def main():
savepath = './save_point'
filepath = './save_point/keras_example_checkpoint.h5'
# Extract MNIST dataset
train_data_filename = maybe_download('train-images-idx3-ubyte.gz')
train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz')
test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz')
test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz')
train_data = extract_data(train_data_filename, 60000, dense=False)
train_data = train_data.reshape((60000, NUM_CHANNELS, IMG_SIZE, IMG_SIZE))
train_labels = extract_labels(train_labels_filename, 60000, one_hot=True)
test_data = extract_data(test_data_filename, 10000, dense=False)
test_data = test_data.reshape((10000, NUM_CHANNELS, IMG_SIZE, IMG_SIZE))
test_labels = extract_labels(test_labels_filename, 10000, one_hot=True)
validation_data = train_data[:VALIDATION_SIZE, ...]
validation_labels = train_labels[:VALIDATION_SIZE, :]
validation_set = (validation_data, validation_labels)
train_data = train_data[VALIDATION_SIZE:, ...]
train_labels = train_labels[VALIDATION_SIZE:, ...]
# Model construction
model = Sequential()
model.add(Convolution2D(32, 3, 3, border_mode='same',
input_shape=(1, 28, 28)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Convolution2D(64, 3, 3, border_mode='same'))
model.add(Flatten())
model.add(Dense(256))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(10))
model.add(Activation('softmax'))
# Define optimizer and configure training process
sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9)
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=["accuracy"])
model.fit(
train_data,
train_labels,
nb_epoch=NUM_EPOCHS,
batch_size=1000,
validation_data=validation_set)
print 'Save model weights'
if not os.path.isdir (savepath):
os.mkdir (savepath)
model.save_weights(filepath, overwrite=True)
predict = model.predict(test_data, batch_size=1000)
print 'Test err: %.1f%%' % error_rate(predict, test_labels)
print 'Test loss: %1.f%%, accuracy: %1.f%%', \
tuple(model.evaluate(test_data, test_labels, batch_size=1000))
def get_label_count(dataset):
"""Computes the label count dictionary: labels to their occurrence counts
inside the dataset.
Args:
dataset (list(dict(str))): A dataset, i.e., a list of json objects fitting
the Golden Dataset Schema.
alpha (float): Alpha parameter in balance-aware data augmentation.
beta (float): Beta parameter in balance-aware data augmentation.
Returns:
dict(str): Labels mapped to the amount of times they occur in the dataset.
"""
label_count = defaultdict(int)
for label_list in [extract_labels(js) for js in dataset]:
for label in label_list:
label_count[label] += 1
return label_count
def rareness_score(paper, label_count, alpha, beta):
'''
Computes a rareness score for a json paper.
/ sum_{label in paper_labels} (1 / count(label)) ^ ALPHA \ ^ BETA
Score = | ------------------------------------------------------ |
\ number of labels in paper /
Args:
paper (dict): a json dictionary, a paper from the golden dataset (or formatted as such)
label_count (dict(str)): Labels mapped to the amount of times they occur in the dataset.
alpha (float): Alpha parameter in balance-aware data augmentation.
beta (float): Beta parameter in balance-aware data augmentation.
Returns:
score (float): computed rareness score
'''
paper_labels = extract_labels(paper)
if not paper_labels:
return 0.0
score = 0.0
for paper_label in paper_labels:
score += (1 / label_count[paper_label])**alpha
return (score / len(paper_labels))**beta
def load_data(dataset_path, data_out, batch_size):
face_detector = ort.InferenceSession('models/version-RFB-320.onnx')
face_detector_input = face_detector.get_inputs()[0].name
paths = pathlib.Path(dataset_path).glob('*.jpg')
paths = sorted([x for x in paths])
random.shuffle(paths)
faces = []
bmis = []
if os.path.exists(data_out):
shutil.rmtree(data_out)
os.mkdir(data_out)
current_batch = 0
for image_path in paths:
face = extract_face(image_path, face_detector, face_detector_input)
if face is None:
continue
faces.append(face)
bmi = extract_labels(image_path)
bmis.append(bmi)
if len(faces) == batch_size:
write_batch(current_batch, faces, bmis, data_out)
current_batch += 1
faces.clear()
bmis.clear()
# Write any remaining data to the batch.
if len(faces) > 0:
write_batch(current_batch, faces, bmis, data_out)
def split(prefix='MATCH/PeTaL',
dataset='cleaned_lens_output.json',
train=0.8,
dev=0.1,
skip=0,
tot=0,
infer_mode=False,
verbose=False):
"""Performs train-test split on newline-delimited json file.
Args:
prefix (string): Path from current working directory to directory containing dataset.
dataset (string): Filename of newline-delimited json dataset.
train (float): Proportion, from 0.0 to 1.0, of dataset used for training.
dev (float): Proportion, from 0.0 to 1.0, of dataset used for validation.
skip (int): Number of training examples by which to rotate the dataset (e.g., for cross-validation).
tot (int): Total number of training examples.
infer_mode (bool): Whether to run in inference mode.
verbose (bool): Verbose output.
"""
logging.basicConfig(level=logging.DEBUG,
format="[%(asctime)s:%(name)s] %(message)s")
logger = logging.getLogger("Split")
########################################
#
# INPUT VALIDATION CHECKS
#
########################################
dataset_path = os.path.join(prefix, dataset)
if not os.path.exists(dataset_path):
logger.error(
f"ERROR: Unable to find dataset json file {os.path.join(os.getcwd(), dataset_path)}."
)
return
if train < 0:
logger.error(f"ERROR: train proportion {train} is less than 0.")
return
elif dev < 0:
logger.error(f"ERROR: dev proportion {dev} is less than 0.")
return
elif train + dev >= 1:
logger.error(
f"ERROR: train proportion {train} + dev proportion {dev} exceeds 1."
)
return
elif tot < 0:
logger.error(f"ERROR: total parameter (tot) {tot} is less than 0.")
return
########################################
#
# INFERENCE MODE
# just copy the whole dataset into test.json
#
########################################
if infer_mode:
infer_path = os.path.join(prefix, 'test.json')
if verbose:
logger.info(
f"Copying from {dataset_path} to {infer_path} for inference mode."
)
with open(dataset_path) as fin, open(infer_path, 'w') as fout:
golden = json.loads(fin.read())
for js in golden:
fout.write(json.dumps(js) + '\n')
########################################
#
# NOT INFERENCE MODE:
# do train-validation-test split
#
########################################
else:
train_path = os.path.join(prefix, 'train.json')
dev_path = os.path.join(prefix, 'dev.json')
test_path = os.path.join(prefix, 'test.json')
if verbose:
logger.info(
f"Transforming from {dataset_path} to {train_path}, {dev_path}, and {test_path}."
)
train_proportion, dev_proportion = train, dev
train_labels = set()
########################################
# If the default value 0 was passed in as tot,
# count the number of examples in the dataset
# and use that as the total number of examples.
# Otherwise stick with the passed-in total argument,
# unless that exceeds the actual number of examples in the dataset.
########################################
with open(dataset_path) as fin:
num_examples = sum(1 for _ in json.loads(fin.read()))
if tot == 0 or tot > num_examples:
tot = num_examples
if verbose:
logger.info(f"{tot} total examples in dataset.")
########################################
# This part figures out what labels appear in the training set
# and then filters out all other labels from the dev and testing sets
# while constructing train.json, dev.json, and test.json.
########################################
with open(dataset_path) as fin, open(train_path, 'w') as fou1, open(
dev_path, 'w') as fou2, open(test_path, 'w') as fou3:
golden = json.loads(fin.read())
# for idx, line in enumerate(fin):
for idx, js in enumerate(golden):
if idx < skip or idx >= tot:
continue
# js = json.loads(line)
######################################## moved to utils.extract_labels
# level1Labels = js['level1'] if js['level1'] else []
# level2Labels = js['level2'] if js['level2'] else []
# level3Labels = js['level3'] if js['level3'] else []
# all_labels = level1Labels + level2Labels + level3Labels
########################################
all_labels = extract_labels(js)
if (idx - skip) % tot < tot * train_proportion:
for l in all_labels:
train_labels.add(l)
js['label'] = all_labels
fou1.write(json.dumps(js) + '\n')
else:
label_new = []
for l in all_labels:
if l in train_labels:
label_new.append(l)
if len(label_new) == 0:
continue
js['label'] = label_new
# print(js['label'])
# print([x for x in js.keys()])
if (idx - skip) % tot < tot * (train_proportion +
dev_proportion):
fou2.write(json.dumps(js) + '\n')
else:
fou3.write(json.dumps(js) + '\n')
# fin.seek(0)
# for idx, line in enumerate(fin):
for idx, js in enumerate(golden):
if idx >= skip:
break
# js = json.loads(line)
######################################## moved to utils.extract_labels
# level1Labels = js['level1'] if js['level1'] else []
# level2Labels = js['level2'] if js['level2'] else []
# level3Labels = js['level3'] if js['level3'] else []
# all_labels = level1Labels + level2Labels + level3Labels
########################################
all_labels = extract_labels(js)
if (idx - skip) % tot < tot * train_proportion:
for l in all_labels:
train_labels.add(l)
js['label'] = all_labels
fou1.write(json.dumps(js) + '\n')
else:
label_new = []
for l in all_labels:
if l in train_labels:
label_new.append(l)
if len(label_new) == 0:
continue
js['label'] = label_new
if (idx - skip) % tot < tot * (train_proportion +
dev_proportion):
fou2.write(json.dumps(js) + '\n')
else:
fou3.write(json.dumps(js) + '\n')
if verbose:
logger.info(f"Number of train labels: {len(train_labels)}")
'''
for vector in direction_vectors
])
print X.shape
y = np.concatenate([y for _ in range(len(direction_vectors) + 1)], axis=0)
print y.shape
return X, y
# Extract data
train_data_filename = maybe_download('train-images-idx3-ubyte.gz')
train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz')
test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz')
test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz')
X_train = extract_data(train_data_filename, 60000, dense=True)
y_train = extract_labels(train_labels_filename, 60000, one_hot=False)
X_test = extract_data(test_data_filename, 10000, dense=True)
y_test = extract_labels(test_labels_filename, 10000, one_hot=False)
#################################################
# Test for decision tree classifier without dimensionality reduction
Tree = DecisionTreeClassifier()
Tree.fit(X_train, y_train)
print 'Without dimenstionality reduction: ', Tree.score(X_test, y_test)
# Dimensionality reduction using PCA (784 -> 64)
pca = PCA(n_components=64)
pca.fit(X_train)
X_train_reduce = pca.transform(X_train)
Tree_2 = DecisionTreeClassifier()
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Parser")
parser.add_argument("--image_path",
type=str,
required=True,
help="path to image")
parser.add_argument("--segmentation_path",
type=str,
required=True,
help="path to semantic segmentation")
parser.add_argument("--label_path",
type=str,
required=True,
help="path to text file with label names")
args = parser.parse_args()
image = load_image(args.image_path)
semseg = load_segmentation(args.segmentation_path)
labels = extract_labels(args.label_path)
print(f"+ Segmentation shape: {semseg.shape}")
# Extract image segments from semantic segmentation
segments = extract_segments(semseg)
print(f"+ Number of segments: {len(segments)}")
# Trigger GUI and visualize segments
visualize(image, segments, labels)
[np.apply_along_axis(shift, 1, X, vector)
for vector in direction_vectors])
print X.shape
y = np.concatenate([y for _ in range(len(direction_vectors) + 1)], axis=0)
print y.shape
return X, y
# Extract data
train_data_filename = maybe_download('train-images-idx3-ubyte.gz')
train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz')
test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz')
test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz')
X_train = extract_data(train_data_filename, 60000, dense=True)
y_train = extract_labels(train_labels_filename, 60000, one_hot=False)
X_test = extract_data(test_data_filename, 10000, dense=True)
y_test = extract_labels(test_labels_filename, 10000, one_hot=False)
#################################################
# Test for decision tree classifier without dimensionality reduction
Tree = DecisionTreeClassifier()
Tree.fit(X_train, y_train)
print 'Without dimenstionality reduction: ', Tree.score(X_test, y_test)
# Dimensionality reduction using PCA (784 -> 64)
pca = PCA(n_components=64)
pca.fit(X_train)
X_train_reduce = pca.transform(X_train)
def main():
if (len(sys.argv) != 5):
sys.exit("invalid command-line arguments format")
# handling command-line arguments
data = DataReader2(sys.argv[1])
data.init_examples()
training_set_size = int(sys.argv[2])
num_trials = int(sys.argv[3])
verbose = int(sys.argv[4])
if (verbose != 1 and verbose != 0):
sys.exit("invalid command-line argument")
if (num_trials < 1):
sys.exit("invalid command-line argument")
# extract examples and attributes
# an example = a feature vector + a label (represented by a tuple)
examples = data.get_examples() # a list of examples
attributes = data.get_attributes() # a list of attribute names
if (training_set_size >= len(examples)):
sys.exit("invalid command-line argument")
# lists of classification performances (correct rates)
# e.x.: [1.0, 0.95, 0.83, ...]
correct_rates_id3 = []
correct_rates_prior = []
for i in range(0, num_trials): # a single trial
print 'TRIAL NUMBER:', i + 1
print '-' * 30
# randomly pick a training set of size *training_set_size*
random.shuffle(examples)
training_examples = examples[0:training_set_size]
testing_examples = examples[training_set_size:]
# a list of actual labels of testing examples
actuals = utils.extract_labels(testing_examples)
# build a decision tree based on these training examples
tree = id3.DTL(training_examples, range(0, len(attributes)), True)
# print the structure of the decision tree built from the training set
print 'DECISION TREE STRUCTURE'
utils.print_tree(tree, attributes)
# list of predicted labels using id3
output_id3_1 = utils.trial_id3(tree, testing_examples)
# list of predicted labels using prior probability
output_prior_1 = utils.trial_priorprob(training_examples,
testing_examples)
# computes and prints correct rate of this trial
correct_rate_id3 = utils.correct_rate(output_id3_1, actuals)
correct_rates_id3.append(correct_rate_id3)
correct_rate_prior = utils.correct_rate(output_prior_1, actuals)
correct_rates_prior.append(correct_rate_prior)
print '\n'
print 'proportion of correct classification'
print 'decision tree:', correct_rate_id3
print 'prior probability:', correct_rate_prior
print '\n'
if (verbose == 1):
output_id3_2 = list(testing_examples)
output_prior_2 = list(testing_examples)
for j in range(0, len(output_id3_2)):
output_id3_2[j][-1] = output_id3_1[j]
output_prior_2[j][-1] = output_prior_1[j]
print '*' * 10, 'examples in the training set: ', '*' * 10
utils.print_dataset(training_examples, attributes)
print '*' * 10, 'examples in the testing set: ', '*' * 10
utils.print_dataset(testing_examples, attributes)
print '*' * 10, 'classification by the decision tree: ', '*' * 10
utils.print_dataset(output_id3_2, attributes)
print '*' * 10, 'classification by prior probability: ', '*' * 10
utils.print_dataset(output_prior_2, attributes)
# other outputs
print '*' * 5, 'information', '*' * 5
print 'file:' + sys.argv[1]
print 'training set size:', sys.argv[2]
print 'testing set size:', len(examples) - int(sys.argv[2])
print 'number of trials:', num_trials
mean_tree = utils.mean(correct_rates_id3)
mean_prior = utils.mean(correct_rates_prior)
print 'mean classification performance (decision tree):', mean_tree
print 'mean classification performance (prior probability):', mean_prior
def main(argv=None): # pylint: disable=unused-argument
# Get the data.
train_data_filename = maybe_download('train-images-idx3-ubyte.gz')
train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz')
test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz')
test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz')
# Extract it into numpy arrays.
train_data = extract_data(train_data_filename, 60000, dense=False)
train_labels = extract_labels(train_labels_filename, 60000, one_hot=True)
test_data = extract_data(test_data_filename, 10000, dense=False )
test_labels = extract_labels(test_labels_filename, 10000, one_hot=True)
# 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, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
train_labels_node = tf.placeholder(tf.float32,
shape=(BATCH_SIZE, NUM_LABELS))
eval_data = tf.placeholder(
tf.float32,
shape=(EVAL_BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
# The variables below hold all the trainable weights. They are passed an
# initial value which will be assigned when when we call:
# {tf.initialize_all_variables().run()}
# First convolutional layer
conv1_weights = tf.Variable(
tf.truncated_normal([3, 3, NUM_CHANNELS, 32], # 5x5 filter, depth 32.
stddev=0.1,
seed=SEED))
conv1_biases = tf.Variable(tf.zeros([32]))
# Two second convolutional layers 5 x 5 filter, and 3 x 3 filters.
conv2_weights = tf.Variable(
tf.truncated_normal([5, 5, 32, 64],
stddev=0.1,
seed=SEED))
conv2_biases = tf.Variable(tf.constant(0.01, shape=[64]))
conv2_weights2 = tf.Variable(
tf.truncated_normal([3, 3, 32, 64],
stddev=0.1,
seed=SEED))
conv2_biases2 = tf.Variable(tf.constant(0.01, shape=[64]))
# First fully connected layer after conv layer
fc1_weights = tf.Variable( # fully connected, depth 512.
tf.truncated_normal(
[IMAGE_SIZE // 4 * IMAGE_SIZE // 4 * 128, 512],
stddev=0.05,
seed=SEED))
fc1_biases = tf.Variable(tf.constant(0.01, shape=[512]))
# Second fully connected layer
fc2_weights = tf.Variable(
tf.truncated_normal([512, 256],
stddev=0.05,
seed=SEED))
fc2_biases = tf.Variable(tf.constant(0.1, shape=[256]))
# Output layer
fc3_weights = tf.Variable(
tf.truncated_normal([256, NUM_LABELS],
stddev=0.04,
seed=SEED))
fc3_biases = tf.Variable(tf.constant(0.1, shape=[NUM_LABELS]))
# We will replicate the model structure for the training subgraph, as well
# as the evaluation subgraphs, while sharing the trainable parameters.
def model(data, train=False):
"""The Model 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: [image index, y, x, depth].
conv = tf.nn.conv2d(data,
conv1_weights,
strides=[1, 1, 1, 1],
padding='SAME')
# Bias and rectified linear non-linearity.
relu = tf.nn.relu(tf.nn.bias_add(conv, conv1_biases))
if train:
relu = tf.nn.dropout(relu, .5)
# 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.
pool = tf.nn.max_pool(relu,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
conv = tf.nn.conv2d(pool,
conv2_weights,
strides=[1, 1, 1, 1],
padding='SAME')
relu = tf.nn.relu(tf.nn.bias_add(conv, conv2_biases))
conv2 = tf.nn.conv2d(pool,
conv2_weights2,
strides=[1, 1, 1, 1],
padding='SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases2))
pool = tf.nn.max_pool(relu,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
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 = tf.concat(3, [pool, pool2])
pool_shape = pool.get_shape().as_list()
reshape = tf.reshape(
pool,
[pool_shape[0], pool_shape[1] * pool_shape[2] * pool_shape[3]])
# Fully connected layer. Note that the '+' operation automatically
# broadcasts the biases.
hidden = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases)
hidden = tf.nn.relu(tf.matmul(hidden, fc2_weights) + fc2_biases)
# Add a 50% dropout during training only. Dropout also scales
# activations such that no rescaling is needed at evaluation time.
if train:
hidden = tf.nn.dropout(hidden, 0.5, seed=SEED)
return tf.matmul(hidden, fc3_weights) + fc3_biases
def extract_filter (data):
conv = 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(conv, 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.
pool = tf.nn.max_pool(relu1,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME')
conv = tf.nn.conv2d(pool,
conv2_weights,
strides=[1, 1, 1, 1],
padding='SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(conv, conv2_biases))
conv2 = tf.nn.conv2d(pool,
conv2_weights2,
strides=[1, 1, 1, 1],
padding='SAME')
relu3 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases2))
return relu1, relu2, relu3
# Training computation: logits + cross-entropy loss.
logits = model(train_data_node, True)
loss = tf.reduce_mean(tf.nn.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) +
tf.nn.l2_loss(fc3_weights) + tf.nn.l2_loss(fc3_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(model(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 = numpy.ndarray(shape=(size, NUM_LABELS), dtype=numpy.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.
saver = tf.train.Saver()
start_time = time.time()
with tf.Session() as sess:
# Run all the initializers to prepare the trainable parameters.
if FLAGS.model:
saver.restore(sess, FLAGS.model) # If model exists, load it
else:
sess.run(tf.initialize_all_variables()) # If there is no model randomly initialize
if FLAGS.train:
# 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 is 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()
print('Step %d (epoch %.2f), %.1f ms' %
(step, float(step) * BATCH_SIZE / train_size,
1000 * elapsed_time / EVAL_FREQUENCY))
print('Minibatch loss: %.3f, learning rate: %.6f' % (l, lr))
print('Minibatch error: %.1f%%' % error_rate(predictions, batch_labels))
print('Validation error: %.1f%%' % error_rate(
eval_in_batches(validation_data, sess), validation_labels))
sys.stdout.flush()
# Finally print the result!
test_error = error_rate(eval_in_batches(test_data, sess), test_labels)
print('Test error: %.1f%%' % test_error)
print ('Optimization done')
print ('Save models')
if not tf.gfile.Exists("./conv_save"):
tf.gfile.MakeDirs("./conv_save")
saver_path = saver.save(sess, "./conv_save/model.ckpt")
print ('Successfully saved file: %s' % saver_path)
else: # If train flag is false, execute image extraction routine
print ("Filter extraction routine")
aa = train_data[1:2, :, :, :]
print (aa.shape)
# Run extract filter operations (conv1, conv2 and conv3 layers)
images = sess.run(extract_filter(train_data[1:2, :, :, :]))
print (images[2].shape)
plt.imshow (images[2][0, :, :, 32] * 255 + 255 / 2, cmap='gray')
# plt.imshow (images[2][0, :, :, 32], cmap='gray')
plt.show ()
# Save all outputs
for i in range (3):
filter_shape = images[i].shape
img_size = [filter_shape[1], filter_shape[2]]
print (img_size)
def main():
savepath = './save_point'
filepath = './save_point/model_api_checkpoint.h5'
train_data_filename = maybe_download('train-images-idx3-ubyte.gz')
train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz')
test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz')
test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz')
train_data = extract_data(train_data_filename, 60000, dense=False)
train_data = train_data.reshape((60000, NUM_CHANNELS, IMG_SIZE, IMG_SIZE))
train_labels = extract_labels(train_labels_filename, 60000, one_hot=True)
test_data = extract_data(test_data_filename, 10000, dense=False)
test_data = test_data.reshape((10000, NUM_CHANNELS, IMG_SIZE, IMG_SIZE))
test_labels = extract_labels(test_labels_filename, 10000, one_hot=True)
validation_data = train_data[:VALIDATION_SIZE, ...]
validation_labels = train_labels[:VALIDATION_SIZE, :]
validation_set = (validation_data, validation_labels)
train_data = train_data[VALIDATION_SIZE:, ...]
train_labels = train_labels[VALIDATION_SIZE:, ...]
img = Input(shape=(1, 28, 28))
conv1 = Convolution2D(32, 3, 3, border_mode='same')(img)
conv1 = Activation('relu')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2_1 = Convolution2D(64, 3, 3, border_mode='same')(pool1)
conv2_2 = Convolution2D(64, 5, 5, border_mode='same')(pool1)
conv2_1 = Activation('relu')(conv2_1)
conv2_2 = Activation('relu')(conv2_2)
pool2_1 = MaxPooling2D(pool_size=(2, 2))(conv2_1)
pool2_2 = MaxPooling2D(pool_size=(2, 2))(conv2_2)
dense1 = Flatten()(pool2_1)
dense2 = Flatten()(pool2_2)
dense = merge([dense1, dense2], mode='concat', concat_axis=1)
dense = Dense(512)(dense)
dense = Activation('relu')(dense)
dense = Dense(256)(dense)
dense = Activation('relu')(dense)
dense = Dense(10)(dense)
output = Activation('softmax')(dense)
model = Model(input=[img], output=[output])
sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9)
model.compile(
optimizer=sgd,
loss=['categorical_crossentropy'],
metrics=["accuracy"])
model.fit(
[train_data],
[train_labels],
nb_epoch=1,
verbose=1,
batch_size=1000,
validation_data=validation_set)
print 'Save model weights'
if not os.path.isdir (savepath):
os.mkdir (savepath)
model.save_weights(filepath, overwrite=True)
predictions = model.predict([test_data],
batch_size=1000)
print 'Test error: %.1f%%' % error_rate(predictions, test_labels)
print 'Test loss: %.14f, Test accurracy %.4f' % \
tuple(model.evaluate([test_data], [test_labels], batch_size=1000))
