# if len(past_losses) == loss_history and np.std(past_losses) <= 0.1:
# rate = rate * 10.0
return max(rate, min_rate)
last_loss = None
saver = tf.train.Saver()
save_step = 10
past_losses = []
loss_history = 5
with tf.Session() as sess:
print 'initializing'
sess.run(init)
print 'session launched'
for step in range(training_iters):
features, targets = batch(batch_size, mode='camera')
sess.run(optimizer,
feed_dict={
x: features,
y: targets,
keep_prob: dropout,
learning_rate: get_learning_rate(last_loss, past_losses)
})
if step % display_step == 0:
loss = sess.run(cost,
feed_dict={
x: features,
y: targets,
keep_prob: 1.
})
last_loss = loss
last_loss = None
saver = tf.train.Saver()
save_step = 100
past_losses = []
loss_history = 5
lowest_loss = 1000.0
lowest_iter = 0
loss_history = []
name = 'basic'
with tf.Session() as sess:
print 'initializing'
sess.run(init)
print 'session launched'
for step in range(training_iters):
features, targets = batch(batch_size)
sess.run(optimizer,
feed_dict={
x: features,
y: targets,
keep_prob: dropout
})
if step % display_step == 0:
loss = sess.run(cost,
feed_dict={
x: features,
y: targets,
keep_prob: 1.
})
print("Iter {}, Minibatch avg error={}".format(
step, np.sqrt(loss / (ackermann_scale**2.0))))
last_loss = None
saver = tf.train.Saver()
save_step = 100
past_losses = []
loss_history = 5
lowest_loss = 1000.0
lowest_iter = 0
loss_history = []
name = 'next_three'
with tf.Session() as sess:
print 'initializing'
sess.run(init)
print 'session launched'
for step in range(training_iters):
features, targets = batch(batch_size, next_x=3)
sess.run(optimizer,
feed_dict={
x: features,
y: targets,
keep_prob: dropout
})
if step % display_step == 0:
loss = sess.run(cost,
feed_dict={
x: features,
y: targets,
keep_prob: 1.
})
print("Iter {}, Minibatch avg error={}".format(
step, np.sqrt(loss / (ackermann_scale**2.0))))
last_loss = None
saver = tf.train.Saver()
save_step = 100
past_losses = []
loss_history = 5
lowest_loss = 1000.0
lowest_iter = 0
loss_history = []
name = 'no_cdf_next_three'
with tf.Session() as sess:
print 'initializing'
sess.run(init)
print 'session launched'
for step in range(training_iters):
features, targets = batch(batch_size, next_x=3, cdf=False)
sess.run(optimizer,
feed_dict={
x: features,
y: targets,
keep_prob: dropout
})
if step % display_step == 0:
loss = sess.run(cost,
feed_dict={
x: features,
y: targets,
keep_prob: 1.
})
print("Iter {}, Minibatch avg error={}".format(
step, np.sqrt(loss / (ackermann_scale**2.0))))
return max(rate, min_rate)
last_loss = None
saver = tf.train.Saver()
save_step = 10
past_losses = []
loss_history = 5
lowest_loss = 1000.0
lowest_iter = 0
with tf.Session() as sess:
print 'initializing'
sess.run(init)
print 'session launched'
for step in range(training_iters):
features, targets = batch(batch_size, mode='laser', laser_cdf=False)
sess.run(optimizer,
feed_dict={
x: features,
y: targets,
keep_prob: 1.0,
learning_rate: get_learning_rate(last_loss)
})
if step % display_step == 0:
loss = sess.run(cost,
feed_dict={
x: features,
y: targets,
keep_prob: 1.
})
last_loss = loss
max_rate = 1
min_rate = 0.001
def get_learning_rate(last_loss):
if last_loss is None:
return 1
rate = min(last_loss / 1000.0, max_rate)
# if len(past_losses) == loss_history and np.std(past_losses) <= 0.1:
# rate = rate * 10.0
return max(rate, min_rate)
# Launch the graph
last_loss = 0
with tf.Session() as sess:
sess.run(init)
# Training cycle
for epoch in range(training_epochs):
features, targets = batch(batch_size, mode='laser')
# Run optimization op (backprop) and cost op (to get loss value)
_, c = sess.run([optimizer, cost],
feed_dict={
x: features,
y: targets,
learning_rate: get_learning_rate(last_loss)
})
if epoch % display_step == 0:
print("Epoch: {}, cost={}".format(epoch, c))
print("Optimization Finished!")
pred = conv_net(x, weights, biases, keep_prob)
print 'model constructed'
cost = tf.reduce_mean(tf.squared_difference(pred, y))
learning_rate = 0.001
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
print 'optimizer created'
init = tf.global_variables_initializer()
print 'init created'
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, 'models/test.ckpt')
features, targets = batch(1000000)
predictions = sess.run(pred, feed_dict={x: features})
x = [t[0]/(100) for t in targets] # remember the division by 100 for denormalization!
y = [float(p[0])/100.0 for p in predictions]
def bound(v):
v = max(0.0, v)
return min(1.0, v)
fig, ax = plt.subplots()
y = [bound(v) for v in y]
x = [bound(v) for v in x]
ax.scatter(x, y)
lims = [
0, 1