In [2]:
from __future__ import print_function
import numpy as np
import tensorflow as tf
from six.moves import cPickle as pickle
import matplotlib.pyplot as plt
import random
import operator
import time
import os
In [3]:
file_name = 'depth_data'
with open(file_name, 'rb') as f:
save = pickle.load(f)
dataset = save['dataset']
names = save['names']
orientations = save['orientations']
del save
In [4]:
# generate labels
# for 10 objectives
image_size = 34
num_labels = 24
num_channels = 1
num_images = dataset.shape[0]
num_train = round(num_images*0.7)
num_valid = round(num_images*0.15)
num_test = round(num_images*0.15)
name2value = {'v8':0,'ducky':1,'stapler':2,'pball':3,'tball':4,'sponge':5,'bclip':6,'tape':7,'gstick':8,'cup':9,
'pen':10,'calc':11,'tmeas':12,'bottle':13,'cpin':14,'scissors':15,'stape':16,'gball':17,'orwidg':18,
'glue':19,'spoon':20,'fork':21,'nerf':22,'eraser':23}
value2name = dict((value,name) for name,value in name2value.items())
labels = np.ndarray(num_images, dtype=np.int32)
index = 0
for name in names:
labels[index] = name2value[name]
index += 1
def randomize(dataset, labels):
permutation = np.random.permutation(labels.shape[0])
shuffled_dataset = dataset[permutation,:,:]
shuffled_labels = labels[permutation]
return shuffled_dataset, shuffled_labels
rdataset, rlabels = randomize(dataset, labels)
train_dataset = rdataset[0:num_train,:,:]
train_labels = rlabels[0:num_train]
valid_dataset = rdataset[num_train:(num_train+num_valid),:,:]
valid_labels = rlabels[num_train:(num_train+num_valid)]
test_dataset = rdataset[(num_train+num_valid):,:,:]
test_labels = rlabels[(num_train+num_valid):]
print('Training:', train_dataset.shape, train_labels.shape)
print('Validation:', valid_dataset.shape, valid_labels.shape)
print('Testing:', test_dataset.shape, test_labels.shape)
display some images to verify that the data is still correct
In [5]:
indices = [random.randint(0,train_dataset.shape[0]) for x in range(3)]
for index in indices:
image = train_dataset[index,:,:]
print(value2name[train_labels[index]])
plt.imshow(image,cmap='Greys_r')
plt.show()
In [6]:
print('......Reformatting......')
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 [7]:
def accuracy(predictions, labels):
return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))
/ predictions.shape[0])
In [8]:
batch_size = 16
patch_size = 5
kernel_size = 2
depth1 = 6 #the depth of 1st convnet
depth2 = 16 #the depth of 2nd convnet
C5_units = 120
F6_units = 84
F7_units = 10
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))
# convolution's input is a tensor of shape [batch,in_height,in_width,in_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(weights and biases)
C1_weights = tf.Variable(tf.truncated_normal([patch_size, patch_size, num_channels, depth1], stddev=0.1))
# convolution's weights are called filter in tensorflow
# it is a tensor of shape [kernel_hight,kernel_width,in_channels,out_channels]
C1_biases = tf.Variable(tf.zeros([depth1]))
# S1_weights # Sub-sampling doesn't need weights and biases
# S1_biases
C3_weights = tf.Variable(tf.truncated_normal([patch_size, patch_size, depth1, depth2], stddev=0.1))
C3_biases = tf.Variable(tf.constant(1.0, shape=[depth2]))
# S4_weights
# S4_biases
# C5 actually is a fully-connected layer
C5_weights = tf.Variable(tf.truncated_normal([6 * 6 * depth2, C5_units], stddev=0.1))
C5_biases = tf.Variable(tf.constant(1.0, shape=[C5_units]))
F6_weights = tf.Variable(tf.truncated_normal([C5_units,F6_units], stddev=0.1))
F6_biases = tf.Variable(tf.constant(1.0, shape=[F6_units]))
# FC and logistic regression replace RBF
F7_weights = tf.Variable(tf.truncated_normal([F6_units,num_labels], stddev=0.1))
F7_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
saver = tf.train.Saver()
# Model
def model(data):
conv = tf.nn.conv2d(data, C1_weights, [1, 1, 1, 1], padding='SAME')
hidden = tf.nn.relu(conv + C1_biases) # relu is better than tanh
max_pool = tf.nn.max_pool(hidden,[1,kernel_size,kernel_size,1],[1,2,2,1],'VALID')
hidden = tf.nn.relu(max_pool)
conv = tf.nn.conv2d(hidden, C3_weights, [1, 1, 1, 1], padding='VALID')
hidden = tf.nn.relu(conv + C3_biases)
max_pool = tf.nn.max_pool(hidden,[1,kernel_size,kernel_size,1],[1,2,2,1],'VALID')
hidden = tf.nn.relu(max_pool)
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, C5_weights) + C5_biases)
fc = tf.matmul(hidden,F6_weights)
hidden = tf.nn.relu(fc + F6_biases)
fc = tf.matmul(hidden,F7_weights)
output = fc + F7_biases
return output
# Training computation.
tf_train_dataset = tf.nn.dropout(tf_train_dataset,0.8) # input dropout
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(0.0008).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]:
## training
start_time = time.time()
num_steps = 70000
config = tf.ConfigProto()
config.gpu_options.allocator_type = 'BFC'
config.gpu_options.allow_growth = True
config.log_device_placement = True
with tf.Session(graph=graph, config = config) as session:
#tf.initialize_all_variables().run()
saver.restore(session, "model.ckpt")
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 % 1000 == 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('--------------------------------------')
print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))
end_time = time.time()
duration = (end_time - start_time)/60
print("Excution time: %0.2fmin" % duration)
save_path = saver.save(session, "new_model.ckpt")
#print("Model saved in file: %s" % save_path)
i_test = 0
while(i_test!=''):
i_test = input("Input an index of test image (or Enter to quit): ")
label = test_labels[int(i_test),:].tolist()
#print("Correct label: "+value2name[label.index(1)])
image = test_dataset[int(i_test),:,:,:].reshape((-1,image_size,image_size,num_channels)).astype(np.float32)
prediction = tf.nn.softmax(model(image))
pre_dict = dict(zip(list(range(num_labels)),prediction.eval()[0]))
sorted_pre_dict = sorted(pre_dict.items(), key=operator.itemgetter(1))
name1 = value2name[sorted_pre_dict[-1][0]]
value1 = str(sorted_pre_dict[-1][1])
name2 = value2name[sorted_pre_dict[-2][0]]
value2 = str(sorted_pre_dict[-2][1])
tile = name1+': '+value1+'\n'+name2+': '+value2
image = image.reshape((image_size,image_size)).astype(np.float32)
plt.imshow(image,cmap='Greys_r')
plt.suptitle(tile, fontsize=12)
plt.xlabel(value2name[label.index(1)], fontsize=12)
plt.show()
In [7]:
batch_size = 16
patch_size = 5
kernel_size = 2
depth1 = 6 #the depth of 1st convnet
depth2 = 16 #the depth of 2nd convnet
C5_units = 120
F6_units = 84
F7_units = 10
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))
# convolution's input is a tensor of shape [batch,in_height,in_width,in_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(weights and biases)
C1_weights = tf.Variable(tf.truncated_normal([patch_size, patch_size, num_channels, depth1], stddev=0.1))
# convolution's weights are called filter in tensorflow
# it is a tensor of shape [kernel_hight,kernel_width,in_channels,out_channels]
C1_biases = tf.Variable(tf.zeros([depth1]))
# S1_weights # Sub-sampling doesn't need weights and biases
# S1_biases
C3_weights = tf.Variable(tf.truncated_normal([patch_size, patch_size, depth1, depth2], stddev=0.1))
C3_biases = tf.Variable(tf.constant(1.0, shape=[depth2]))
# S4_weights
# S4_biases
# C5 actually is a fully-connected layer
C5_weights = tf.Variable(tf.truncated_normal([6 * 6 * depth2, C5_units], stddev=0.1))
C5_biases = tf.Variable(tf.constant(1.0, shape=[C5_units]))
F6_weights = tf.Variable(tf.truncated_normal([C5_units,F6_units], stddev=0.1))
F6_biases = tf.Variable(tf.constant(1.0, shape=[F6_units]))
# FC and logistic regression replace RBF
F7_weights = tf.Variable(tf.truncated_normal([F6_units,num_labels], stddev=0.1))
F7_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
saver = tf.train.Saver()
# Model
def model(data):
conv = tf.nn.conv2d(data, C1_weights, [1, 1, 1, 1], padding='SAME')
hidden = tf.nn.relu(conv + C1_biases) # relu is better than tanh
max_pool = tf.nn.max_pool(hidden,[1,kernel_size,kernel_size,1],[1,2,2,1],'VALID')
hidden = tf.nn.relu(max_pool)
conv = tf.nn.conv2d(hidden, C3_weights, [1, 1, 1, 1], padding='VALID')
hidden = tf.nn.relu(conv + C3_biases)
max_pool = tf.nn.max_pool(hidden,[1,kernel_size,kernel_size,1],[1,2,2,1],'VALID')
hidden = tf.nn.relu(max_pool)
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, C5_weights) + C5_biases)
fc = tf.matmul(hidden,F6_weights)
hidden = tf.nn.relu(fc + F6_biases)
fc = tf.matmul(hidden,F7_weights)
output = fc + F7_biases
return output
# Training computation.
tf_train_dataset = tf.nn.dropout(tf_train_dataset,0.8) # input dropout
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(0.0008).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))
config = tf.ConfigProto()
config.log_device_placement = True
#wd = os.getcwd()
with tf.Session(graph=graph, config = config) as session:
saver.restore(session, "model.ckpt")
print("Model restored.")
print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))
i_test = 0
while(i_test!=''):
i_test = input("Input an index of test image (or Enter to quit): ")
label = test_labels[int(i_test),:].tolist()
#print("Correct label: "+value2name[label.index(1)])
image = test_dataset[int(i_test),:,:,:].reshape((-1,image_size,image_size,num_channels)).astype(np.float32)
prediction = tf.nn.softmax(model(image))
pre_dict = dict(zip(list(range(num_labels)),prediction.eval()[0]))
sorted_pre_dict = sorted(pre_dict.items(), key=operator.itemgetter(1))
name1 = value2name[sorted_pre_dict[-1][0]]
value1 = str(sorted_pre_dict[-1][1])
name2 = value2name[sorted_pre_dict[-2][0]]
value2 = str(sorted_pre_dict[-2][1])
tile = name1+': '+value1+'\n'+name2+': '+value2
image = image.reshape((image_size,image_size)).astype(np.float32)
plt.imshow(image,cmap='Greys_r')
plt.suptitle(tile, fontsize=12)
plt.xlabel(value2name[label.index(1)], fontsize=12)
plt.show()