extract.py报错

[vk-rrr]

运行extract.py报错
Traceback (most recent call last):
File "/home/hzh/Github/LLCM-main/DEEN/extract.py", line 268, in
query_feat_pool, query_feat_fc = extract_query_feat(query_loader)
File "/home/hzh/Github/LLCM-main/DEEN/extract.py", line 138, in extract_query_feat
query_feat_pool[ptr:ptr + batch_num, :] = feat_pool.detach().cpu().numpy()
ValueError: could not broadcast input array from shape (192,2048) into shape (64,2048)

把通道维度连接成一个特征，但是可视化的图不对，请问可以给出符合DEEN的extract.py详细代码吗？

def extract_query_feat(query_loader):
net.eval()
print('Extracting Query Feature...')
start = time.time()
ptr = 0
query_feat_pool = np.zeros((nquery, pool_dim * 3))
query_feat_fc = np.zeros((nquery, pool_dim * 3))
with torch.no_grad():
for batch_idx, (input, label) in enumerate(query_loader):
batch_num = input.size(0)
input = Variable(input.cuda())
feat_pool, feat_fc = net(input, input, label, test_mode[1]) # [192, 2048], [192, 2048]

        feat_pool1, feat_pool2, feat_pool3 = torch.chunk(feat_pool, 3, dim=0)
        feat_fc1, feat_fc2, feat_fc3 = torch.chunk(feat_fc, 3, dim=0)
        feat_pool = torch.cat((feat_pool1, feat_pool2, feat_pool3), 1)
        feat_fc = torch.cat((feat_fc1, feat_fc2, feat_fc3), 1)

        query_feat_pool[ptr:ptr + batch_num, :] = feat_pool.detach().cpu().numpy()
        query_feat_fc[ptr:ptr + batch_num, :] = feat_fc.detach().cpu().numpy()
        ptr = ptr + batch_num
print('Extracting Time:\t {:.3f}'.format(time.time() - start))
return query_feat_pool, query_feat_fc

[vk-rrr]

help!help!

[ccccxxxhhh299]

解决了吗？ 救命！

[vk-rrr]

我是这么修改的，不知道对不对
def extract_query_feat(query_loader):
net.eval()
print('Extracting Query Feature...')
start = time.time()
ptr = 0
# query_feat_pool = np.zeros((nquery, pool_dim * 3))
# query_feat_fc = np.zeros((nquery, pool_dim * 3))
query_feat_pool1 = np.zeros((nquery, pool_dim))
query_feat_fc1 = np.zeros((nquery, pool_dim))
query_feat_pool2 = np.zeros((nquery, pool_dim))
query_feat_fc2 = np.zeros((nquery, pool_dim))
query_feat_pool3 = np.zeros((nquery, pool_dim))
query_feat_fc3 = np.zeros((nquery, pool_dim))
with torch.no_grad():
for batch_idx, (input, label) in enumerate(query_loader):
batch_num = input.size(0)
input = Variable(input.cuda())
feat_pool, feat_fc = net(input, input, label, test_mode[1]) # [192, 2048], [192, 2048]
#
# feat_pool1, feat_pool2, feat_pool3 = torch.chunk(feat_pool, 3, dim=0)
# feat_fc1, feat_fc2, feat_fc3 = torch.chunk(feat_fc, 3, dim=0)
# feat_pool = torch.cat((feat_pool1, feat_pool2, feat_pool3), 1)
# feat_fc = torch.cat((feat_fc1, feat_fc2, feat_fc3), 1)

        query_feat_pool1[ptr:ptr + batch_num, :] = feat_pool[:batch_num].detach().cpu().numpy()
        query_feat_fc1[ptr:ptr + batch_num, :] = feat_fc[:batch_num].detach().cpu().numpy()
        query_feat_pool2[ptr:ptr + batch_num, :] = feat_pool[batch_num:2 * batch_num].detach().cpu().numpy()
        query_feat_fc2[ptr:ptr + batch_num, :] = feat_fc[batch_num:2 * batch_num].detach().cpu().numpy()
        query_feat_pool3[ptr:ptr + batch_num, :] = feat_pool[batch_num * 2:].detach().cpu().numpy()
        query_feat_fc3[ptr:ptr + batch_num, :] = feat_fc[batch_num * 2:].detach().cpu().numpy()
        ptr = ptr + batch_num
print('Extracting Time:\t {:.3f}'.format(time.time() - start))
query_feat_pool = np.concatenate((query_feat_pool1, query_feat_pool2, query_feat_pool3), axis=0)
query_feat_fc = np.concatenate((query_feat_fc1, query_feat_fc2, query_feat_fc3), axis=0)
return query_feat_pool, query_feat_fc

[vk-rrr]

tsne.py可视化里的代码怎么改呀，救救