def train_joint(self, sess, image_data, reward_data, y_data, num_epochs, logs):
"""Trains the planner module given MDPs with reward functions and the
corresponding policies.
Essentially performs train_reward, but also passes planner graph
Validation can be turned off by explicitly passing None.
"""
if self.config.verbosity >= 3:
print(fmt_row(10, ["Iteration", "Train Cost", "Train Err", "Iter Time"]))
if reward_data is None:
# Initialize the reward array to random values
reward_data = np.random.randn(*image_data.shape)
planner_ops = [self.planner_optimize_op, self.step1_cost]
batch_size = self.config.batchsize
num_batches = int(image_data.shape[0] / batch_size)
errors, costs, times = [], [], []
for epoch in range(num_epochs):
tstart = time.time()
errors_per_batch, costs_per_batch = [], []
for batch_num in range(num_batches):
start, end = batch_num * batch_size, (batch_num + 1) * batch_size
# We can't feed in reward_data directly to self.reward.
# See train_reward for the explanation.
fd = {
"reward_input:0": reward_data[start:end],
}
sess.run([self.assign_reward.op], feed_dict=fd)
fd = {
"image:0": image_data[start:end],
"y:0": y_data[start:end]
}
# Run both step1 & step2 ops, report only error & step2 cost
_, err, cost, _, _ = sess.run(
[self.reward_optimize_op, self.err, self.step2_cost] + planner_ops,
feed_dict=fd)
errors_per_batch.append(err)
costs_per_batch.append(cost)
epoch_error = sum(errors_per_batch) / len(errors_per_batch)
errors.append(epoch_error)
epoch_cost = sum(costs_per_batch) / len(costs_per_batch)
costs.append(epoch_cost)
elapsed = time.time() - tstart
times.append(elapsed)
if self.config.verbosity >= 3:
print(fmt_row(10, [epoch, epoch_cost, epoch_error, elapsed]))
reward_data[start:end] = self.reward.eval()
logs['train_joint_errs'].append(errors)
logs['train_joint_costs'].append(costs)
logs['train_joint_times'].append(times)
self.accuracy = 100 * (1 - errors[-1])
logs['accuracy'] = self.accuracy
return reward_data
def train_reward(self, sess, image_data, reward_data, y_data, num_epochs, logs):
"""Infers the reward using backprop, holding the planner fixed.
Due to Tensorflow constraints, image_data must contain exactly
batch_size number of MDPs on which the reward should be inferred.
The rewards are initialized to the values in reward_data. If reward_data
is None, the rewards are initialized to all zeroes.
"""
if self.config.verbosity >= 3:
print(fmt_row(10, ["Iteration", "Train Cost", "Train Err", "Iter Time"]))
if reward_data is None:
reward_data = np.random.randn(*image_data.shape)
batch_size = self.config.batchsize
num_batches = int(image_data.shape[0] / batch_size)
costs, errs = [], []
for batch_num in range(num_batches):
if self.config.verbosity >= 2 and batch_num % 10 == 0:
print('Batch {} of {}'.format(batch_num, num_batches))
start, end = batch_num * batch_size, (batch_num + 1) * batch_size
# We can't feed in reward_data directly to self.reward, because then
# it will treat it as a constant and will not be able to update it
# with backprop. Instead, we first run an op that assigns the
# reward, and only then do the backprop.
fd = {
"reward_input:0": reward_data[start:end],
}
sess.run([self.assign_reward.op], feed_dict=fd)
if batch_num % 10 == 0:
costs.append([])
errs.append([])
for epoch in range(num_epochs):
tstart = time.time()
fd = {
"image:0": image_data[start:end],
"y:0": y_data[start:end]
}
_, e_, c_ = sess.run(
[self.reward_optimize_op, self.err, self.step2_cost],
feed_dict=fd)
elapsed = time.time() - tstart
if self.config.verbosity >= 3 and batch_num % 10 == 0:
print(fmt_row(10, [epoch, c_, e_, elapsed]))
costs[-1].append(c_)
errs[-1].append(e_)
reward_data[start:end] = self.reward.eval()
logs['train_reward_costs'].append(costs)
logs['train_reward_errs'].append(errs)
return reward_data
TrA = []
TeA = []
# Launch the graph
with tf.Session() as sess:
if config.log:
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name, var)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(config.logdir, sess.graph)
sess.run(init)
if Have_trained == True:
model_file = tf.train.latest_checkpoint('ckpt/')
saver.restore(sess, model_file)
batch_size = config.batchsize
print(fmt_row(10, ["Epoch", "Train Cost",
"Train Acc", "Epoch Time", "Test Acc"]))
for epoch in range(int(config.epochs)):
learning_rate = learning_rate * 0.95
tstart = time.time()
avg_err, avg_cost = 0.0, 0.0
avg_err_L = []
avg_cost_L = []
num_batches = int(Xtrain.shape[0] / batch_size)
# Loop over all batches
for i in range(0, Xtrain.shape[0], batch_size):
j = i + batch_size
if j <= Xtrain.shape[0]:
# Run optimization op (backprop) and cost op (to get loss value)
fd = {X: Xtrain[i:j], S1: S1train[i:j], S2: S2train[i:j],
y: ytrain[i * config.statebatchsize:j * config.statebatchsize], LR: learning_rate, keep_drop: 0.5}
_, e_, c_ = sess.run([optimizer, err, cost], feed_dict=fd)
saver = tf.train.Saver()
Xtrain, S1train, S2train, ytrain, Xtest, S1test, S2test, ytest = process_gridworld_data(
input=config.input, imsize=config.imsize)
# Launch the graph
with tf.Session() as sess:
if config.log:
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name, var)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(config.logdir, sess.graph)
sess.run(init)
batch_size = config.batchsize
print(fmt_row(10, ["Epoch", "Train Cost", "Train Err", "Epoch Time"]))
for epoch in range(int(config.epochs)):
tstart = time.time()
avg_err, avg_cost = 0.0, 0.0
num_batches = int(Xtrain.shape[0] / batch_size)
# Loop over all batches
for i in range(0, Xtrain.shape[0], batch_size):
j = i + batch_size
if j <= Xtrain.shape[0]:
# Run optimization op (backprop) and cost op (to get loss value)
fd = {
X: Xtrain[i:j],
S1: S1train[i:j],
S2: S2train[i:j],
y:
ytrain[i * config.statebatchsize:j * config.statebatchsize]
def train_planner(self, sess, train_data, validation_data, num_epochs, logs):
"""Trains the planner module given MDPs with reward functions and the
corresponding policies.
Validation can be turned off by explicitly passing None.
"""
if self.config.verbosity >= 3:
print(fmt_row(10, ["Epoch", "Train Cost", "Train Err", "Valid Err", "Epoch Time"]))
avg_costs, train_errs, validation_errs, times = [], [], [], []
for epoch in range(int(num_epochs)):
_, (avg_cost, avg_err), elapsed, epoch_dist = self.run_epoch(
sess, train_data, [self.planner_optimize_op],
[self.step1_cost, self.err], [self.pred_dist, self.y_dist])
# Display logs per epoch step
if self.config.verbosity >= 3 and epoch % self.config.display_step == 0:
if validation_data is not None:
_, (validation_err,), _, _ = self.run_epoch(sess, validation_data, [], [self.err])
else:
validation_err = 'N/A'
print(fmt_row(10, [epoch, avg_cost, avg_err, validation_err, elapsed]))
avg_costs.append(avg_cost)
train_errs.append(avg_err)
validation_errs.append(validation_err)
times.append(elapsed)
elif self.config.verbosity >= 2 and epoch % self.config.display_step == 0:
print('Epoch {} of {}'.format(epoch, num_epochs))
# if self.config.log:
# summary = tf.Summary()
# summary.ParseFromString(sess.run(summary_op))
# summary.value.add(tag='Average error', simple_value=float(avg_err))
# summary.value.add(tag='Average cost', simple_value=float(avg_cost))
# summary_writer.add_summary(summary, epoch)
logs['train_planner_costs'].append(avg_costs)
logs['train_planner_train_errs'].append(train_errs)
logs['train_planner_validation_errs'].append(validation_errs)
logs['train_planner_times'].append(times)
# TODO(rohinmshah): This seems redundant
num_actions = self.config.num_actions
action_dists = [np.zeros(num_actions), np.zeros(num_actions)]
action_dists = [d + b_d for d, b_d in zip(action_dists, epoch_dist)]
action_dists = [d / (np.sum(d)) for d in action_dists]
pred = action_dists[0].tolist()
actual = action_dists[1].tolist()
logs['train_planner_predicted_action_dists'].append(pred)
logs['train_planner_actual_action_dists'].append(actual)
if self.config.verbosity >= 3:
print("Action Distribution Comparison")
print("------------------------------")
print(fmt_row(10, ["Predicted"] + pred))
print(fmt_row(10, ["Actual"]+ actual))
if validation_data is not None:
_, (err,), _, _ = self.run_epoch(sess, validation_data, [], [self.err])
logs['accuracy'].append(100 * (1 - err))
if self.config.verbosity >= 1:
print('Validation Accuracy: ' + str(100 * (1 - err)))
# Saving SavedModel instance
if self.config.savemodel:
saver = tf.train.Saver()
# This allows for the model to perform reward inference
saver.save(sess, "model_save_sess_0/")
init = tf.global_variables_initializer()
saver = tf.train.Saver()
Xtrain, S1train, S2train, ytrain, Xtest, S1test, S2test, ytest = process_gridworld_data(input=config.input, imsize=config.imsize)
# Launch the graph
with tf.Session() as sess:
if config.log:
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name, var)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(config.logdir, sess.graph)
sess.run(init)
batch_size = config.batchsize
print(fmt_row(10, ["Epoch", "Train Cost", "Train Err", "Epoch Time"]))
for epoch in range(int(config.epochs)):
tstart = time.time()
avg_err, avg_cost = 0.0, 0.0
num_batches = int(Xtrain.shape[0]/batch_size)
# Loop over all batches
for i in range(0, Xtrain.shape[0], batch_size):
j = i + batch_size
if j <= Xtrain.shape[0]:
# Run optimization op (backprop) and cost op (to get loss value)
fd = {X: Xtrain[i:j], S1: S1train[i:j], S2: S2train[i:j],
y: ytrain[i * config.statebatchsize:j * config.statebatchsize]}
_, e_, c_ = sess.run([optimizer, err, cost], feed_dict=fd)
avg_err += e_
avg_cost += c_
# Display logs per epoch step
