Skip to content

wagamamaz/tensorlayer-tricks

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

76 Commits
 
 

Repository files navigation

How to use TensorLayer

While research in Deep Learning continues to improve the world, we use a bunch of tricks to implement algorithms with TensorLayer day to day.

Here are a summary of the tricks to use TensorLayer. If you find a trick that is particularly useful in practice, please open a Pull Request to add it to the document. If we find it to be reasonable and verified, we will merge it in.

1. Installation

  • To keep your TL version and edit the source code easily, you can download the whole repository by excuting git clone https://github.com/zsdonghao/tensorlayer.git in your terminal, then copy the tensorlayer folder into your project
  • As TL is growing very fast, if you want to use pip install, we suggest you to install the master version
  • For NLP application, you will need to install NLTK and NLTK data

2. Interaction between TF and TL

3. Training/Testing switching

def mlp(x, is_train=True, reuse=False):
    with tf.variable_scope("MLP", reuse=reuse):
      net = InputLayer(x, name='in')
      net = DropoutLayer(net, 0.8, True, is_train, name='drop1')
      net = DenseLayer(net, n_units=800, act=tf.nn.relu, name='dense1')
      net = DropoutLayer(net, 0.8, True, is_train, name='drop2')
      net = DenseLayer(net, n_units=800, act=tf.nn.relu, name='dense2')
      net = DropoutLayer(net, 0.8, True, is_train, name='drop3')
      net = DenseLayer(net, n_units=10, act=tf.identity, name='out')
      logits = net.outputs
      net.outputs = tf.nn.sigmoid(net.outputs)
      return net, logits
x = tf.placeholder(tf.float32, shape=[None, 784], name='x')
y_ = tf.placeholder(tf.int64, shape=[None, ], name='y_')
net_train, logits = mlp(x, is_train=True, reuse=False)
net_test, _ = mlp(x, is_train=False, reuse=True)
cost = tl.cost.cross_entropy(logits, y_, name='cost')

More in here.

4. Get variables and outputs

train_vars = tl.layers.get_variables_with_name('MLP', True, True)
train_op = tf.train.AdamOptimizer(learning_rate=0.0001).minimize(cost, var_list=train_vars)
layers = tl.layers.get_layers_with_name(network, "MLP", True)
  • This method usually be used for activation regularization.

5. Data augmentation for large dataset

If your dataset is large, data loading and data augmentation will become the bottomneck and slow down the training. To speed up the data processing you can:

6. Data augmentation for small dataset

If your data size is small enough to feed into the memory of your machine, and data augmentation is simple. To debug easily, you can:

7. Pre-trained CNN and Resnet

8. Using tl.models

  • Use pretrained VGG16 for ImageNet classification
x = tf.placeholder(tf.float32, [None, 224, 224, 3])
# get the whole model
vgg = tl.models.VGG16(x)
# restore pre-trained VGG parameters
sess = tf.InteractiveSession()
vgg.restore_params(sess)
# use for inferencing
probs = tf.nn.softmax(vgg.outputs)
  • Extract features with VGG16 and retrain a classifier with 100 classes
x = tf.placeholder(tf.float32, [None, 224, 224, 3])
# get VGG without the last layer
vgg = tl.models.VGG16(x, end_with='fc2_relu')
# add one more layer
net = tl.layers.DenseLayer(vgg, 100, name='out')
# initialize all parameters
sess = tf.InteractiveSession()
tl.layers.initialize_global_variables(sess)
# restore pre-trained VGG parameters
vgg.restore_params(sess)
# train your own classifier (only update the last layer)
train_params = tl.layers.get_variables_with_name('out')
  • Reuse model
x1 = tf.placeholder(tf.float32, [None, 224, 224, 3])
x2 = tf.placeholder(tf.float32, [None, 224, 224, 3])
# get VGG without the last layer
vgg1 = tl.models.VGG16(x1, end_with='fc2_relu')
# reuse the parameters of vgg1 with different input
vgg2 = tl.models.VGG16(x2, end_with='fc2_relu', reuse=True)
# restore pre-trained VGG parameters (as they share parameters, we don’t need to restore vgg2)
sess = tf.InteractiveSession()
vgg1.restore_params(sess)

9. Customized layer

    1. Write a TL layer directly
    1. Use LambdaLayer, it can also accept functions with new variables. With this layer you can connect all third party TF libraries and your customized function to TL. Here is an example of using Keras and TL together.
import tensorflow as tf
import tensorlayer as tl
from keras.layers import *
from tensorlayer.layers import *
def my_fn(x):
    x = Dropout(0.8)(x)
    x = Dense(800, activation='relu')(x)
    x = Dropout(0.5)(x)
    x = Dense(800, activation='relu')(x)
    x = Dropout(0.5)(x)
    logits = Dense(10, activation='linear')(x)
    return logits

network = InputLayer(x, name='input')
network = LambdaLayer(network, my_fn, name='keras')
...

10. Sentences tokenization

>>> captions = ["one two , three", "four five five"] # 2个 句 子 
>>> processed_capts = []
>>> for c in captions:
>>>    c = tl.nlp.process_sentence(c, start_word="<S>", end_word="</S>")
>>>    processed_capts.append(c)
>>> print(processed_capts)
... [['<S>', 'one', 'two', ',', 'three', '</S>'],
... ['<S>', 'four', 'five', 'five', '</S>']]
>>> tl.nlp.create_vocab(processed_capts, word_counts_output_file='vocab.txt', min_word_count=1)
... [TL] Creating vocabulary.
... Total words: 8
... Words in vocabulary: 8
... Wrote vocabulary file: vocab.txt
  • Finally use tl.nlp.Vocabulary to create a vocabulary object from the txt vocabulary file created by tl.nlp.create_vocab
>>> vocab = tl.nlp.Vocabulary('vocab.txt', start_word="<S>", end_word="</S>", unk_word="<UNK>")
... INFO:tensorflow:Initializing vocabulary from file: vocab.txt
... [TL] Vocabulary from vocab.txt : <S> </S> <UNK>
... vocabulary with 10 words (includes start_word, end_word, unk_word)
...   start_id: 2
...   end_id: 3
...   unk_id: 9
...   pad_id: 0

Then you can map word to ID or vice verse as follow:

>>> vocab.id_to_word(2)
... 'one'
>>> vocab.word_to_id('one')
... 2
>>> vocab.id_to_word(100)
... '<UNK>'
>>> vocab.word_to_id('hahahaha')
... 9

11. Dynamic RNN and sequence length

  • Apply zero padding on a batch of tokenized sentences as follow:
>>> sequences = [[1,1,1,1,1],[2,2,2],[3,3]]
>>> sequences = tl.prepro.pad_sequences(sequences, maxlen=None, 
...         dtype='int32', padding='post', truncating='pre', value=0.)
... [[1 1 1 1 1]
...  [2 2 2 0 0]
...  [3 3 0 0 0]]
>>> data = [[1,2,0,0,0], [1,2,3,0,0], [1,2,6,1,0]]
>>> o = tl.layers.retrieve_seq_length_op2(data)
>>> sess = tf.InteractiveSession()
>>> tl.layers.initialize_global_variables(sess)
>>> print(o.eval())
... [2 3 4]

12. Save models

    1. tl.files.save_npz save all model parameters (weights) into a a list of array, restore using tl.files.load_and_assign_npz
    1. tl.files.save_npz_dict save all model parameters (weights) into a dictionary of array, key is the parameter name, restore using tl.files.load_and_assign_npz_dict
    1. tl.files.save_ckpt save all model parameters (weights) into TensorFlow ckpt file, restore using tl.files.load_ckpt.

13. Compatibility with other TF wrappers

TL can interact with other TF wrappers, which means if you find some codes or models implemented by other wrappers, you can just use it !

  • Other TensorFlow layer implementations can be connected into TensorLayer via LambdaLayer, see example here)
  • TF-Slim to TL: SlimNetsLayer (you can use all Google's pre-trained convolutional models with this layer !!!)

14. Others

  • BatchNormLayer's decay default is 0.9, set to 0.999 for large dataset.
  • Matplotlib issue arise when importing TensorLayer issues, see FQA

Useful links

Author

  • Zhang Rui
  • Hao Dong