def show_histogram(image: tf.Tensor, channel: str, nbins: int,
source_range: str, normalize: bool) -> None:
"""
Given an image tensor, display it side-by-side with it's normalized histogram
"""
values, centers = histogram(image,
nbins=nbins,
source_range=source_range,
normalize=normalize)
fig, ax = plt.subplots(ncols=2, figsize=(10, 5))
# Lookup table so that matplotlib displays the channel aswell
cmap = dict(grey=plt.cm.gray,
r=plt.cm.Reds,
g=plt.cm.Greens,
b=plt.cm.Blues)
ax[0].imshow(image.eval(), cmap=cmap[channel])
ax[0].axis('off')
ax[1].plot(centers.eval(), values.eval(), lw=2)
# Lookup table for nice titles
names = dict(grey="grey", r="red", g="green", b="blue")
ax[1].set_title(f'Histogram of {names[channel]} values')
plt.tight_layout()
def update_prototype(feature_map: tf.Tensor,
y: tf.Tensor,
center: tf.Tensor,
ld=5e-2):
v_feature_map = feature_map.eval()
v_y = np.argmax(y.eval(), axis=1)
v_center = center.eval()
df = pd.DataFrame({
'y':
v_y,
"feature_map":
[v_feature_map[i] for i in range(v_feature_map.shape[0])]
})
df2 = df.groupby("y")["feature_map"].apply(np.mean)
update_ops = []
for i in df2.index:
update_ops.append(
tf.assign(center[i], (1 - ld) * v_center[i] + ld * df2.loc[i]))
return update_ops
def _get_tensor_value(tensor_or_eager_tensor: tf.Tensor) -> Any:
if ops.executing_eagerly_outside_functions():
return np.asarray(tensor_or_eager_tensor)
else:
with tf.compat.v1.Session():
return tensor_or_eager_tensor.eval()
def print_layer(layer: tf.Tensor):
values = layer.eval()
print('Layer: {:s}'.format(layer.name))
print(values)
def train_and_test_model(training_op: tf.Operation, test_tensor: tf.Tensor,
global_step: tf.Tensor,
batch: IsingFileRecord.IsingBatch,
config: GlobalConfig):
"""
Remember to scope under a default graph
:param training_op: Runs a training step in the graph.
Can itself be a no-op
:param test_tensor: 2D tensor, the direct output layer of the NN
:param batch: the IsingBatch tensor fed into the *TEST* tensor
:param config: GlobalConfig
:return:
"""
#EVAL_FREQUENCY = int(0.05*FLAGS.max_steps)
#g = tf.Graph()
#with g.as_default():
#global_training_step = tf.Variable(
# 0, trainable=False, name='global_step')
ch1 = logging.StreamHandler()
ch2 = logging.FileHandler('run.log')
logging.basicConfig(format='%(asctime)s %(message)s',
level=logging.DEBUG,
handlers=[ch1, ch2])
last_eval_accuracy = tf.Variable(0,
trainable=False,
name='eval_acc',
dtype=tf.float32)
next_acc = tf.placeholder(tf.float32, [])
update_acc = last_eval_accuracy.assign(next_acc)
acc_sum_op = tf.summary.scalar('Eval_Accuracy', last_eval_accuracy)
chk_saver = tf.train.Saver()
sum_writer = tf.summary.FileWriter(FLAGS.train_dir, tf.get_default_graph())
sum_writer.add_session_log(tf.SessionLog(status=tf.SessionLog.START))
merged_sum = tf.summary.merge_all()
sess = tf.Session()
_load_chk_or_init(chk_saver, sess)
log_freq = config.max_steps // 20
if config.log_freq is not None:
log_freq = config.log_freq
acc = 0.0
with sess.as_default():
#initialization
coord = tf.train.Coordinator(
) # start coordinator and threads for queues
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
#training loop
for step in range(config.max_steps):
i = global_step.eval() #the current step
# evaluate
if i % config.eval_freq == 0:
eval_result = _eval_once(test_tensor, batch, config, sess)
acc = eval_result.overall_accuracy
# update accuracy tensor
sess.run(update_acc, feed_dict={next_acc: acc})
# run training and summary
_, summary = sess.run([training_op, merged_sum])
#sess.run(training_op)
if i % config.sum_save_freq == 0:
sum_writer.add_summary(summary, i)
#log every so often
if (step) % log_freq == 0:
logging.info('Step %d: %f', i, last_eval_accuracy.eval())
#print("Step ", i, ": ", last_eval_accuracy.eval())
chk_saver.save(sess, FLAGS.train_dir + "/chk")
#final evaluation
eval_result = _eval_once(test_tensor, batch, config, sess)
sess.run(update_acc,
feed_dict={next_acc: eval_result.overall_accuracy})
acc_sum = sess.run(acc_sum_op)
sum_writer.add_summary(acc_sum, config.max_steps)
#final checkpoint
chk_saver.save(sess, FLAGS.train_dir + "/chk")
coord.request_stop()
coord.join(threads, stop_grace_period_secs=10)
sum_writer.add_session_log(tf.SessionLog(status=tf.SessionLog.STOP))
sum_writer.flush()
sum_writer.close()
sess.close()
return eval_result
