def test_ctdet_decode():
np.random.seed(32)
hm = np.random.randn(2, 64, 64, 80)
reg = np.random.randn(2, 64, 64, 2) * 10.0
wh = np.random.randn(2, 64, 64, 2) * 20.0
keras_hm = K.constant(hm)
keras_reg = K.constant(reg)
keras_wh = K.constant(wh)
keras_detections = K.eval(
_ctdet_decode(keras_hm, keras_reg, keras_wh, output_stride=1))
gold_fn = 'path_for_colab/tests/data/ctdet_decode_gold.p'
if not os.path.exists(gold_fn):
import torch as th
import sys
sys.path.append(os.path.expanduser('~/Pytorch/CenterNet/src'))
from lib.models.decode import ctdet_decode # noqa
py_hm = th.from_numpy(hm.transpose(0, 3, 1, 2)).float()
py_hm.sigmoid_()
py_reg = th.from_numpy(reg.transpose(0, 3, 1, 2)).float()
py_wh = th.from_numpy(wh.transpose(0, 3, 1, 2)).float()
py_detections = ctdet_decode(py_hm, py_reg, py_wh).detach().numpy()
with open(gold_fn, 'wb') as f:
pickle.dump(py_detections, f)
else:
with open(gold_fn, 'rb') as f:
py_detections = pickle.load(f)
assert np.allclose(keras_detections, py_detections)
def save_result1(self, outputs, batch,
results): # 这里的outputs包含model输出的所有特征图（3个特征图）
# 对每一个output特征图都计算dets_out，然后将3个dets_out整合起来，根据score来排序，然后用NMS或者其他方法来过滤！！！
dets_list = []
dets_score_list = []
for idx in range(len(outputs)):
output = outputs[idx]
reg = output['reg'] if self.opt.reg_offset else None
dets = ctdet_decode(output['hm'],
output['wh'],
reg=reg,
cat_spec_wh=self.opt.cat_spec_wh,
K=self.opt.K)
dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2])
# print(type(dets), dets.shape)
dets_list.append(dets)
dets_300 = np.concatenate(
(dets_list[0], dets_list[1], dets_list[2]),
axis=1) # 将3个dets:array[1,100,6]连接起来成为dets_300:array[1,300,6]
dets_300_sort = dets_300.copy()
array_for_sort = dets_300.reshape(
-1, dets.shape[2])[:, :5] # 将dets_300的最后一列改变成score，方便下面排序
sort_idx = np.lexsort(
-array_for_sort.T) # sort_idx是dets_300按照score从大到小排序的索引
# print(array_for_sort.shape, sort_idx)
for key, item in enumerate(sort_idx):
dets_300_sort[0][key] = dets_300[0][
item] # 将排序后的numpy数组保存到 dets_300_sort
# print(dets_300_sort.shape, dets_300_sort[0][0][4],dets_300_sort[0][101][4])
# dets_300_sort = dets_300_sort[:,:100,:] # 检测结果只取前100？？
dets = dets_300_sort # 这个命名只是为了不改下面的两个语句中的变量名
dets_outs = ctdet_post_process(dets.copy(),
batch['meta']['c'].cpu().numpy(),
batch['meta']['s'].cpu().numpy(),
output['hm'].shape[2],
output['hm'].shape[3],
output['hm'].shape[1])
for j in range(1, self.opt.num_classes +
1): # 给数组reshape一下，要不然不符合NMS需要输入2维数组的输入要求
dets_outs[0][j] = np.array(dets_outs[0][j],
dtype=np.float32).reshape(-1, 5)
# print(type(dets_outs),len(dets_outs[0]))
results_nms = {}
for j in range(1, self.opt.num_classes + 1):
results_nms[j] = np.concatenate(
[dets_out[j] for dets_out in dets_outs],
axis=0).astype(np.float32)
# print(j, results_nms[j])
soft_nms(results_nms[j], Nt=0.5, method=2)
# print(111, type(dets_outs[0]), dets_outs[0].keys())
# print(222, type(results_nms), results_nms.keys()) # results_nms是dets_outs[0]经过NMS之后得到的结果
# results[batch['meta']['img_id'].cpu().numpy()[0]] = dets_outs[0]
results[batch['meta']['img_id'].cpu().numpy()[0]] = results_nms
def save_result(self, output, batch, results):
reg = output['reg'] if self.opt.reg_offset else None
dets = ctdet_decode(output['hm'],
output['wh'],
reg=reg,
cat_spec_wh=self.opt.cat_spec_wh,
K=self.opt.K)
dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2])
dets_out = ctdet_post_process(dets.copy(),
batch['meta']['c'].cpu().numpy(),
batch['meta']['s'].cpu().numpy(),
output['hm'].shape[2],
output['hm'].shape[3],
output['hm'].shape[1])
results[batch['meta']['img_id'].cpu().numpy()[0]] = dets_out[0]
def process(self, images, return_time=False):
with torch.no_grad():
output = self.model(images)[-1]
hm = output['hm'].sigmoid_()
wh = output['wh']
# reg = output['reg']
torch.cuda.synchronize()
forward_time = time.time()
# dets = ctdet_decode(hm, wh, reg=reg)
dets = ctdet_decode(hm, wh)
if return_time:
return output, dets, forward_time
else:
return output, dets
def debug(self, batch, output, iter_id):
opt = self.opt
reg = output['reg'] if opt.reg_offset else None
dets = ctdet_decode(output['hm'],
output['wh'],
reg=reg,
cat_spec_wh=opt.cat_spec_wh,
K=opt.K)
dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2])
dets[:, :, :4] *= opt.down_ratio
dets_gt = batch['meta']['gt_det'].numpy().reshape(1, -1, dets.shape[2])
dets_gt[:, :, :4] *= opt.down_ratio
for i in range(1):
debugger = Debugger(dataset=opt.dataset,
ipynb=(opt.debug == 3),
theme=opt.debugger_theme)
img = batch['input'][i].detach().cpu().numpy().transpose(1, 2, 0)
img = np.clip(((img * opt.std + opt.mean) * 255.), 0,
255).astype(np.uint8)
pred = debugger.gen_colormap(
output['hm'][i].detach().cpu().numpy())
gt = debugger.gen_colormap(batch['hm'][i].detach().cpu().numpy())
debugger.add_blend_img(img, pred, 'pred_hm')
debugger.add_blend_img(img, gt, 'gt_hm')
debugger.add_img(img, img_id='out_pred')
for k in range(len(dets[i])):
if dets[i, k, 4] > opt.center_thresh:
# add_coco_bbox(bbox, cat, conf=1, show_txt=True, img_id='default'):
debugger.add_coco_bbox(dets[i, k, :4],
dets[i, k, -1],
dets[i, k, 4],
img_id='out_pred')
debugger.add_img(img, img_id='out_gt')
for k in range(len(dets_gt[i])):
if dets_gt[i, k, 4] > opt.center_thresh:
# add_coco_bbox(bbox, cat, conf=1, show_txt=True, img_id='default'):
debugger.add_coco_bbox(dets_gt[i, k, :4],
dets_gt[i, k, -1],
dets_gt[i, k, 4],
img_id='out_gt')
if opt.debug == 4:
debugger.save_all_imgs(opt.debug_dir,
prefix='{}'.format(iter_id))
else:
debugger.show_all_imgs(pause=True)
def process(self, images, return_time=False):
with torch.no_grad():
output = self.model(images)[-1]
hm = output['hm'].sigmoid_()
wh = output['wh']
reg = output['reg'] if self.opt.reg_offset else None
if self.opt.flip_test:
hm = (hm[0:1] + flip_tensor(hm[1:2])) / 2
wh = (wh[0:1] + flip_tensor(wh[1:2])) / 2
reg = reg[0:1] if reg is not None else None
torch.cuda.synchronize()
forward_time = time.time()
dets = ctdet_decode(hm, wh, reg=reg, K=self.opt.K)
if return_time:
return output, dets, forward_time
else:
return output, dets
def save_result(self, output, batch,
results): # 能不能尽量不改这里面的小函数，只改变针对函数的输入？？？？？
# print(output['hm'].size(), output['wh'].size(), output['reg'].size())
# [1, 20, 128, 128] [1, 2, 128, 128] [1, 2, 128, 128]
reg = output[
'reg'] if self.opt.reg_offset else None # self.opt.reg_offset = True
dets = ctdet_decode(output['hm'],
output['wh'],
reg=reg,
cat_spec_wh=self.opt.cat_spec_wh,
K=self.opt.K)
# print(dets.size()) #[1, 100, 6] ，其中的6代表了 [tlx, tly, brx, bry, scores, clses]， 按照score的高低排序
dets = dets.detach().cpu().numpy().reshape(
1, -1, dets.shape[2]) # detach()用来切断梯度的传播
# print(dets.shape) #(1, 100, 6)
dets_out = ctdet_post_process(dets.copy(),
batch['meta']['c'].cpu().numpy(),
batch['meta']['s'].cpu().numpy(),
output['hm'].shape[2],
output['hm'].shape[3],
output['hm'].shape[1])
# print(len(dets_out), dets_out) # len(dets_out) = 1 ； dets_out = [{1: [...], 2: [], 3: [[130.917, 491.190, 285.635, 504.33, 0.797]], 4: [...],...}, 20: [...]]
results[batch['meta']['img_id'].cpu().numpy()[0]] = dets_out[0]
def val(self, epoch, data_loader):
model_with_loss = self.model_with_loss
if len(self.opt.gpus) > 1:
model_with_loss = self.model_with_loss.module
model_with_loss.eval()
torch.cuda.empty_cache()
opt = self.opt
results = {}
data_time, batch_time = AverageMeter(), AverageMeter()
avg_loss_stats = {l: AverageMeter() for l in self.loss_stats}
num_iters = len(data_loader) if opt.num_iters < 0 else opt.num_iters
bar = Bar('{}/{}'.format(opt.task, opt.exp_id), max=num_iters)
end = time.time()
for iter_id, batch in enumerate(data_loader):
if iter_id >= num_iters:
break
data_time.update(time.time() - end)
for k in batch:
if k != 'meta':
batch[k] = batch[k].to(device=opt.device,
non_blocking=True)
output, loss, loss_stats = model_with_loss(batch)
results.setdefault('img_id', []).extend(batch['meta']['img_id'])
boxes = ctdet_decode(output['hm'],
output['wh'],
output['reg'],
K=self.opt.K)
meta = batch['meta']
c = meta['c'].numpy()
s = meta['s'].numpy()
_, _, out_h, out_w = output['hm'].shape
dets = ctdet_post_process(boxes.cpu().numpy(), [c], [s], out_h,
out_w, opt.num_classes)
# breakpoint()
results.setdefault('detections', []).extend(dets)
loss = loss.mean()
batch_time.update(time.time() - end)
end = time.time()
Bar.suffix = '{phase}: [{0}][{1}/{2}]|Tot: {total:} |ETA: {eta:} '.format(
epoch,
iter_id,
num_iters,
phase='val',
total=bar.elapsed_td,
eta=bar.eta_td)
for l in avg_loss_stats:
# breakpoint()
loss_t = loss_stats[l] or torch.zeros(1)
avg_loss_stats[l].update(loss_t.mean().item(),
batch['input'].size(0))
Bar.suffix += '|{} {:.4f} '.format(l, avg_loss_stats[l].avg)
self.val_writer.add_scalar(l, avg_loss_stats[l].avg,
self.global_steps, 1)
if not opt.hide_data_time:
Bar.suffix = Bar.suffix + '|Data {dt.val:.3f}s({dt.avg:.3f}s) ' \
'|Net {bt.avg:.3f}s'.format(dt=data_time, bt=batch_time)
if opt.print_iter > 0:
if iter_id % opt.print_iter == 0:
print('{}/{}| {}'.format(opt.task, opt.exp_id, Bar.suffix))
else:
bar.next()
if opt.debug > 0:
self.debug(batch, output, iter_id)
if opt.test:
self.save_result(output, batch, results)
del output, loss, loss_stats
# break
ret = {k: v.avg for k, v in avg_loss_stats.items()}
for l in avg_loss_stats:
self.val_writer.add_scalar(l, avg_loss_stats[l].avg,
self.global_steps)
bar.next()
bar.finish()
ret['time'] = bar.elapsed_td.total_seconds() / 60.
return ret, results