def predict_single_sample():
mnist = input_data.read_data_sets('data/mnist-data', one_hot=False)
sess = tf.InteractiveSession()
# setup siamese network
siamese = inference.siamese(tf.estimator.ModeKeys.PREDICT)
saver = tf.train.Saver()
print('Restore parameters from model {}'.format(model_snapshot_path))
saver.restore(sess, save_path=model_snapshot_path)
test_images = mnist.test.images
test_labels = mnist.test.labels
print('test_images:', test_images.shape, test_images.dtype)
print('test_labels:', test_labels.shape, test_labels.dtype)
gallery_image = []
gallery_label = []
for i in range(100):
if len(gallery_image) != 10:
if test_labels[i] not in gallery_label:
gallery_label.append(test_labels[i])
gallery_image.append(test_images[i])
else:
break
# print(type(gallery_image[1]))
# gallery_image = np.array(gallery_image)
# print(i, len(gallery_label), type(gallery_label[0]))
# print(gallery_label)
# print(type(gallery_image[1]))
# print(gallery_image.shape)
# visualize.show_data(gallery_image.reshape([-1, 28, 28]), 'gallery_image')
# plt.show()
tm = test_images[121]
distance = siamese.distance.eval({
siamese.x1:
inference.format_single_sample(tm),
siamese.x2:
gallery_image
})
idn = siamese.single_sample_identity.eval({
siamese.x1:
inference.format_single_sample(tm),
siamese.x2:
gallery_image
})
print(idn, type(idn))
print('predict label:', gallery_label[idn])
print('true label:', test_labels[121])
print('distance:', distance)
print(tm.reshape([28, 28])[14, :])
plt.imshow((tm.reshape([28, 28]) * 255).astype('uint8'), cmap='gray')
plt.show()
def validate_accuracy():
mnist = input_data.read_data_sets('data/mnist-data', one_hot=False)
sess = tf.InteractiveSession()
# setup siamese network
siamese = inference.siamese(tf.estimator.ModeKeys.EVAL)
saver = tf.train.Saver()
print('Restore parameters from model {}'.format(model_snapshot_path))
saver.restore(sess, save_path=model_snapshot_path)
test_images = mnist.test.images
test_labels = mnist.test.labels
print('test_images:', test_images.shape, test_images.dtype)
print('test_labels:', test_labels.shape, test_labels.dtype)
gallery_image = []
gallery_label = []
for i in range(100):
if len(gallery_image) != 10:
if test_labels[i] not in gallery_label:
gallery_label.append(test_labels[i])
gallery_image.append(test_images[i])
else:
break
correct_count = 0
for i in range(100, len(test_images)):
tm = test_images[i]
idn = siamese.single_sample_identity.eval({
siamese.x1:
inference.format_single_sample(tm),
siamese.x2:
gallery_image
})
if gallery_label[idn] == test_labels[i]:
correct_count += 1
if (i + 1) % 1000 == 0:
print('Test {:d} images'.format(i + 1))
accuracy = correct_count / (len(test_images) - 100)
print('accuracy:', accuracy)
def train_siamese():
# prepare data and tf.session
global_step = tf.Variable(0, name='global_step', trainable=False)
lr = tf.train.piecewise_constant(global_step, boundaries, learning_rates)
siamese = inference.siamese(tf.estimator.ModeKeys.TRAIN)
train_step = tf.train.GradientDescentOptimizer(lr).minimize(
siamese.loss, global_step=global_step)
saver = tf.train.Saver()
mnist = input_data.read_data_sets('data/mnist-data', one_hot=False)
test_images = mnist.test.images
test_labels = mnist.test.labels
gallery_image = []
gallery_label = []
for i in range(100):
if len(gallery_image) != 10:
if test_labels[i] not in gallery_label:
gallery_label.append(test_labels[i])
gallery_image.append(test_images[i])
else:
break
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
batch_x1, batch_y1 = mnist.train.next_batch(128)
batch_x2, batch_y2 = mnist.train.next_batch(128)
batch_y = (batch_y1 == batch_y2).astype('float')
initial_loss = sess.run(siamese.loss,
feed_dict={
siamese.x1: batch_x1,
siamese.x2: batch_x2,
siamese.y_: batch_y
})
print('The initial loss:', initial_loss)
print('Global step:', sess.run(global_step))
print('Initial learning rate:', sess.run(lr))
print('Start train...')
for step in range(80000):
iterations = step + 1
batch_x1, batch_y1 = mnist.train.next_batch(128)
batch_x2, batch_y2 = mnist.train.next_batch(128)
batch_y = (batch_y1 == batch_y2).astype('float')
_, loss_v, gs_v, lr_v = sess.run(
[train_step, siamese.loss, global_step, lr],
feed_dict={
siamese.x1: batch_x1,
siamese.x2: batch_x2,
siamese.y_: batch_y
})
if np.isnan(loss_v):
print('Model diverged with loss = NaN')
quit()
if iterations % 100 == 0:
# print('step %d: loss %.3f' % (iterations, loss_v))
print(
'Global step: {:d}, iterations: {:d}, learning rate: {:.5f}, loss: {:.4f}'
.format(gs_v, iterations, lr_v, loss_v))
if iterations % 1000 == 0:
saver.save(sess=sess,
save_path=model_save_path,
global_step=iterations)
print('Start test...')
correct_count = 0
for i in range(100, len(test_images)):
tm = test_images[i]
idn = siamese.single_sample_identity.eval({
siamese.x1:
inference.format_single_sample(tm),
siamese.x2:
gallery_image
})
if gallery_label[idn] == test_labels[i]:
correct_count += 1
accuracy = correct_count / (len(test_images) - 100)
print('Test accuracy: {:.4f}'.format(accuracy))