def train_edge(conv_weight_decay=REGULARIZATION_STRENGTH):
training_queue = ShuffleQueue(cleaneval_train)
data_shape = [BATCH_SIZE, PATCH_SIZE - 1, 1, N_EDGE_FEATURES]
labs_shape = [BATCH_SIZE, PATCH_SIZE - 1, 1, 1]
train_features = tf.placeholder(tf.float32, shape=data_shape)
train_labels = tf.placeholder(tf.int64, shape=labs_shape)
logits = edge(train_features,
is_training=True,
conv_weight_decay=conv_weight_decay,
dropout_keep_prob=DROPOUT_KEEP_PROB)
l = loss(tf.reshape(logits, [-1, 4]), tf.reshape(train_labels, [-1]))
train_op = tf.train.AdamOptimizer(LEARNING_RATE).minimize(l)
test_features = tf.placeholder(tf.float32)
tf.get_variable_scope().reuse_variables()
test_logits = edge(test_features, is_training=False)
saver = tf.train.Saver(tf.get_collection(EDGE_VARIABLES))
init_op = tf.global_variables_initializer()
with tf.Session() as session:
# Initialize
session.run(init_op)
def prediction(features):
features = features[np.newaxis, :, np.newaxis, :]
logits = session.run(test_logits,
feed_dict={test_features: features})
return np.argmax(logits, axis=-1).flatten()
BEST_VAL_SO_FAR = 0
for step in range(TRAIN_STEPS + 1):
# Construct a bs-length numpy array
_, edge_features, labels, edge_labels = get_batch(training_queue)
# Run a training step
loss_val, _ = session.run([l, train_op],
feed_dict={
train_features: edge_features,
train_labels: edge_labels
})
if step % 100 == 0:
accuracy_validation = evaluate_edge(cleaneval_validation,
prediction)
accuracy_train = evaluate_edge(cleaneval_train, prediction)
if accuracy_validation > BEST_VAL_SO_FAR:
best = True
saver.save(session,
os.path.join(CHECKPOINT_DIR, 'edge.ckpt'))
BEST_VAL_SO_FAR = accuracy_validation
else:
best = False
print("%10d: train=%.4f, val=%.4f %s" %
(step, accuracy_train, accuracy_validation,
'*' if best else ''))
# saver.save(session, os.path.join(CHECKPOINT_DIR, 'edge.ckpt'))
return accuracy_validation
def test_structured(lamb=EDGE_LAMBDA):
from data import cleaneval_test, cleaneval_train, cleaneval_validation
unary_features = tf.placeholder(tf.float32)
edge_features = tf.placeholder(tf.float32)
# hack to get the right shape weights
_ = unary(tf.placeholder(tf.float32, shape=[1,PATCH_SIZE,1,N_FEATURES]), False)
_ = edge(tf.placeholder(tf.float32, shape=[1,PATCH_SIZE,1,N_EDGE_FEATURES]), False)
tf.get_variable_scope().reuse_variables()
unary_logits = unary(unary_features, is_training=False)
edge_logits = edge(edge_features, is_training=False)
unary_saver = tf.train.Saver(tf.get_collection(UNARY_VARIABLES))
edge_saver = tf.train.Saver(tf.get_collection(EDGE_VARIABLES))
init_op = tf.global_variables_initializer()
with tf.Session() as session:
session.run(init_op)
unary_saver.restore(session, os.path.join(CHECKPOINT_DIR, "unary.ckpt"))
edge_saver.restore(session, os.path.join(CHECKPOINT_DIR, "edge.ckpt"))
from time import time
start = time()
def prediction_structured(features, edge_feat):
features = features[np.newaxis, :, np.newaxis, :]
edge_feat = edge_feat[np.newaxis, :, np.newaxis, :]
unary_lgts = session.run(unary_logits, feed_dict={unary_features: features})
edge_lgts = session.run(edge_logits, feed_dict={edge_features: edge_feat})
return viterbi(unary_lgts.reshape([-1,2]), edge_lgts.reshape([-1,4]), lam=lamb)
def prediction_unary(features, _):
features = features[np.newaxis, :, np.newaxis, :]
logits = session.run(unary_logits, feed_dict={unary_features: features})
return np.argmax(logits, axis=-1).flatten()
accuracy, precision, recall, f1 = evaluate_unary(cleaneval_test, prediction_structured)
accuracy_u, precision_u, recall_u, f1_u = evaluate_unary(cleaneval_test, prediction_unary)
end = time()
print('duration', end-start)
print('size', len(cleaneval_test))
print("Structured: Accuracy=%.5f, precision=%.5f, recall=%.5f, F1=%.5f" % (accuracy, precision, recall, f1))
print("Just unary: Accuracy=%.5f, precision=%.5f, recall=%.5f, F1=%.5f" % (accuracy_u, precision_u, recall_u, f1_u))
def classify(block_features_file,
edge_features_file,
labels_output_file,
lamb=EDGE_LAMBDA):
block_features = np.genfromtxt(block_features_file, delimiter=',')
edge_features = np.genfromtxt(edge_features_file, delimiter=',')
# Reshape
block_features = block_features.T[np.newaxis, :,
np.newaxis, :].astype(np.float32)
edge_features = edge_features.T[np.newaxis, :,
np.newaxis, :].astype(np.float32)
unary_features = tf.constant(block_features)
edge_features = tf.constant(edge_features)
unary_logits = unary(unary_features, is_training=False)
edge_logits = edge(edge_features, is_training=False)
unary_saver = tf.train.Saver(tf.get_collection(UNARY_VARIABLES))
edge_saver = tf.train.Saver(tf.get_collection(EDGE_VARIABLES))
init_op = tf.global_variables_initializer()
with tf.Session() as session:
session.run(init_op)
unary_saver.restore(session, os.path.join(CHECKPOINT_DIR,
"unary.ckpt"))
edge_saver.restore(session, os.path.join(CHECKPOINT_DIR, "edge.ckpt"))
from time import time
start = time()
unary_lgts = session.run(unary_logits)
edge_lgts = session.run(edge_logits)
labels = viterbi(unary_lgts.reshape([-1, 2]),
edge_lgts.reshape([-1, 4]),
lam=lamb).astype(np.int32)
duration = time() - start
print("Done. Classification took %.2f seconds " % duration)
with open(labels_output_file, 'w') as fp:
fp.write(",".join('%d' % label for label in labels))
print('CSV labels written to %s' % labels_output_file)
