import tensorflow as tf
from notmnist.dataset import loadDataset
from notmnist.settings import *
from notmnist.utils import accuracy
ds = loadDataset('notMNIST_reformatted_conv.pickle')
train_dataset = ds['train_dataset']
train_labels = ds['train_labels']
valid_dataset = ds['valid_dataset']
valid_labels = ds['valid_labels']
test_dataset = ds['test_dataset']
test_labels = ds['test_labels']
exampleWidth = image_size * image_size
num_channels = 1
def maxpoolConvNet(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(
'''
Further performance optimizations of previous networks by adding
more layers and learning rate decay.
'''
import tensorflow as tf
import numpy as np
from tensorflow.python.framework import dtypes
from notmnist.dataset import loadDataset
from notmnist.settings import *
ds = loadDataset('notMNIST_reformatted_1d_images.pickle')
train_dataset = ds['train_dataset']
train_labels = ds['train_labels']
valid_dataset = ds['valid_dataset']
valid_labels = ds['valid_labels']
test_dataset = ds['test_dataset']
test_labels = ds['test_labels']
exampleWidth = image_size * image_size
import attr
@attr.s
class MultilayerNN():
''' Fully connected multilayer NN. '''
input = attr.ib()
inputSize = attr.ib()
outputSize = attr.ib()
hiddenLayers = attr.ib([10]) # sizes of hidden layers
Turn lables from integers to 1-hot encoding.
:return:
'''
dataset = dataset.reshape((-1, image_size * image_size)).astype(np.float32)
# Map 0 to [1.0, 0.0, 0.0 ...], 1 to [0.0, 1.0, 0.0 ...]
labels = (np.arange(num_labels) == labels[:,None]).astype(np.float32)
return dataset, labels
def reformatDataset(ds, reformatter):
new = {}
new['train_dataset'], new['train_labels'] = reformatter(ds['train_dataset'], ds['train_labels'])
new['valid_dataset'], new['valid_labels'] = reformatter(ds['valid_dataset'], ds['valid_labels'])
new['test_dataset'], new['test_labels'] = reformatter(ds['test_dataset'], ds['test_labels'])
return new
def reformatConv(dataset, labels, image_size = 28, num_labels = 10, num_channels = 1):
'''
Reformat MNIST dataset for convolutional networks, turning images to 3d
arrays by adding additional 'channel' dimension.
'''
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
if __name__ == '__main__':
ds = loadDataset()
#ds = loadDataset('notMNIST_reformatted_1d_images.pickle')
#printDsetShapes(ds)
# nds = reformatDataset(ds, reformatConv)
# printDsetShapes(nds)
# saveDataset(nds, 'notMNIST_reformatted_conv.pickle')