def conv_train(train_dataset, train_labels, valid_dataset, valid_labels, test_dataset, test_labels, image_size,
num_labels, basic_hps, stride_ps, drop=False, lrd=False, get_grad=False, norm_list=None):
batch_size = basic_hps['batch_size']
patch_size = basic_hps['patch_size']
depth = basic_hps['depth']
num_hidden = basic_hps['num_hidden']
num_channels = 1
layer_cnt = basic_hps['layer_sum']
loss_collect = list()
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.
input_weights = tf.Variable(tf.truncated_normal(
[patch_size, patch_size, num_channels, depth], stddev=0.1))
input_biases = tf.Variable(tf.zeros([depth]))
mid_layer_cnt = layer_cnt - 1
layer_weights = list()
layer_biases = [tf.Variable(tf.constant(1.0, shape=[depth])) for _ in range(mid_layer_cnt)]
output_weights = list()
output_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
final_weights = tf.Variable(tf.truncated_normal(
[num_hidden, num_labels], stddev=0.1))
final_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
weight_set_done = False
# Model.
def model(data):
if not large_data_size(data) or not large_data_size(input_weights):
stride_ps[0] = [1, 1, 1, 1]
conv = tf.nn.conv2d(data, input_weights, stride_ps[0], use_cudnn_on_gpu=True, padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + input_biases)
if drop:
hidden = tf.nn.dropout(hidden, 0.5)
for i in range(mid_layer_cnt):
# print(hidden)
if not weight_set_done:
# avoid filter shape larger than input shape
hid_shape = hidden.get_shape()
# print(hid_shape)
filter_w = patch_size / (i + 1)
filter_h = patch_size / (i + 1)
# print(filter_w)
# print(filter_h)
if filter_w > hid_shape[1]:
filter_w = int(hid_shape[1])
if filter_h > hid_shape[2]:
filter_h = int(hid_shape[2])
layer_weight = tf.Variable(tf.truncated_normal(shape=[filter_w, filter_h, depth, depth],
stddev=0.1))
layer_weights.append(layer_weight)
if not large_data_size(hidden) or not large_data_size(layer_weights[i]):
# print("is not large data")
stride_ps[i + 1] = [1, 1, 1, 1]
# print(stride_ps[i + 1])
# print(len(stride_ps))
# print(i + 1)
conv = tf.nn.conv2d(hidden, layer_weights[i], stride_ps[i + 1], use_cudnn_on_gpu=True, padding='SAME')
if not large_data_size(conv):
conv = maxpool2d(conv, 1, 1)
else:
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer_biases[i])
if drop:
hidden = tf.nn.dropout(hidden, 0.7)
shapes = hidden.get_shape().as_list()
shape_mul = 1
for s in shapes[1:]:
shape_mul *= s
if not weight_set_done:
output_size = shape_mul
output_weights.append(tf.Variable(tf.truncated_normal([output_size, num_hidden], stddev=0.1)))
reshape = tf.reshape(hidden, [shapes[0], shape_mul])
hidden = tf.nn.relu(tf.matmul(reshape, output_weights[0]) + output_biases)
if drop:
hidden = tf.nn.dropout(hidden, 0.8)
return tf.matmul(hidden, final_weights) + final_biases
# Training computation.
logits = model(tf_train_dataset)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))
# Optimizer.
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
starter_learning_rate = 0.1
learning_rate = tf.train.exponential_decay(starter_learning_rate, cur_step, 10000, 0.96, staircase=True)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=cur_step)
else:
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))
num_steps = 3001
start_fit = 600
init_loss = []
with tf.Session(graph=graph) as session:
tf.initialize_all_variables().run()
print('Initialized')
end_train = False
mean_loss = 0
for step in range(num_steps):
if end_train:
break
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)
mean_loss += l
if step % 5 == 0:
mean_loss /= 5.0
loss_collect.append(mean_loss)
mean_loss = 0
if norm_list is None:
return [1 for _ in range(len(basic_hps))]
if step >= start_fit:
# print(loss_collect)
if step == start_fit:
res = fit_loss(1,
[batch_size / norm_list[0], depth / norm_list[1], num_hidden / norm_list[2],
layer_cnt / norm_list[3], patch_size / norm_list[4]],
loss_collect)
else:
res = fit_loss(0,
[batch_size / norm_list[0], depth / norm_list[1], num_hidden / norm_list[2],
layer_cnt / norm_list[3], patch_size / norm_list[4]],
loss_collect)
if get_grad:
better_hyper([batch_size / norm_list[0], depth / norm_list[1], num_hidden / norm_list[2],
layer_cnt / norm_list[3], patch_size / norm_list[4]],
loss_collect)
loss_collect.remove(loss_collect[0])
ret = res['ret']
if ret == 1 and not get_grad:
print('ret is end train when step is {step}'.format(step=step))
init_loss.append(loss_collect)
end_train = True
if step % 50 == 0:
print('Minibatch loss at step %d: %f' % (step, l))
print('Validation accuracy: %.1f%%' % accuracy(
valid_prediction.eval(), valid_labels))
print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))
if end_train:
hypers = better_hyper(
[batch_size / norm_list[0], depth / norm_list[1], num_hidden / norm_list[2],
layer_cnt / norm_list[3], patch_size / norm_list[4]],
init_loss[0])
print(hypers)
hypers = [hyper * norm_list[i] for i, hyper in enumerate(hypers)]
print(norm_list)
print(hypers)
for i in range(len(hypers)):
if hypers[i] <= 1.0:
hypers[i] = 1
else:
hypers[i] = int(hypers[i])
else:
hypers = [batch_size, depth, num_hidden, layer_cnt, patch_size]
return end_train, hypers
def tf_deep_nn(regular=False, drop_out=False, lrd=False, layer_cnt=2):
batch_size = 128
graph = tf.Graph()
with graph.as_default():
tf_train_dataset = tf.placeholder(tf.float32, shape=(batch_size, feature_dim))
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)
hidden_node_count = 32
# start weight
hidden_stddev = np.sqrt(2.0 / 100)
weights1 = tf.Variable(tf.truncated_normal([feature_dim, hidden_node_count], stddev=hidden_stddev))
biases1 = tf.Variable(tf.zeros([hidden_node_count]))
# middle weight
weights = []
biases = []
hidden_cur_cnt = hidden_node_count
for i in range(layer_cnt - 2):
if hidden_cur_cnt > 2:
hidden_next_cnt = int(hidden_cur_cnt / 2)
else:
hidden_next_cnt = 2
hidden_stddev = np.sqrt(2.0 / hidden_cur_cnt / 10)
weights.append(tf.Variable(tf.truncated_normal([hidden_cur_cnt, hidden_next_cnt], stddev=hidden_stddev)))
biases.append(tf.Variable(tf.zeros([hidden_next_cnt])))
hidden_cur_cnt = hidden_next_cnt
# first wx + b
y0 = tf.matmul(tf_train_dataset, weights1) + biases1
# first sigmoid
hidden = tf.nn.sigmoid(y0)
# hidden = y0
hidden_drop = hidden
# first DropOut
keep_prob = 0.5
if drop_out:
hidden_drop = tf.nn.dropout(hidden, keep_prob)
# first wx+b for valid
valid_y0 = tf.matmul(tf_valid_dataset, weights1) + biases1
valid_hidden = tf.nn.sigmoid(valid_y0)
# valid_hidden = valid_y0
# first wx+b for test
test_y0 = tf.matmul(tf_test_dataset, weights1) + biases1
test_hidden = tf.nn.sigmoid(test_y0)
# test_hidden = test_y0
# middle layer
for i in range(layer_cnt - 2):
y1 = tf.matmul(hidden_drop, weights[i]) + biases[i]
hidden_drop = tf.nn.sigmoid(y1)
if drop_out:
keep_prob += 0.5 * i / (layer_cnt + 1)
hidden_drop = tf.nn.dropout(hidden_drop, keep_prob)
y0 = tf.matmul(hidden, weights[i]) + biases[i]
hidden = tf.nn.sigmoid(y0)
# hidden = y0
valid_y0 = tf.matmul(valid_hidden, weights[i]) + biases[i]
valid_hidden = tf.nn.sigmoid(valid_y0)
# valid_hidden = valid_y0
test_y0 = tf.matmul(test_hidden, weights[i]) + biases[i]
test_hidden = tf.nn.sigmoid(test_y0)
# test_hidden = test_y0
# last weight
weights2 = tf.Variable(tf.truncated_normal([hidden_cur_cnt, num_labels], stddev=hidden_stddev / 2))
biases2 = tf.Variable(tf.zeros([num_labels]))
# last wx + b
logits = tf.matmul(hidden_drop, weights2) + biases2
# predicts
logits_predict = tf.matmul(hidden, weights2) + biases2
valid_predict = tf.matmul(valid_hidden, weights2) + biases2
test_predict = tf.matmul(test_hidden, weights2) + biases2
l2_loss = 0
# enable regularization
if regular:
l2_loss = tf.nn.l2_loss(weights1) + tf.nn.l2_loss(weights2)
for i in range(len(weights)):
l2_loss += tf.nn.l2_loss(weights[i])
# l2_loss += tf.nn.l2_loss(biases[i])
beta = 1e-2
l2_loss *= beta
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=tf_train_labels)) + l2_loss
# Optimizer.
if lrd:
cur_step = tf.Variable(0, trainable=False) # count the number of steps taken.
starter_learning_rate = 0.4
learning_rate = tf.train.exponential_decay(starter_learning_rate, cur_step, 500, 0.75, staircase=True)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=cur_step)
else:
optimizer = tf.train.AdamOptimizer(0.5).minimize(loss)
# Predictions for the training, validation, and test data.
train_prediction = tf.nn.softmax(logits_predict)
valid_prediction = tf.nn.softmax(valid_predict)
test_prediction = tf.nn.softmax(test_predict)
num_steps = 8001
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
print("Initialized")
for step in range(num_steps):
offset_range = train_labels.shape[0] - batch_size
offset = (step * batch_size) % offset_range
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))
def conv_train(train_dataset,
train_labels,
valid_dataset,
valid_labels,
test_dataset,
test_labels,
image_size,
num_labels,
basic_hps,
stride_ps,
lrd=False,
get_grad=False):
batch_size = basic_hps['batch_size']
patch_size = basic_hps['patch_size']
depth = basic_hps['depth']
num_hidden = basic_hps['num_hidden']
num_channels = 1
layer_cnt = basic_hps['layer_sum']
loss_collect = list()
first_hidden_num = basic_hps['num_hidden']
second_hidden_num = first_hidden_num / 2 + 1
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)
input_weights = tf.Variable(
tf.truncated_normal([patch_size, patch_size, num_channels, depth],
stddev=0.1))
input_biases = tf.Variable(tf.zeros([depth]))
mid_layer_cnt = layer_cnt - 1
layer_weights = list()
layer_biases = [
tf.Variable(tf.constant(1.0, shape=[depth / (i + 2)]))
for i in range(mid_layer_cnt)
]
output_weights = list()
output_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
first_nn_weights = tf.Variable(
tf.truncated_normal([first_hidden_num, second_hidden_num],
stddev=0.1))
second_nn_weights = tf.Variable(
tf.truncated_normal([second_hidden_num, num_labels], stddev=0.1))
first_nn_biases = tf.Variable(
tf.constant(1.0, shape=[second_hidden_num]))
second_nn_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
# Model.
def model(data, init=False):
# Variables.
if not large_data_size(data) or not large_data_size(input_weights):
stride_ps[0] = [1, 1, 1, 1]
conv = tf.nn.conv2d(data,
input_weights,
stride_ps[0],
use_cudnn_on_gpu=True,
padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + input_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.8)
for i in range(mid_layer_cnt):
# print(hidden)
if init:
# avoid filter shape larger than input shape
hid_shape = hidden.get_shape()
# print(hid_shape)
filter_w = patch_size / (i + 1)
filter_h = patch_size / (i + 1)
# print(filter_w)
# print(filter_h)
if filter_w > hid_shape[1]:
filter_w = int(hid_shape[1])
if filter_h > hid_shape[2]:
filter_h = int(hid_shape[2])
layer_weight = tf.Variable(
tf.truncated_normal(shape=[
filter_w, filter_h, depth / (i + 1),
depth / (i + 2)
],
stddev=0.1))
layer_weights.append(layer_weight)
if not large_data_size(hidden) or not large_data_size(
layer_weights[i]):
# print("is not large data")
stride_ps[i + 1] = [1, 1, 1, 1]
# print(stride_ps[i + 1])
# print(len(stride_ps))
# print(i + 1)
conv = tf.nn.conv2d(hidden,
layer_weights[i],
stride_ps[i + 1],
use_cudnn_on_gpu=True,
padding='SAME')
if not large_data_size(conv):
conv = maxpool2d(conv, 1, 1)
else:
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer_biases[i])
if init:
hidden = tf.nn.dropout(hidden, 0.8)
shapes = hidden.get_shape().as_list()
shape_mul = 1
for s in shapes[1:]:
shape_mul *= s
if init:
output_size = shape_mul
output_weights.append(
tf.Variable(
tf.truncated_normal([output_size, num_hidden],
stddev=0.1)))
reshape = tf.reshape(hidden, [shapes[0], shape_mul])
hidden = tf.nn.relu6(
tf.matmul(reshape, output_weights[0]) + output_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, first_nn_weights) + first_nn_biases
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, second_nn_weights) + second_nn_biases
return hidden
# Training computation.
logits = model(tf_train_dataset, init=True)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))
# Optimizer.
starter_learning_rate = 0.1
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
learning_rate = tf.train.exponential_decay(starter_learning_rate,
cur_step,
10000,
0.96,
staircase=True)
optimizer = tf.train.GradientDescentOptimizer(
learning_rate).minimize(loss, global_step=cur_step)
else:
optimizer = tf.train.AdagradOptimizer(
starter_learning_rate).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))
num_steps = 3001
start_fit = 600
init_loss = []
with tf.Session(graph=graph) as session:
tf.initialize_all_variables().run()
print('Initialized')
end_train = False
mean_loss = 0
for step in range(num_steps):
if end_train:
break
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)
mean_loss += l
if step % 5 == 0:
mean_loss /= 5.0
loss_collect.append(mean_loss)
mean_loss = 0
if step >= start_fit:
# print(loss_collect)
if step == start_fit:
res = fit_loss(1, [
batch_size, depth, num_hidden, layer_cnt,
patch_size
], loss_collect)
else:
res = fit_loss(0, [
batch_size, depth, num_hidden, layer_cnt,
patch_size
], loss_collect)
if get_grad:
better_hyper([
batch_size, depth, num_hidden, layer_cnt,
patch_size
], loss_collect)
loss_collect.remove(loss_collect[0])
ret = res['ret']
if ret == 1 and not get_grad:
print('ret is end train when step is {step}'.format(
step=step))
init_loss.append(loss_collect)
end_train = True
if step % 50 == 0:
print('Minibatch loss at step %d: %f' % (step, l))
print('Validation accuracy: %.1f%%' % accuracy(
valid_prediction.eval(), valid_labels))
print('Test accuracy: %.1f%%' %
accuracy(test_prediction.eval(), test_labels))
if end_train:
hypers = better_hyper(
[batch_size, depth, num_hidden, layer_cnt, patch_size],
init_loss[0])
print(hypers)
for i in range(len(hypers)):
if hypers[i] <= 1.0:
hypers[i] = 1
else:
hypers[i] = int(hypers[i])
else:
hypers = [batch_size, depth, num_hidden, layer_cnt, patch_size]
return end_train, hypers
def conv_train(train_dataset, train_labels, valid_dataset, valid_labels,
test_dataset, test_labels, image_size, num_labels, basic_hps,
stride_ps):
batch_size = basic_hps['batch_size']
patch_size = basic_hps['patch_size']
depth = basic_hps['depth']
num_hidden = basic_hps['num_hidden']
num_channels = 1
layer_cnt = basic_hps['layer_sum']
starter_learning_rate = basic_hps['start_learning_rate']
loss_collect = list()
first_hidden_num = basic_hps['num_hidden']
second_hidden_num = first_hidden_num / 2 + 1
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)
input_weights = tf.Variable(
tf.truncated_normal([patch_size, patch_size, num_channels, depth],
stddev=0.1))
input_biases = tf.Variable(tf.zeros([depth]))
mid_layer_cnt = layer_cnt - 1
layer_weights = list()
layer_biases = [
tf.Variable(tf.constant(1.0, shape=[depth / (i + 2)]))
for i in range(mid_layer_cnt)
]
output_weights = list()
output_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
first_nn_weights = tf.Variable(
tf.truncated_normal([first_hidden_num, second_hidden_num],
stddev=0.1))
second_nn_weights = tf.Variable(
tf.truncated_normal([second_hidden_num, num_labels], stddev=0.1))
first_nn_biases = tf.Variable(
tf.constant(1.0, shape=[second_hidden_num]))
second_nn_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
# Model.
def model(data, init=False):
# Variables.
if not large_data_size(data) or not large_data_size(input_weights):
stride_ps[0] = [1, 1, 1, 1]
conv = tf.nn.conv2d(data,
input_weights,
stride_ps[0],
use_cudnn_on_gpu=True,
padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + input_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.8)
for i in range(mid_layer_cnt):
# print(hidden)
if init:
hid_shape = hidden.get_shape()
filter_w = patch_size / (i + 1)
filter_h = patch_size / (i + 1)
if filter_w > hid_shape[1]:
filter_w = int(hid_shape[1])
if filter_h > hid_shape[2]:
filter_h = int(hid_shape[2])
layer_weight = tf.Variable(
tf.truncated_normal(shape=[
filter_w, filter_h, depth / (i + 1),
depth / (i + 2)
],
stddev=0.1))
layer_weights.append(layer_weight)
if not large_data_size(hidden) or not large_data_size(
layer_weights[i]):
stride_ps[i + 1] = [1, 1, 1, 1]
conv = tf.nn.conv2d(hidden,
layer_weights[i],
stride_ps[i + 1],
use_cudnn_on_gpu=True,
padding='SAME')
if not large_data_size(conv):
conv = maxpool2d(conv, 1, 1)
else:
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer_biases[i])
if init:
hidden = tf.nn.dropout(hidden, 0.8)
shapes = hidden.get_shape().as_list()
shape_mul = 1
for s in shapes[1:]:
shape_mul *= s
if init:
output_size = shape_mul
output_weights.append(
tf.Variable(
tf.truncated_normal([output_size, num_hidden],
stddev=0.1)))
reshape = tf.reshape(hidden, [shapes[0], shape_mul])
hidden = tf.nn.relu6(
tf.matmul(reshape, output_weights[0]) + output_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, first_nn_weights) + first_nn_biases
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, second_nn_weights) + second_nn_biases
return hidden
# Training computation.
logits = model(tf_train_dataset, init=True)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))
optimizer = tf.train.AdagradOptimizer(starter_learning_rate).minimize(
loss)
train_prediction = tf.nn.softmax(logits)
valid_prediction = tf.nn.softmax(model(tf_valid_dataset))
test_prediction = tf.nn.softmax(model(tf_test_dataset))
num_steps = 1001
with tf.Session(graph=graph) as session:
tf.initialize_all_variables().run()
print('Initialized')
mean_loss = 0
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)
mean_loss += l
if step % 5 == 0:
mean_loss /= 5.0
loss_collect.append(mean_loss)
mean_loss = 0
if step % 50 == 0:
print('Minibatch loss at step %d: %f' % (step, l))
print('Validation accuracy: %.1f%%' %
accuracy(valid_prediction.eval(), valid_labels))
print('Test accuracy: %.1f%%' %
accuracy(test_prediction.eval(), test_labels))
hypers = better_trend_hyper(
[batch_size, depth, num_hidden, layer_cnt, patch_size],
loss_collect)
print(hypers)
for i in range(len(hypers)):
if hypers[i] <= 1.0:
hypers[i] = 1
else:
hypers[i] = int(hypers[i])
return hypers
def conv_train(train_dataset,
train_labels,
valid_dataset,
valid_labels,
test_dataset,
test_labels,
image_size,
num_labels,
basic_hps,
stride_ps,
drop=False,
lrd=False):
batch_size = basic_hps['batch_size']
patch_size = basic_hps['patch_size']
depth = basic_hps['depth']
first_hidden_num = basic_hps['num_hidden']
second_hidden_num = first_hidden_num / 2 + 1
num_channels = 1
layer_cnt = basic_hps['layer_sum']
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.
# the third parameter must be same as the last layer depth
input_weights = tf.Variable(
tf.truncated_normal([patch_size, patch_size, num_channels, depth],
stddev=0.1))
input_biases = tf.Variable(tf.zeros([depth]))
mid_layer_cnt = layer_cnt - 1
layer_weights = list()
layer_biases = [
tf.Variable(tf.constant(1.0, shape=[depth * (i + 2)]))
for i in range(mid_layer_cnt)
]
output_weights = list()
output_biases = tf.Variable(tf.constant(1.0, shape=[first_hidden_num]))
first_nn_weights = tf.Variable(
tf.truncated_normal([first_hidden_num, second_hidden_num],
stddev=0.1))
second_nn_weights = tf.Variable(
tf.truncated_normal([second_hidden_num, num_labels], stddev=0.1))
first_nn_biases = tf.Variable(
tf.constant(1.0, shape=[second_hidden_num]))
second_nn_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
# Model.
def model(data, model_drop=True, init=True):
if not large_data_size(data) or not large_data_size(input_weights):
stride_ps[0] = [1, 1, 1, 1]
conv = tf.nn.conv2d(data,
input_weights,
stride_ps[0],
use_cudnn_on_gpu=True,
padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu6(conv + input_biases)
if drop and model_drop:
hidden = tf.nn.dropout(hidden, 0.8)
for i in range(mid_layer_cnt):
print(hidden)
if init:
# avoid filter shape larger than input shape
hid_shape = hidden.get_shape()
# print(hid_shape)
filter_w = patch_size / (i + 1)
filter_h = patch_size / (i + 1)
# print(filter_w)
# print(filter_h)
if filter_w > hid_shape[1]:
filter_w = int(hid_shape[1])
if filter_h > hid_shape[2]:
filter_h = int(hid_shape[2])
layer_weight = tf.Variable(
tf.truncated_normal(shape=[
filter_w, filter_h, depth * (i + 1),
depth * (i + 2)
],
stddev=0.1))
layer_weights.append(layer_weight)
if not large_data_size(hidden) or not large_data_size(
layer_weights[i]):
# print("is not large data")
stride_ps[i + 1] = [1, 1, 1, 1]
# print(stride_ps[i + 1])
# print(len(stride_ps))
# print(i + 1)
conv = tf.nn.conv2d(hidden,
layer_weights[i],
stride_ps[i + 1],
use_cudnn_on_gpu=True,
padding='SAME')
if not large_data_size(conv):
print('not large')
conv = maxpool2d(conv, 1, 1)
else:
conv = maxpool2d(conv)
hidden = tf.nn.relu6(conv + layer_biases[i])
shapes = hidden.get_shape().as_list()
shape_mul = 1
for s in shapes[1:]:
shape_mul *= s
if init:
output_size = shape_mul
output_weights.append(
tf.Variable(
tf.truncated_normal([output_size, first_hidden_num],
stddev=0.1)))
reshape = tf.reshape(hidden, [shapes[0], shape_mul])
hidden = tf.nn.relu6(
tf.matmul(reshape, output_weights[0]) + output_biases)
if drop and model_drop:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, first_nn_weights) + first_nn_biases
if drop and model_drop:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, second_nn_weights) + second_nn_biases
return hidden
# Training computation.
logits = model(tf_train_dataset)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=logits,
labels=tf_train_labels))
# Optimizer.
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
starter_learning_rate = 0.1
learning_rate = tf.train.exponential_decay(starter_learning_rate,
cur_step,
600,
0.1,
staircase=True)
optimizer = tf.train.GradientDescentOptimizer(
learning_rate).minimize(loss, global_step=cur_step)
else:
optimizer = tf.train.AdagradOptimizer(0.06).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, model_drop=False, init=False))
test_prediction = tf.nn.softmax(
model(tf_test_dataset, model_drop=False, init=False))
saver = tf.train.Saver()
# on step 1750, run over 55000 train images
num_steps = 1750 * 3
save_path = 'conv_mnist'
save_flag = True
with tf.Session(graph=graph) as session:
if os.path.exists(save_path) and save_flag:
# Restore variables from disk.
saver.restore(session, save_path)
else:
tf.global_variables_initializer().run()
print('Initialized')
end_train = False
mean_loss = 0
for step in range(num_steps):
if end_train:
break
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)
mean_loss += l
if step % 10 == 0:
mean_loss /= 10.0
if step % 200 == 0:
print('Minibatch loss at step %d: %f' % (step, mean_loss))
print('Validation accuracy: %.1f%%' %
accuracy(valid_prediction.eval(), valid_labels))
mean_loss = 0
if save_flag:
saver.save(session, save_path)
print('Test accuracy: %.1f%%' %
accuracy(test_prediction.eval(), test_labels))
def conv_train():
batch_size = 16
patch_size = 5
depth = 16
num_hidden = 64
num_channels = 1
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(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))
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))
def conv_train(train_dataset, train_labels, valid_dataset, valid_labels, test_dataset, test_labels, image_size,
num_labels, basic_hps, stride_ps):
batch_size = basic_hps['batch_size']
patch_size = basic_hps['patch_size']
depth = basic_hps['depth']
num_hidden = basic_hps['num_hidden']
num_channels = 1
layer_cnt = basic_hps['layer_sum']
starter_learning_rate = basic_hps['starter_learning_rate']
loss_collect = list()
first_hidden_num = basic_hps['num_hidden']
second_hidden_num = first_hidden_num / 2 + 1
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)
input_weights = tf.Variable(tf.truncated_normal(
[patch_size, patch_size, num_channels, depth], stddev=0.1))
input_biases = tf.Variable(tf.zeros([depth]))
mid_layer_cnt = layer_cnt - 1
layer_weights = list()
layer_biases = [tf.Variable(tf.constant(1.0, shape=[depth / (i + 2)])) for i in range(mid_layer_cnt)]
output_weights = list()
output_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
first_nn_weights = tf.Variable(tf.truncated_normal(
[first_hidden_num, second_hidden_num], stddev=0.1))
second_nn_weights = tf.Variable(tf.truncated_normal(
[second_hidden_num, num_labels], stddev=0.1))
first_nn_biases = tf.Variable(tf.constant(1.0, shape=[second_hidden_num]))
second_nn_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
# Model.
def model(data, init=False):
# Variables.
if not large_data_size(data) or not large_data_size(input_weights):
stride_ps[0] = [1, 1, 1, 1]
conv = tf.nn.conv2d(data, input_weights, stride_ps[0], use_cudnn_on_gpu=True, padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + input_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.8)
for i in range(mid_layer_cnt):
# print(hidden)
if init:
hid_shape = hidden.get_shape()
filter_w = patch_size / (i + 1)
filter_h = patch_size / (i + 1)
if filter_w > hid_shape[1]:
filter_w = int(hid_shape[1])
if filter_h > hid_shape[2]:
filter_h = int(hid_shape[2])
layer_weight = tf.Variable(tf.truncated_normal(shape=[filter_w, filter_h, depth / (i + 1), depth / (i + 2)],
stddev=0.1))
layer_weights.append(layer_weight)
if not large_data_size(hidden) or not large_data_size(layer_weights[i]):
stride_ps[i + 1] = [1, 1, 1, 1]
conv = tf.nn.conv2d(hidden, layer_weights[i], stride_ps[i + 1], use_cudnn_on_gpu=True, padding='SAME')
if not large_data_size(conv):
conv = maxpool2d(conv, 1, 1)
else:
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer_biases[i])
if init:
hidden = tf.nn.dropout(hidden, 0.8)
shapes = hidden.get_shape().as_list()
shape_mul = 1
for s in shapes[1:]:
shape_mul *= s
if init:
output_size = shape_mul
output_weights.append(tf.Variable(tf.truncated_normal([output_size, num_hidden], stddev=0.1)))
reshape = tf.reshape(hidden, [shapes[0], shape_mul])
hidden = tf.nn.relu6(tf.matmul(reshape, output_weights[0]) + output_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, first_nn_weights) + first_nn_biases
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, second_nn_weights) + second_nn_biases
return hidden
# Training computation.
logits = model(tf_train_dataset, init=True)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=tf_train_labels))
optimizer = tf.train.AdagradOptimizer(starter_learning_rate).minimize(loss)
train_prediction = tf.nn.softmax(logits)
valid_prediction = tf.nn.softmax(model(tf_valid_dataset))
num_steps = 1001
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
print('Initialized')
end_train = False
mean_loss = 0
for step in range(num_steps):
if end_train:
break
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)
mean_loss += l
if step % 10 == 0:
mean_loss /= 5.0
mean_loss = 0
if step % 100 == 0:
loss_collect.append(mean_loss)
print('Minibatch loss at step %d: %f' % (step, l))
print('Validation accuracy: %.1f%%' % accuracy(
valid_prediction.eval(), valid_labels))
def better_conv_train(drop=False, lrd=False):
batch_size = 16
patch_size = 5
depth = 16
num_hidden = 64
num_channels = 1
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(
[64, 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')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer1_biases)
if drop:
hidden = tf.nn.dropout(hidden, 0.5)
conv = tf.nn.conv2d(hidden, layer2_weights, [1, 2, 2, 1], padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer2_biases)
if drop:
hidden = tf.nn.dropout(hidden, 0.7)
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)
if drop:
hidden = tf.nn.dropout(hidden, 0.8)
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=logits, labels=tf_train_labels))
# Optimizer.
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
starter_learning_rate = 0.1
learning_rate = tf.train.exponential_decay(starter_learning_rate, cur_step, 10000, 0.96, staircase=True)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=cur_step)
else:
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))
num_steps = 5001
losses = []
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().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)
losses.append(l)
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))
print(losses)
def tf_deep_nn(regular=False, drop_out=False, lrd=False, layer_cnt=2):
batch_size = 128
graph = tf.Graph()
with graph.as_default():
tf_train_dataset = tf.placeholder(tf.float32,
shape=(batch_size, feature_dim))
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)
hidden_node_count = 32
# start weight
hidden_stddev = np.sqrt(2.0 / 100)
weights1 = tf.Variable(
tf.truncated_normal([feature_dim, hidden_node_count],
stddev=hidden_stddev))
biases1 = tf.Variable(tf.zeros([hidden_node_count]))
# middle weight
weights = []
biases = []
hidden_cur_cnt = hidden_node_count
for i in range(layer_cnt - 2):
if hidden_cur_cnt > 2:
hidden_next_cnt = int(hidden_cur_cnt / 2)
else:
hidden_next_cnt = 2
hidden_stddev = np.sqrt(2.0 / hidden_cur_cnt / 10)
weights.append(
tf.Variable(
tf.truncated_normal([hidden_cur_cnt, hidden_next_cnt],
stddev=hidden_stddev)))
biases.append(tf.Variable(tf.zeros([hidden_next_cnt])))
hidden_cur_cnt = hidden_next_cnt
# first wx + b
y0 = tf.matmul(tf_train_dataset, weights1) + biases1
# first sigmoid
hidden = tf.nn.sigmoid(y0)
# hidden = y0
hidden_drop = hidden
# first DropOut
keep_prob = 0.5
if drop_out:
hidden_drop = tf.nn.dropout(hidden, keep_prob)
# first wx+b for valid
valid_y0 = tf.matmul(tf_valid_dataset, weights1) + biases1
valid_hidden = tf.nn.sigmoid(valid_y0)
# valid_hidden = valid_y0
# first wx+b for test
test_y0 = tf.matmul(tf_test_dataset, weights1) + biases1
test_hidden = tf.nn.sigmoid(test_y0)
# test_hidden = test_y0
# middle layer
for i in range(layer_cnt - 2):
y1 = tf.matmul(hidden_drop, weights[i]) + biases[i]
hidden_drop = tf.nn.sigmoid(y1)
if drop_out:
keep_prob += 0.5 * i / (layer_cnt + 1)
hidden_drop = tf.nn.dropout(hidden_drop, keep_prob)
y0 = tf.matmul(hidden, weights[i]) + biases[i]
hidden = tf.nn.sigmoid(y0)
# hidden = y0
valid_y0 = tf.matmul(valid_hidden, weights[i]) + biases[i]
valid_hidden = tf.nn.sigmoid(valid_y0)
# valid_hidden = valid_y0
test_y0 = tf.matmul(test_hidden, weights[i]) + biases[i]
test_hidden = tf.nn.sigmoid(test_y0)
# test_hidden = test_y0
# last weight
weights2 = tf.Variable(
tf.truncated_normal([hidden_cur_cnt, num_labels],
stddev=hidden_stddev / 2))
biases2 = tf.Variable(tf.zeros([num_labels]))
# last wx + b
logits = tf.matmul(hidden_drop, weights2) + biases2
# predicts
logits_predict = tf.matmul(hidden, weights2) + biases2
valid_predict = tf.matmul(valid_hidden, weights2) + biases2
test_predict = tf.matmul(test_hidden, weights2) + biases2
l2_loss = 0
# enable regularization
if regular:
l2_loss = tf.nn.l2_loss(weights1) + tf.nn.l2_loss(weights2)
for i in range(len(weights)):
l2_loss += tf.nn.l2_loss(weights[i])
# l2_loss += tf.nn.l2_loss(biases[i])
beta = 1e-2
l2_loss *= beta
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits,
tf_train_labels)) + l2_loss
# Optimizer.
if lrd:
cur_step = tf.Variable(
0, trainable=False) # count the number of steps taken.
starter_learning_rate = 0.4
learning_rate = tf.train.exponential_decay(starter_learning_rate,
cur_step,
500,
0.75,
staircase=True)
optimizer = tf.train.GradientDescentOptimizer(
learning_rate).minimize(loss, global_step=cur_step)
else:
optimizer = tf.train.AdamOptimizer(0.5).minimize(loss)
# Predictions for the training, validation, and test data.
train_prediction = tf.nn.softmax(logits_predict)
valid_prediction = tf.nn.softmax(valid_predict)
test_prediction = tf.nn.softmax(test_predict)
num_steps = 8001
with tf.Session(graph=graph) as session:
tf.initialize_all_variables().run()
print("Initialized")
for step in range(num_steps):
offset_range = train_labels.shape[0] - batch_size
offset = (step * batch_size) % offset_range
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))
def better_conv_train(drop=False, lrd=False):
batch_size = 12
patch_size = 2
depth = 12
num_hidden = 64
num_channels = 1
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(
[48, 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')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer1_biases)
if drop:
hidden = tf.nn.dropout(hidden, 0.5)
conv = tf.nn.conv2d(hidden, layer2_weights, [1, 2, 2, 1], padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer2_biases)
if drop:
hidden = tf.nn.dropout(hidden, 0.7)
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)
if drop:
hidden = tf.nn.dropout(hidden, 0.8)
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.
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
starter_learning_rate = 0.1
learning_rate = tf.train.exponential_decay(starter_learning_rate, cur_step, 10000, 0.96, staircase=True)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=cur_step)
else:
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))
num_steps = 5001
losses = []
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)
losses.append(l)
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))
print(losses)
def conv_train():
batch_size = 16
patch_size = 5
depth = 16
num_hidden = 64
num_channels = 1
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(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))
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))
def tf_better_nn(offset_range=-1, regular=False, drop_out=False, lrd=False):
batch_size = 128
graph = tf.Graph()
with graph.as_default():
# Input data. For the training data, we use a placeholder that will be fed
# at run time with a training minibatch.
tf_train_dataset = tf.placeholder(tf.float32,
shape=(batch_size,
image_size * image_size))
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)
hidden_node_count = 1024
# Variables.
weights1 = tf.Variable(
tf.truncated_normal([image_size * image_size, hidden_node_count]))
biases1 = tf.Variable(tf.zeros([hidden_node_count]))
weights2 = tf.Variable(
tf.truncated_normal([hidden_node_count, num_labels]))
biases2 = tf.Variable(tf.zeros([num_labels]))
# Training computation. right most
ys = tf.matmul(tf_train_dataset, weights1) + biases1
hidden = tf.nn.relu(ys)
h_fc = hidden
valid_y0 = tf.matmul(tf_valid_dataset, weights1) + biases1
valid_hidden1 = tf.nn.relu(valid_y0)
test_y0 = tf.matmul(tf_test_dataset, weights1) + biases1
test_hidden1 = tf.nn.relu(test_y0)
# enable DropOut
keep_prob = tf.placeholder(tf.float32)
if drop_out:
hidden_drop = tf.nn.dropout(hidden, keep_prob)
h_fc = hidden_drop
# left most
logits = tf.matmul(h_fc, weights2) + biases2
# only drop out when train
logits_predict = tf.matmul(hidden, weights2) + biases2
valid_predict = tf.matmul(valid_hidden1, weights2) + biases2
test_predict = tf.matmul(test_hidden1, weights2) + biases2
# loss
l2_loss = tf.nn.l2_loss(weights1) + tf.nn.l2_loss(
biases1) + tf.nn.l2_loss(weights2) + tf.nn.l2_loss(biases2)
# enable regularization
if not regular:
l2_loss = 0
beta = 0.002
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(
logits=logits, labels=tf_train_labels)) + beta * l2_loss
# Optimizer.
optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
starter_learning_rate = 0.1
learning_rate = tf.train.exponential_decay(starter_learning_rate,
cur_step,
10000,
0.96,
staircase=True)
optimizer = tf.train.GradientDescentOptimizer(
learning_rate).minimize(loss, global_step=cur_step)
# Predictions for the training, validation, and test data.
train_prediction = tf.nn.softmax(logits_predict)
valid_prediction = tf.nn.softmax(valid_predict)
test_prediction = tf.nn.softmax(test_predict)
num_steps = 30001
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
print("Initialized")
for step in range(num_steps):
# Pick an offset within the training data, which has been randomized.
# Note: we could use better randomization across epochs.
if offset_range == -1:
offset_range = train_labels.shape[0] - batch_size
offset = (step * batch_size) % offset_range
# Generate a minibatch.
batch_data = train_dataset[offset:(offset + batch_size), :]
batch_labels = train_labels[offset:(offset + batch_size), :]
# Prepare a dictionary telling the session where to feed the minibatch.
# The key of the dictionary is the placeholder node of the graph to be fed,
# and the value is the numpy array to feed to it.
feed_dict = {
tf_train_dataset: batch_data,
tf_train_labels: batch_labels,
keep_prob: 0.5
}
_, l, predictions = session.run(
[optimizer, loss, train_prediction], feed_dict=feed_dict)
if step % 500 == 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))
def tf_better_nn(offset_range=-1, regular=False, drop_out=False, lrd=False):
batch_size = 128
graph = tf.Graph()
with graph.as_default():
# Input data. For the training data, we use a placeholder that will be fed
# at run time with a training minibatch.
tf_train_dataset = tf.placeholder(tf.float32,
shape=(batch_size, image_size * image_size))
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)
hidden_node_count = 1024
# Variables.
weights1 = tf.Variable(
tf.truncated_normal([image_size * image_size, hidden_node_count]))
biases1 = tf.Variable(tf.zeros([hidden_node_count]))
weights2 = tf.Variable(
tf.truncated_normal([hidden_node_count, num_labels]))
biases2 = tf.Variable(tf.zeros([num_labels]))
# Training computation. right most
ys = tf.matmul(tf_train_dataset, weights1) + biases1
hidden = tf.nn.relu(ys)
h_fc = hidden
valid_y0 = tf.matmul(tf_valid_dataset, weights1) + biases1
valid_hidden1 = tf.nn.relu(valid_y0)
test_y0 = tf.matmul(tf_test_dataset, weights1) + biases1
test_hidden1 = tf.nn.relu(test_y0)
# enable DropOut
keep_prob = tf.placeholder(tf.float32)
if drop_out:
hidden_drop = tf.nn.dropout(hidden, keep_prob)
h_fc = hidden_drop
# left most
logits = tf.matmul(h_fc, weights2) + biases2
# only drop out when train
logits_predict = tf.matmul(hidden, weights2) + biases2
valid_predict = tf.matmul(valid_hidden1, weights2) + biases2
test_predict = tf.matmul(test_hidden1, weights2) + biases2
# loss
l2_loss = tf.nn.l2_loss(weights1) + tf.nn.l2_loss(biases1) + tf.nn.l2_loss(weights2) + tf.nn.l2_loss(biases2)
# enable regularization
if not regular:
l2_loss = 0
beta = 0.002
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=tf_train_labels)) + beta * l2_loss
# Optimizer.
optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
starter_learning_rate = 0.1
learning_rate = tf.train.exponential_decay(starter_learning_rate, cur_step, 10000, 0.96, staircase=True)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=cur_step)
# Predictions for the training, validation, and test data.
train_prediction = tf.nn.softmax(logits_predict)
valid_prediction = tf.nn.softmax(valid_predict)
test_prediction = tf.nn.softmax(test_predict)
num_steps = 30001
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
print("Initialized")
for step in range(num_steps):
# Pick an offset within the training data, which has been randomized.
# Note: we could use better randomization across epochs.
if offset_range == -1:
offset_range = train_labels.shape[0] - batch_size
offset = (step * batch_size) % offset_range
# Generate a minibatch.
batch_data = train_dataset[offset:(offset + batch_size), :]
batch_labels = train_labels[offset:(offset + batch_size), :]
# Prepare a dictionary telling the session where to feed the minibatch.
# The key of the dictionary is the placeholder node of the graph to be fed,
# and the value is the numpy array to feed to it.
feed_dict = {tf_train_dataset: batch_data, tf_train_labels: batch_labels, keep_prob: 0.5}
_, l, predictions = session.run(
[optimizer, loss, train_prediction], feed_dict=feed_dict)
if step % 500 == 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))
def conv_train(train_dataset, train_labels, valid_dataset, valid_labels, test_dataset, test_labels, image_size,
num_labels, basic_hps, stride_ps, drop=False, lrd=False):
batch_size = basic_hps['batch_size']
patch_size = basic_hps['patch_size']
depth = basic_hps['depth']
first_hidden_num = basic_hps['num_hidden']
second_hidden_num = first_hidden_num / 2 + 1
num_channels = 1
layer_cnt = basic_hps['layer_sum']
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.
# the third parameter must be same as the last layer depth
input_weights = tf.Variable(tf.truncated_normal(
[patch_size, patch_size, num_channels, depth], stddev=0.1))
input_biases = tf.Variable(tf.zeros([depth]))
mid_layer_cnt = layer_cnt - 1
layer_weights = list()
layer_biases = [tf.Variable(tf.constant(1.0, shape=[depth * (i + 2)])) for i in range(mid_layer_cnt)]
output_weights = list()
output_biases = tf.Variable(tf.constant(1.0, shape=[first_hidden_num]))
first_nn_weights = tf.Variable(tf.truncated_normal(
[first_hidden_num, second_hidden_num], stddev=0.1))
second_nn_weights = tf.Variable(tf.truncated_normal(
[second_hidden_num, num_labels], stddev=0.1))
first_nn_biases = tf.Variable(tf.constant(1.0, shape=[second_hidden_num]))
second_nn_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
# Model.
def model(data, model_drop=True, init=True):
if not large_data_size(data) or not large_data_size(input_weights):
stride_ps[0] = [1, 1, 1, 1]
conv = tf.nn.conv2d(data, input_weights, stride_ps[0], use_cudnn_on_gpu=True, padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu6(conv + input_biases)
if drop and model_drop:
hidden = tf.nn.dropout(hidden, 0.8)
for i in range(mid_layer_cnt):
print(hidden)
if init:
# avoid filter shape larger than input shape
hid_shape = hidden.get_shape()
# print(hid_shape)
filter_w = patch_size / (i + 1)
filter_h = patch_size / (i + 1)
# print(filter_w)
# print(filter_h)
if filter_w > hid_shape[1]:
filter_w = int(hid_shape[1])
if filter_h > hid_shape[2]:
filter_h = int(hid_shape[2])
layer_weight = tf.Variable(tf.truncated_normal(
shape=[filter_w, filter_h, depth * (i + 1), depth * (i + 2)], stddev=0.1))
layer_weights.append(layer_weight)
if not large_data_size(hidden) or not large_data_size(layer_weights[i]):
# print("is not large data")
stride_ps[i + 1] = [1, 1, 1, 1]
# print(stride_ps[i + 1])
# print(len(stride_ps))
# print(i + 1)
conv = tf.nn.conv2d(hidden, layer_weights[i], stride_ps[i + 1], use_cudnn_on_gpu=True, padding='SAME')
if not large_data_size(conv):
print('not large')
conv = maxpool2d(conv, 1, 1)
else:
conv = maxpool2d(conv)
hidden = tf.nn.relu6(conv + layer_biases[i])
shapes = hidden.get_shape().as_list()
shape_mul = 1
for s in shapes[1:]:
shape_mul *= s
if init:
output_size = shape_mul
output_weights.append(tf.Variable(tf.truncated_normal([output_size, first_hidden_num], stddev=0.1)))
reshape = tf.reshape(hidden, [shapes[0], shape_mul])
hidden = tf.nn.relu6(tf.matmul(reshape, output_weights[0]) + output_biases)
if drop and model_drop:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, first_nn_weights) + first_nn_biases
if drop and model_drop:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, second_nn_weights) + second_nn_biases
return hidden
# Training computation.
logits = model(tf_train_dataset)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=tf_train_labels))
# Optimizer.
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
starter_learning_rate = 0.1
learning_rate = tf.train.exponential_decay(starter_learning_rate, cur_step, 600, 0.1, staircase=True)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=cur_step)
else:
optimizer = tf.train.AdagradOptimizer(0.06).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, model_drop=False, init=False))
test_prediction = tf.nn.softmax(model(tf_test_dataset, model_drop=False, init=False))
saver = tf.train.Saver()
# on step 1750, run over 55000 train images
num_steps = 1750 * 3
save_path = 'conv_mnist'
save_flag = True
with tf.Session(graph=graph) as session:
if os.path.exists(save_path) and save_flag:
# Restore variables from disk.
saver.restore(session, save_path)
else:
tf.global_variables_initializer().run()
print('Initialized')
end_train = False
mean_loss = 0
for step in range(num_steps):
if end_train:
break
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)
mean_loss += l
if step % 10 == 0:
mean_loss /= 10.0
if step % 200 == 0:
print('Minibatch loss at step %d: %f' % (step, mean_loss))
print('Validation accuracy: %.1f%%' % accuracy(
valid_prediction.eval(), valid_labels))
mean_loss = 0
if save_flag:
saver.save(session, save_path)
print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))
def conv_train(basic_hps, stride_ps, layer_cnt=3, drop=False, lrd=False):
batch_size = basic_hps['batch_size']
patch_size = basic_hps['patch_size']
depth = basic_hps['depth']
num_hidden = basic_hps['num_hidden']
num_channels = basic_hps['num_channels']
loss_collect = list()
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.
input_weights = tf.Variable(tf.truncated_normal(
[patch_size, patch_size, num_channels, depth], stddev=0.1))
input_biases = tf.Variable(tf.zeros([depth]))
mid_layer_cnt = layer_cnt - 1
layer_weights = [tf.Variable(tf.truncated_normal(
[patch_size, patch_size, depth, depth], stddev=0.1)) for _ in range(mid_layer_cnt)]
layer_biases = [tf.Variable(tf.constant(1.0, shape=[depth])) for _ in range(mid_layer_cnt)]
output_size = size_by_conv(stride_ps, [batch_size, image_size, image_size, num_channels], layer_cnt)
output_weights = tf.Variable(tf.truncated_normal([output_size, num_hidden], stddev=0.1))
output_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
final_weights = tf.Variable(tf.truncated_normal(
[num_hidden, num_labels], stddev=0.1))
final_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
# Model.
def model(data):
conv = tf.nn.conv2d(data, input_weights, stride_ps[0], use_cudnn_on_gpu=True, padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + input_biases)
if drop:
hidden = tf.nn.dropout(hidden, 0.5)
for i in range(mid_layer_cnt):
print(i)
conv = tf.nn.conv2d(hidden, layer_weights[i], stride_ps[i + 1], use_cudnn_on_gpu=True, padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer_biases[i])
if drop:
hidden = tf.nn.dropout(hidden, 0.7)
shape = hidden.get_shape().as_list()
reshape = tf.reshape(hidden, [shape[0], output_size])
hidden = tf.nn.relu(tf.matmul(reshape, output_weights) + output_biases)
if drop:
hidden = tf.nn.dropout(hidden, 0.8)
return tf.matmul(hidden, final_weights) + final_biases
# Training computation.
logits = model(tf_train_dataset)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits, tf_train_labels))
# Optimizer.
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
starter_learning_rate = 0.1
learning_rate = tf.train.exponential_decay(starter_learning_rate, cur_step, 10000, 0.96, staircase=True)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=cur_step)
else:
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))
num_steps = 3001
fit_frep = 100
with tf.Session(graph=graph) as session:
tf.initialize_all_variables().run()
print('Initialized')
end_train = False
for step in range(num_steps):
if end_train:
break
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)
loss_collect.append(l)
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))
if step == fit_frep:
res = fit_loss([batch_size, depth, num_hidden], loss_collect)
ret = res['ret']
if ret == 1:
print('ret is end train when step is {step}'.format(step=step))
elif step % fit_frep == 0 and step != 0:
for i in range(fit_frep):
res = fit_loss(
[batch_size, depth, num_hidden],
loss_collect[i + step - fit_frep * 2 + 1: i + step - fit_frep + 2])
ret = res['ret']
if i == 0:
print(res)
if ret == 1:
print('ret is end train when step is {step}'.format(step=step))
end_train = True
break
print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))
for loss in loss_collect:
print(loss)
def conv_train(train_dataset, train_labels, valid_dataset, valid_labels, test_dataset, test_labels, image_size,
num_labels, basic_hps, stride_ps, lrd=False):
batch_size = basic_hps['batch_size']
patch_size = basic_hps['patch_size']
depth = basic_hps['depth']
num_hidden = basic_hps['num_hidden']
num_channels = 1
layer_cnt = basic_hps['layer_sum']
loss_collect = list()
first_hidden_num = basic_hps['num_hidden']
second_hidden_num = first_hidden_num / 2 + 1
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)
input_weights = tf.Variable(tf.truncated_normal(
[patch_size, patch_size, num_channels, depth], stddev=0.1))
input_biases = tf.Variable(tf.zeros([depth]))
mid_layer_cnt = layer_cnt - 1
layer_weights = list()
layer_biases = [tf.Variable(tf.constant(1.0, shape=[depth / (i + 2)])) for i in range(mid_layer_cnt)]
output_weights = list()
output_biases = tf.Variable(tf.constant(1.0, shape=[num_hidden]))
first_nn_weights = tf.Variable(tf.truncated_normal(
[first_hidden_num, second_hidden_num], stddev=0.1))
second_nn_weights = tf.Variable(tf.truncated_normal(
[second_hidden_num, num_labels], stddev=0.1))
first_nn_biases = tf.Variable(tf.constant(1.0, shape=[second_hidden_num]))
second_nn_biases = tf.Variable(tf.constant(1.0, shape=[num_labels]))
# Model.
def model(data, init=False):
# Variables.
if not large_data_size(data) or not large_data_size(input_weights):
stride_ps[0] = [1, 1, 1, 1]
conv = tf.nn.conv2d(data, input_weights, stride_ps[0], use_cudnn_on_gpu=True, padding='SAME')
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + input_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.8)
for i in range(mid_layer_cnt):
# print(hidden)
if init:
# avoid filter shape larger than input shape
hid_shape = hidden.get_shape()
# print(hid_shape)
filter_w = patch_size / (i + 1)
filter_h = patch_size / (i + 1)
# print(filter_w)
# print(filter_h)
if filter_w > hid_shape[1]:
filter_w = int(hid_shape[1])
if filter_h > hid_shape[2]:
filter_h = int(hid_shape[2])
layer_weight = tf.Variable(tf.truncated_normal(shape=[filter_w, filter_h, depth / (i + 1), depth / (i + 2)],
stddev=0.1))
layer_weights.append(layer_weight)
if not large_data_size(hidden) or not large_data_size(layer_weights[i]):
# print("is not large data")
stride_ps[i + 1] = [1, 1, 1, 1]
# print(stride_ps[i + 1])
# print(len(stride_ps))
# print(i + 1)
conv = tf.nn.conv2d(hidden, layer_weights[i], stride_ps[i + 1], use_cudnn_on_gpu=True, padding='SAME')
if not large_data_size(conv):
conv = maxpool2d(conv, 1, 1)
else:
conv = maxpool2d(conv)
hidden = tf.nn.relu(conv + layer_biases[i])
if init:
hidden = tf.nn.dropout(hidden, 0.8)
shapes = hidden.get_shape().as_list()
shape_mul = 1
for s in shapes[1:]:
shape_mul *= s
if init:
output_size = shape_mul
output_weights.append(tf.Variable(tf.truncated_normal([output_size, num_hidden], stddev=0.1)))
reshape = tf.reshape(hidden, [shapes[0], shape_mul])
hidden = tf.nn.relu6(tf.matmul(reshape, output_weights[0]) + output_biases)
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, first_nn_weights) + first_nn_biases
if init:
hidden = tf.nn.dropout(hidden, 0.5)
hidden = tf.matmul(hidden, second_nn_weights) + second_nn_biases
return hidden
# Training computation.
logits = model(tf_train_dataset, init=True)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=tf_train_labels))
# Optimizer.
starter_learning_rate = 0.1
if lrd:
cur_step = tf.Variable(0) # count the number of steps taken.
learning_rate = tf.train.exponential_decay(starter_learning_rate, cur_step, 10000, 0.96, staircase=True)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=cur_step)
else:
optimizer = tf.train.AdagradOptimizer(starter_learning_rate).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))
num_steps = 3001
start_fit = 600
init_loss = []
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
print('Initialized')
end_train = False
mean_loss = 0
for step in range(num_steps):
if end_train:
break
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)
mean_loss += l
if step % 5 == 0:
mean_loss /= 5.0
loss_collect.append(mean_loss)
mean_loss = 0
if step >= start_fit:
# print(loss_collect)
if step == start_fit:
res = fit_more(1, [batch_size, depth, num_hidden, layer_cnt, patch_size], loss_collect)
else:
res = fit_more(0, [batch_size, depth, num_hidden, layer_cnt, patch_size], loss_collect)
loss_collect.remove(loss_collect[0])
ret = res['ret']
if ret == 1:
print('ret is end train when step is {step}'.format(step=step))
init_loss.append(loss_collect)
more_index = predict_future([batch_size, depth, num_hidden, layer_cnt, patch_size], init_loss[0])
print('more index: %d' % more_index)
for i in range(more_index):
offset = ((step + i + 1) * 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)
loss_collect.append(l)
file_helper.write('/home/cwh/coding/python/NN/line.txt', str(loss_collect[20]))
loss_collect.remove(loss_collect[0])
for loss in loss_collect[21:]:
file_helper.write('/home/cwh/coding/python/NN/line.txt', str(loss))
end_train = True
file_helper.write('/home/cwh/coding/python/NN/line.txt', '===')
if step % 50 == 0:
print('Minibatch loss at step %d: %f' % (step, l))
print('Validation accuracy: %.1f%%' % accuracy(
valid_prediction.eval(), valid_labels))
print('Test accuracy: %.1f%%' % accuracy(test_prediction.eval(), test_labels))
