def testOReval_one_epoch(sess,
ops,
ops_data,
testOReval_writer,
file_lenth,
batch_size=BATCH_SIZE):
is_training = False
loss_sum = 0.
total_seen = 0
total_ave_error = 0.
cnt_loop = 0
for i in range(file_lenth): #tfrecords
with tf.variable_scope('data_packer'):
list_rgb, list_srgb = sess.run([
ops_data['list_rgb_%s' % (i)], ops_data['list_srgb_%s' % (i)]
])
current_rgb, current_srgb = tl.data_loader(list_rgb, list_srgb)
idx = 0
num = int(current_rgb.shape[0])
with tf.variable_scope('testORvali_layer') as sc:
while idx < num:
log_string('----' + 'testORvali-batch-count=' + str(cnt_loop) +
'-----')
rgb = current_rgb[idx:idx + batch_size]
srgb = current_srgb[idx:idx + batch_size]
idx += batch_size
#feeding
feed_dict = {
ops['rgb_pl']: rgb,
ops['srgb_pl']: srgb,
ops['is_training_pl']: is_training,
}
#run the session and get all the results we need
summary, step, accuracy, loss_val = sess.run(
[ops['merged'], ops['step'], ops['accuracy'], ops['loss']],
feed_dict=feed_dict)
#then cook
testORval_writer.add_summary(summary, step)
print('accuracy(test):', accuracy)
print('loss(test):', loss_val)
total_ave_error += accuracy
total_seen += BATCH_SIZE
loss_sum += (loss_val)
cnt_loop += 1
log_string('mean Lab-accuracy/pair:%f' %
(total_ave_error / float(cnt_loop)))
log_string('mean loss/pair: %f' % (loss_sum / float(cnt_loop))) #()
def train_one_epoch(sess,ops,ops_data,train_writer,file_lenth,batch_size=BATCH_SIZE):
is_training = True
loss_sum = 0.
total_seen = 0
total_correct=0.
cnt_loop=0
for i in range(file_lenth): #tfrecords
with tf.variable_scope('data_packer'):
list_cloud,list_label=sess.run([ops_data['list_cloud_%s'%(i)],ops_data['list_label_%s'%(i)]])
current_cloud,current_label=tl.data_loader(list_cloud,list_label)
idx=0
num=int(current_cloud.shape[0])
with tf.variable_scope('train_layer') as sc:
while idx<num:
log_string('----' + 'train-batch-count='+str(cnt_loop) + '-----')
cloud=current_cloud[idx:idx+batch_size]
label=current_label[idx:idx+batch_size]
idx+=batch_size
#feeding
feed_dict = {ops['pointclouds_pl']: cloud, #(B,N,c)
ops['labels_pl']: label, #(B) or (B,N)
ops['is_training_pl']: is_training,}
#run the session and get all the results we need
summary, step, _, loss_val, pred_val = sess.run([ops['merged'], ops['step'],
ops['train_op'], ops['loss'], ops['pred']], feed_dict=feed_dict)
#then cook
'''
train_writer.add_summary(summary, step)
'''
pred_val = np.argmax(pred_val, 1) #(B,class_num)->(B,1) #supervised&hard learning
correct = np.sum(pred_val == label[start_idx:end_idx]) #a bool tensor,sum it then get a number represents the correct_predicted results
total_correct += correct
total_seen += BATCH_SIZE
loss_sum += loss_val
log_string('mean loss: %f' % (loss_sum / float(total_seen)))
log_string('accuracy: %f' % (total_correct / float(total_seen)))
def train_one_epoch(sess,ops,ops_data,train_writer,file_lenth,batch_size=BATCH_SIZE):
is_training = True
loss_sum = 0.
total_seen = 0
total_correct=0.
cnt_loop=0
for i in range(file_lenth): #tfrecords
with tf.variable_scope('data_packer'):
list_img,list_label=sess.run([ops_data['list_img_%s'%(i)],ops_data['list_label_%s'%(i)]])
current_img,current_label=tl.data_loader(list_img,list_label)
idx=0
num=int(current_img.shape[0])
with tf.variable_scope('train_layer') as sc:
while idx+batch_size<num:
log_string('----' + 'train-batch-count='+str(cnt_loop) + '-----')
imgs=current_img[idx:idx+batch_size]
labels=current_label[idx:idx+batch_size]
idx+=batch_size
#feeding
feed_dict = {ops['img_pl']: imgs,
ops['label_pl']: labels,
ops['is_training_pl']: is_training}
#run the session and get all the results we need
summary, step,_,pred_val,loss_val= sess.run([ops['merged'], ops['step'],
ops['train_op'],ops['pred'],ops['loss']], feed_dict=feed_dict)
#then cook
train_writer.add_summary(summary, step)
print('loss(train):',loss_val)
pred_val = np.argmax(pred_val[:,4:], 1) #(B,class_num)->(B,1) #supervised&hard learning
correct = np.sum(pred_val == labels[:,4])
total_correct+=correct
total_seen += BATCH_SIZE
loss_sum += loss_val
cnt_loop+=1
log_string('accuracy: %f' % (total_correct / float(total_seen)))
log_string('mean loss: %f' % (loss_sum / float(cnt_loop))) #()
def testOReval_one_epoch(sess,
ops,
ops_data,
testOReval_writer,
file_lenth,
batch_size=BATCH_SIZE):
is_training = False
loss_sum = 0.
total_seen = 0
total_correct = 0.
total_seen_class = [0 for _ in range(CLASS_NUM)]
total_correct_class = [0 for _ in range(CLASS_NUM)]
cnt_loop = 0
for i in range(file_lenth): #tfrecords
with tf.variable_scope('data_packer'):
list_img, list_label = sess.run([
ops_data['list_img_%s' % (i)], ops_data['list_label_%s' % (i)]
])
current_img, current_label = tl.data_loader(list_img, list_label)
#print(current_label)
idx = 0
num = int(current_img.shape[0])
with tf.variable_scope('test/vali_layer') as sc:
while idx + batch_size < num:
log_string('----' + 'test/vali-batch-count=' + str(cnt_loop) +
'-----')
imgs = current_img[idx:idx + batch_size]
labels = current_label[idx:idx + batch_size]
labels = np.squeeze(labels)
idx += batch_size
#feeding
feed_dict = {
ops['img_pl']: imgs,
ops['label_pl']: labels,
ops['is_training_pl']: is_training,
}
#run the session and get all the results we need
summary, step, pred_lb, loss_val = sess.run(
[
ops['merged'], ops['step'], ops['pred_label'],
ops['loss']
],
feed_dict=feed_dict)
testOReval_writer.add_summary(summary, step)
print('loss(test):', loss_val)
#then cook
#pred_val = np.argmax(pred_val, 1)
correct = np.sum(
pred_lb ==
labels) #in this line ,labels have been changed to float32
total_correct += correct
total_seen += BATCH_SIZE
loss_sum += (loss_val * BATCH_SIZE)
#extra loop:
#count how many times each class has occured in prediction,and numbering the correct ones
templ = labels #(B)
for i in range(batch_size):
l = int(templ[i]) #(1)
total_correct_class[l] += (pred_lb[i] == l)
total_seen_class[l] += 1
cnt_loop += 1
acc = np.mean(
np.array(total_correct_class) /
np.array(total_seen_class, dtype=np.float))
log_string('test mean loss: %f' % (loss_sum / float(cnt_loop)))
log_string('test accuracy: %f' % (total_correct / float(total_seen)))
log_string('test mean class acc: %f' % (acc))
return acc