def train_net():
#------进入计算图--------
x_train, y_train, x_val, y_val = input_data.read_img(
FILEPATH, WIDTH, HEIGHT, CHANNELS, ratio)
x_train_batch, y_train_batch = input_data.bulid_batch(
x_train, y_train, BATCH_SIZE)
x_val_batch, y_val_batch = input_data.bulid_batch(x_val, y_val, BATCH_SIZE)
batch_train_len = x_train_batch.shape[0]
batch_val_len = x_val_batch.shape[0]
#定义网络 x为输入占位符 y为输出占位符
#image_max = tf.reduce_max(x_train, name='image_max')
#image_min = tf.reduce_min(x_train,name='image_min')
x = tf.placeholder(tf.float32,
shape=[BATCH_SIZE, HEIGHT, WIDTH, CHANNELS],
name='input')
y = tf.placeholder(tf.int64, shape=[BATCH_SIZE], name='labels_placeholder')
_, _, softmax_linear = model.build_network(x, True, False)
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits\
(logits=softmax_linear, labels=y, name='xentropy_per_example')
train_loss = tf.reduce_mean(cross_entropy, name='loss')
if modeltype == 'NOQUANT': goto.end
#fake quant插入到op之前
tf.contrib.quantize.create_training_graph(
input_graph=tf.get_default_graph(), quant_delay=2000)
label.end
train_step = trainning(train_loss, LEARNING_RATE)
#准确略计算
correct = tf.nn.in_top_k(softmax_linear, y, 1)
correct = tf.cast(correct, tf.float16)
train_acc = tf.reduce_mean(correct)
#------------结束计算图-------------
with tf.Session() as sess:
saver = tf.compat.v1.train.Saver()
sess.run(tf.global_variables_initializer())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
valstep = 0
#max = sess.run(image_max)
#min = sess.run(image_min)
#训练
try:
ckpt = tf.train.get_checkpoint_state(TRAIN_LOGS_DIR)
global_step = 0
if ckpt and ckpt.model_checkpoint_path:
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
saver.restore(sess, ckpt.model_checkpoint_path)
print('Loading success, global_step is %s' % global_step)
for i in range(MAX_STEP + 1):
#if_train = True
pos = i % batch_train_len
_, acc, loss = sess.run([train_step, train_acc, train_loss],
feed_dict={
x: x_train_batch[pos],
y: y_train_batch[pos]
})
#每50步打印一次准确率和损失函数
if i % 50 == 0:
print(
'Step %d, train loss = %.2f, train accuracy = %.2f%%' %
(i, loss, acc * 100.0))
#每200步用验证集的数据进行验证
if i % 200 == 0:
#if_train = False #量化模式下用变量替代占位符.注意 如果要用tflite的话,if_train不要用占位符！
vpos = valstep % batch_val_len
val_loss, val_acc = sess.run([train_loss, train_acc],
feed_dict={
x: x_val_batch[vpos],
y: y_val_batch[vpos]
})
valstep = valstep + 1
print(
'** Step %d, val loss = %.2f, val accuracy = %.2f%% **'
% (i, val_loss, val_acc * 100.0))
#每500步保存一次变量值
if i % 500 == 0:
checkpoint_path = os.path.join(TRAIN_LOGS_DIR,
'saved_model.ckpt')
tmpstep = i + int(global_step)
saver.save(sess, checkpoint_path, global_step=tmpstep)
except tf.errors.OutOfRangeError:
print('Done training -- epoch limit reached')
finally:
coord.request_stop()
coord.join(threads)
def deep_maxent_irl():
# hyper parameters
H = 48
W = 128
N_STATES = H * W
N_ACTIONS = 8
SHAPE = [H, W, N_STATES, N_ACTIONS]
DISCOUNT = 1
LEARNING_RATE_BASE = 0.001
DECAY_STEPS = 500
DECAY_RATE = 0.99
GRAPH_SAVE_INTERVAL = 50
IMG_PATH = '/home/zhuzeyu/real_datasets/DIRL_DataSets/orig_img/'
REF_PATH = '/home/zhuzeyu/real_datasets/DIRL_DataSets/track_ref/'
# create model directory
MODEL_DIR = "model"
if not tf.gfile.Exists(MODEL_DIR):
tf.gfile.MakeDirs(MODEL_DIR)
# placeholders
input_img = tf.placeholder(tf.float32, [None, H, W, 3], name='input_img')
grad_r_placeholder = tf.placeholder(tf.float32, [H*W, 1])
# define reward and loss
rewards = inference.inference(input_img)
rewards_flattened = tf.reshape(rewards, [N_STATES, 1])
theta = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
l2_loss = tf.reduce_mean([tf.nn.l2_loss(v) for v in theta])
l2_loss = l2_loss/100000.0
loss = tf.multiply(grad_r_placeholder, rewards_flattened)
loss = tf.reduce_sum(loss, name='loss') #+ l2_loss
# define training
global_step = tf.Variable(0, trainable=False)
lr = tf.train.exponential_decay(
LEARNING_RATE_BASE,
global_step,
DECAY_STEPS,
DECAY_RATE
)
optimizer = tf.train.GradientDescentOptimizer(lr)
train_step = optimizer.minimize(loss, global_step=global_step)
init = tf.global_variables_initializer()
saver = tf.train.Saver(max_to_keep=100)
extracts = get_queue(IMG_PATH)
#path_orig = IMG_PATH + extracts[447]
#path_ref = REF_PATH + extracts[447]
#terminal, start, img, traj, ref = input_data.read_img(SHAPE, path_orig, path_ref)
with tf.Session() as sess:
init.run()
for epoch in range(20):
for iteration in range(100):
path_orig = IMG_PATH + extracts[iteration]
path_ref = REF_PATH + extracts[iteration]
terminal, start, img, traj, ref = input_data.read_img(SHAPE, path_orig, path_ref)
# get rewards
r = sess.run(rewards, feed_dict={input_img: img})
r_np = np.reshape(r, [-1, ])
# get policy
_, policy = value_iteration.value_iteration(SHAPE, r_np, DISCOUNT, terminal)
# compute expected svf
mu_exp = compute_state_visitation_freq2(SHAPE, traj, start, policy)
# compute expert svf
# mu_D = demo_svf(traj, N_STATES)
mu_D = field_svf(ref)
# compute loss
grad_r = mu_exp - mu_D
index = np.sum(np.abs(grad_r))
grad_r = np.reshape(grad_r, [-1, 1]) # 原来是一维
# train
sess.run(train_step, feed_dict={grad_r_placeholder: grad_r, input_img: img})
lss = sess.run(loss, feed_dict={grad_r_placeholder: grad_r, input_img: img})
print(index)
print(lss)
# print(sess.run(l2_loss))
# save graph
if (iteration + 1) % GRAPH_SAVE_INTERVAL == 0:
MODEL_NAME = 'model' + str(epoch) + str(iteration) + '.ckpt'
saver.save(sess, os.path.join(MODEL_DIR, MODEL_NAME))
import numpy as np import tensorflow as tf from input_data import read_img from input_data import set_val from model import inference train_dir = 'train/' logs_train_dir = 'save/model' #将所有的图片resize成100*100 w=100 h=100 c=3 data,label,num_classes=read_img(train_dir) x_train,y_train,x_val,y_val=set_val(data,label,0.7) #0.7为划分比例 # 占位符 x = tf.placeholder(tf.float32, shape=[None, w, h, c], name='x') y_ = tf.placeholder(tf.int32, shape=[None, ], name='y_') regularizer = tf.contrib.layers.l2_regularizer(0.0001) # 返回一个执行L2正则化的函数.在损失函数上加上正则项是防止过拟合的一个重要方法 logits = inference(x, False, regularizer,num_classes) # (小处理)将logits乘以1赋值给logits_eval，定义name，方便在后续调用模型时通过tensor名字调用输出tensor b = tf.constant(value=1, dtype=tf.float32) logits_eval = tf.multiply(logits, b, name='logits_eval') loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y_) train_op = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss) correct_prediction = tf.equal(tf.cast(tf.argmax(logits, 1), tf.int32), y_) # tf.equal Returns:A `Tensor` of type `bool`.
def epoch(weight=None,
height=None,
dimension=None,
class_size=None,
train=False,
path=None,
ratio=None,
n_epoch=None,
batch_size=None,
model_path=None):
regularizer = tf.contrib.layers.l2_regularizer(0.001)
x = tf.placeholder(tf.float32,
shape=[None, weight, height, dimension],
name='x')
y_ = tf.placeholder(tf.int32, shape=[
None,
], name='y_')
# todo
# 选用不同的 CNN
logits, pred = interface(input_tensor=x,
regularizer=regularizer,
train=train,
class_size=class_size)
# logits, pred = cnn(input_sensor=x, class_size=class_size)
loss = tf.losses.sparse_softmax_cross_entropy(labels=y_, logits=logits)
train_op = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
correct_prediction = tf.equal(tf.cast(tf.argmax(logits, 1), tf.int32), y_)
acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# 输入数据
data, label = input_data.read_img(path, weight, height, dimension)
x_train, y_train, x_val, y_val = input_data.shuffle_and_period(
data, label, ratio)
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
for epoch in range(n_epoch):
train_loss, train_acc, n_batch = 0, 0, 0
for x_train_a, y_train_a in input_data.mini_batches(x_train,
y_train,
batch_size,
shuffle=True):
_, err, ac = sess.run([train_op, loss, acc],
feed_dict={
x: x_train_a,
y_: y_train_a
})
train_loss += err
train_acc += ac
n_batch += 1
print('Epoch %d - train loss: %f' % (epoch, (train_loss / n_batch)))
print('Epoch %d - train acc: %f' % (epoch, train_acc / n_batch))
# validation
val_loss, val_acc, n_batch = 0, 0, 0
for x_val_a, y_val_a in input_data.mini_batches(x_val,
y_val,
batch_size,
shuffle=False):
err, ac = sess.run([loss, acc],
feed_dict={
x: x_val_a,
y_: y_val_a
})
val_loss += err
val_acc += ac
n_batch += 1
print('Epoch %d - Validation loss: %f' % (epoch, val_loss / n_batch))
print('Epoch %d - Validation Accuracy: %f' % (epoch,
(val_acc / n_batch)))
if epoch % 5 == 0:
saver.save(sess, model_path + "save_net.ckpt", epoch)
print('Trained Model Saved.')
def main(unused_argv):
if FLAGS.job_name is None or FLAGS.job_name == "":
raise ValueError("Must specify an explicit `job_name`")
if FLAGS.task_index is None or FLAGS.task_index == "":
raise ValueError("Must specify an explicit `task_index`")
print("job name = %s" % FLAGS.job_name)
print("task index = %d" % FLAGS.task_index)
#Construct the cluster and start the server
# 读取集群描述信息
ps_spec = FLAGS.ps_hosts.split(",")
worker_spec = FLAGS.worker_hosts.split(",")
# Get the number of workers.
num_workers = len(worker_spec)
# 创建TensorFlow集群描述对象
cluster = tf.train.ClusterSpec({"ps": ps_spec, "worker": worker_spec})
# 为本地执行任务创建TensorFlow Server对象。
if not FLAGS.existing_servers:
# Not using existing servers. Create an in-process server.
# 创建本地Sever对象，从tf.train.Server这个定义开始，每个节点开始不同
# 根据执行的命令的参数(作业名字)不同，决定这个任务是哪个任务
# 如果作业名字是ps，进程就加入这里，作为参数更新的服务，等待其他工作节点给它提交参数更新的数据
# 如果作业名字是worker，就执行后面的计算任务
server = tf.train.Server(cluster,
job_name=FLAGS.job_name,
task_index=FLAGS.task_index)
# 如果是参数服务器，直接启动即可。这里，进程就会阻塞在这里
# 下面的tf.train.replica_device_setter代码会将参数批定给ps_server保管
if FLAGS.job_name == "ps":
server.join()
# 处理工作节点
# 找出worker的主节点，即task_index为0的点
is_chief = (FLAGS.task_index == 0)
# 如果使用gpu
if FLAGS.num_gpus > 0:
# Avoid gpu allocation conflict: now allocate task_num -> #gpu
# for each worker in the corresponding machine
gpu = (FLAGS.task_index % FLAGS.num_gpus)
# 分配worker到指定gpu上运行
worker_device = "/job:worker/task:%d/gpu:%d" % (FLAGS.task_index, gpu)
# 如果使用cpu
elif FLAGS.num_gpus == 0:
# Just allocate the CPU to worker server
# 把cpu分配给worker
cpu = 0
worker_device = "/job:worker/task:%d/cpu:%d" % (FLAGS.task_index, cpu)
# The device setter will automatically place Variables ops on separate
# parameter servers (ps). The non-Variable ops will be placed on the workers.
# The ps use CPU and workers use corresponding GPU
# 用tf.train.replica_device_setter将涉及变量操作分配到参数服务器上，使用CPU。将涉及非变量操作分配到工作节点上，使用上一步worker_device值。
# 在这个with语句之下定义的参数，会自动分配到参数服务器上去定义。如果有多个参数服务器，就轮流循环分配
with tf.device(
tf.train.replica_device_setter(worker_device=worker_device,
ps_device="/job:ps/cpu:0",
cluster=cluster)):
with tf.variable_scope('inputdata') as scope:
# 获取图片和标签集
train, train_label = input_data.read_img(FLAGS.train_dir)
# 生成批次
train_batch, train_label_batch = input_data.get_batch(
train, train_label, FLAGS.IMG_W, FLAGS.IMG_H, FLAGS.BATCH_SIZE,
FLAGS.CAPACITY)
# 定义全局步长，默认值为0
global_step = tf.Variable(0, name="global_step", trainable=False)
train_logits = model.inference(train_batch, FLAGS.BATCH_SIZE,
FLAGS.N_CLASSES)
train_loss = model.losses(train_logits, train_label_batch)
accuracy = model.evaluation(train_logits, train_label_batch)
# merge all summaries into a single "operation" which we can execute in a session
summary_op = tf.summary.merge_all()
init_op = tf.global_variables_initializer()
print("Variables initialized ...")
# 异步训练模式：自己计算完成梯度就去更新参数，不同副本之间不会去协调进度
opt = tf.train.AdamOptimizer(FLAGS.learning_rate)
# 同步训练模式
if FLAGS.sync_replicas:
if FLAGS.replicas_to_aggregate is None:
replicas_to_aggregate = num_workers
else:
replicas_to_aggregate = FLAGS.replicas_to_aggregate
# 使用SyncReplicasOptimizer作优化器，并且是在图间复制情况下
# 在图内复制情况下将所有梯度平均
opt = tf.train.SyncReplicasOptimizer(
opt,
replicas_to_aggregate=replicas_to_aggregate,
total_num_replicas=num_workers,
name="mnist_sync_replicas")
train_step = opt.minimize(train_loss, global_step=global_step)
if FLAGS.sync_replicas:
local_init_op = opt.local_step_init_op
if is_chief:
# 所有进行计算工作节点里一个主工作节点(chief)
# 主节点负责初始化参数、模型保存、概要保存
local_init_op = opt.chief_init_op
ready_for_local_init_op = opt.ready_for_local_init_op
# Initial token and chief queue runners required by the sync_replicas mode
# 同步训练模式所需初始令牌、主队列
chief_queue_runner = opt.get_chief_queue_runner()
sync_init_op = opt.get_init_tokens_op()
init_op = tf.global_variables_initializer()
if FLAGS.sync_replicas:
# 创建一个监管程序，用于统计训练模型过程中的信息
# lodger 是保存和加载模型路径
# 启动就会去这个logdir目录看是否有检查点文件，有的话就自动加载
# 没有就用init_op指定初始化参数
# 主工作节点(chief)负责模型参数初始化工作
# 过程中，其他工作节点等待主节眯完成初始化工作，初始化完成后，一起开始训练数据
# global_step值是所有计算节点共享的
# 在执行损失函数最小值时自动加1，通过global_step知道所有计算节点一共计算多少步
sv = tf.train.Supervisor(
is_chief=is_chief,
logdir=FLAGS.logs_train_dir,
init_op=init_op,
local_init_op=local_init_op,
ready_for_local_init_op=ready_for_local_init_op,
recovery_wait_secs=1,
global_step=global_step)
else:
sv = tf.train.Supervisor(is_chief=is_chief,
logdir=FLAGS.logs_train_dir,
init_op=init_op,
recovery_wait_secs=1,
global_step=global_step)
# 创建会话，设置属性allow_soft_placement为True
# 所有操作默认使用被指定设置，如GPU
# 如果该操作函数没有GPU实现，自动使用CPU设备
sess_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False,
device_filters=[
"/job:ps",
"/job:worker/task:%d" %
FLAGS.task_index
])
# The chief worker (task_index==0) session will prepare the session,
# while the remaining workers will wait for the preparation to complete.
# 主工作节点(chief)，task_index为0节点初始化会话
# 其余工作节点等待会话被初始化后进行计算
if is_chief:
print("Worker %d: Initializing session..." % FLAGS.task_index)
else:
print("Worker %d: Waiting for session to be initialized..." %
FLAGS.task_index)
if FLAGS.existing_servers:
server_grpc_url = "grpc://" + worker_spec[FLAGS.task_index]
print("Using existing server at: %s" % server_grpc_url)
# 创建TensorFlow会话对象，用于执行TensorFlow图计算
# prepare_or_wait_for_session需要参数初始化完成且主节点准备好后，才开始训练
sess = sv.prepare_or_wait_for_session(server_grpc_url,
config=sess_config)
else:
sess = sv.prepare_or_wait_for_session(server.target,
config=sess_config)
print("Worker %d: Session initialization complete." % FLAGS.task_index)
if FLAGS.sync_replicas and is_chief:
# Chief worker will start the chief queue runner and call the init op.
sess.run(sync_init_op)
global threads
threads = sv.start_queue_runners(sess, [chief_queue_runner])
else:
threads = sv.start_queue_runners(sess)
# Perform training
# 执行分布式模型训练
time_begin = time.time()
coord = tf.train.Coordinator()
print("Training begins @ %f" % time_begin)
local_step = 0
try:
for step in np.arange(FLAGS.MAX_STEP):
if coord.should_stop():
break
_, tra_loss, tra_acc = sess.run(
[train_step, train_loss, accuracy])
if step % 50 == 0:
print(
'Step %d, train loss = %.2f, train accuracy = %.2f%%' %
(step, tra_loss, tra_acc * 100.0))
except tf.errors.OutOfRangeError:
print('Done training -- epoch limit reached')
finally:
coord.request_stop()
coord.join(threads)
sess.close()
time_end = time.time()
print("Training ends @ %f" % time_end)
training_time = time_end - time_begin
print("Training elapsed time: %f s" % training_time)
import img_utils
import numpy as np
import mdp.value_iteration4 as vi
import mdp.gridworld2 as gridworld
H = 24
W = 64
N_STATES = H * W
TRAJ_LEN = 4
def int_to_point(i):
return i % W, i // W
__, img, _ = input_data.read_img(H, W)
input_img = tf.placeholder(tf.float32, [None, H, W, 3])
sess = tf.Session()
# load meta graph and restore weights
saver = tf.train.import_meta_graph(
'/Users/David/Desktop/model/model24*64/ckpt/model119.ckpt.meta')
saver.restore(sess, '/Users/David/Desktop/model/model24*64/ckpt/model119.ckpt')
# get all tensors. not necessary
graph = tf.get_default_graph()
tensor_name_list = [
tensor.name for tensor in tf.get_default_graph().as_graph_def().node
]
print(tensor_name_list)