpy()
else:
hashcode[i*test_batch_size:(i+1)*test_batch_size,:] = BinaryCode.data.cpu().numpy()
test_hashcode = np.matrix(hashcode)
time1 = time.time()
print 'retrieval costs: ',time1-time0
Hamming_distance = 0.5*(-np.dot(test_hashcode,test_hashcode.transpose())+nbits)
time2 = time.time()
print 'hamming distance computation costs: ',time2-time1
HammingRank = np.argsort(Hamming_distance, axis=0)
time3 = time.time()
print 'hamming ranking costs: ',time3-time2
labels = label_array.getmatrics()
print 'labels shape: ',labels.shape
sim_labels = np.dot(labels, labels.transpose())
time6 = time.time()
print 'similarity labels generation costs: ', time6 - time3
records = open('./results/64_9288_2021.txt','w+')
maps = []
map_list = [5,10,20,40,60,80,100]
for i in map_list:
map,_,_ = tools.mAP(sim_labels, HammingRank,i)
maps.append(map)
records.write('topK: '+str(i)+'\tmap: '+str(map)+'\n')
print 'i: ',i,' map: ', map,'\n'
time7 = time.time()
records.close()
def save_nf(model):
'''
To prepare latent video features, you can first train BTH model
with only mask_loss and save features with this function.
'''
num_sample = 45585 # number of training videos
new_feats = np.zeros((num_sample,hidden_size),dtype = np.float32)
rem = num_sample%test_batch_size
eval_loader = get_eval_loader(train_feat_path,batch_size=test_batch_size)
batch_num = len(eval_loader)
for i, data in enumerate(eval_loader):
data = {key: value.cuda() for key, value in data.items()}
_,_,x = model.forward(data["visual_word"])
feat = torch.mean(x,1)
if i == batch_num-1:
new_feats[i*test_batch_size:,:] = feat[:rem,:].data.cpu().numpy()
else:
new_feats[i*test_batch_size:(i+1)*test_batch_size,:] = feat.data.cpu().numpy()
h5 = h5py.File(latent_feat_path, 'w')
h5.create_dataset('feats', data = new_feats)
h5.close()
img_names = os.path.join(video_path, str(j) + '.jpg')
im = Image.open(img_names)
im_re = im.resize((image_width, image_width),
Image.ANTIALIAS)
imm = np.array(im_re).astype(np.float32) / 255
img_array_test[j - ii_num] = imm
array_test_all[array_id] = img_array_test
array_id += 1
print('\n')
return array_test_all
if data_type == 'cholec80':
seq_test = get_seq_path_test(data_root_80)
test_loader = get_eval_loader(seq_test, test_batch_size)
# --------- training funtions ------------------------------------
class Round3(Function):
@staticmethod
def forward(ctx, input, training=False, inplace=False):
output = torch.round(input)
ctx.input = input
return output
@staticmethod
def backward(ctx, grad_output):
mask = 1 - (ctx.input == 0)
mask = Variable(mask).cuda().float()
grad_output = grad_output * mask
