def main():
opt = arg_parse()
cluster_tracks = pickle_load(opt.track_path)
score = purity_score_for_track(cluster_tracks)
print('score %f' % score)
write(
opt.result_path, 'cluster task: %s \t\tscore: %f\n' %
(opt.track_path.split('/')[-1], score))
def random_hp():
image_size = 28
num_labels = 10
train_dataset, train_labels, valid_dataset, valid_labels, test_dataset, test_labels = \
format_mnist()
pick_size = 2048
valid_dataset = valid_dataset[0:pick_size, :, :, :]
valid_labels = valid_labels[0:pick_size, :]
test_dataset = test_dataset[0:pick_size, :, :, :]
test_labels = test_labels[0:pick_size, :]
basic_hypers = {
'batch_size': 10,
'patch_size': 10,
'depth': 10,
'num_hidden': 10,
'layer_sum': 3,
'starter_learning_rate': 0.1
}
if basic_hypers['patch_size'] > 28:
basic_hypers['patch_size'] = 28
print('=' * 80)
print(basic_hypers)
for _ in range(200):
random_hypers = {
'batch_size':
basic_hypers['batch_size'] + random.randint(-9, 30),
'patch_size':
basic_hypers['patch_size'] + random.randint(-5, 15),
'depth':
basic_hypers['depth'] + random.randint(-9, 10),
'num_hidden':
basic_hypers['num_hidden'] * random.randint(1, 10),
'layer_sum':
basic_hypers['layer_sum'] + random.randint(-1, 2),
'starter_learning_rate':
basic_hypers['starter_learning_rate'] * random.uniform(0, 5)
}
# if basic_hypers['batch_size'] < 10:
# basic_hypers['batch_size'] = 10
if random_hypers['patch_size'] > 28:
random_hypers['patch_size'] = 28
print('=' * 80)
print(random_hypers)
stride_params = [[1, 2, 2, 1]
for _ in range(random_hypers['layer_sum'])]
hps, loss_es = conv_train(train_dataset, train_labels, valid_dataset,
valid_labels, test_dataset, test_labels,
image_size, num_labels, random_hypers,
stride_params)
file_helper.write('hps', str(hps))
file_helper.write('losses', str(loss_es))
def write_with_confidence(confidence_bound):
cross_score_path = 'data/top10-test/cross_filter_score.log'
cross_pid_path = 'data/top10-test/cross_filter_pid.log'
filter_pid_path = 'data/top10/conf_filter_pid.log'
filter_score_path = 'data/top10/conf_filter_score.log'
safe_remove(filter_pid_path)
safe_remove(filter_score_path)
score_lines = read_lines(cross_score_path)
persons_scores = [[float(score) for score in score_line.split()] for score_line in score_lines]
max_score = max([max(predict_scores) for predict_scores in persons_scores])
pid_lines = read_lines(cross_pid_path)
write_line_cnt = 0
for i in range(len(persons_scores)):
scores = persons_scores[i]
pids = pid_lines[i].split()
has_write = False
for j in range(len(persons_scores[0])):
confidence = max(scores[j]/max_score, 1 - scores[j]/max_score)
if confidence < confidence_bound:
# print(confidence)
if not has_write:
write_line_cnt += 1
has_write = True
write(filter_pid_path, pids[j] + ' ')
write(filter_score_path, '%f ' % scores[j])
write(filter_pid_path, '\n')
write(filter_score_path, '\n')
return write_line_cnt
def avg_acc2(grid_eval_path):
grid_infos = read_lines(grid_eval_path)
before_vision_accs = [0.0, 0.0, 0.0, 0.0]
before_fusion_accs = [0.0, 0.0, 0.0]
after_vision_accs = [0.0, 0.0, 0.0]
after_fusion_accs = [0.0, 0.0, 0.0]
i_cv_cnt = 0
for i, grid_info in enumerate(grid_infos):
if i % 2 != 0:
accs = grid_info.split()
for j in range(4):
before_vision_accs[j] += float(accs[j])
i_cv_cnt += 1
write('grid_eval.log', '\n' + grid_eval_path + '\n')
write(
'grid_eval.log', '& %.2f & %.2f & %.2f & %.2f\n' %
(before_vision_accs[0] * 10, before_vision_accs[1] * 10,
before_vision_accs[2] * 10, before_vision_accs[3] * 10))
def write_unequal_rand_st_model(fusion_param):
# fusion_param = get_fusion_param()
rand_answer_path = fusion_param['answer_path'].replace(ctrl_msg['data_folder_path'],
ctrl_msg['data_folder_path'] + '_uerand')
rand_folder_path = folder(rand_answer_path)
safe_mkdir(rand_folder_path)
# although copy all info including pid info, but not use in later training
shutil.copy(fusion_param['answer_path'], rand_answer_path)
rand_path = rand_folder_path + '/renew_pid.log'
safe_remove(rand_path)
origin_tracks = get_tracks(fusion_param)
pid_cnt = len(origin_tracks)
origin_pids = map(lambda x: x + 1, range(pid_cnt))
persons_rand_predict_idx_s = [random.sample(origin_pids, pid_cnt) for _ in range(pid_cnt)]
viz_pid_path = fusion_param['renew_pid_path']
viz_score_path = fusion_param['renew_ac_path']
viz_pids = read_lines(viz_pid_path)
viz_pids = [per_viz_pids.split() for per_viz_pids in viz_pids]
viz_scores = read_lines(viz_score_path)
viz_scores = [per_viz_scores.split() for per_viz_scores in viz_scores]
viz_same_pids = [
[
int(viz_pid) for viz_pid, viz_score in zip(per_viz_pids, per_viz_scores) if float(viz_score) > 0.7
] for per_viz_scores, per_viz_pids in zip(viz_scores, viz_pids)
]
persons_unequal_rand_predict_idx_s = list()
for i in range(pid_cnt):
diff_persons = list(set(persons_rand_predict_idx_s[i]) ^ set(viz_same_pids[i]))
diff_cnt = len(diff_persons)
persons_unequal_rand_predict_idx_s.append(
random.sample(diff_persons, diff_cnt)
)
write_content = ''
for rand_predict_idx_s in persons_unequal_rand_predict_idx_s:
for rand_predict_idx in rand_predict_idx_s:
write_content += str(rand_predict_idx) + ' '
write_content += '\n'
write(rand_path, write_content)
def duke_tracks_realloc(tracks, track_name, data_type):
target_list_path = '../data/%s/%s.list' % (track_name, data_type)
target_track_info_dir_path = '/home/cwh/coding/TrackViz/data/%s' % track_name
safe_mkdir(target_track_info_dir_path)
target_track_dir_path = '/home/cwh/coding/%s' % track_name
safe_mkdir(target_track_dir_path)
target_track_type_dir_path = '/home/cwh/coding/%s/%s' % (track_name,
data_type)
safe_mkdir(target_track_type_dir_path)
names = list()
for track in tracks:
img_name = '%04d_c%d_f%07d.jpg' % (int(track[0]), track[1], track[2])
shutil.copyfile(
'/home/cwh/coding/DukeMTMC-reID/%s/%s' % (data_type, img_name),
'%s/%s' % (target_track_type_dir_path, img_name))
names.append(img_name)
names.append('\n')
list_stmt = ''.join(names)
safe_remove(target_list_path)
write(target_list_path, list_stmt)
def avg_acc(grid_eval_path):
grid_infos = read_lines(grid_eval_path)
before_vision_accs = [0.0, 0.0, 0.0]
before_fusion_accs = [0.0, 0.0, 0.0]
after_vision_accs = [0.0, 0.0, 0.0]
after_fusion_accs = [0.0, 0.0, 0.0]
i_cv_cnt = 0
for i, grid_info in enumerate(grid_infos):
if i % 2 != 0:
accs = grid_info.split()
if i_cv_cnt % 4 == 0:
for j in range(3):
before_vision_accs[j] += float(accs[j])
if i_cv_cnt % 4 == 1:
for j in range(3):
before_fusion_accs[j] += float(accs[j])
if i_cv_cnt % 4 == 2:
for j in range(3):
after_vision_accs[j] += float(accs[j])
if i_cv_cnt % 4 == 3:
for j in range(3):
after_fusion_accs[j] += float(accs[j])
i_cv_cnt += 1
write('grid_eval.log', '\n' + grid_eval_path + '\n')
write(
'grid_eval.log', 'before_retrain_vision\n%f %f %f\n' %
(before_vision_accs[0] / 10, before_vision_accs[1] / 10,
before_vision_accs[2] / 10))
write(
'grid_eval.log', 'before_retrain_fusion\n%f %f %f\n' %
(before_fusion_accs[0] / 10, before_fusion_accs[1] / 10,
before_fusion_accs[2] / 10))
write(
'grid_eval.log', 'after_retrain_vision\n%f %f %f\n' %
(after_vision_accs[0] / 10, after_vision_accs[1] / 10,
after_vision_accs[2] / 10))
write(
'grid_eval.log', 'after_retrain_fusion\n%f %f %f\n' %
(after_fusion_accs[0] / 10, after_fusion_accs[1] / 10,
after_fusion_accs[2] / 10))
def write_rand_pid(fusion_param):
# 对每张左图, 随机生成250个右图的pid，相当于一个随机的renew_pid.log，rand_path也会每次都删除，所以不存在缓存
# todo 不一定需要生成250个右图
# fusion_param = get_fusion_param()
rand_answer_path = fusion_param['answer_path'].replace(ctrl_msg['data_folder_path'], ctrl_msg['data_folder_path'] + '_rand')
rand_folder_path = folder(rand_answer_path)
safe_mkdir(rand_folder_path)
# although copy all info including pid info, but not use in later training
shutil.copy(fusion_param['answer_path'], rand_answer_path)
rand_path = rand_folder_path + '/renew_pid.log'
safe_remove(rand_path)
origin_tracks = get_tracks(fusion_param)
pid_cnt = len(origin_tracks)
origin_pids = map(lambda x: x+1, range(pid_cnt))
persons_rand_predict_idx_s = [random.sample(origin_pids, pid_cnt) for _ in range(pid_cnt)]
write_content = ''
for rand_predict_idx_s in persons_rand_predict_idx_s:
for rand_predict_idx in rand_predict_idx_s:
write_content += str(rand_predict_idx) + ' '
write_content += '\n'
write(rand_path, write_content)
def simple_fusion_st_img_ranker(fusion_param):
ep = fusion_param['ep']
en = fusion_param['en']
# 从renew_pid和renew_ac获取预测的人物id和图像分数
persons_ap_scores = predict_img_scores(fusion_param)
persons_ap_pids = predict_pids(fusion_param)
# 从磁盘获取之前建立的时空模型，以及随机时空模型
camera_delta_s = pickle_load(fusion_param['distribution_pickle_path'])
real_tracks = train_tracks(fusion_param)
# 计算时空评分和随机时空评分
persons_track_scores = predict_track_scores(real_tracks, camera_delta_s, fusion_param)
persons_cross_scores = list()
log_path = fusion_param['eval_fusion_path']
map_score_path = fusion_param['fusion_normal_score_path']
safe_remove(map_score_path)
safe_remove(log_path)
line_log_cnt = 10
for i, person_ap_pids in enumerate(persons_ap_pids):
cross_scores = list()
for j, person_ap_pid in enumerate(person_ap_pids):
cross_score = persons_track_scores[i][j] * persons_ap_scores[i][j]
cross_scores.append(cross_score)
persons_cross_scores.append(cross_scores)
print 'img score ready'
max_score = max([max(predict_cross_scores) for predict_cross_scores in persons_cross_scores])
for i, person_cross_scores in enumerate(persons_cross_scores):
for j, person_cross_score in enumerate(person_cross_scores):
if person_cross_score > 0:
# diff seq not sort and not normalize
persons_cross_scores[i][j] /= max_score
else:
persons_cross_scores[i][j] *= -1.
# if real_tracks[i][1] == real_tracks[persons_ap_pids[i][j]][1]:
# # print '%d, %d' % (i, j)
# persons_cross_scores[i][j] = 0
person_score_idx_s = list()
top1_scores = list()
print 'above person score ready'
for i, person_cross_scores in enumerate(persons_cross_scores):
# 单个probe的预测结果中按score排序，得到index，用于对pid进行排序
sort_score_idx_s = sorted(range(len(person_cross_scores)), key=lambda k: -person_cross_scores[k])
person_score_idx_s.append(sort_score_idx_s)
# 统计top1分布，后面计算中位数用
top1_scores.append(person_cross_scores[sort_score_idx_s[0]])
# 降序排，取前60%处的分数
sorted_top1_scores = sorted(top1_scores, reverse=True)
mid_score = sorted_top1_scores[int(len(sorted_top1_scores) * 0.5)]
mid_score_path = fusion_param['mid_score_path']
safe_remove(mid_score_path)
write(mid_score_path, '%f\n' % mid_score)
print(str(mid_score))
sorted_persons_ap_pids = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
sorted_persons_ap_scores = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
for i, person_ap_pids in enumerate(persons_ap_pids):
for j in range(len(person_ap_pids)):
sorted_persons_ap_pids[i][j] = persons_ap_pids[i][person_score_idx_s[i][j]]
sorted_persons_ap_scores[i][j] = persons_cross_scores[i][person_score_idx_s[i][j]]
np.savetxt(log_path, sorted_persons_ap_pids, fmt='%d')
np.savetxt(map_score_path, sorted_persons_ap_scores, fmt='%f')
def fusion_st_img_ranker(fusion_param):
ep = fusion_param['ep']
en = fusion_param['en']
# 从renew_pid和renew_ac获取预测的人物id和图像分数
persons_ap_scores = predict_img_scores(fusion_param)
persons_ap_pids = predict_pids(fusion_param)
# 从磁盘获取之前建立的时空模型，以及随机时空模型
camera_delta_s = pickle_load(fusion_param['distribution_pickle_path'])
rand_delta_s = pickle_load(fusion_param['rand_distribution_pickle_path'])
# diff_delta_s = pickle_load(fusion_param['rand_distribution_pickle_path'].replace('rand', 'diff'))
real_tracks = train_tracks(fusion_param)
# 计算时空评分和随机时空评分
persons_track_scores = predict_track_scores(real_tracks, camera_delta_s, fusion_param)
rand_track_scores = predict_track_scores(real_tracks, rand_delta_s, fusion_param)
diff_track_scores = rand_track_scores #predict_track_scores(real_tracks, diff_delta_s, fusion_param)
persons_cross_scores = list()
log_path = fusion_param['eval_fusion_path']
map_score_path = fusion_param['fusion_normal_score_path']
safe_remove(map_score_path)
safe_remove(log_path)
line_log_cnt = 10
for i, person_ap_scores in enumerate(persons_ap_scores):
cur_max_vision = max(person_ap_scores)
cur_min_vision = min(person_ap_scores)
persons_ap_scores[i] = (persons_ap_scores[i] - cur_min_vision) / (cur_max_vision - cur_min_vision)
persons_ap_scores[i] = np.exp(persons_ap_scores[i] * 3)
cur_max_vision = max(persons_ap_scores[i])
cur_min_vision = min(persons_ap_scores[i])
persons_ap_scores[i] = (persons_ap_scores[i] - cur_min_vision) / (cur_max_vision - cur_min_vision)
for i, person_ap_pids in enumerate(persons_ap_pids):
cross_scores = list()
for j, person_ap_pid in enumerate(person_ap_pids):
if rand_track_scores[i][j] < 0.00002:
cross_score = (persons_track_scores[i][j]*(1-ep) - en*diff_track_scores[i][j]) * (persons_ap_scores[i][j]+ep/(1-ep-en)) / 0.00002
else:
cross_score = (persons_track_scores[i][j] * (1 - ep) - en * diff_track_scores[i][j]) * (
persons_ap_scores[i][j] + ep / (1 - ep - en)) / rand_track_scores[i][j]
cross_scores.append(cross_score)
persons_cross_scores.append(cross_scores)
print 'img score ready'
max_score = max([max(predict_cross_scores) for predict_cross_scores in persons_cross_scores])
print 'max_cross_score %f' % max_score
for i, person_cross_scores in enumerate(persons_cross_scores):
for j, person_cross_score in enumerate(person_cross_scores):
if person_cross_score > 0: # diff camera
# diff seq not sort and not normalize
persons_cross_scores[i][j] /= max_score
else: # same camera and (diff camere && rand score == -1 )
persons_cross_scores[i][j] = persons_ap_scores[i][j]
# if real_tracks[i][1] == real_tracks[persons_ap_pids[i][j]][1]:
# # print '%d, %d' % (i, j)
# persons_cross_scores[i][j] = 0
if i == 0 and j % 100 == 0:
print 'st:%f st_diff%f vis:%f fus:%f' % (
persons_track_scores[i][j], diff_track_scores[i][j], persons_ap_scores[i][j], persons_cross_scores[i][j])
person_score_idx_s = list()
top1_scores = list()
print 'above person score ready'
for i, person_cross_scores in enumerate(persons_cross_scores):
# 单个probe的预测结果中按score排序，得到index，用于对pid进行排序
sort_score_idx_s = sorted(range(len(person_cross_scores)), key=lambda k: -person_cross_scores[k])
person_score_idx_s.append(sort_score_idx_s)
# 统计top1分布，后面计算中位数用
top1_scores.append(person_cross_scores[sort_score_idx_s[0]])
# 降序排，取前60%处的分数
sorted_top1_scores = sorted(top1_scores, reverse=True)
mid_score = sorted_top1_scores[int(len(sorted_top1_scores) * 0.5)]
mid_score_path = fusion_param['mid_score_path']
safe_remove(mid_score_path)
write(mid_score_path, '%f\n' % mid_score)
print(str(mid_score))
sorted_persons_ap_pids = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
sorted_persons_ap_scores = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
for i, person_ap_pids in enumerate(persons_ap_pids):
for j in range(len(person_ap_pids)):
sorted_persons_ap_pids[i][j] = persons_ap_pids[i][person_score_idx_s[i][j]]
sorted_persons_ap_scores[i][j] = persons_cross_scores[i][person_score_idx_s[i][j]]
np.savetxt(log_path, sorted_persons_ap_pids, fmt='%d')
np.savetxt(map_score_path, sorted_persons_ap_scores, fmt='%f')
def conv_train(train_dataset,
train_labels,
valid_dataset,
valid_labels,
test_dataset,
test_labels,
image_size,
num_labels,
basic_hps,
stride_ps,
lrd=False):
batch_size = basic_hps['batch_size']
patch_size = basic_hps['patch_size']
depth = basic_hps['depth']
num_hidden = basic_hps['num_hidden']
num_channels = 1
layer_cnt = basic_hps['layer_sum']
loss_collect = list()
first_hidden_num = basic_hps['num_hidden']
second_hidden_num = first_hidden_num / 2 + 1
graph = tf.Graph()
with graph.as_default():
# Input data.
tf_train_dataset = tf.placeholder(tf.float32,
shape=(batch_size, image_size,
image_size, num_channels))
tf_train_labels = tf.placeholder(tf.float32,
shape=(batch_size, num_labels))
tf_valid_dataset = tf.constant(valid_dataset)
tf_test_dataset = tf.constant(test_dataset)
input_weights = tf.Variable(
tf.truncated_normal([patch_size, patch_size, num_channels, depth],
stddev=0.1))
input_biases = tf.Variable(tf.zeros([depth]))
mid_layer_cnt = layer_cnt - 1
layer_weights = list()
layer_biases = [
tf.Variable(tf.constant(1.0, shape=[depth / (i + 2)]))
for i in range(mid_layer_cnt)
]
output_weights = list()
output_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
first_nn_weights = tf.Variable(
tf.truncated_normal([first_hidden_num, second_hidden_num],
stddev=0.1))
second_nn_weights = tf.Variable(
tf.truncated_normal([second_hidden_num, num_labels], stddev=0.1))
first_nn_biases = tf.Variable(
tf.constant(1.0, shape=[second_hidden_num]))
second_nn_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
# Model.
def model(data, init=False):
# Variables.
if not large_data_size(data) or not large_data_size(input_weights):
stride_ps[0] = [1, 1, 1, 1]
conv = tf.nn.conv2d(data,
input_weights,
stride_ps[0],
use_cudnn_on_gpu=True,
padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + input_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.8)
for i in range(mid_layer_cnt):
# print(hidden)
if init:
# avoid filter shape larger than input shape
hid_shape = hidden.get_shape()
# print(hid_shape)
filter_w = patch_size / (i + 1)
filter_h = patch_size / (i + 1)
# print(filter_w)
# print(filter_h)
if filter_w > hid_shape[1]:
filter_w = int(hid_shape[1])
if filter_h > hid_shape[2]:
filter_h = int(hid_shape[2])
layer_weight = tf.Variable(
tf.truncated_normal(shape=[
filter_w, filter_h, depth / (i + 1),
depth / (i + 2)
],
stddev=0.1))
layer_weights.append(layer_weight)
if not large_data_size(hidden) or not large_data_size(
layer_weights[i]):
# print("is not large data")
stride_ps[i + 1] = [1, 1, 1, 1]
# print(stride_ps[i + 1])
# print(len(stride_ps))
# print(i + 1)
conv = tf.nn.conv2d(hidden,
layer_weights[i],
stride_ps[i + 1],
use_cudnn_on_gpu=True,
padding='SAME')
if not large_data_size(conv):
conv = maxpool2d(conv, 1, 1)
else:
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer_biases[i])
if init:
hidden = tf.nn.dropout(hidden, 0.8)
shapes = hidden.get_shape().as_list()
shape_mul = 1
for s in shapes[1:]:
shape_mul *= s
if init:
output_size = shape_mul
output_weights.append(
tf.Variable(
tf.truncated_normal([output_size, num_hidden],
stddev=0.1)))
reshape = tf.reshape(hidden, [shapes[0], shape_mul])
hidden = tf.nn.relu6(
tf.matmul(reshape, output_weights[0]) + output_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, first_nn_weights) + first_nn_biases
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, second_nn_weights) + second_nn_biases
return hidden
# Training computation.
logits = model(tf_train_dataset, init=True)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))
# Optimizer.
starter_learning_rate = 0.1
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
learning_rate = tf.train.exponential_decay(starter_learning_rate,
cur_step,
10000,
0.96,
staircase=True)
optimizer = tf.train.GradientDescentOptimizer(
learning_rate).minimize(loss, global_step=cur_step)
else:
optimizer = tf.train.AdagradOptimizer(
starter_learning_rate).minimize(loss)
# Predictions for the training, validation, and test data.
train_prediction = tf.nn.softmax(logits)
valid_prediction = tf.nn.softmax(model(tf_valid_dataset))
test_prediction = tf.nn.softmax(model(tf_test_dataset))
num_steps = 3001
start_fit = 600
init_loss = []
with tf.Session(graph=graph) as session:
tf.initialize_all_variables().run()
print('Initialized')
end_train = False
mean_loss = 0
for step in range(num_steps):
if end_train:
break
offset = (step * batch_size) % (train_labels.shape[0] - batch_size)
batch_data = train_dataset[offset:(offset + batch_size), :, :, :]
batch_labels = train_labels[offset:(offset + batch_size), :]
feed_dict = {
tf_train_dataset: batch_data,
tf_train_labels: batch_labels
}
_, l, predictions = session.run(
[optimizer, loss, train_prediction], feed_dict=feed_dict)
mean_loss += l
if step % 5 == 0:
mean_loss /= 5.0
loss_collect.append(mean_loss)
mean_loss = 0
if step >= start_fit:
# print(loss_collect)
if step == start_fit:
res = fit_more(1, [
batch_size, depth, num_hidden, layer_cnt,
patch_size
], loss_collect)
else:
res = fit_more(0, [
batch_size, depth, num_hidden, layer_cnt,
patch_size
], loss_collect)
loss_collect.remove(loss_collect[0])
ret = res['ret']
if ret == 1:
print('ret is end train when step is {step}'.format(
step=step))
init_loss.append(loss_collect)
more_index = predict_future([
batch_size, depth, num_hidden, layer_cnt,
patch_size
], init_loss[0])
print('more index: %d' % more_index)
for i in range(more_index):
offset = ((step + i + 1) * batch_size) % (
train_labels.shape[0] - batch_size)
batch_data = train_dataset[offset:(
offset + batch_size), :, :, :]
batch_labels = train_labels[offset:(offset +
batch_size), :]
feed_dict = {
tf_train_dataset: batch_data,
tf_train_labels: batch_labels
}
_, l, predictions = session.run(
[optimizer, loss, train_prediction],
feed_dict=feed_dict)
loss_collect.append(l)
file_helper.write(
'/home/cwh/coding/python/NN/line.txt',
str(loss_collect[20]))
loss_collect.remove(loss_collect[0])
for loss in loss_collect[21:]:
file_helper.write(
'/home/cwh/coding/python/NN/line.txt',
str(loss))
end_train = True
file_helper.write(
'/home/cwh/coding/python/NN/line.txt', '===')
if step % 50 == 0:
print('Minibatch loss at step %d: %f' % (step, l))
print('Validation accuracy: %.1f%%' %
accuracy(valid_prediction.eval(), valid_labels))
print('Test accuracy: %.1f%%' %
accuracy(test_prediction.eval(), test_labels))
def save_fusion_param():
write('data/fusion_param.json', fusion_param.__str__)
def conv_train(train_dataset, train_labels, valid_dataset, valid_labels, test_dataset, test_labels, image_size,
num_labels, basic_hps, stride_ps, lrd=False):
batch_size = basic_hps['batch_size']
patch_size = basic_hps['patch_size']
depth = basic_hps['depth']
num_hidden = basic_hps['num_hidden']
num_channels = 1
layer_cnt = basic_hps['layer_sum']
loss_collect = list()
first_hidden_num = basic_hps['num_hidden']
second_hidden_num = first_hidden_num / 2 + 1
graph = tf.Graph()
with graph.as_default():
# Input data.
tf_train_dataset = tf.placeholder(
tf.float32, shape=(batch_size, image_size, image_size, num_channels))
tf_train_labels = tf.placeholder(tf.float32, shape=(batch_size, num_labels))
tf_valid_dataset = tf.constant(valid_dataset)
tf_test_dataset = tf.constant(test_dataset)
input_weights = tf.Variable(tf.truncated_normal(
[patch_size, patch_size, num_channels, depth], stddev=0.1))
input_biases = tf.Variable(tf.zeros([depth]))
mid_layer_cnt = layer_cnt - 1
layer_weights = list()
layer_biases = [tf.Variable(tf.constant(1.0, shape=[depth / (i + 2)])) for i in range(mid_layer_cnt)]
output_weights = list()
output_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
first_nn_weights = tf.Variable(tf.truncated_normal(
[first_hidden_num, second_hidden_num], stddev=0.1))
second_nn_weights = tf.Variable(tf.truncated_normal(
[second_hidden_num, num_labels], stddev=0.1))
first_nn_biases = tf.Variable(tf.constant(1.0, shape=[second_hidden_num]))
second_nn_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
# Model.
def model(data, init=False):
# Variables.
if not large_data_size(data) or not large_data_size(input_weights):
stride_ps[0] = [1, 1, 1, 1]
conv = tf.nn.conv2d(data, input_weights, stride_ps[0], use_cudnn_on_gpu=True, padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + input_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.8)
for i in range(mid_layer_cnt):
# print(hidden)
if init:
# avoid filter shape larger than input shape
hid_shape = hidden.get_shape()
# print(hid_shape)
filter_w = patch_size / (i + 1)
filter_h = patch_size / (i + 1)
# print(filter_w)
# print(filter_h)
if filter_w > hid_shape[1]:
filter_w = int(hid_shape[1])
if filter_h > hid_shape[2]:
filter_h = int(hid_shape[2])
layer_weight = tf.Variable(tf.truncated_normal(shape=[filter_w, filter_h, depth / (i + 1), depth / (i + 2)],
stddev=0.1))
layer_weights.append(layer_weight)
if not large_data_size(hidden) or not large_data_size(layer_weights[i]):
# print("is not large data")
stride_ps[i + 1] = [1, 1, 1, 1]
# print(stride_ps[i + 1])
# print(len(stride_ps))
# print(i + 1)
conv = tf.nn.conv2d(hidden, layer_weights[i], stride_ps[i + 1], use_cudnn_on_gpu=True, padding='SAME')
if not large_data_size(conv):
conv = maxpool2d(conv, 1, 1)
else:
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer_biases[i])
if init:
hidden = tf.nn.dropout(hidden, 0.8)
shapes = hidden.get_shape().as_list()
shape_mul = 1
for s in shapes[1:]:
shape_mul *= s
if init:
output_size = shape_mul
output_weights.append(tf.Variable(tf.truncated_normal([output_size, num_hidden], stddev=0.1)))
reshape = tf.reshape(hidden, [shapes[0], shape_mul])
hidden = tf.nn.relu6(tf.matmul(reshape, output_weights[0]) + output_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, first_nn_weights) + first_nn_biases
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, second_nn_weights) + second_nn_biases
return hidden
# Training computation.
logits = model(tf_train_dataset, init=True)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=tf_train_labels))
# Optimizer.
starter_learning_rate = 0.1
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
learning_rate = tf.train.exponential_decay(starter_learning_rate, cur_step, 10000, 0.96, staircase=True)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=cur_step)
else:
optimizer = tf.train.AdagradOptimizer(starter_learning_rate).minimize(loss)
# Predictions for the training, validation, and test data.
train_prediction = tf.nn.softmax(logits)
valid_prediction = tf.nn.softmax(model(tf_valid_dataset))
test_prediction = tf.nn.softmax(model(tf_test_dataset))
num_steps = 3001
start_fit = 600
init_loss = []
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
print('Initialized')
end_train = False
mean_loss = 0
for step in range(num_steps):
if end_train:
break
offset = (step * batch_size) % (train_labels.shape[0] - batch_size)
batch_data = train_dataset[offset:(offset + batch_size), :, :, :]
batch_labels = train_labels[offset:(offset + batch_size), :]
feed_dict = {tf_train_dataset: batch_data, tf_train_labels: batch_labels}
_, l, predictions = session.run(
[optimizer, loss, train_prediction], feed_dict=feed_dict)
mean_loss += l
if step % 5 == 0:
mean_loss /= 5.0
loss_collect.append(mean_loss)
mean_loss = 0
if step >= start_fit:
# print(loss_collect)
if step == start_fit:
res = fit_more(1, [batch_size, depth, num_hidden, layer_cnt, patch_size], loss_collect)
else:
res = fit_more(0, [batch_size, depth, num_hidden, layer_cnt, patch_size], loss_collect)
loss_collect.remove(loss_collect[0])
ret = res['ret']
if ret == 1:
print('ret is end train when step is {step}'.format(step=step))
init_loss.append(loss_collect)
more_index = predict_future([batch_size, depth, num_hidden, layer_cnt, patch_size], init_loss[0])
print('more index: %d' % more_index)
for i in range(more_index):
offset = ((step + i + 1) * batch_size) % (train_labels.shape[0] - batch_size)
batch_data = train_dataset[offset:(offset + batch_size), :, :, :]
batch_labels = train_labels[offset:(offset + batch_size), :]
feed_dict = {tf_train_dataset: batch_data, tf_train_labels: batch_labels}
_, l, predictions = session.run(
[optimizer, loss, train_prediction], feed_dict=feed_dict)
loss_collect.append(l)
file_helper.write('/home/cwh/coding/python/NN/line.txt', str(loss_collect[20]))
loss_collect.remove(loss_collect[0])
for loss in loss_collect[21:]:
file_helper.write('/home/cwh/coding/python/NN/line.txt', str(loss))
end_train = True
file_helper.write('/home/cwh/coding/python/NN/line.txt', '===')
if step % 50 == 0:
print('Minibatch loss at step %d: %f' % (step, l))
print('Validation accuracy: %.1f%%' % accuracy(
valid_prediction.eval(), valid_labels))
print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))
from util.file_helper import read_lines, write
import seaborn as sns
import numpy as np
if __name__ == '__main__':
# for i in range(10):
for i in range(1):
person_lines = read_lines('data/top-m2g-std%d-train/cross_filter_pid.log' % i)
predict_persons = [predict_line.split() for predict_line in person_lines]
score_lines = read_lines('data/top-m2g-std%d-train/cross_filter_score.log' % i)
predict_scores = [score_line.split() for score_line in score_lines]
write('data/top1.txt', 'std %d top1\n' % i)
scores = list()
pos_cnt = 0
in_cnt = 0
up_b = 1.
down_b = 0.16
for k in range(10):
for j in range(len(predict_persons)):
score = float(predict_scores[j][0])
r = int(predict_persons[j][0]) - j - 1
if abs(r) == 1:
write('data/top1.txt', 'left %d, right %d, score %f\n' % (j + 1, int(predict_persons[j][0]), score))
scores.append(score)
if down_b < score < up_b:
in_cnt += 1
if abs(r) == 1:
pos_cnt += 1
if in_cnt == 0:
in_cnt = 1