#!/usr/bin/env python
# coding: utf-8
# # TensorFlow__MLP_ladder-network_MNIST-labeled100__pollenjp
#
# このノートブックは [nbviewer](http://nbviewer.jupyter.org/) を介して読むことをおすすめします。
# ( I recommend you to see this notebook on [nbviewer](http://nbviewer.jupyter.org/) .)
#
#
# - 論文
# - [Semi-Supervised Learning with Ladder Networks - Antti Rasmus, Harri Valpola, Mikko Honkala, Mathias Berglund, Tapani Raiko](https://arxiv.org/abs/1507.02672)
#
# - code
# - [rinuboney/ladder](https://github.com/rinuboney/ladder)
# - [tarvaina/tensorflow-ladder](https://github.com/tarvaina/tensorflow-ladder)
#
#
Table of Contents
#
# ## Config
# ### Import
# In[1]:
import os,sys
print(sys.version)
import re
from pathlib import Path
import math
# In[2]:
# OPTIONAL: Load the "autoreload" extension so that code can change
get_ipython().run_line_magic('load_ext', 'autoreload')
# OPTIONAL: always reload modules so that as you change code in src, it gets loaded
get_ipython().run_line_magic('autoreload', '2')
# If you want to reload manually, add a below line head.
get_ipython().run_line_magic('aimport', '')
# ref: https://ipython.org/ipython-doc/3/config/extensions/autoreload.html
import matplotlib.pyplot as plt
get_ipython().run_line_magic('matplotlib', 'inline')
import numpy as np
import sklearn
from sklearn import datasets
import tqdm
import csv
import pandas as pd
seed = None
np.random.seed(seed=seed)
print("numpy ver: {}".format(np.__version__))
print("scikit-learn ver: {}".format(sklearn.__version__))
print("pandas ver: {}".format(pd.__version__))
# ### TensorFlow
# In[3]:
#____________________________________________________________________________________________________
# TensorFlow and Keras GPU configures
##________________________________________________________________________________
## OPTIONAL : set a GPU viewed by TensorFlow
###____________________________________________________________
### - https://stackoverflow.com/questions/37893755/tensorflow-set-cuda-visible-devices-within-jupyter
import os
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" # see issue #152
os.environ["CUDA_VISIBLE_DEVICES"]="0"
##________________________________________________________________________________
##________________________________________________________________________________
## TensorFlow
###____________________________________________________________
import tensorflow as tf
print("tensorflow ver: {}".format(tf.__version__))
### eager mode
#tf.enable_eager_execution()
print("tf.executing_eagerly(): {}".format(tf.executing_eagerly()))
# You can double check that you have the correct devices visible to TF
# - https://stackoverflow.com/questions/37893755/tensorflow-set-cuda-visible-devices-within-jupyter
from tensorflow.python.client import device_lib
print("""
________________________________________
Visible GPUs from TensorFlow
________________________________________""")
for _device in device_lib.list_local_devices():
match = re.search(pattern=r'name: "/device:(?P[A-Z]{3}):(?P\d{1})*',
string=str(_device))
if match is None:
print("Not Match")
continue
if match.group("name") == "CPU":
name, device_num = match.group("name", "device_num")
print()
print("({}:{})".format(name, device_num))
continue
name, device_num = match.group("name", "device_num")
match = re.search(pattern=r'.*pci bus id: (?P\d{4}:\d{2}:\d{2}.\d{1}).*',
string=str(_device))
if match is None:
print("No GPUs")
continue
print("({}:{}: pci_bus_id: {})".format(name, device_num, match.group("pci_bus_id")))
print("________________________________________")
###____________________________________________________________
### sessioin
global _SESSION
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)
config.gpu_options.allow_growth = True
#_SESSION = tf.Session(config=config)
###____________________________________________________________
##________________________________________________________________________________
#____________________________________________________________________________________________________
# ### Path
# In[4]:
HOME = Path(os.getcwd()).parent
print(HOME)
# In[5]:
path_list = []
data_Path = HOME / "data"
path_list.append(data_Path)
raw_Path = data_Path / "raw"
path_list.append(raw_Path)
plot_images_Path = data_Path / "plot_images"
path_list.append(plot_images_Path)
src_Path = HOME / "src"
path_list.append(src_Path)
for _Path in path_list:
_path = str(_Path)
if not os.path.exists(_path):
os.makedirs(name=_path)
print("make a directory: \n\t", _path)
else:
print(os.path.exists(_path), ": ", _path)
# ### Import from Local
# In[6]:
sys.path.append(str(src_Path))
# In[7]:
from data import mnist_input_data
from utils_tensorflow.tensorflow_graph_in_jupyer import show_computational_graph
# ## Ladder Networks
# In[8]:
tf.reset_default_graph()
# In[9]:
layer_sizes = [784, 1000, 500, 250, 250, 250, 10]
L = len(layer_sizes) - 1 # number of layers
num_examples = 60000
num_epochs = 150
num_labeled = 100
starter_learning_rate = 0.02
decay_after = 15 # epoch after which to begin learning rate decay
batch_size = 100
num_iter = (num_examples//batch_size) * num_epochs # number of loop iterations
# In[10]:
with tf.name_scope(name="PLACEHOLDER"):
inputs = tf.placeholder(tf.float32, shape=(None, layer_sizes[0]))
outputs = tf.placeholder(tf.float32)
# ### Initialize Layer's and Batch Normalization's Weights
# #### [DEF] b_i, w_i
# In[11]:
def b_i(inits, size, name):
return tf.Variable(inits * tf.ones([size]), name=name)
def w_i(shape, name):
return tf.Variable(tf.random_normal(shape, name=name)) / math.sqrt(shape[0])
# In[12]:
shapes = list(zip(list(layer_sizes)[:-1], list(layer_sizes[1:])))
weights = {
'W': [w_i(s, "W") for s in shapes], # Encoder weights
'V': [w_i(s[::-1], "V") for s in shapes], # Decoder weights
# batch normalization parameter to shift the normalized value
'beta': [b_i(0.0, layer_sizes[l+1], "beta") for l in range(L)],
# batch normalization parameter to scale the normalized value
'gamma': [b_i(1.0, layer_sizes[l+1], "beta") for l in range(L)],
}
# In[13]:
noise_std = 0.3 # scaling factor for noise used in corrupted encoder
# hyperparameters that denote the importance of each layer
denoising_cost = [1000.0, 10.0, 0.10, 0.10, 0.10, 0.10, 0.10]
# In[14]:
join = lambda l, u: tf.concat([l, u], 0)
labeled = lambda x : tf.slice(x, [0, 0], [batch_size, -1]) if x is not None else x
unlabeled = lambda x : tf.slice(x, [batch_size, 0], [-1, -1]) if x is not None else x
split_lu = lambda x : (labeled(x), unlabeled(x))
# In[15]:
training = tf.placeholder(tf.bool)
ewma = tf.train.ExponentialMovingAverage(decay=0.99) # to calculate the moving averages of mean and variance
bn_assigns= [] # this list stores the updates to be made to average mean and variance
# ### batch_normalization
# #### [DEF] batch_normalization
# In[16]:
def batch_normalization(batch, mean=None, var=None):
"""
Parameters
----------
batch :
mean :
var :
Returns
-------
normalized batch :
"""
if mean is None or var is None: # まだ平均と分散を計算していない場合
mean, var = tf.nn.moments(batch, axes=[0])
return (batch - mean) / tf.sqrt(var + tf.constant(1e-10))
# In[17]:
# average mean and variance of all layers
running_mean = [tf.Variable(tf.constant(0.0, shape=[l]), trainable=False) for l in layer_sizes[1:]]
running_var = [tf.Variable(tf.constant(1.0, shape=[l]), trainable=False) for l in layer_sizes[1:]]
# #### [DEF] update_batch_normalization
# In[18]:
def update_batch_normalization(batch, l):
"""
batch normalize + update average mean and variance of layer l
Parameters
----------
batch :
l : layer
Globals
-------
running_mean, running_var : list, These list stores average mean and variance of all layers
ewma : tf.train.ExponentialMovingAverage, Calculate the moving averages of mean and variance
bn_assigns : list, This list stores the updates to be made to average mean and variance
Returns
-------
normalized batch
"""
mean, var = tf.nn.moments(batch, axes=[0])
assign_mean = running_mean[l-1].assign(mean) # Update
assign_var = running_var[l-1].assign(var) # Update
bn_assigns.append(ewma.apply([running_mean[l-1], running_var[l-1]])) # Store moving averages
with tf.control_dependencies([assign_mean, assign_var]): # return after assign
return (batch - mean) / tf.sqrt(var + 1e-10)
# ### Encoder
# In[19]:
def encoder(inputs, noise_std):
"""
Parameters
----------
inputs :
noised_std : float,
noised_std != 0.0 --> Corrupted Encoder
noised_std == 0.0 --> Clean Encoder
Globals
-------
split_lu : func
layer_sizes : list
weights : dict
join : func
batch_normalization : func
running_mean, running_var : list, These list stores average mean and variance of all layers
Returns
-------
"""
h = inputs + tf.random_normal(tf.shape(inputs)) * noise_std # add noise to input
d = {} # to store the pre-activation, activation, mean and variance for each layer
# The data for labeled and unlabeled examples are stored separately
d['labeled'] = {'z': {}, 'm': {}, 'v': {}, 'h': {}} # m=mean, v=variance
d['unlabeled'] = {'z': {}, 'm': {}, 'v': {}, 'h': {}} # m=mean, v=variance
d['labeled']['z'][0], d['unlabeled']['z'][0] = split_lu(h)
for l in range(1, L+1):
print( "Layer {:>3}: {:>5} -> {:>5}".format(l,layer_sizes[l-1], layer_sizes[l]) )
d['labeled']['h'][l-1], d['unlabeled']['h'][l-1] = split_lu(h)
z_pre = tf.matmul(h, weights['W'][l-1]) # pre-activation
z_pre_l, z_pre_u = split_lu(z_pre) # split labeled and unlabeled examples
m, v = tf.nn.moments(z_pre_u, axes=[0]) # compute mean, variance using twice later (efficiency)
#----------------------------------------
# if training:
def training_batch_norm():
# Training batch normalization
# batch normalization for labeled and unlabeled examples is performed separately
if noise_std > 0: # Corrupted Encoder
# Corrupted encoder
# batch normalization + noise
z = join(batch_normalization(z_pre_l), batch_normalization(z_pre_u, m, v))
z += tf.random_normal(tf.shape(z_pre)) * noise_std
else: # Clean Encoder
# Clean encoder
# batch normalization + update the average mean and variance using batch mean and variance of labeled examples
z = join(update_batch_normalization(z_pre_l, l), batch_normalization(z_pre_u, m, v))
return z
# else:
def eval_batch_norm():
# Evaluation batch normalization
# obtain average mean and variance and use it to normalize the batch
mean, var = ewma.average(running_mean[l-1]), ewma.average(running_var[l-1])
z = batch_normalization(z_pre, mean, var)
# Instead of the above statement, the use of the following 2 statements containing a typo
# consistently produces a 0.2% higher accuracy for unclear reasons.
# m_l, v_l = tf.nn.moments(z_pre_l, axes=[0])
# z = join(batch_normalization(z_pre_l, m_l, mean, var), batch_normalization(z_pre_u, mean, var))
return z
# perform batch normalization according to value of boolean "training" placeholder:
z = tf.cond(pred=training, true_fn=training_batch_norm, false_fn=eval_batch_norm)
#----------------------------------------
if l == L:
# use softmax activation in output layer
h = tf.nn.softmax(weights['gamma'][l-1] * (z + weights["beta"][l-1]))
else:
# use ReLU activation in hidden layers
h = tf.nn.relu(z + weights["beta"][l-1])
d['labeled']['z'][l] , d['unlabeled']['z'][l] = split_lu(z)
d['unlabeled']['m'][l], d['unlabeled']['v'][l] = m, v # save mean and variance of unlabeled examples for decoding
d['labeled']['h'][l], d['unlabeled']['h'][l] = split_lu(h)
return h, d
# #### Encode
# In[20]:
with tf.name_scope(name="Corrupted_Encoder"):
print( "=== Corrupted Encoder ===")
y_c, corr = encoder(inputs, noise_std)
with tf.name_scope(name="Clean_Encoder"):
print( "=== Clean Encoder ===" )
y, clean = encoder(inputs, 0.0) # 0.0 -> do not add noise
# ### Decoder
# #### [DEF] g_gauss
# In[21]:
def g_gauss(z_c, u, size):
"""
gaussian denoising function proposed in the original paper
Parameters
----------
z_c : z in Corrupted Layer
u : batch normalized h~(l) (l=0,...,L)
size :
Returns
-------
"""
w_i = lambda inits, name: tf.Variable(inits * tf.ones([size]), name=name)
a1 = w_i(0., 'a1')
a2 = w_i(1., 'a2')
a3 = w_i(0., 'a3')
a4 = w_i(0., 'a4')
a5 = w_i(0., 'a5')
a6 = w_i(0., 'a6')
a7 = w_i(1., 'a7')
a8 = w_i(0., 'a8')
a9 = w_i(0., 'a9')
a10 = w_i(0., 'a10')
mu = a1 * tf.sigmoid(a2 * u + a3) + a4 * u + a5
v = a6 * tf.sigmoid(a7 * u + a8) + a9 * u + a10
z_est = (z_c - mu) * v + mu
return z_est
# #### Decode
# In[22]:
# Decoder
print( "=== Decoder ===" )
with tf.name_scope(name="Decoder"):
z_est = {}
d_cost = [] # to store the denoising cost of all layers
for l in range(L, -1, -1):
print( "Layer {:>2}: {:>5} -> {:>5}, denoising cost: {:>7.1f}".format(l, layer_sizes[l+1] if l+1 < len(layer_sizes) else "None", layer_sizes[l], denoising_cost[l]))
z, z_c = clean['unlabeled']['z'][l], corr['unlabeled']['z'][l]
m, v = clean['unlabeled']['m'].get(l, 0), clean['unlabeled']['v'].get(l, 1-1e-10)
if l == L:
u = unlabeled(y_c)
else:
u = tf.matmul(z_est[l+1], weights['V'][l])
u = batch_normalization(u)
z_est[l] = g_gauss(z_c, u, layer_sizes[l])
z_est_bn = (z_est[l] - m) / v
# append the cost of this layer to d_cost
d_cost.append((tf.reduce_mean(tf.reduce_sum(tf.square(z_est_bn - z), 1)) / layer_sizes[l]) * denoising_cost[l])
# ### parameter
# #### Cost, Loss
# In[23]:
# calculate total unsupervised cost by adding the denoising cost of all layers
with tf.name_scope(name="Cost"):
u_cost = tf.add_n(d_cost)
y_N = labeled(y_c)
cost = -tf.reduce_mean(tf.reduce_sum(outputs*tf.log(y_N), 1)) # supervised cost
loss = cost + u_cost # total cost
# #### accuracy
# In[24]:
with tf.name_scope(name="pred_cost"):
pred_cost = -tf.reduce_mean(tf.reduce_sum(outputs*tf.log(y), 1)) # cost used for prediction
with tf.name_scope(name="accuracy"):
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(outputs, 1)) # no of correct predictions
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) * tf.constant(100.0)
# #### optimizer
# In[25]:
with tf.name_scope(name="Optimizer"):
learning_rate = tf.Variable(starter_learning_rate, trainable=False)
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
# In[26]:
# add the updates of batch normalization statistics to train_step
bn_updates = tf.group(*bn_assigns)
with tf.control_dependencies([train_step]):
train_step = tf.group(bn_updates)
# ### Computational Graph
# In[27]:
show_computational_graph(tf.get_default_graph())
# ### Load MNIST
# In[28]:
print( "=== Loading Data ===" )
mnist = mnist_input_data.read_data_sets(train_dir=str(raw_Path / "MNIST_data"), n_labeled=num_labeled,
fake_data=False, one_hot=True)
# In[29]:
saver = tf.train.Saver()
# ### Session
# In[30]:
print( "=== Starting Session ===" )
sess = tf.Session()
# In[31]:
i_iter = 0
# #### chackpoints
# In[32]:
ckpt = tf.train.get_checkpoint_state('checkpoints/') # get latest checkpoint (if any)
if ckpt and ckpt.model_checkpoint_path:
# if checkpoint exists, restore the parameters and set epoch_n and i_iter
saver.restore(sess, ckpt.model_checkpoint_path)
epoch_n = int(ckpt.model_checkpoint_path.split('-')[1])
i_iter = (epoch_n+1) * (num_examples//batch_size)
print( "Restored Epoch ", epoch_n )
else:
# no checkpoint exists. create checkpoints directory if it does not exist.
if not os.path.exists('checkpoints'):
os.makedirs('checkpoints')
init = tf.global_variables_initializer()
sess.run(init)
# In[33]:
print( "=== Training ===" )
print( "Initial Accuracy: ", sess.run(accuracy,
feed_dict={
inputs: mnist.semi_test.images,
outputs: mnist.semi_test.labels,
training: False}), "%" )
# #### TRAINING
# In[34]:
for i in tqdm.tqdm(range(i_iter, num_iter)):
images, labels = mnist.semi_train.next_batch(batch_size)
sess.run(train_step, feed_dict={inputs: images,
outputs: labels,
training: True})
if (i > 1) and ((i+1) % (num_iter//num_epochs) == 0):
epoch_n = i//(num_examples//batch_size)
if (epoch_n+1) >= decay_after:
# decay learning rate
# learning_rate = starter_learning_rate * ((num_epochs - epoch_n) / (num_epochs - decay_after))
ratio = 1.0 * (num_epochs - (epoch_n+1)) # epoch_n + 1 because learning rate is set for next epoch
ratio = max(0, ratio / (num_epochs - decay_after))
sess.run(learning_rate.assign(starter_learning_rate * ratio))
print( "iter {}: test_acc:{}%".format(i, sess.run(accuracy,
feed_dict={ inputs: mnist.semi_test.images,
outputs: mnist.semi_test.labels,
training: False}) ) )
#saver.save(sess, 'checkpoints/model.ckpt', epoch_n)
# print( "Epoch ", epoch_n, ", Accuracy: ", sess.run(accuracy, feed_dict={inputs: mnist.test.images, outputs:mnist.test.labels, training: False}), "%" )
#with open('train_log', 'a') as train_log:
# # write test accuracy to file "train_log"
# train_log_w = csv.writer(train_log)
# log_i = [epoch_n] + sess.run([accuracy], feed_dict={inputs: mnist.test.images, outputs: mnist.test.labels, training: False})
# train_log_w.writerow(log_i)
# #### Final Accuracy
# In[35]:
print( "Final Accuracy: {}".format(sess.run(accuracy,
feed_dict={
inputs: mnist.semi_test.images,
outputs: mnist.semi_test.labels,
training: False}), "%" ) )
sess.close()
# ## show Result
# ### Computational Graph
# In[36]:
show_computational_graph(tf.get_default_graph())
# In[ ]:
# In[ ]:
# In[ ]:
# In[ ]:
# In[ ]:
# ## End