def debug_data(self, data):
for i, t in enumerate(torch.unbind(data['image'])):
a = t.cpu().numpy()
logging.info('image%d: %f %s' % (i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
for i, t in enumerate(torch.unbind(data['tensor'])):
a = t.cpu().numpy()
logging.info('tensor%d: %f %s' % (i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
def stat_ap(self):
cls_num = [0 for _ in self.category]
cls_score = [np.array([], dtype=np.float32) for _ in self.category]
cls_tp = [np.array([], dtype=np.bool) for _ in self.category]
for data in tqdm.tqdm(self.loader):
for key in data:
t = data[key]
if torch.is_tensor(t):
data[key] = utils.ensure_device(t)
tensor = torch.autograd.Variable(data['tensor'], volatile=True)
batch_size = tensor.size(0)
pred = pybenchmark.profile('inference')(model._inference)(
self.inference, tensor)
_prob, pred['cls'] = conv_logits(pred)
pred['iou'] = pred['iou'].contiguous()
pred['prob'] = pred['iou'] * _prob
for key in pred:
pred[key] = pred[key].data
if self.config.getboolean('eval', 'debug'):
self.debug_data(data)
self.debug_pred(pred)
norm_bbox(data, pred)
for path, size, difficult, image, data_yx_min, data_yx_max, data_cls, yx_min, yx_max, iou, prob, cls in zip(
*(data[key]
for key in 'path, size, difficult'.split(', ')),
*(torch.unbind(data[key])
for key in 'image, yx_min, yx_max, cls'.split(', ')),
*(torch.unbind(pred[key].view(batch_size, -1, 2))
for key in 'yx_min, yx_max'.split(', ')),
*(torch.unbind(pred[key].view(batch_size, -1))
for key in 'iou, prob, cls'.split(', '))):
data_yx_min, data_yx_max, data_cls = filter_valid(
data_yx_min, data_yx_max, data_cls, difficult)
for c in data_cls.cpu().numpy():
cls_num[c] += 1
yx_min, yx_max, cls, score = self.filter_visible(
yx_min, yx_max, iou, prob, cls)
keep = pybenchmark.profile('nms')(utils.postprocess.nms)(
yx_min, yx_max, score,
self.config.getfloat('detect', 'overlap'))
if keep:
keep = utils.ensure_device(torch.LongTensor(keep))
yx_min, yx_max, cls, score = (t[keep]
for t in (yx_min, yx_max,
cls, score))
for c in set(cls.cpu().numpy()):
c = int(c) # PyTorch's bug
_score, tp = self.filter_cls(c, path, data_yx_min,
data_yx_max, data_cls,
yx_min, yx_max, cls,
score)
cls_score[c] = np.append(cls_score[c],
_score.cpu().numpy())
cls_tp[c] = np.append(cls_tp[c], tp)
return cls_num, cls_score, cls_tp
def debug_data(self, data):
for i, t in enumerate(torch.unbind(data['image'])):
a = t.cpu().numpy()
logging.info(
'image%d: %f %s' %
(i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
for i, t in enumerate(torch.unbind(data['tensor'])):
a = t.cpu().numpy()
logging.info(
'tensor%d: %f %s' %
(i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
def debug_pred(self, pred):
for i, t in enumerate(torch.unbind(pred['iou'])):
a = t.cpu().numpy()
logging.info('iou%d: %f %s' % (i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
for i, t in enumerate(torch.unbind(pred['center_offset'])):
a = t.cpu().numpy()
logging.info('center_offset%d: %f %s' % (i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
for i, t in enumerate(torch.unbind(pred['size_norm'])):
a = t.cpu().numpy()
logging.info('size_norm%d: %f %s' % (i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
for i, t in enumerate(torch.unbind(pred['logits'])):
a = t.cpu().numpy()
logging.info('logits%d: %f %s' % (i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
def stat_ap(self):
cls_num = [0 for _ in self.category]
cls_score = [np.array([], dtype=np.float32) for _ in self.category]
cls_tp = [np.array([], dtype=np.bool) for _ in self.category]
for data in tqdm.tqdm(self.loader):
for key in data:
t = data[key]
if torch.is_tensor(t):
data[key] = utils.ensure_device(t)
tensor = torch.autograd.Variable(data['tensor'], volatile=True)
pred = pybenchmark.profile('inference')(model._inference)(
self.inference, tensor)
pred['iou'] = pred['iou'].contiguous()
logits = get_logits(pred)
pred['prob'] = F.softmax(logits, -1)
for key in pred:
pred[key] = pred[key].data
if self.config.getboolean('eval', 'debug'):
self.debug_data(data)
self.debug_pred(pred)
norm_bbox_data(data)
norm_bbox_pred(pred)
for path, difficult, image, data_yx_min, data_yx_max, data_cls, iou, yx_min, yx_max, prob in zip(
*(data[key] for key in 'path, difficult'.split(', ')),
*(torch.unbind(data[key])
for key in 'image, yx_min, yx_max, cls'.split(', ')),
*(torch.unbind(pred[key])
for key in 'iou, yx_min, yx_max, prob'.split(', '))):
data_yx_min, data_yx_max, data_cls = filter_valid(
data_yx_min, data_yx_max, data_cls, difficult)
for c in data_cls.cpu().numpy():
cls_num[c] += 1
iou = iou.view(-1)
yx_min, yx_max, prob = (t.view(-1, t.size(-1))
for t in (yx_min, yx_max, prob))
ret = postprocess(self.config, iou, yx_min, yx_max, prob)
if ret is not None:
iou, yx_min, yx_max, cls, score = ret
for c in set(cls.cpu().numpy()):
c = int(c) # PyTorch's bug
_score, tp = self.filter_cls(c, path, data_yx_min,
data_yx_max, data_cls,
yx_min, yx_max, cls,
score)
cls_score[c] = np.append(cls_score[c],
_score.cpu().numpy())
cls_tp[c] = np.append(cls_tp[c], tp)
return cls_num, cls_score, cls_tp
def debug_pred(self, pred):
for i, t in enumerate(torch.unbind(pred['iou'])):
a = t.cpu().numpy()
logging.info(
'iou%d: %f %s' %
(i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
for i, t in enumerate(torch.unbind(pred['center_offset'])):
a = t.cpu().numpy()
logging.info(
'center_offset%d: %f %s' %
(i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
for i, t in enumerate(torch.unbind(pred['size_norm'])):
a = t.cpu().numpy()
logging.info(
'size_norm%d: %f %s' %
(i, utils.abs_mean(a), hashlib.md5(a.tostring()).hexdigest()))
def __call__(self):
changed = np.zeros([self.height, self.width], np.bool)
for yx in tqdm.tqdm(self.loader):
batch_size = yx.size(0)
tensor = torch.zeros(batch_size, 3, self.height, self.width)
for i, _yx in enumerate(torch.unbind(yx)):
y, x = torch.unbind(_yx)
tensor[i, :, y, x] = 1
tensor = utils.ensure_device(tensor)
output = self.dnn(torch.autograd.Variable(tensor))
output = output[:, :, self.i, self.j]
cmp = output == self.output
cmp = torch.prod(cmp, -1).data
for _yx, c in zip(torch.unbind(yx), torch.unbind(cmp)):
y, x = torch.unbind(_yx)
changed[y, x] = c
return changed
def __call__(self):
changed = np.zeros([self.height, self.width], np.bool)
for yx in tqdm.tqdm(self.loader):
batch_size = yx.size(0)
tensor = torch.zeros(batch_size, 3, self.height, self.width)
for i, _yx in enumerate(torch.unbind(yx)):
y, x = torch.unbind(_yx)
tensor[i, :, y, x] = 1
tensor = utils.ensure_device(tensor)
output = self.dnn(torch.autograd.Variable(tensor, volatile=True))
output = output[:, :, self.i, self.j]
cmp = output == self.output
cmp = torch.prod(cmp, -1).data
for _yx, c in zip(torch.unbind(yx), torch.unbind(cmp)):
y, x = torch.unbind(_yx)
changed[y, x] = c
return changed
def stat_ap(self):
cls_num = [0 for _ in self.category]
cls_score = [np.array([], dtype=np.float32) for _ in self.category]
cls_tp = [np.array([], dtype=np.bool) for _ in self.category]
for data in tqdm.tqdm(self.loader):
for key in data:
t = data[key]
if torch.is_tensor(t):
data[key] = utils.ensure_device(t)
tensor = torch.autograd.Variable(data['tensor'], volatile=True)
pred = pybenchmark.profile('inference')(model._inference)(self.inference, tensor)
pred['iou'] = pred['iou'].contiguous()
logits = get_logits(pred)
pred['prob'] = F.softmax(logits, -1)
for key in pred:
pred[key] = pred[key].data
if self.config.getboolean('eval', 'debug'):
self.debug_data(data)
self.debug_pred(pred)
norm_bbox_data(data)
norm_bbox_pred(pred)
for path, difficult, image, data_yx_min, data_yx_max, data_cls, iou, yx_min, yx_max, prob in zip(*(data[key] for key in 'path, difficult'.split(', ')), *(torch.unbind(data[key]) for key in 'image, yx_min, yx_max, cls'.split(', ')), *(torch.unbind(pred[key]) for key in 'iou, yx_min, yx_max, prob'.split(', '))):
data_yx_min, data_yx_max, data_cls = filter_valid(data_yx_min, data_yx_max, data_cls, difficult)
for c in data_cls.cpu().numpy():
cls_num[c] += 1
iou = iou.view(-1)
yx_min, yx_max, prob = (t.view(-1, t.size(-1)) for t in (yx_min, yx_max, prob))
ret = postprocess(self.config, iou, yx_min, yx_max, prob)
if ret is not None:
iou, yx_min, yx_max, cls, score = ret
for c in set(cls.cpu().numpy()):
c = int(c) # PyTorch's bug
_score, tp = self.filter_cls(c, path, data_yx_min, data_yx_max, data_cls, yx_min, yx_max, cls, score)
cls_score[c] = np.append(cls_score[c], _score.cpu().numpy())
cls_tp[c] = np.append(cls_tp[c], tp)
return cls_num, cls_score, cls_tp
