#!/usr/bin/env python
# coding: utf-8

# Deep Learning
# =============
# 
# Assignment 4
# ------------
# 
# Previously in `2_fullyconnected.ipynb` and `3_regularization.ipynb`, we trained fully connected networks to classify [notMNIST](http://yaroslavvb.blogspot.com/2011/09/notmnist-dataset.html) characters.
# 
# The goal of this assignment is make the neural network convolutional.

# In[1]:


# These are all the modules we'll be using later. Make sure you can import them
# before proceeding further.
from __future__ import print_function
import numpy as np
import tensorflow as tf
from six.moves import cPickle as pickle
from six.moves import range


# In[2]:


pickle_file = 'notMNIST.pickle'

with open(pickle_file, 'rb') as f:
  save = pickle.load(f)
  train_dataset = save['train_dataset']
  train_labels = save['train_labels']
  valid_dataset = save['valid_dataset']
  valid_labels = save['valid_labels']
  test_dataset = save['test_dataset']
  test_labels = save['test_labels']
  del save  # hint to help gc free up memory
  print('Training set', train_dataset.shape, train_labels.shape)
  print('Validation set', valid_dataset.shape, valid_labels.shape)
  print('Test set', test_dataset.shape, test_labels.shape)


# Reformat into a TensorFlow-friendly shape:
# - convolutions need the image data formatted as a cube (width by height by #channels)
# - labels as float 1-hot encodings.

# In[3]:


image_size = 28
num_labels = 10
num_channels = 1 # grayscale

import numpy as np

def reformat(dataset, labels):
  dataset = dataset.reshape(
    (-1, image_size, image_size, num_channels)).astype(np.float32)
  labels = (np.arange(num_labels) == labels[:,None]).astype(np.float32)
  return dataset, labels
train_dataset, train_labels = reformat(train_dataset, train_labels)
valid_dataset, valid_labels = reformat(valid_dataset, valid_labels)
test_dataset, test_labels = reformat(test_dataset, test_labels)
print('Training set', train_dataset.shape, train_labels.shape)
print('Validation set', valid_dataset.shape, valid_labels.shape)
print('Test set', test_dataset.shape, test_labels.shape)


# In[4]:


def accuracy(predictions, labels):
  return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))
          / predictions.shape[0])


# Let's build a small network with two convolutional layers, followed by one fully connected layer. Convolutional networks are more expensive computationally, so we'll limit its depth and number of fully connected nodes.

# In[5]:


batch_size = 16
patch_size = 5
depth = 16
num_hidden = 64

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)
  
  # Variables.
  layer1_weights = tf.Variable(tf.truncated_normal(
      [patch_size, patch_size, num_channels, depth], stddev=0.1))
  layer1_biases = tf.Variable(tf.zeros([depth]))
  layer2_weights = tf.Variable(tf.truncated_normal(
      [patch_size, patch_size, depth, depth], stddev=0.1))
  layer2_biases = tf.Variable(tf.constant(1.0, shape=[depth]))
  layer3_weights = tf.Variable(tf.truncated_normal(
      [image_size // 4 * image_size // 4 * depth, num_hidden], stddev=0.1))
  layer3_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
  layer4_weights = tf.Variable(tf.truncated_normal(
      [num_hidden, num_labels], stddev=0.1))
  layer4_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
  
  # Model.
  def model(data):
    conv = tf.nn.conv2d(data, layer1_weights, [1, 2, 2, 1], padding='SAME')
    hidden = tf.nn.relu(conv + layer1_biases)
    conv = tf.nn.conv2d(hidden, layer2_weights, [1, 2, 2, 1], padding='SAME')
    hidden = tf.nn.relu(conv + layer2_biases)
    shape = hidden.get_shape().as_list()
    reshape = tf.reshape(hidden, [shape[0], shape[1] * shape[2] * shape[3]])
    hidden = tf.nn.relu(tf.matmul(reshape, layer3_weights) + layer3_biases)
    return tf.matmul(hidden, layer4_weights) + layer4_biases
  
  # Training computation.
  logits = model(tf_train_dataset)
  loss = tf.reduce_mean(
    tf.nn.softmax_cross_entropy_with_logits(labels=tf_train_labels, logits=logits))
    
  # Optimizer.
  optimizer = tf.train.GradientDescentOptimizer(0.05).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))


# In[7]:


num_steps = 1001

with tf.Session(graph=graph) as session:
  tf.initialize_all_variables().run()
  print('Initialized')
  for step in range(num_steps):
    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)
    if (step % 50 == 0):
      print('Minibatch loss at step %d: %f' % (step, l))
      print('Minibatch accuracy: %.1f%%' % accuracy(predictions, batch_labels))
      print('Validation accuracy: %.1f%%' % accuracy(
        valid_prediction.eval(), valid_labels))
  print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))


# ---
# Problem 1
# ---------
# 
# The convolutional model above uses convolutions with stride 2 to reduce the dimensionality. Replace the strides by a max pooling operation (`nn.max_pool()`) of stride 2 and kernel size 2.
# 
# ---

# In[8]:


batch_size = 16
patch_size = 5
depth = 16
num_hidden = 64

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)
  
  # Variables.
  layer1_weights = tf.Variable(tf.truncated_normal(
      [patch_size, patch_size, num_channels, depth], stddev=0.1))
  layer1_biases = tf.Variable(tf.zeros([depth]))
  layer2_weights = tf.Variable(tf.truncated_normal(
      [patch_size, patch_size, depth, depth], stddev=0.1))
  layer2_biases = tf.Variable(tf.constant(1.0, shape=[depth]))
  layer3_weights = tf.Variable(tf.truncated_normal(
      [image_size // 4 * image_size // 4 * depth, num_hidden], stddev=0.1))
  layer3_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
  layer4_weights = tf.Variable(tf.truncated_normal(
      [num_hidden, num_labels], stddev=0.1))
  layer4_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
  
  # Model.
  def model(data):
    conv1 = tf.nn.conv2d(data, layer1_weights, [1, 1, 1, 1], padding='SAME')
    bias1 = tf.nn.relu(conv1 + layer1_biases)
    pool1 = tf.nn.max_pool(bias1, [1, 2, 2, 1], [1, 2, 2, 1], padding='SAME')
    conv2 = tf.nn.conv2d(pool1, layer2_weights, [1, 1, 1, 1], padding='SAME')
    bias2 = tf.nn.relu(conv2 + layer2_biases)
    pool2 = tf.nn.max_pool(bias2, [1, 2, 2, 1], [1, 2, 2, 1], padding='SAME')
    shape = pool2.get_shape().as_list()
    reshape = tf.reshape(pool2, [shape[0], shape[1] * shape[2] * shape[3]])
    hidden = tf.nn.relu(tf.matmul(reshape, layer3_weights) + layer3_biases)
    return tf.matmul(hidden, layer4_weights) + layer4_biases
  
  # Training computation.
  logits = model(tf_train_dataset)
  loss = tf.reduce_mean(
    tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))
    
  # Optimizer.
  optimizer = tf.train.GradientDescentOptimizer(0.05).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))


# In[9]:


num_steps = 1001

with tf.Session(graph=graph) as session:
  tf.initialize_all_variables().run()
  print('Initialized')
  for step in range(num_steps):
    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)
    if (step % 50 == 0):
      print('Minibatch loss at step %d: %f' % (step, l))
      print('Minibatch accuracy: %.1f%%' % accuracy(predictions, batch_labels))
      print('Validation accuracy: %.1f%%' % accuracy(
        valid_prediction.eval(), valid_labels))
  print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))


# ---
# Problem 2
# ---------
# 
# Try to get the best performance you can using a convolutional net. Look for example at the classic [LeNet5](http://yann.lecun.com/exdb/lenet/) architecture, adding Dropout, and/or adding learning rate decay.
# 
# ---

# The CNN below is loosely inspired by the LeNet5 architecture.

# In[10]:


batch_size = 16
patch_size = 5
depth = 16
num_hidden = 64

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)
  
  # Variables.
  layer1_weights = tf.Variable(tf.truncated_normal(
      [patch_size, patch_size, num_channels, depth], stddev=0.1))
  layer1_biases = tf.Variable(tf.zeros([depth]))
  layer2_weights = tf.Variable(tf.truncated_normal(
      [patch_size, patch_size, depth, depth], stddev=0.1))
  layer2_biases = tf.Variable(tf.constant(1.0, shape=[depth]))
  size3 = ((image_size - patch_size + 1) // 2 - patch_size + 1) // 2
  layer3_weights = tf.Variable(tf.truncated_normal(
      [size3 * size3 * depth, num_hidden], stddev=0.1))
  layer3_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
  layer4_weights = tf.Variable(tf.truncated_normal(
      [num_hidden, num_labels], stddev=0.1))
  layer4_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
  
  # Model.
  def model(data):
    # C1 input 28 x 28
    conv1 = tf.nn.conv2d(data, layer1_weights, [1, 1, 1, 1], padding='VALID')
    bias1 = tf.nn.relu(conv1 + layer1_biases)
    # S2 input 24 x 24
    pool2 = tf.nn.avg_pool(bias1, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')
    # C3 input 12 x 12
    conv3 = tf.nn.conv2d(pool2, layer2_weights, [1, 1, 1, 1], padding='VALID')
    bias3 = tf.nn.relu(conv3 + layer2_biases)
    # S4 input 8 x 8
    pool4 = tf.nn.avg_pool(bias3, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')
    # F6 input 4 x 4
    shape = pool4.get_shape().as_list()
    reshape = tf.reshape(pool4, [shape[0], shape[1] * shape[2] * shape[3]])
    hidden = tf.nn.relu(tf.matmul(reshape, layer3_weights) + layer3_biases)
    return tf.matmul(hidden, layer4_weights) + layer4_biases
  
  # Training computation.
  logits = model(tf_train_dataset)
  loss = tf.reduce_mean(
    tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))
    
  # Optimizer.
  optimizer = tf.train.GradientDescentOptimizer(0.05).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))


# In[11]:


num_steps = 20001

with tf.Session(graph=graph) as session:
  tf.initialize_all_variables().run()
  print('Initialized')
  for step in range(num_steps):
    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)
    if (step % 50 == 0):
      print('Minibatch loss at step %d: %f' % (step, l))
      print('Minibatch accuracy: %.1f%%' % accuracy(predictions, batch_labels))
      print('Validation accuracy: %.1f%%' % accuracy(
        valid_prediction.eval(), valid_labels))
  print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))


# The accuracy is good, but not as good as the 3-layer network from the previous assignment.

# The next version of the net uses dropout and learning rate decay:

# In[13]:


batch_size = 16
patch_size = 5
depth = 16
num_hidden = 64
beta_regul = 1e-3
drop_out = 0.5

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)
  global_step = tf.Variable(0)
  
  # Variables.
  layer1_weights = tf.Variable(tf.truncated_normal(
      [patch_size, patch_size, num_channels, depth], stddev=0.1))
  layer1_biases = tf.Variable(tf.zeros([depth]))
  layer2_weights = tf.Variable(tf.truncated_normal(
      [patch_size, patch_size, depth, depth], stddev=0.1))
  layer2_biases = tf.Variable(tf.constant(1.0, shape=[depth]))
  size3 = ((image_size - patch_size + 1) // 2 - patch_size + 1) // 2
  layer3_weights = tf.Variable(tf.truncated_normal(
      [size3 * size3 * depth, num_hidden], stddev=0.1))
  layer3_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
  layer4_weights = tf.Variable(tf.truncated_normal(
      [num_hidden, num_hidden], stddev=0.1))
  layer4_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
  layer5_weights = tf.Variable(tf.truncated_normal(
      [num_hidden, num_labels], stddev=0.1))
  layer5_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
  
  # Model.
  def model(data, keep_prob):
    # C1 input 28 x 28
    conv1 = tf.nn.conv2d(data, layer1_weights, [1, 1, 1, 1], padding='VALID')
    bias1 = tf.nn.relu(conv1 + layer1_biases)
    # S2 input 24 x 24
    pool2 = tf.nn.avg_pool(bias1, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')
    # C3 input 12 x 12
    conv3 = tf.nn.conv2d(pool2, layer2_weights, [1, 1, 1, 1], padding='VALID')
    bias3 = tf.nn.relu(conv3 + layer2_biases)
    # S4 input 8 x 8
    pool4 = tf.nn.avg_pool(bias3, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')
    # F5 input 4 x 4
    shape = pool4.get_shape().as_list()
    reshape = tf.reshape(pool4, [shape[0], shape[1] * shape[2] * shape[3]])
    hidden5 = tf.nn.relu(tf.matmul(reshape, layer3_weights) + layer3_biases)
    # F6
    drop5 = tf.nn.dropout(hidden5, keep_prob)
    hidden6 = tf.nn.relu(tf.matmul(hidden5, layer4_weights) + layer4_biases)
    drop6 = tf.nn.dropout(hidden6, keep_prob)
    return tf.matmul(drop6, layer5_weights) + layer5_biases
  
  # Training computation.
  logits = model(tf_train_dataset, drop_out)
  loss = tf.reduce_mean(
    tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))
    
  # Optimizer.
  learning_rate = tf.train.exponential_decay(0.05, global_step, 1000, 0.85, staircase=True)
  optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
  
  # Predictions for the training, validation, and test data.
  train_prediction = tf.nn.softmax(logits)
  valid_prediction = tf.nn.softmax(model(tf_valid_dataset, 1.0))
  test_prediction = tf.nn.softmax(model(tf_test_dataset, 1.0))


# In[14]:


num_steps = 5001

with tf.Session(graph=graph) as session:
  tf.initialize_all_variables().run()
  print('Initialized')
  for step in range(num_steps):
    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)
    if (step % 50 == 0):
      print('Minibatch loss at step %d: %f' % (step, l))
      print('Minibatch accuracy: %.1f%%' % accuracy(predictions, batch_labels))
      print('Validation accuracy: %.1f%%' % accuracy(
        valid_prediction.eval(), valid_labels))
  print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))


# Well, the accuracy is worst. This net has many meta parameters and I don't feel comfortable in tuning them randomly. I should probably change the depth and make it different between the layers, since it looks like the increasing number of feature maps is a key design item.