def calculate_error_rate_variance_from_table(
tables=['../data/tfidf-pairs.npy'],
split=0.0,
deviation=0.025,
verbose=True,
normalize=(0.0, 1.0)):
data = []
for t in tables:
f = open(t, 'rb')
r = np.load(f)
data.append(r)
for d in xrange(len(data)):
data[d] = (data[d] - normalize[0]) / (normalize[1] - normalize[0])
if verbose:
print 'Calculating error rate variance with a deviation of %.5f...' % deviation
splits = np.linspace(split - deviation, split + deviation, 100)
errors = []
for s in splits:
errors.append(metrics.error_rate(data[0], data[1], s))
errors = np.asarray(errors)
v = np.std(errors)
m = np.max(errors)
if verbose:
print 'Stdev: %.5f' % v
print 'Max value: %.5f' % m
print 'Used split point: %.5f' % split
return v, m
def calculate_error_rate_from_table(tables=['../data/tfidf-pairs.npy'], split=0.0, verbose=True, normalize=(0.0, 1.0)):
data = []
for t in tables:
f = open(t, 'rb')
r = np.load(f)
data.append(r)
for d in xrange(len(data)):
data[d] = (data[d] - normalize[0]) / (normalize[1] - normalize[0])
if verbose:
print 'Calculating optimal error rate...'
error = metrics.error_rate(data[0], data[1], split)
if verbose:
print 'Error: %.5f' % error
print 'Used split point: %.5f' % split
return error
def calculate_error_rate_from_table(tables=['../data/tfidf-pairs.npy'],
split=0.0,
verbose=True,
normalize=(0.0, 1.0)):
data = []
for t in tables:
f = open(t, 'rb')
r = np.load(f)
data.append(r)
for d in xrange(len(data)):
data[d] = (data[d] - normalize[0]) / (normalize[1] - normalize[0])
if verbose:
print 'Calculating optimal error rate...'
error = metrics.error_rate(data[0], data[1], split)
if verbose:
print 'Error: %.5f' % error
print 'Used split point: %.5f' % split
return error
def calculate_error_rate_variance_from_table(tables=['../data/tfidf-pairs.npy'], split=0.0, deviation=0.025, verbose=True, normalize=(0.0, 1.0)):
data = []
for t in tables:
f = open(t, 'rb')
r = np.load(f)
data.append(r)
for d in xrange(len(data)):
data[d] = (data[d] - normalize[0]) / (normalize[1] - normalize[0])
if verbose:
print 'Calculating error rate variance with a deviation of %.5f...' % deviation
splits = np.linspace(split - deviation, split + deviation, 100)
errors = []
for s in splits:
errors.append(metrics.error_rate(data[0], data[1], s))
errors = np.asarray(errors)
v = np.std(errors)
m = np.max(errors)
if verbose:
print 'Stdev: %.5f' % v
print 'Max value: %.5f' % m
print 'Used split point: %.5f' % split
return v, m
def main(argv = None): # pylint: disable=unused-argument
# load imageset
train_set_folder = os.path.join(module_dir, os.path.pardir, os.path.pardir, 'data/ocr/train')
test_set_folder = os.path.join(module_dir, os.path.pardir, os.path.pardir, 'data/ocr/test')
# Extract it into numpy arrays.
train_data, train_labels = load_imageset(train_set_folder, to_img_size = (28, 28, 1), ext = 'png')
test_data, test_labels = load_imageset(test_set_folder, to_img_size = (28, 28, 1), ext = 'png')
height = train_data.shape[1]
width = train_data.shape[2]
channel = (train_data.shape[3] if train_data.ndim > 3 else 1)
label_max = np.amax(train_labels)
label_min = np.amin(train_labels)
num_labels = label_max - label_min + 1
# Generate a validation set.
train_data, train_labels, validation_data, validation_labels = split_cv(train_data, train_labels, 0.1)
num_epochs = NUM_EPOCHS
train_size = train_labels.shape[0]
# This is where training samples and labels are fed to the graph.
# These placeholder nodes will be fed a batch of training data at each
# training step using the {feed_dict} argument to the Run() call below.
train_data_node = tf.placeholder(
tf.float32,
shape = (BATCH_SIZE, height, width, channel))
train_labels_node = tf.placeholder(tf.int64, shape = (BATCH_SIZE,))
eval_data = tf.placeholder(
tf.float32,
shape=(EVAL_BATCH_SIZE, height, width, channel))
# The variables below hold all the trainable weights. They are passed an
# initial value which will be assigned when we call:
# {tf.initialize_all_variables().run()}
conv1_weights = tf.Variable(
tf.truncated_normal([5, 5, channel, 32], # 5x5 filter, depth 32.
stddev = 0.1,
seed = SEED),
name="conv1_weights")
conv1_biases = tf.Variable(tf.zeros([32]), name = "conv1_biases")
conv2_weights = tf.Variable(
tf.truncated_normal([5, 5, 32, 64],
stddev = 0.1,
seed = SEED),
name="conv2_weights")
conv2_biases = tf.Variable(tf.constant(0.1, shape = [64]), name = "conv2_biases")
fc1_weights = tf.Variable( # fully connected, depth 512.
tf.truncated_normal(
[height // 4 * width // 4 * 64, 512],
stddev = 0.1,
seed = SEED),
name = "fc1_weights")
fc1_biases = tf.Variable(tf.constant(0.1, shape = [512]), name = "fc1_biases")
fc2_weights = tf.Variable(
tf.truncated_normal([512, num_labels],
stddev = 0.1,
seed = SEED),
name = "fc2_weights")
fc2_biases = tf.Variable(tf.constant(0.1, shape = [num_labels]), name = "fc2_biases")
# We will replicate the model structure for the training subgraph, as well
# as the evaluation subgraphs, while sharing the trainable parameters.
def lenet2(data, train = False):
"""LeNet2 definition."""
# 2D convolution, with 'SAME' padding (i.e. the output feature map has
# the same size as the input). Note that {strides} is a 4D array whose
# shape matches the data layout: [n, h, w, c].
conv1 = tf.nn.conv2d(data,
conv1_weights,
strides = [1, 1, 1, 1],
padding = 'SAME')
# Bias and rectified linear non-linearity.
relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases))
# Max pooling. The kernel size spec {ksize} also follows the layout of
# the data. Here we have a pooling window of 2, and a stride of 2.
pool1 = tf.nn.max_pool(relu1,
ksize = [1, 2, 2, 1],
strides = [1, 2, 2, 1],
padding = 'SAME')
conv2 = tf.nn.conv2d(pool1,
conv2_weights,
strides = [1, 1, 1, 1],
padding = 'SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))
pool2 = tf.nn.max_pool(relu2,
ksize = [1, 2, 2, 1],
strides = [1, 2, 2, 1],
padding = 'SAME')
# Reshape the feature map cuboid into a 2D matrix to feed it to the
# fully connected layers.
pool_shape = pool2.get_shape().as_list()
reshape = tf.reshape(pool2,
[pool_shape[0], pool_shape[1] * pool_shape[2] * pool_shape[3]])
# Fully connected layer. Note that the '+' operation automatically
# broadcasts the biases.
fc1 = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases)
# Add a 50% dropout during training only. Dropout also scales
# activations such that no rescaling is needed at evaluation time.
if train:
fc1 = tf.nn.dropout(fc1, 0.5, seed = SEED)
return tf.matmul(fc1, fc2_weights) + fc2_biases
# Training computation: logits + cross-entropy loss.
logits = lenet2(train_data_node, True)
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
logits, train_labels_node))
# L2 regularization for the fully connected parameters.
regularizers = (tf.nn.l2_loss(fc1_weights) + tf.nn.l2_loss(fc1_biases) +
tf.nn.l2_loss(fc2_weights) + tf.nn.l2_loss(fc2_biases))
# Add the regularization term to the loss.
loss += 5e-4 * regularizers
# Optimizer: set up a variable that's incremented once per batch and
# controls the learning rate decay.
batch = tf.Variable(0)
# Decay once per epoch, using an exponential schedule starting at 0.01.
learning_rate = tf.train.exponential_decay(
0.01, # Base learning rate.
batch * BATCH_SIZE, # Current index into the dataset.
train_size, # Decay step.
0.95, # Decay rate.
staircase=True)
# Use simple momentum for the optimization.
optimizer = tf.train.MomentumOptimizer(learning_rate,
0.9).minimize(loss,
global_step = batch)
# Predictions for the current training minibatch.
train_prediction = tf.nn.softmax(logits)
# Predictions for the test and validation, which we'll compute less often.
eval_prediction = tf.nn.softmax(lenet2(eval_data))
# Small utility function to evaluate a dataset by feeding batches of data to
# {eval_data} and pulling the results from {eval_predictions}.
# Saves memory and enables this to run on smaller GPUs.
def eval_in_batches(data, sess):
"""Get all predictions for a dataset by running it in small batches."""
size = data.shape[0]
if size < EVAL_BATCH_SIZE:
raise ValueError("batch size for evals larger than dataset: %d" % size)
predictions = np.ndarray(shape = (size, num_labels), dtype = np.float32)
for begin in xrange(0, size, EVAL_BATCH_SIZE):
end = begin + EVAL_BATCH_SIZE
if end <= size:
predictions[begin:end, :] = sess.run(
eval_prediction,
feed_dict={eval_data: data[begin:end, ...]})
else:
batch_predictions = sess.run(
eval_prediction,
feed_dict={eval_data: data[-EVAL_BATCH_SIZE:, ...]})
predictions[begin:, :] = batch_predictions[begin - size:, :]
return predictions
# Create a local session to run the training.
start_time = time.time()
model_dir = os.path.join(module_dir, os.path.pardir, os.path.pardir, 'models')
with tf.Session() as sess:
# Run all the initializers to prepare the trainable parameters.
tf.initialize_all_variables().run()
# Import base model weights
saver = tf.train.Saver([conv1_weights, conv1_biases, conv2_weights, conv2_biases, fc1_weights, fc1_biases])
ckpt = tf.train.get_checkpoint_state(os.path.join(model_dir, 'base'))
if ckpt and ckpt.model_checkpoint_path:
logger.info("Continue training from the model {}".format(ckpt.model_checkpoint_path))
saver.restore(sess, ckpt.model_checkpoint_path)
# for var in tf.trainable_variables():
# logger.info(var.eval())
logger.info('Initialized!')
# Loop through training steps.
for step in xrange(int(num_epochs * train_size) // BATCH_SIZE):
# Compute the offset of the current minibatch in the data.
# Note that we could use better randomization across epochs.
offset = (step * BATCH_SIZE) % (train_size - BATCH_SIZE)
batch_data = train_data[offset:(offset + BATCH_SIZE), ...]
batch_labels = train_labels[offset:(offset + BATCH_SIZE)]
# This dictionary maps the batch data (as a numpy array) to the
# node in the graph it should be fed to.
feed_dict = {train_data_node: batch_data,
train_labels_node: batch_labels}
# Run the graph and fetch some of the nodes.
_, l, lr, predictions = sess.run(
[optimizer, loss, learning_rate, train_prediction],
feed_dict=feed_dict)
if step % EVAL_FREQUENCY == 0:
elapsed_time = time.time() - start_time
start_time = time.time()
logger.info('Step %d (epoch %.2f), %.1f ms' %
(step, float(step) * BATCH_SIZE / train_size,
1000 * elapsed_time / EVAL_FREQUENCY))
logger.info('Minibatch loss: %.3f, learning rate: %.6f' % (l, lr))
logger.info('Minibatch error: %.1f%%' % error_rate(predictions, batch_labels))
logger.info('Validation error: %.1f%%' % error_rate(
eval_in_batches(validation_data, sess), validation_labels))
sys.stdout.flush()
# Finally print the result!
test_precision = precision(eval_in_batches(test_data, sess), test_labels)
logger.info('Test precision: %.1f%%' % test_precision)
# Model persistence
saver = tf.train.Saver([conv1_weights, conv1_biases, conv2_weights, conv2_biases, fc1_weights, fc1_biases, fc2_weights, fc2_biases])
model_path = os.path.join(model_dir, "finetuned", "lenet_finetuned.ckpt")
save_path = saver.save(sess, model_path)
logger.info("Model saved in file: %s" % save_path)
def main(argv = None):
# load mnist into numpy arrays.
train_data, train_labels = load_minst_data(t = 'train')
test_data, test_labels = load_minst_data(t = 'test')
height = train_data.shape[1]
width = train_data.shape[2]
channel = (train_data.shape[3] if train_data.ndim > 3 else 1)
label_max = np.amax(train_labels)
label_min = np.amin(train_labels)
num_labels = label_max - label_min + 1
# Generate a validation set.
validation_data = train_data[:VALIDATION_SIZE, ...]
validation_labels = train_labels[:VALIDATION_SIZE]
train_data = train_data[VALIDATION_SIZE:, ...]
train_labels = train_labels[VALIDATION_SIZE:]
num_epochs = NUM_EPOCHS
train_size = train_labels.shape[0]
# This is where training samples and labels are fed to the graph.
# These placeholder nodes will be fed a batch of training data at each
# training step using the {feed_dict} argument to the Run() call below.
train_data_node = tf.placeholder(
tf.float32,
shape = (BATCH_SIZE, height, width, channel))
train_labels_node = tf.placeholder(tf.int64, shape = (BATCH_SIZE,))
eval_data = tf.placeholder(
tf.float32,
shape = (EVAL_BATCH_SIZE, height, width, channel))
# The variables below hold all the trainable weights. They are passed an
# initial value which will be assigned when we call:
# {tf.initialize_all_variables().run()}
conv1_weights = tf.Variable(
tf.truncated_normal([5, 5, channel, 32], # 5x5 filter, depth 32.
stddev = 0.1,
seed = SEED),
name = "conv1_weights")
conv1_biases = tf.Variable(tf.zeros([32]), name = "conv1_biases")
conv2_weights = tf.Variable(
tf.truncated_normal([5, 5, 32, 64],
stddev = 0.1,
seed = SEED),
name = "conv2_weights")
conv2_biases = tf.Variable(tf.constant(0.1, shape = [64]), name = "conv2_biases")
fc1_weights = tf.Variable( # fully connected, depth 512.
tf.truncated_normal(
[height // 4 * width // 4 * 64, 512],
stddev = 0.1,
seed = SEED),
name = "fc1_weights")
fc1_biases = tf.Variable(tf.constant(0.1, shape = [512]), name = "fc1_biases")
fc2_weights = tf.Variable(
tf.truncated_normal([512, num_labels],
stddev = 0.1,
seed = SEED),
name = "fc2_weights")
fc2_biases = tf.Variable(tf.constant(0.1, shape = [num_labels]), name = "fc2_biases")
# We will replicate the model structure for the training subgraph, as well
# as the evaluation subgraphs, while sharing the trainable parameters.
def lenet(data, train = False):
"""LeNet definition."""
# 2D convolution, with 'SAME' padding (i.e. the output feature map has
# the same size as the input). Note that {strides} is a 4D array whose
# shape matches the data layout: [n, h, w, c].
conv1 = tf.nn.conv2d(data,
conv1_weights,
strides = [1, 1, 1, 1],
padding = 'SAME')
# Bias and rectified linear non-linearity.
relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases))
# Max pooling. The kernel size spec {ksize} also follows the layout of
# the data. Here we have a pooling window of 2, and a stride of 2.
pool1 = tf.nn.max_pool(relu1,
ksize = [1, 2, 2, 1],
strides = [1, 2, 2, 1],
padding = 'SAME')
conv2 = tf.nn.conv2d(pool1,
conv2_weights,
strides = [1, 1, 1, 1],
padding = 'SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))
pool2 = tf.nn.max_pool(relu2,
ksize = [1, 2, 2, 1],
strides = [1, 2, 2, 1],
padding = 'SAME')
# Reshape the feature map cuboid into a 2D matrix to feed it to the
# fully connected layers.
pool_shape = pool2.get_shape().as_list()
reshape = tf.reshape(pool2,
[pool_shape[0], pool_shape[1] * pool_shape[2] * pool_shape[3]])
# Fully connected layer. Note that the '+' operation automatically
# broadcasts the biases.
fc1 = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases)
# Add a 50% dropout during training only. Dropout also scales
# activations such that no rescaling is needed at evaluation time.
if train:
fc1 = tf.nn.dropout(fc1, 0.5, seed = SEED)
return tf.matmul(fc1, fc2_weights) + fc2_biases
# Training computation: logits + cross-entropy loss.
logits = lenet(train_data_node, True)
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits, train_labels_node))
# L2 regularization for the fully connected parameters.
regularizers = (tf.nn.l2_loss(fc1_weights) + tf.nn.l2_loss(fc1_biases) +
tf.nn.l2_loss(fc2_weights) + tf.nn.l2_loss(fc2_biases))
# Add the regularization term to the loss.
loss += 5e-4 * regularizers
# Optimizer: set up a variable that's incremented once per batch and
# controls the learning rate decay.
batch = tf.Variable(0)
# Decay once per epoch, using an exponential schedule starting at 0.01.
learning_rate = tf.train.exponential_decay(
0.01, # Base learning rate.
batch * BATCH_SIZE, # Current index into the dataset.
train_size, # Decay step.
0.95, # Decay rate.
staircase = True)
# Use simple momentum for the optimization.
optimizer = tf.train.MomentumOptimizer(learning_rate,
0.9).minimize(loss,
global_step = batch)
# Predictions for the current training minibatch.
train_prediction = tf.nn.softmax(logits)
# Predictions for the test and validation, which we'll compute less often.
eval_prediction = tf.nn.softmax(lenet(eval_data))
# Small utility function to evaluate a dataset by feeding batches of data to
# {eval_data} and pulling the results from {eval_predictions}.
# Saves memory and enables this to run on smaller GPUs.
def eval_in_batches(data, sess):
"""Get all predictions for a dataset by running it in small batches."""
size = data.shape[0]
if size < EVAL_BATCH_SIZE:
raise ValueError("batch size for evals larger than dataset: %d" % size)
predictions = np.ndarray(shape = (size, num_labels), dtype = np.float32)
for begin in xrange(0, size, EVAL_BATCH_SIZE):
end = begin + EVAL_BATCH_SIZE
if end <= size:
predictions[begin:end, :] = sess.run(
eval_prediction,
feed_dict={eval_data: data[begin:end, ...]})
else:
batch_predictions = sess.run(
eval_prediction,
feed_dict={eval_data: data[-EVAL_BATCH_SIZE:, ...]})
predictions[begin:, :] = batch_predictions[begin - size:, :]
return predictions
# Create a local session to run the training.
start_time = time.time()
model_dir = os.path.join(module_dir, os.path.pardir, os.path.pardir, 'models')
with tf.Session() as sess:
# Run all the initializers to prepare the trainable parameters.
tf.initialize_all_variables().run()
logger.info('Initialized!')
# Loop through training steps.
for step in xrange(int(num_epochs * train_size) // BATCH_SIZE):
# Compute the offset of the current minibatch in the data.
# Note that we could use better randomization across epochs.
offset = (step * BATCH_SIZE) % (train_size - BATCH_SIZE)
batch_data = train_data[offset:(offset + BATCH_SIZE), ...]
batch_labels = train_labels[offset:(offset + BATCH_SIZE)]
# This dictionary maps the batch data (as a numpy array) to the
# node in the graph it should be fed to.
feed_dict = {train_data_node: batch_data,
train_labels_node: batch_labels}
# Run the graph and fetch some of the nodes.
_, l, lr, predictions = sess.run(
[optimizer, loss, learning_rate, train_prediction],
feed_dict=feed_dict)
if step % EVAL_FREQUENCY == 0:
elapsed_time = time.time() - start_time
start_time = time.time()
logger.info('Step %d (epoch %.2f), %.1f ms' %
(step, float(step) * BATCH_SIZE / train_size,
1000 * elapsed_time / EVAL_FREQUENCY))
logger.info('Minibatch loss: %.3f, learning rate: %.6f' % (l, lr))
logger.info('Minibatch training error: %.1f%%' % error_rate(predictions, batch_labels))
logger.info('Validation error: %.1f%%' % error_rate(eval_in_batches(validation_data, sess), validation_labels))
sys.stdout.flush()
# Finally print the result!
test_precision = precision(eval_in_batches(test_data, sess), test_labels)
logger.info('Test precision: %.1f%%' % test_precision)
# Model persistence
saver = tf.train.Saver([conv1_weights, conv1_biases, conv2_weights, conv2_biases, fc1_weights, fc1_biases])
model_path = os.path.join(model_dir, "base", "lenet_base.ckpt")
save_path = saver.save(sess, model_path)
logger.info("Model saved in file: %s" % save_path)
