tensorlayer / tensorlayer-chinese Goto Github PK

tutoiral*.py utils.py cost.py 等更新了

Conv2d(n, 64, (5, 5), (1, 1), padding='SAME', name='cnn2')
这个函数Conv2d的W_init_args = {}, b_init_args = {}参数没有根据输入加入dtype，而参数初始化本身是做float32的，导致了无法运行。
需加入W_init_args={"dtype":tf.float16},b_init_args={"dtype":tf.float16}参数

with tl.distributed.DistributedSession(task_spec=task_spec) as sess:
tl.files.load_and_assign_npz(sess=sess, name='n.npz', network=net)
这样写会报错：
RuntimeError: Graph is finalized and cannot be modified.

帅哥，我电脑wechat掉线了，在公海上手机没法scan QR code了。。突然发现没有你的邮箱，有事Line或邮件联系。。

tl.prepro.obj_box_crop
when I set “is_center=False, is_random=True ”
the changed label is not in the correct position
I just found.... did I use this function wrong?
I cannot post pic...

是否需要自定义输入层来实现？
参考例子：Keras/多输入和多输出模型

/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/importlib/_bootstrap.py:205: RuntimeWarning: compiletime version 3.5 of module 'tensorflow.python.framework.fast_tensor_util' does not match runtime version 3.6
return f(*args, **kwds)

是3.6版本的python不兼容吗

@lgarithm @luomai

• 与英文文档同步，换成 numpydoc 格式
• 去除ops
• 欢迎补充

When I look at the source code of tl.layers.seq2seq, something confused me. In traditional seq2seq model, the encode part should return the last output instead of all output. But in the "network_encode" part, “return_last” has been set to "False", that means encode part will return all outputs. Why? Isn't it better to set "return_last=True"? Thanks!
（英文水平不怎么好，所以还是对问题加一段中文翻译好了：经典seq2seq模型的encode部分应该是前面隐藏单元的加权，并输出最后一个unit。但tl.layers.seq2seq源代码中network_encode部分，用DynamicRNNLayer函数进行编码，return_last函数被设置为False，是否意味着encode部分会输出所有单元呢？为什么不是设为True只输出最后一个呢？）

我在最新的官方文档上看到 RNN 的输入形状是这样的 [time_steps, batch_size, num_features]，它的 batch 在第二维，因为同时给出了一个例子，应该不是输入错误；
然后我看到 TL 的文档里输入是 [batch_size, n_steps, n_features]
是TL在内部重新做了处理吗，还是两种都可以?

num of params: 1276810
layer 0: Tensor("drop1/mul:0", shape=(?, 784), dtype=float32)
layer 1: Tensor("relu1/Relu:0", shape=(?, 800), dtype=float32)
layer 2: Tensor("drop2/mul:0", shape=(?, 800), dtype=float32)
layer 3: Tensor("relu2/Relu:0", shape=(?, 800), dtype=float32)
layer 4: Tensor("drop3/mul:0", shape=(?, 800), dtype=float32)
layer 5: Tensor("output_layer/Identity:0", shape=(?, 10), dtype=float32)
learning_rate: 0.000100
batch_size: 128
Epoch 1 of 1 took 1.700945s
train loss: 0.439682
val loss: 0.384359
val acc: 0.880900
Traceback (most recent call last):
File "tutorial_mnist.py", line 562, in
main_test_layers(model='relu') # model = relu, dropconnect
File "tutorial_mnist.py", line 154, in main_test_layers
to save the feature images for different dataset")
Exception: You should change visualize_W(), if you want to save the feature images for different dataset

当我用文档提供的例子tutorial_mnist_distributed做实验的时候，我发现当我把创建Session的代码改为：
with tl.distributed.DistributedSession(task_spec=task_spec,checkpoint_dir="checkpoint",save_checkpoint_secs=1000,save_summaries_steps=100,save_summaries_secs=None) as sess:
这时候会出现RuntimeError:"Global step should be created to use StepCounterHook."
另外单独指定task_spec和checkpoint_dir属性也会报错，
tl.distributed.DistributedSession(task_spec=task_spec,checkpoint_dir="checkpoint") as sess:
TypeError: unorderable types: object() > int()
这个错误可以通过指定save_summaries_steps和save_summaries_secs解决。
感觉tensorlayer对分布式的支持并不成熟，有很多问题。

• Autoencoder is a unsupervised learning models which able to extract representative features, it has become more widely used for learning generative models of data and Greedy layer-wise pre-train.
• 自编码器是一种能够提取具有代表性特征的无监督(Unsupervised)学习模型，
  它已经广泛使用于数据生成模式的学习与逐层贪婪的预训练(Greedy layer-wise pre-train)。
• Since every layer knows its immediate incoming layers, the output layer (or output layers) of a network double as a handle to the network as a whole, so usually this is the only thing we will pass on to the rest of the code.
• 由于每一层都知道它在一个网络中的直接输入层和（多个）输出接收层，**网络的输出层 double as a handle 对于网络作为一个整体，**所以通常这是我们唯一要传递给其他代码的内容。

重点在加粗的部分，有没有更加合适的翻译?

以TF为参考:

• tf.Variable
• tf.nn.bias_add
• tf.matmul
• tf.nn.relu
• tf.nn.dropout
• tf.reduce_mean
• tf.nn.sparse_softmax_cross_entropy_with_logits
• tf.optimizers.Adam

TF求导例子：

optimizer = tf.optimizers.Adam(learning_rate=0.0001)
with tf.GradientTape() as tape:
            ## compute outputs
            _logits = MLP(X_batch, foo=1)
            ## compute loss and update model
            _loss = tl.cost.cross_entropy(_logits, y_batch, name='train_loss')

grad = tape.gradient(_loss, MLP.weights)
optimizer.apply_gradients(zip(grad, MLP.weights))

我在 windows下通过 xshell 访问 centos，在 centos 中运行 example 中的 tutorial_mnist.py 就报异常了：QXcbConnection: Could not connect to display
原因应该是服务器中不能运行 GUI 程序
错误的位置应该是 tutorial_mnist.py 中第 390 行左右用到的 tl.visualize ，再进去就是 tl.visualize 这个文件中的 import matplotlib.pyplot as plt
我的解决方法：在 tl.visualize 源码中把被注释掉的 matplotlib.use('Agg') 去掉注释，就可以了。
现在我想请问一下有没有其他途径可以解决这个问题呢？

def train(x_data, y_data):

x_train, y_train, x_valid, y_valid = split_data(x_data, y_data)

# train
#
sess = tf.InteractiveSession()
x = tf.placeholder(tf.float32, (None, 227, 227, 3))
y_ = tf.placeholder(tf.int64, (None, 17))
net = model_alexnet(x, 17)
y = net.outputs

# define cost and metric
cost = tl.cost.cross_entropy(y, tf.argmax(y_, 1), name='cost')
correct_predic = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
acc = tf.reduce_mean(tf.cast(correct_predic, tf.float32))
# y_op = tf.argmax(tf.nn.softmax(y))
train_op = tf.train.AdamOptimizer(1e-5).minimize(cost)

tl.layers.initialize_global_variables(sess)

# 导入参数 --> 用同一段代码训练过得参数继续训练，提示load success但是fit的时候正确率还是从很
# 低开始，像是从头开始训练了
#
if os.path.isfile('net_params.npz'):
    print('load pre-train params...')
    tl.files.load_and_assign_npz(sess, 'net_params.npz', net)

tl.utils.fit(sess, net, train_op, cost, x_train, y_train, x, y_,
             acc, batch_size=32, n_epoch=100, print_freq=10,
             X_val=x_valid, y_val=y_valid, eval_train=True, tensorboard=True)
tl.files.save_npz(net.all_params, name='net_params.npz', sess=sess)
sess.close()

如上注释...

加了 DeConvxxLayer, SlimNetsLayer etc.