def EvalOneEpoch(self, Loader):
batch_cnt = 1
samp_cnt = 0
avg_loss = 0.
avg_correct_rate = 0.
avg_iou = 0.
while True:
## get next batch
SuccessFlag, data, seg, weak_seg_onehot, mb_size = Loader.NextBatch_TestSet(
)
if not SuccessFlag:
break
## Normalize Validation Points
# data = PointNet_CAM.pc_max_normalize(data)
## Dummy Variables
Mask_bin = np.ones(shape=[mb_size, data.shape[1]],
dtype=np.float32)
loss_mb, Z_prob_mb = \
self.sess.run([self.loss, self.Z_prob],
feed_dict={self.X_ph: data,
self.Y_ph: seg,
self.Is_Training_ph: True,
self.Mask_ph: Mask_bin})
## Calculate loss and correct rate
pred_mb = np.argmax(Z_prob_mb, axis=-1)
correct = np.mean(pred_mb == seg)
m_iou = np.mean(Tool.IoU(pred_mb, seg, Loader.numParts))
# avg_loss = (batch_cnt - 1) / batch_cnt * avg_loss + (loss_mb / mb_size) / batch_cnt
# avg_correct_rate = (batch_cnt - 1) / batch_cnt * avg_correct_rate + correct / batch_cnt
avg_loss = (avg_loss * samp_cnt + loss_mb) / (samp_cnt + mb_size)
avg_correct_rate = (avg_correct_rate * samp_cnt +
correct * mb_size) / (samp_cnt + mb_size)
avg_iou = (avg_iou * samp_cnt + m_iou * mb_size) / (samp_cnt +
mb_size)
samp_cnt += mb_size
print(
'\rBatch {:d} EvaluatedSamp {:d} Avg Loss {:.4f} Avg Correct Rate {:.3f}% Avg IoU {:.3f}%'
.format(batch_cnt, samp_cnt, avg_loss, 100 * avg_correct_rate,
100 * avg_iou),
end='')
batch_cnt += 1
Loader.ResetLoader_TestSet()
return avg_loss, avg_correct_rate, avg_iou
def EvalOneEpoch_Full(self, Loader, Eval):
batch_cnt = 1
samp_cnt = 0
avg_loss = 0.
avg_correct_rate = 0.
avg_iou = 0.
while True:
## get next batch
SuccessFlag, data, seg, weak_seg_onehot, mb_size = Loader.NextBatch_TestSet(
)
if not SuccessFlag:
break
if mb_size < Loader.batchsize:
data_feed = np.concatenate([
data,
np.tile(data[np.newaxis, 0, ...],
[Loader.batchsize - mb_size, 1, 1])
],
axis=0)
seg_feed = np.concatenate([
seg,
np.tile(seg[np.newaxis, 0],
[Loader.batchsize - mb_size, 1])
],
axis=0)
seg_Onehot_feed = Tool.OnehotEncode(seg_feed, 13)
else:
data_feed = data
seg_Onehot_feed = Tool.OnehotEncode(seg, 13)
Mask_bin_feed = np.ones(shape=[Loader.batchsize, data.shape[1]],
dtype=np.float32)
## Replicate for Siamese Network Input
data_feed_rep = []
seg_Onehot_feed_rep = []
Mask_bin_feed_rep = []
for b_i in range(self.BATCH_SIZE):
data_feed_rep.append(data_feed[b_i])
data_feed_rep.append(data_feed[b_i])
seg_Onehot_feed_rep.append(seg_Onehot_feed[b_i])
seg_Onehot_feed_rep.append(seg_Onehot_feed[b_i])
Mask_bin_feed_rep.append(Mask_bin_feed[b_i])
Mask_bin_feed_rep.append(Mask_bin_feed[b_i])
data_feed_rep = np.stack(data_feed_rep, axis=0)
seg_Onehot_feed_rep = np.stack(seg_Onehot_feed_rep, axis=0)
Mask_bin_feed_rep = np.stack(Mask_bin_feed_rep, axis=0)
loss_mb, Z_prob_mb = \
self.sess.run([self.loss, self.Z_prob],
feed_dict={self.X_ph: data_feed_rep,
self.Y_ph: seg_Onehot_feed_rep,
self.Is_Training_ph: False,
self.Mask_ph: Mask_bin_feed_rep})
Z_prob_mb = Z_prob_mb[0:2 * mb_size:2, ...]
## Calculate loss and correct rate
pred_mb = np.argmax(Z_prob_mb, axis=-1)
correct = np.mean(pred_mb == seg)
m_iou = np.mean(Tool.IoU(pred_mb, seg, Loader.numParts))
avg_loss = (avg_loss * samp_cnt + loss_mb) / (samp_cnt + mb_size)
avg_correct_rate = (avg_correct_rate * samp_cnt +
correct * mb_size) / (samp_cnt + mb_size)
avg_iou = (avg_iou * samp_cnt + m_iou * mb_size) / (samp_cnt +
mb_size)
samp_cnt += mb_size
print(
'\rBatch {:d} EvaluatedSamp {:d} Avg Loss {:.4f} Avg Correct Rate {:.3f}% Avg IoU {:.3f}%'
.format(batch_cnt, samp_cnt, avg_loss, 100 * avg_correct_rate,
100 * avg_iou),
end='')
batch_cnt += 1
Loader.ResetLoader_TestSet()
return avg_loss, avg_correct_rate, avg_iou