def test_compute_error(self, classes):
mnist = MNIST()
xin, _ = mnist.get_batch(30, dataset='testing', classes=classes)
xout, _ = mnist.get_batch(
6,
dataset='testing',
classes=[x for x in range(10) if x not in classes])
x = np.concatenate((xin, xout))
np.random.shuffle(x)
errors = []
for img in x:
errors.append(self.compute_error(img, expand_dims=True))
top_k_error = np.sort(errors)[-6]
errors = np.reshape(errors, (6, 6))
input_images = self.weights_to_grid(x.transpose((1, 2, 3, 0)), 6, 6)
fig, (ax0, ax1, ax2) = plt.subplots(ncols=3, figsize=(10, 5))
fig.suptitle("Trained with classes")
ax0.imshow(input_images, cmap=plt.cm.gray, interpolation='nearest')
ax0.set_title('input images')
shown = ax1.imshow(errors,
cmap=plt.cm.inferno,
interpolation='nearest')
plt.colorbar(shown, ax=ax1)
ax2.imshow(errors >= top_k_error,
plt.cm.inferno,
interpolation="nearest")
ax1.set_title('Errors')
plt.show()
def train(self,
batch_size,
passes,
new_training=True,
classes=list(range(10))):
"""
:param batch_size:
:param passes:
:param new_training:
:return:
"""
mnist = MNIST()
with tf.Session() as sess:
# prepare session
if new_training:
saver, global_step = Model.start_new_session(sess)
else:
saver, global_step = Model.continue_previous_session(
sess, ckpt_file='saver/checkpoint')
# start training
for step in range(1 + global_step, 1 + passes + global_step):
x, y = mnist.get_batch(batch_size, classes=classes)
self.training.run(feed_dict={self.x: x})
if step % 10 == 0:
loss = self.loss.eval(feed_dict={self.x: x})
print("pass {}, training loss {}".format(step, loss))
if step % 100 == 0: # save weights
saver.save(sess, 'saver/cnn', global_step=step)
print('checkpoint saved')
def lime(model):
#process image
mnist = MNIST()
image, label = mnist.get_batch(1, dataset='testing')
image = image[0]
image = np.concatenate((image, image, image), axis=2)
print(image.shape)
print('Print image:')
plt.imshow(np.squeeze(image) / 2 + 0.5)
plt.show()
print(model.compute_error(image, expand_dims=True))
#Explain
explainer = lime_image.LimeImageExplainer()
explanation = explainer.explain_instance(image,
model.compute_error,
top_labels=2)
print(explanation)
#Show superpixels
temp, mask = explanation.get_image_and_mask(explanation.top_labels[0],
positive_only=True,
hide_rest=True)
#pdb.set_trace()
print('Print superpixels:')
marked = seg.mark_boundaries(temp / 2 + 0.5, mask).astype(np.uint8)
plt.imshow(marked)
plt.show()
def reconstruct(self):
"""
"""
def weights_to_grid(weights, rows, cols):
"""convert the weights tensor into a grid for visualization"""
height, width, in_channel, out_channel = weights.shape
padded = np.pad(weights, [(1, 1), (1, 1), (0, 0),
(0, rows * cols - out_channel)],
mode='constant',
constant_values=0)
transposed = padded.transpose((3, 1, 0, 2))
reshaped = transposed.reshape((rows, -1))
grid_rows = [
row.reshape((-1, height + 2, in_channel)).transpose((1, 0, 2))
for row in reshaped
]
grid = np.concatenate(grid_rows, axis=0)
return grid.squeeze()
mnist = MNIST()
with tf.Session() as sess:
saver, global_step = Model.continue_previous_session(
sess, ckpt_file='saver/checkpoint')
# visualize weights
first_layer_weights = tf.get_default_graph().get_tensor_by_name(
"conv_1/kernel:0").eval()
grid_image = weights_to_grid(first_layer_weights, 4, 8)
fig, ax0 = plt.subplots(ncols=1, figsize=(8, 4))
ax0.imshow(grid_image, cmap=plt.cm.gray, interpolation='nearest')
ax0.set_title('first conv layers weights')
plt.show()
# visualize results
batch_size = 32
x, y = mnist.get_batch(batch_size, dataset='testing')
org, recon = sess.run((self.x, self.reconstruction),
feed_dict={self.x: x})
input_images = weights_to_grid(org.transpose((1, 2, 3, 0)), 6, 6)
recon_images = weights_to_grid(recon.transpose((1, 2, 3, 0)), 6, 6)
fig, (ax0, ax1) = plt.subplots(ncols=2, figsize=(10, 5))
ax0.imshow(input_images, cmap=plt.cm.gray, interpolation='nearest')
ax0.set_title('input images')
ax1.imshow(recon_images, cmap=plt.cm.gray, interpolation='nearest')
ax1.set_title('reconstructed images')
plt.show()
def train(self, batch_size, passes, new_training=True):
"""
:param batch_size:
:param passes:
:param new_training:
:return:
"""
mnist = MNIST()
with tf.Session() as sess:
# prepare session
if new_training:
saver, global_step = Model.start_new_session(sess)
else:
saver, global_step = Model.continue_previous_session(
sess, ckpt_file='./saver/checkpoint')
print("mnist mid point")
# start training
for step in range(1 + global_step, 1 + passes + global_step):
print(str(step) + " entered")
x, y = mnist.get_batch(batch_size)
print(x.shape)
#print(y.shape)
print("got x and y")
self.training.run(feed_dict={self.x: x})
print("imported from mnist.py")
if step % 10 == 0:
loss = self.loss.eval(feed_dict={self.x: x})
print("pass {}, training loss {}".format(step, loss))
if step % 5000 == 0: # save weights
saver.save(sess, './saver/cnn', global_step=step)
print('checkpoint saved')
print(str(step) + " exited")
del x
def reconstruct(self, vis_weights=True, classes="all"):
"""
"""
def weights_to_grid(weights, rows, cols):
"""convert the weights tensor into a grid for visualization"""
height, width, in_channel, out_channel = weights.shape
padded = np.pad(weights, [(1, 1), (1, 1), (0, 0),
(0, rows * cols - out_channel)],
mode='constant',
constant_values=0)
transposed = padded.transpose((3, 1, 0, 2))
reshaped = transposed.reshape((rows, -1))
grid_rows = [
row.reshape((-1, height + 2, in_channel)).transpose((1, 0, 2))
for row in reshaped
]
grid = np.concatenate(grid_rows, axis=0)
return grid.squeeze()
mnist = MNIST()
with tf.Session() as sess:
saver, global_step = Model.continue_previous_session(
sess, ckpt_file='saver/checkpoint')
# visualize weights
if vis_weights:
first_layer_weights = tf.get_default_graph(
).get_tensor_by_name("conv_1/kernel:0").eval()
grid_image = weights_to_grid(first_layer_weights, 4, 8)
fig, ax0 = plt.subplots(ncols=1, figsize=(8, 4))
ax0.imshow(grid_image,
cmap=plt.cm.gray,
interpolation='nearest')
ax0.set_title('first conv layers weights')
plt.show()
# visualize results
batch_size = 36
x, y = mnist.get_batch(batch_size, dataset='testing')
org, recon = sess.run((self.x, self.reconstruction),
feed_dict={self.x: x})
diff = (org - recon).squeeze()
magnitude = np.linalg.norm(org.squeeze(), axis=(1, 2))
error = np.linalg.norm(diff, axis=(1, 2))
print(np.mean(error))
error = error / magnitude
input_images = weights_to_grid(org.transpose((1, 2, 3, 0)), 6, 6)
recon_images = weights_to_grid(recon.transpose((1, 2, 3, 0)), 6, 6)
errors = np.reshape(error, (6, 6))
fig, (ax0, ax1, ax2) = plt.subplots(ncols=3, figsize=(10, 5))
fig.suptitle("Trained with classes: {}".format(classes))
ax0.imshow(input_images, cmap=plt.cm.gray, interpolation='nearest')
ax0.set_title('input images')
ax1.imshow(recon_images, cmap=plt.cm.gray, interpolation='nearest')
ax1.set_title('reconstructed images')
shown = ax2.imshow(errors,
cmap=plt.cm.inferno,
interpolation='nearest')
plt.colorbar(shown, ax=ax2)
ax2.set_title('Errors')
plt.show()
