def train_with_goals(noise=0, iterations=10000, learning_rate=0.1):
model = nn.ElmanGoalNet(size_hidden=15,
size_observation=9,
size_action=8,
size_goal1=2,
size_goal2=0)
num_episodes = iterations
model.learning_rate = learning_rate
model.L2_regularization = 0.
rng_avg_loss = 0.
rng_avg_actions = 0.
rng_avg_goals = 0.
for episode in range(num_episodes):
seqid = utils.idx_from_probabilities(
pnas2018task.sequence_probabilities)
goal = pnas2018task.goals[seqid]
sequence = pnas2018task.seqs[seqid]
inputs = utils.liststr_to_onehot(sequence[:-1],
pnas2018task.all_inputs)
targets = utils.liststr_to_onehot(sequence[1:],
pnas2018task.all_outputs)
model.action = np.zeros((1, model.size_action), dtype=np.float32)
# run the network
with tf.GradientTape() as tape:
# Initialize context with random/uniform values.
model.context = np.zeros((1, model.size_hidden), dtype=np.float32)
model.goal1 = goal[0]
for i in range(len(targets)):
model.action = np.zeros((1, model.size_action),
dtype=np.float32)
# Add noise
model.context += np.float32(
np.random.normal(0., noise, size=(1, model.size_hidden)))
observation = inputs[i].reshape(1, -1)
model.feedforward(observation)
# Get some statistics about what was correct and what wasn't
tchoices = np.array(model.h_action_wta).reshape(
(-1, len(targets[0])))
ratios = scripts.evaluate([tchoices], [targets])
loss, _ = model.train_obsolete(targets, goal, None, tape)
# Monitor progress using rolling averages.
speed = 2. / (
episode + 2
) if episode < 1000 else 0.001 # enables more useful evaluations for early trials
rng_avg_loss = utils.rolling_avg(rng_avg_loss, loss, speed)
rng_avg_actions = utils.rolling_avg(rng_avg_actions, ratios[0], speed)
rng_avg_goals = utils.rolling_avg(
rng_avg_goals, ratios[0] == 1,
speed) # whole action sequence correct ?
# Display on the console at regular intervals
if (episode < 1000 and episode in [3 ** n for n in range(50)]) or episode % 1000 == 0 \
or episode + 1 == num_episodes:
print(
"{0}: avg loss={1}, \tactions={2}, \tfull_sequence={3}".format(
episode, rng_avg_loss, rng_avg_actions, rng_avg_goals))
return model
def train(model=None, noise=0., iterations=5000, l1reg=0.0, l2reg= 0.0, algorithm=nn.SGD,
size_hidden=15, learning_rate=None, loss_type='cross_entropy',
initial_context=pnas2018.ZEROS):
if model is None:
model = nn.ElmanGoalNet(size_hidden=size_hidden, size_observation=len(rewardtask.all_inputs),
size_action=len(rewardtask.all_outputs), size_goal1=0, size_goal2=0,
algorithm=algorithm, initialization="normal")
num_episodes = iterations
if learning_rate is not None: # Else keep the model's learning rate
model.learning_rate = learning_rate
model.L1_regularization = l1reg
model.L2_regularization = l2reg
rng_avg_loss = 0.
rng_avg_actions = 0.
rng_avg_sequence = 0.
for episode in range(num_episodes):
model.new_episode(initial_context=initial_context)
seqid = utils.idx_from_probabilities(rewardtask.sequence_probabilities)
sequence = rewardtask.seqs[seqid]
inputs = utils.liststr_to_onehot(sequence[:-1], rewardtask.all_inputs)
targets = utils.liststr_to_onehot(sequence[1:], rewardtask.all_outputs)
# run the network
with tf.GradientTape(persistent=True) as tape:
for i in range(len(targets)):
model.action = np.zeros((1, model.size_action), dtype=np.float32)
model.context += np.float32(np.random.normal(0., noise, size=(1, model.size_hidden)))
observation = inputs[i].reshape(1, -1)
model.feedforward(observation)
#if episode % 2 == 0:
# Get some statistics about what was correct and what wasn't
tchoices = np.array(model.h_action_wta).reshape((-1, len(targets[0])))
ratios = scripts.evaluate([tchoices], [targets])
# Train model, record loss.
if loss_type==pnas2018.MSE:
loss, _ = model.train_MSE(targets, None, None, tape)
elif loss_type==pnas2018.CROSS_ENTROPY:
loss, _ = model.train_obsolete(targets, None, None, tape)
else:
loss, _ = model.train(tape, targets)
del tape
#if episode % 2 == 0:
# Monitor progress using rolling averages.
speed = 2. / (episode + 2) if episode < 1000 else 0.001 # enables more useful evaluations for early trials
rng_avg_loss = utils.rolling_avg(rng_avg_loss, loss, speed)
rng_avg_actions = utils.rolling_avg(rng_avg_actions, ratios[0], speed)
rng_avg_sequence = utils.rolling_avg(rng_avg_sequence, ratios[0] == 1,
speed) # whole action sequence correct ?
# Display on the console at regular intervals
if (episode < 1000 and episode in [3 ** n for n in range(50)]) or episode % 1000 == 0 \
or episode + 1 == num_episodes:
print(
"{0}: avg loss={1}, \tactions={2}, \tfull_sequence={3}".format(
episode, rng_avg_loss, rng_avg_actions, rng_avg_sequence))
return model, rng_avg_sequence
def train_predictive_net(model=None, iterations=5000, learning_rate=0.1, algorithm=nn.RMSPROP, hidden_units=15):
if model is None:
model = nn.PredictiveNet(size_hidden=hidden_units, algorithm=algorithm,
size_observation=len(pnas2018task.all_inputs),
size_action=len(pnas2018task.all_outputs))
num_episodes = iterations
model.learning_rate = learning_rate
rng_avg_loss = 0.
rng_avg_actions = 0.
rng_avg_full_seq = 0.
rng_avg_preds = 0.
for episode in range(num_episodes):
seqid = utils.idx_from_probabilities(pnas2018task.sequence_probabilities)
sequence = pnas2018task.seqs[seqid]
inputs = utils.liststr_to_onehot(sequence[:-1], pnas2018task.all_inputs)
action_targets = utils.liststr_to_onehot(sequence[1:], pnas2018task.all_outputs)
prediction_targets = utils.liststr_to_onehot(sequence[1:], pnas2018task.all_inputs)
model.action = np.zeros((1, model.size_action), dtype=np.float32)
model.prediction_linear = np.zeros((1, model.size_observation), dtype=np.float32) #initial prediction = 0
# run the network
# Initialize context with random/uniform values.
with tf.GradientTape() as tape:
model.context = np.zeros((1, model.size_hidden), dtype=np.float32)
for i in range(len(action_targets)):
model.action = np.zeros((1, model.size_action), dtype=np.float32)
#model.context += np.float32(np.random.normal(0., noise, size=(1, model.size_hidden)))
observation = inputs[i].reshape(1, -1)
model.feedforward(observation)
# Get some statistics about what was correct and what wasn't
tchoices = np.array(model.h_action_wta).reshape((-1, len(action_targets[0]))) # reshape to (x, 8)
ratios = scripts.evaluate([tchoices], [action_targets])
tpreds = np.array(model.h_prediction_wta).reshape((-1, len(prediction_targets[0])))
ratios_predictions = scripts.evaluate([tpreds], [prediction_targets])
# Train model, record loss. NOTE: targets and predictions are identical for this task!!!
loss, gradients = model.train(tape, [action_targets, prediction_targets])
# Monitor progress using rolling averages.
speed = 2. / (episode + 2) if episode < 1000 else 0.001 # enables more useful evaluations for early trials
rng_avg_loss = utils.rolling_avg(rng_avg_loss, loss, speed)
rng_avg_actions = utils.rolling_avg(rng_avg_actions, ratios[0], speed)
rng_avg_preds = utils.rolling_avg(rng_avg_preds, ratios_predictions[0], speed)
rng_avg_full_seq = utils.rolling_avg(rng_avg_full_seq, ratios[0] == 1, speed) # whole action sequence correct ?
# Display on the console at regular intervals
if (episode < 1000 and episode in [3 ** n for n in range(50)]) or episode % 1000 == 0 \
or episode + 1 == num_episodes:
grad_avg = sum([np.sum(tf.reduce_sum(tf.abs(gradient)).numpy()) for gradient in gradients])/sum([tf.size(gradient).numpy() for gradient in gradients])
grad_max = max([np.max(tf.reduce_max(tf.abs(gradient)).numpy()) for gradient in gradients])
print("{0}: avg loss={1}, \tactions={2}, \tfull_seq={3}, \tpredictions={4}".format(
episode, rng_avg_loss, rng_avg_actions, rng_avg_full_seq, rng_avg_preds))
return model
def train_with_goals(model=None,
mse=False,
learning_rate=0.1,
noise=0.,
iterations=5000,
l2reg=0.0,
algorithm=nn.SGD,
hidden_units=15,
reg_strength=0.,
reg_increase="square"):
num_goals = 2
if model is None:
model = nn.ElmanGoalNet(size_hidden=hidden_units,
algorithm=algorithm,
size_observation=len(all_inputs),
size_action=len(all_inputs),
size_goal1=num_goals,
size_goal2=0)
num_episodes = iterations
model.learning_rate = 0.5 if mse else learning_rate
model.L2_regularization = l2reg
rng_avg_loss = 0.
rng_avg_actions = 0.
rng_avg_goals = 0.
for episode in range(num_episodes):
decider = np.random.uniform()
if decider < 0.6:
seqid = 0
elif decider < 0.8:
seqid = 1
else:
seqid = 2
sequence = seqs[seqid]
goal = goals[seqid]
inputs = utils.liststr_to_onehot(sequence[:-1], all_inputs)
targets = utils.liststr_to_onehot(sequence[1:], all_outputs)
targets_goal1 = goal
model.action = np.zeros((1, model.size_action), dtype=np.float32)
# run the network
with tf.GradientTape() as tape:
# Initialize context with random/uniform values.
model.context = np.zeros((1, model.size_hidden), dtype=np.float32)
model.goal1 = goal[0]
for i in range(len(targets)):
model.action = np.zeros((1, model.size_action),
dtype=np.float32)
model.context += np.float32(
np.random.normal(0., noise, size=(1, model.size_hidden)))
observation = inputs[i].reshape(1, -1)
model.feedforward(observation)
# Get some statistics about what was correct and what wasn't
tchoices = np.array(model.h_action_wta).reshape(
(-1, len(targets[0])))
ratios = scripts.evaluate([tchoices], [targets])
cols = model.size_hidden
# Regularization in the hidden layer weights
# Recurrent hidden to hidden connections
extra_loss = pnashierarchy.weight_regularization_calculator(
model.hidden_layer.w, [0, model.size_hidden], [0, cols],
reg_strength,
reg_type="recurrent",
reg_increase=reg_increase)
# Prev action to hidden
# extra_loss += weight_regularization_calculator(model.hidden_layer.w,
# [model.size_hidden+9, model.size_hidden+9+model.size_action],
# [0, cols],
# reg_strength, reg_type="input_right", reg_increase=reg_increase)
# Prev goal to hidden
extra_loss += pnashierarchy.weight_regularization_calculator(
model.hidden_layer.w, [
model.size_hidden + 9 + model.size_action,
model.size_hidden + 9 + model.size_action + num_goals
], [0, cols],
reg_strength,
reg_type="input_left",
reg_increase=reg_increase)
# SWITCHED OUTPUT LEFT AND OUTPUT RIGHT.
#Regularization in the output layers (goals and actions) weights
# hidden to next action
extra_loss += pnashierarchy.weight_regularization_calculator(
model.action_layer.w, [0, model.size_hidden],
[0, model.size_action],
reg_strength,
reg_type="output_right",
reg_increase=reg_increase)
# Hidden to next goal
extra_loss += pnashierarchy.weight_regularization_calculator(
model.goal1_layer.w, [0, model.size_hidden],
[0, model.size_action],
reg_strength,
reg_type="output_left",
reg_increase=reg_increase)
# Regularization of the observation (only goes to the action side)
#extra_loss += weight_regularization_calculator(model.hidden_layer.w,
# [model.size_hidden, model.size_hidden+model.size_observation],
# [0, cols],
# reg_strength, reg_type="input_right", reg_increase=reg_increase)
loss, _ = model.train_obsolete(targets, goal, None, tape,
extra_loss)
# Train model, record loss.
#if mse:
# loss = model.train_MSE(targets, None, None, tape)
#else:
# loss, gradients = model.train_obsolete(targets, targets_goal1, None, tape)
# Monitor progress using rolling averages.
speed = 2. / (
episode + 2
) if episode < 1000 else 0.001 # enables more useful evaluations for early trials
rng_avg_loss = utils.rolling_avg(rng_avg_loss, loss, speed)
rng_avg_actions = utils.rolling_avg(rng_avg_actions, ratios[0], speed)
rng_avg_goals = utils.rolling_avg(
rng_avg_goals, ratios[0] == 1,
speed) # whole action sequence correct ?
# Display on the console at regular intervals
if (episode < 1000 and episode in [3 ** n for n in range(50)]) or episode % 1000 == 0 \
or episode + 1 == num_episodes:
print(
"{0}: avg loss={1}, \tactions={2}, \tfull_sequence={3}".format(
episode, rng_avg_loss, rng_avg_actions, rng_avg_goals))
return model
def train(model=None,
mse=False,
noise=0.,
iterations=5000,
l2reg=0.0,
learning_rate=0.1,
algorithm=nn.SGD,
hidden_units=15):
if model is None:
model = nn.ElmanGoalNet(size_hidden=hidden_units,
algorithm=algorithm,
size_observation=len(all_inputs),
size_action=len(all_inputs),
size_goal1=0,
size_goal2=0)
num_episodes = iterations
model.learning_rate = learning_rate
model.L2_regularization = l2reg
rng_avg_loss = 0.
rng_avg_actions = 0.
rng_avg_full_seq = 0.
for episode in range(num_episodes):
seqid = utils.idx_from_probabilities(sequence_probabilities)
sequence = seqs[seqid]
inputs = utils.liststr_to_onehot(sequence[:-1], all_inputs)
targets = utils.liststr_to_onehot(sequence[1:], all_outputs)
model.action = np.zeros((1, model.size_action), dtype=np.float32)
# run the network
# Initialize context with random/uniform values.
with tf.GradientTape() as tape:
model.context = np.zeros((1, model.size_hidden), dtype=np.float32)
for i in range(len(targets)):
model.action = np.zeros((1, model.size_action),
dtype=np.float32)
model.context += np.float32(
np.random.normal(0., noise, size=(1, model.size_hidden)))
observation = inputs[i].reshape(1, -1)
model.feedforward(observation)
# Get some statistics about what was correct and what wasn't
tchoices = np.array(model.h_action_wta).reshape(
(-1, len(targets[0])))
ratios = scripts.evaluate([tchoices], [targets])
# Train model, record loss.
if mse:
loss, gradients = model.train_MSE(targets, None, None, tape)
else:
loss, gradients = model.train_obsolete(targets, None, None,
tape)
# Monitor progress using averages
speed = 2. / (
episode + 2
) if episode < 1000 else 0.001 # enables more useful evaluations for early trials
rng_avg_loss = utils.rolling_avg(rng_avg_loss, loss, speed)
rng_avg_actions = utils.rolling_avg(rng_avg_actions, ratios[0], speed)
rng_avg_full_seq = utils.rolling_avg(
rng_avg_full_seq, ratios[0] == 1,
speed) # whole action sequence correct ?
# Display on the console at regular intervals
if (episode < 1000 and episode in [3 ** n for n in range(50)]) or episode % 1000 == 0 \
or episode + 1 == num_episodes:
grad_avg = sum([
np.sum(tf.reduce_sum(tf.abs(gradient)).numpy())
for gradient in gradients
]) / sum([tf.size(gradient).numpy() for gradient in gradients])
grad_max = max([
np.max(tf.reduce_max(tf.abs(gradient)).numpy())
for gradient in gradients
])
print(
"{0}: avg loss={1}, \tactions={2}, \tfull_seq={3}, \tgrad_avg={4}, \tgrad_max={5}"
.format(episode, rng_avg_loss, rng_avg_actions,
rng_avg_full_seq, grad_avg, grad_max))
return model
def train_supervised_teacoffeeenv(model, environment, num_episodes):
env = environment
rng_avg_loss = 0.
rng_avg_actions = 0.
rng_avg_goals = 0.
rng_avg_goal1 = 0.
rng_avg_goal2 = 0.
rng_avg_action1 = 0.
goal_list = [
"g_1_make_coffee", "g_1_make_tea", "g_2_add_grounds", "g_2_add_cream",
"g_2_add_sugar", "g_2_drink", "g_2_dip_teabag"
]
for episode in range(num_episodes):
goal = goal_list[episode % len(goal_list)] # Cycle through the goals
targets = tce.target_list[goal] # Get the target actions and goals
env.state = state.State(tce.TeaCoffeeData()) # Reinitialize the state
env.state.current.set_field(goal, 1.) # Set the goalS as active
# run the network
with tf.GradientTape() as tape:
# Initialize context with random/uniform values.
#model.context = np.zeros((1, model.size_hidden))
model.context = np.float32(
np.random.uniform(0.01, 0.99, (1, model.size_hidden)))
targets_onehot = [[], [], []] # actions, goal1s, goal2s
for i, target in enumerate(targets):
# Set up the input to be the correct actions and goals
targets_onehot[0].append(
utils.str_to_onehot(targets[i][0],
tce.TeaCoffeeData.actions_list))
targets_onehot[1].append(
utils.str_to_onehot(targets[i][1],
tce.TeaCoffeeData.goals1_list))
targets_onehot[2].append(
utils.str_to_onehot(targets[i][2],
tce.TeaCoffeeData.goals2_list))
model.action, model.goal1, model.goal2 = [
targets_onehot[j][-1] for j in range(3)
]
observation = env.observe()
model.feedforward(observation)
env.do_action(
target[0]
) # Transition the MDP according to the *target* action, not the chosen action!
# Get some statistics about what was correct and what wasn't
ratios = evaluate(
[model.h_action_wta, model.h_goal1_wta, model.h_goal2_wta],
targets_onehot)
rng_avg_action1 = utils.rolling_avg(
rng_avg_action1,
ratio_correct([model.h_action_wta[0]], [targets_onehot[0][0]]),
2. / (episode + 2) if episode < 1000 else 0.001)
# Train model, record loss.
loss = model.train_obsolete(targets_onehot[0], targets_onehot[1],
targets_onehot[2], tape)
# Monitor progress using rolling averages.
speed = 2. / (
episode + 2
) if episode < 1000 else 0.001 # enables more useful evaluations for early trials
rng_avg_loss = utils.rolling_avg(rng_avg_loss, loss, speed)
rng_avg_actions = utils.rolling_avg(rng_avg_actions, ratios[0],
speed)
rng_avg_goals = utils.rolling_avg(
rng_avg_goals, ratios[0] == 1,
speed) # whole action sequence correct ?
rng_avg_goal1 = utils.rolling_avg(rng_avg_goal1, ratios[1], speed)
rng_avg_goal2 = utils.rolling_avg(rng_avg_goal2, ratios[2], speed)
# Display on the console at regular intervals
if (episode < 1000 and episode in [3**n for n in range(50)]) or episode % 1000 == 0 \
or episode+1 == num_episodes:
print(
"{0}: avg loss={1}, \tactions={2}, \tfull_sequence={3}\tgoal1={4}\tgoal2={5}\tfirst_action={6}"
.format(episode, rng_avg_loss, rng_avg_actions,
rng_avg_goals, rng_avg_goal1, rng_avg_goal2,
rng_avg_action1))
def train_supervised(model, env, num_episodes):
rng_avg_loss = 0.
rng_avg_actions = 0.
rng_avg_goals = 0.
rng_avg_goal1 = 0.
rng_avg_goal2 = 0.
rng_avg_action1 = 0.
for episode in range(num_episodes):
# run the network
with tf.GradientTape() as tape:
# Initialize model with random/uniform values.
model.context = np.zeros((1, model.size_hidden))
# Initialize environment with a random sequence
sequence = random.choice(env.sequences)
sequence.initialize(
) # This is a bit weird: it's the environment that is getting initialized, not the sequence
for target in sequence.targets:
# Set up the input to be the correct actions and goals
model.action = target.action_one_hot
model.goal1 = target.goal1_one_hot if target.goal1_one_hot is not None else zeros
model.goal1 = target.goal1_one_hot if target.goal1_one_hot is not None else zeros
observation = env.observe()
model.feedforward(observation)
env.do_action(
target[0]
) # Transition the MDP according to the *target* action, not the chosen action!
# Get some statistics about what was correct and what wasn't
ratios = evaluate(
[model.h_action_wta, model.h_goal1_wta, model.h_goal2_wta],
targets_onehot)
rng_avg_action1 = utils.rolling_avg(
rng_avg_action1,
ratio_correct([model.h_action_wta[0]], [targets_onehot[0][0]]),
2. / (episode + 2) if episode < 1000 else 0.001)
# Train model, record loss.
loss = model.train_obsolete(targets_onehot[0], targets_onehot[1],
targets_onehot[2], tape)
# Monitor progress using rolling averages.
speed = 2. / (
episode + 2
) if episode < 1000 else 0.001 # enables more useful evaluations for early trials
rng_avg_loss = utils.rolling_avg(rng_avg_loss, loss, speed)
rng_avg_actions = utils.rolling_avg(rng_avg_actions, ratios[0],
speed)
rng_avg_goals = utils.rolling_avg(
rng_avg_goals, ratios[0] == 1,
speed) # whole action sequence correct ?
rng_avg_goal1 = utils.rolling_avg(rng_avg_goal1, ratios[1], speed)
rng_avg_goal2 = utils.rolling_avg(rng_avg_goal2, ratios[2], speed)
# Display on the console at regular intervals
if (episode < 1000 and episode in [3**n for n in range(50)]) or episode % 1000 == 0 \
or episode+1 == num_episodes:
print(
"{0}: avg loss={1}, \tactions={2}, \tfull_sequence={3}\tgoal1={4}\tgoal2={5}\tfirst_action={6}"
.format(episode, rng_avg_loss, rng_avg_actions,
rng_avg_goals, rng_avg_goal1, rng_avg_goal2,
rng_avg_action1))
def train_hierarchical_nogoals(noise=0,
iterations=10000,
learning_rate=0.1,
reg_strength=0.001,
reg_increase="linear"):
model = nn.ElmanGoalNet(size_hidden=15,
size_observation=9,
size_action=8,
size_goal1=0,
size_goal2=0)
num_episodes = iterations
model.learning_rate = learning_rate
model.L2_regularization = 0.
rng_avg_loss = 0.
rng_avg_actions = 0.
rng_avg_goals = 0.
for episode in range(num_episodes):
model.new_episode()
seqid = utils.idx_from_probabilities(
pnas2018task.sequence_probabilities)
#goal = pnas2018task.goals[seqid]
sequence = pnas2018task.seqs[seqid]
inputs = utils.liststr_to_onehot(sequence[:-1],
pnas2018task.all_inputs)
targets = utils.liststr_to_onehot(sequence[1:],
pnas2018task.all_outputs)
model.action = np.zeros((1, model.size_action), dtype=np.float32)
# run the network
with tf.GradientTape() as tape:
# Initialize context with random/uniform values.
#model.context = np.zeros((1, model.size_hidden), dtype=np.float32)
#model.goal1 = np.zeros_like(goal[0])
for i in range(len(targets)):
#model.action = np.zeros((1, model.size_action), dtype=np.float32)
# Add noise
model.context += np.float32(
np.random.normal(0., noise, size=(1, model.size_hidden)))
observation = inputs[i].reshape(1, -1)
model.feedforward(observation)
# Get some statistics about what was correct and what wasn't
tchoices = np.array(model.h_action_wta).reshape(
(-1, len(targets[0])))
ratios = scripts.evaluate([tchoices], [targets])
# Train model, record loss.
cols = model.size_hidden
# Regularization in the hidden layer weights
# Recurrent hidden to hidden connections
extra_loss = utils.weight_regularization_calculator(
model.hidden_layer.w, [0, model.size_hidden], [0, cols],
reg_strength,
reg_type="recurrent",
reg_increase=reg_increase)
# Prev action to hidden
#extra_loss += weight_regularization_calculator(model.hidden_layer.w,
# [model.size_hidden+9, model.size_hidden+9+model.size_action], [0, cols],
# reg_strength, reg_type="input_right", reg_increase=reg_increase)
# Prev goal to hidden
#extra_loss += weight_regularization_calculator(model.hidden_layer.w,
# [model.size_hidden+9+model.size_action, model.size_hidden+9+model.size_action+2], [0, cols],
# reg_strength, reg_type="input_left", reg_increase=reg_increase)
#Regularization in the output layers (goals and actions) weights
# hidden to next action
extra_loss += utils.weight_regularization_calculator(
model.action_layer.w, [0, model.size_hidden],
[0, model.size_action],
reg_strength,
reg_type="output_right",
reg_increase=reg_increase)
# Hidden to next goal
#extra_loss += weight_regularization_calculator(model.goal1_layer.w,
# [0, model.size_hidden], [0, model.size_action],
# reg_strength, reg_type="output_left", reg_increase=reg_increase)
# Regularization of the observation (only goes to the action side)
#extra_loss += weight_regularization_calculator(model.hidden_layer.w,
# [model.size_hidden, model.size_hidden+model.size_observation],
# [0, cols],
# reg_strength, reg_type="input_right", reg_increase=reg_increase)
loss, _ = model.train_obsolete(targets, None, None, tape,
extra_loss)
#if(episode%100 == 0):
# print(loss.numpy()-extra_loss.numpy(), extra_loss.numpy())
# Monitor progress using rolling averages.
speed = 2. / (
episode + 2
) if episode < 1000 else 0.001 # enables more useful evaluations for early trials
rng_avg_loss = utils.rolling_avg(rng_avg_loss, loss, speed)
rng_avg_actions = utils.rolling_avg(rng_avg_actions, ratios[0], speed)
rng_avg_goals = utils.rolling_avg(
rng_avg_goals, ratios[0] == 1,
speed) # whole action sequence correct ?
# Display on the console at regular intervals
if (episode < 1000 and episode in [3 ** n for n in range(50)]) or episode % 1000 == 0 \
or episode + 1 == num_episodes:
print(
"{0}: avg loss={1}, \tactions={2}, \tfull_sequence={3}".format(
episode, rng_avg_loss, rng_avg_actions, rng_avg_goals))
return model
