def reshape_for_broadcasting(source, target):
"""Reshapes a tensor (source) to have the correct shape and dtype of the target
before broadcasting it with MPI.
"""
dim = len(target.get_shape())
shape = ([1] * (dim - 1)) + [-1]
return tf.reshape(tf.cast(source, target.dtype), shape)
def compute_gradients(self, loss, var_list, **kwargs):
grads_and_vars = tf.train.AdamOptimizer.compute_gradients(self, loss, var_list, **kwargs)
grads_and_vars = [(g, v) for g, v in grads_and_vars if g is not None]
flat_grad = tf.concat([tf.reshape(g, (-1,)) for g, v in grads_and_vars], axis=0)
shapes = [v.shape.as_list() for g, v in grads_and_vars]
sizes = [int(np.prod(s)) for s in shapes]
num_tasks = self.comm.Get_size()
buf = np.zeros(sum(sizes), np.float32)
def _collect_grads(flat_grad):
self.comm.Allreduce(flat_grad, buf, op=MPI.SUM)
np.divide(buf, float(num_tasks), out=buf)
return buf
avg_flat_grad = tf.py_func(_collect_grads, [flat_grad], tf.float32)
avg_flat_grad.set_shape(flat_grad.shape)
avg_grads = tf.split(avg_flat_grad, sizes, axis=0)
avg_grads_and_vars = [(tf.reshape(g, v.shape), v)
for g, (_, v) in zip(avg_grads, grads_and_vars)]
return avg_grads_and_vars
def nn(input, layers_sizes, reuse=None, flatten=False, name=""):
"""Creates a simple neural network
"""
for i, size in enumerate(layers_sizes):
activation = tf.nn.relu if i < len(layers_sizes) - 1 else None
input = tf.layers.dense(
inputs=input,
units=size,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
reuse=reuse,
name=name + '_' + str(i))
if activation:
input = activation(input)
if flatten:
assert layers_sizes[-1] == 1
input = tf.reshape(input, [-1])
return input
def learn(*,
network,
env,
total_timesteps,
timesteps_per_batch=1024, # what to train on
max_kl=0.001,
cg_iters=10,
gamma=0.99,
lam=1.0, # advantage estimation
seed=None,
ent_coef=0.0,
cg_damping=1e-2,
vf_stepsize=3e-4,
vf_iters=3,
max_episodes=0, max_iters=0, # time constraint
callback=None,
load_path=None,
**network_kwargs
):
'''
learn a policy function with TRPO algorithm
Parameters:
----------
network neural network to learn. Can be either string ('mlp', 'cnn', 'lstm', 'lnlstm' for basic types)
or function that takes input placeholder and returns tuple (output, None) for feedforward nets
or (output, (state_placeholder, state_output, mask_placeholder)) for recurrent nets
env environment (one of the gym environments or wrapped via tensorflow_code-pytorch.common.vec_env.VecEnv-type class
timesteps_per_batch timesteps per gradient estimation batch
max_kl max KL divergence between old policy and new policy ( KL(pi_old || pi) )
ent_coef coefficient of policy entropy term in the optimization objective
cg_iters number of iterations of conjugate gradient algorithm
cg_damping conjugate gradient damping
vf_stepsize learning rate for adam optimizer used to optimie value function loss
vf_iters number of iterations of value function optimization iterations per each policy optimization step
total_timesteps max number of timesteps
max_episodes max number of episodes
max_iters maximum number of policy optimization iterations
callback function to be called with (locals(), globals()) each policy optimization step
load_path str, path to load the model from (default: None, i.e. no model is loaded)
**network_kwargs keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
Returns:
-------
learnt model
'''
nworkers = MPI.COMM_WORLD.Get_size()
rank = MPI.COMM_WORLD.Get_rank()
cpus_per_worker = 1
U.get_session(config=tf.ConfigProto(
allow_soft_placement=True,
inter_op_parallelism_threads=cpus_per_worker,
intra_op_parallelism_threads=cpus_per_worker
))
policy = build_policy(env, network, value_network='copy', **network_kwargs)
set_global_seeds(seed)
np.set_printoptions(precision=3)
# Setup losses and stuff
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
ob = observation_placeholder(ob_space)
with tf.variable_scope("pi"):
pi = policy(observ_placeholder=ob)
with tf.variable_scope("oldpi"):
oldpi = policy(observ_placeholder=ob)
atarg = tf.placeholder(dtype=tf.float32, shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
ac = pi.pdtype.sample_placeholder([None])
kloldnew = oldpi.pd.kl(pi.pd)
ent = pi.pd.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = ent_coef * meanent
vferr = tf.reduce_mean(tf.square(pi.vf - ret))
ratio = tf.exp(pi.pd.logp(ac) - oldpi.pd.logp(ac)) # advantage * pnew / pold
surrgain = tf.reduce_mean(ratio * atarg)
optimgain = surrgain + entbonus
losses = [optimgain, meankl, entbonus, surrgain, meanent]
loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy"]
dist = meankl
all_var_list = get_trainable_variables("pi")
# var_list = [v for v in all_var_list if v.name.split("/")[1].startswith("pol")]
# vf_var_list = [v for v in all_var_list if v.name.split("/")[1].startswith("vf")]
var_list = get_pi_trainable_variables("pi")
vf_var_list = get_vf_trainable_variables("pi")
vfadam = MpiAdam(vf_var_list)
get_flat = U.GetFlat(var_list)
set_from_flat = U.SetFromFlat(var_list)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32, shape=[None], name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
sz = U.intprod(shape)
tangents.append(tf.reshape(flat_tangent[start:start + sz], shape))
start += sz
gvp = tf.add_n([tf.reduce_sum(g * tangent) for (g, tangent) in zipsame(klgrads, tangents)]) # pylint: disable=E1111
fvp = U.flatgrad(gvp, var_list)
assign_old_eq_new = U.function([], [], updates=[tf.assign(oldv, newv)
for (oldv, newv) in
zipsame(get_variables("oldpi"), get_variables("pi"))])
compute_losses = U.function([ob, ac, atarg], losses)
compute_lossandgrad = U.function([ob, ac, atarg], losses + [U.flatgrad(optimgain, var_list)])
compute_fvp = U.function([flat_tangent, ob, ac, atarg], fvp)
compute_vflossandgrad = U.function([ob, ret], U.flatgrad(vferr, vf_var_list))
@contextmanager
def timed(msg):
if rank == 0:
print(colorize(msg, color='magenta'))
tstart = time.time()
yield
print(colorize("done in %.3f seconds" % (time.time() - tstart), color='magenta'))
else:
yield
def allmean(x):
assert isinstance(x, np.ndarray)
out = np.empty_like(x)
MPI.COMM_WORLD.Allreduce(x, out, op=MPI.SUM)
out /= nworkers
return out
U.initialize()
if load_path is not None:
pi.load(load_path)
th_init = get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
set_from_flat(th_init)
vfadam.sync()
print("Init param sum", th_init.sum(), flush=True)
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi, env, timesteps_per_batch, stochastic=True)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=40) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=40) # rolling buffer for episode rewards
if sum([max_iters > 0, total_timesteps > 0, max_episodes > 0]) == 0:
# noththing to be done
return pi
assert sum([max_iters > 0, total_timesteps > 0, max_episodes > 0]) < 2, \
'out of max_iters, total_timesteps, and max_episodes only one should be specified'
while True:
if callback: callback(locals(), globals())
if total_timesteps and timesteps_so_far >= total_timesteps:
break
elif max_episodes and episodes_so_far >= max_episodes:
break
elif max_iters and iters_so_far >= max_iters:
break
logger.log("********** Iteration %i ************" % iters_so_far)
with timed("sampling"):
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, gamma, lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
ob, ac, atarg, tdlamret = seg["ob"], seg["ac"], seg["adv"], seg["tdlamret"]
vpredbefore = seg["vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()) / atarg.std() # standardized advantage function estimate
if hasattr(pi, "ret_rms"): pi.ret_rms.update(tdlamret)
if hasattr(pi, "ob_rms"): pi.ob_rms.update(ob) # update running mean/std for policy
args = seg["ob"], seg["ac"], atarg
fvpargs = [arr[::5] for arr in args]
def fisher_vector_product(p):
return allmean(compute_fvp(p, *fvpargs)) + cg_damping * p
assign_old_eq_new() # set old parameter values to new parameter values
with timed("computegrad"):
*lossbefore, g = compute_lossandgrad(*args)
lossbefore = allmean(np.array(lossbefore))
g = allmean(g)
if np.allclose(g, 0):
logger.log("Got zero gradient. not updating")
else:
with timed("cg"):
stepdir = cg(fisher_vector_product, g, cg_iters=cg_iters, verbose=rank == 0)
assert np.isfinite(stepdir).all()
shs = .5 * stepdir.dot(fisher_vector_product(stepdir))
lm = np.sqrt(shs / max_kl)
# logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g))
fullstep = stepdir / lm
expectedimprove = g.dot(fullstep)
surrbefore = lossbefore[0]
stepsize = 1.0
thbefore = get_flat()
for _ in range(10):
thnew = thbefore + fullstep * stepsize
set_from_flat(thnew)
meanlosses = surr, kl, *_ = allmean(np.array(compute_losses(*args)))
improve = surr - surrbefore
logger.log("Expected: %.3f Actual: %.3f" % (expectedimprove, improve))
if not np.isfinite(meanlosses).all():
logger.log("Got non-finite value of losses -- bad!")
elif kl > max_kl * 1.5:
logger.log("violated KL constraint. shrinking step.")
elif improve < 0:
logger.log("surrogate didn't improve. shrinking step.")
else:
logger.log("Stepsize OK!")
break
stepsize *= .5
else:
logger.log("couldn't compute a good step")
set_from_flat(thbefore)
if nworkers > 1 and iters_so_far % 20 == 0:
paramsums = MPI.COMM_WORLD.allgather((thnew.sum(), vfadam.getflat().sum())) # list of tuples
assert all(np.allclose(ps, paramsums[0]) for ps in paramsums[1:])
for (lossname, lossval) in zip(loss_names, meanlosses):
logger.record_tabular(lossname, lossval)
with timed("vf"):
for _ in range(vf_iters):
for (mbob, mbret) in dataset.iterbatches((seg["ob"], seg["tdlamret"]),
include_final_partial_batch=False, batch_size=64):
g = allmean(compute_vflossandgrad(mbob, mbret))
vfadam.update(g, vf_stepsize)
logger.record_tabular("ev_tdlam_before", explained_variance(vpredbefore, tdlamret))
lrlocal = (seg["ep_lens"], seg["ep_rets"]) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples
lens, rews = map(flatten_lists, zip(*listoflrpairs))
lenbuffer.extend(lens)
rewbuffer.extend(rews)
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - tstart)
if rank == 0:
logger.dump_tabular()
return pi
def _create_network(self, reuse=False):
logger.info("Creating a DDPG agent with action space %d x %s..." % (self.dimu, self.max_u))
self.sess = tf.get_default_session()
if self.sess is None:
self.sess = tf.InteractiveSession()
# running averages
with tf.variable_scope('o_stats') as vs:
if reuse:
vs.reuse_variables()
self.o_stats = Normalizer(self.dimo, self.norm_eps, self.norm_clip, sess=self.sess)
with tf.variable_scope('g_stats') as vs:
if reuse:
vs.reuse_variables()
self.g_stats = Normalizer(self.dimg, self.norm_eps, self.norm_clip, sess=self.sess)
# mini-batch sampling.
batch = self.staging_tf.get()
batch_tf = OrderedDict([(key, batch[i])
for i, key in enumerate(self.stage_shapes.keys())])
batch_tf['r'] = tf.reshape(batch_tf['r'], [-1, 1])
#choose only the demo buffer samples
mask = np.concatenate((np.zeros(self.batch_size - self.demo_batch_size), np.ones(self.demo_batch_size)), axis = 0)
# networks
with tf.variable_scope('main') as vs:
if reuse:
vs.reuse_variables()
self.main = self.create_actor_critic(batch_tf, net_type='main', **self.__dict__)
vs.reuse_variables()
with tf.variable_scope('target') as vs:
if reuse:
vs.reuse_variables()
target_batch_tf = batch_tf.copy()
target_batch_tf['o'] = batch_tf['o_2']
target_batch_tf['g'] = batch_tf['g_2']
self.target = self.create_actor_critic(
target_batch_tf, net_type='target', **self.__dict__)
vs.reuse_variables()
assert len(self._vars("main")) == len(self._vars("target"))
# loss functions
target_Q_pi_tf = self.target.Q_pi_tf
clip_range = (-self.clip_return, 0. if self.clip_pos_returns else np.inf)
target_tf = tf.clip_by_value(batch_tf['r'] + self.gamma * target_Q_pi_tf, *clip_range)
self.Q_loss_tf = tf.reduce_mean(tf.square(tf.stop_gradient(target_tf) - self.main.Q_tf))
if self.bc_loss ==1 and self.q_filter == 1 : # train with demonstrations and use bc_loss and q_filter both
maskMain = tf.reshape(tf.boolean_mask(self.main.Q_tf > self.main.Q_pi_tf, mask), [-1]) #where is the demonstrator action better than actor action according to the critic? choose those samples only
#define the cloning loss on the actor's actions only on the samples which adhere to the above masks
self.cloning_loss_tf = tf.reduce_sum(tf.square(tf.boolean_mask(tf.boolean_mask((self.main.pi_tf), mask), maskMain, axis=0) - tf.boolean_mask(tf.boolean_mask((batch_tf['u']), mask), maskMain, axis=0)))
self.pi_loss_tf = -self.prm_loss_weight * tf.reduce_mean(self.main.Q_pi_tf) #primary loss scaled by it's respective weight prm_loss_weight
self.pi_loss_tf += self.prm_loss_weight * self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u)) #L2 loss on action values scaled by the same weight prm_loss_weight
self.pi_loss_tf += self.aux_loss_weight * self.cloning_loss_tf #adding the cloning loss to the actor loss as an auxilliary loss scaled by its weight aux_loss_weight
elif self.bc_loss == 1 and self.q_filter == 0: # train with demonstrations without q_filter
self.cloning_loss_tf = tf.reduce_sum(tf.square(tf.boolean_mask((self.main.pi_tf), mask) - tf.boolean_mask((batch_tf['u']), mask)))
self.pi_loss_tf = -self.prm_loss_weight * tf.reduce_mean(self.main.Q_pi_tf)
self.pi_loss_tf += self.prm_loss_weight * self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u))
self.pi_loss_tf += self.aux_loss_weight * self.cloning_loss_tf
else: #If not training with demonstrations
self.pi_loss_tf = -tf.reduce_mean(self.main.Q_pi_tf)
self.pi_loss_tf += self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u))
self.pi_loss_tf = -tf.reduce_mean(self.main.Q_pi_tf)
self.pi_loss_tf += self.action_l2 * tf.reduce_mean(tf.square(self.main.pi_tf / self.max_u))
Q_grads_tf = tf.gradients(self.Q_loss_tf, self._vars('main/Q'))
pi_grads_tf = tf.gradients(self.pi_loss_tf, self._vars('main/pi'))
assert len(self._vars('main/Q')) == len(Q_grads_tf)
assert len(self._vars('main/pi')) == len(pi_grads_tf)
self.Q_grads_vars_tf = zip(Q_grads_tf, self._vars('main/Q'))
self.pi_grads_vars_tf = zip(pi_grads_tf, self._vars('main/pi'))
self.Q_grad_tf = flatten_grads(grads=Q_grads_tf, var_list=self._vars('main/Q'))
self.pi_grad_tf = flatten_grads(grads=pi_grads_tf, var_list=self._vars('main/pi'))
# optimizers
self.Q_adam = MpiAdam(self._vars('main/Q'), scale_grad_by_procs=False)
self.pi_adam = MpiAdam(self._vars('main/pi'), scale_grad_by_procs=False)
# polyak averaging
self.main_vars = self._vars('main/Q') + self._vars('main/pi')
self.target_vars = self._vars('target/Q') + self._vars('target/pi')
self.stats_vars = self._global_vars('o_stats') + self._global_vars('g_stats')
self.init_target_net_op = list(
map(lambda v: v[0].assign(v[1]), zip(self.target_vars, self.main_vars)))
self.update_target_net_op = list(
map(lambda v: v[0].assign(self.polyak * v[0] + (1. - self.polyak) * v[1]), zip(self.target_vars, self.main_vars)))
# initialize all variables
tf.variables_initializer(self._global_vars('')).run()
self._sync_optimizers()
self._init_target_net()
def flatten_grads(var_list, grads):
"""Flattens a variables and their gradients.
"""
return tf.concat(
[tf.reshape(grad, [U.numel(v)]) for (v, grad) in zip(var_list, grads)],
0)
def multi_modal_network_fp(dim_input=27,
dim_output=7,
batch_size=25,
network_config=None):
"""
An example a network in tf that has both state and image inputs, with the feature
point architecture (spatial softmax + expectation).
Args:
dim_input: Dimensionality of input.
dim_output: Dimensionality of the output.
batch_size: Batch size.
network_config: dictionary of network structure parameters
Returns:
A tfMap object that stores inputs, outputs, and scalar loss.
"""
n_layers = 3
layer_size = 20
dim_hidden = (n_layers - 1) * [layer_size]
dim_hidden.append(dim_output)
pool_size = 2
filter_size = 5
# List of indices for state (vector) data and image (tensor) data in observation.
x_idx, img_idx, i = [], [], 0
for sensor in network_config['obs_include']:
dim = network_config['sensor_dims'][sensor]
if sensor in network_config['obs_image_data']:
img_idx = img_idx + list(range(i, i + dim))
else:
x_idx = x_idx + list(range(i, i + dim))
i += dim
nn_input, action, precision = get_input_layer(dim_input, dim_output)
state_input = nn_input[:, 0:x_idx[-1] + 1]
image_input = nn_input[:, x_idx[-1] + 1:img_idx[-1] + 1]
# image goes through 3 convnet layers
num_filters = network_config['num_filters']
im_height = network_config['image_height']
im_width = network_config['image_width']
num_channels = network_config['image_channels']
image_input = tf.reshape(image_input,
[-1, num_channels, im_width, im_height])
image_input = tf.transpose(image_input, perm=[0, 3, 2, 1])
# we pool twice, each time reducing the image size by a factor of 2.
conv_out_size = int(im_width / (2.0 * pool_size) * im_height /
(2.0 * pool_size) * num_filters[1])
first_dense_size = conv_out_size + len(x_idx)
# Store layers weight & bias
with tf.variable_scope('conv_params'):
weights = {
'wc1':
init_weights(
[filter_size, filter_size, num_channels, num_filters[0]],
name='wc1'), # 5x5 conv, 1 input, 32 outputs
'wc2':
init_weights(
[filter_size, filter_size, num_filters[0], num_filters[1]],
name='wc2'), # 5x5 conv, 32 inputs, 64 outputs
'wc3':
init_weights(
[filter_size, filter_size, num_filters[1], num_filters[2]],
name='wc3'), # 5x5 conv, 32 inputs, 64 outputs
}
biases = {
'bc1': init_bias([num_filters[0]], name='bc1'),
'bc2': init_bias([num_filters[1]], name='bc2'),
'bc3': init_bias([num_filters[2]], name='bc3'),
}
conv_layer_0 = conv2d(img=image_input,
w=weights['wc1'],
b=biases['bc1'],
strides=[1, 2, 2, 1])
conv_layer_1 = conv2d(img=conv_layer_0, w=weights['wc2'], b=biases['bc2'])
conv_layer_2 = conv2d(img=conv_layer_1, w=weights['wc3'], b=biases['bc3'])
_, num_rows, num_cols, num_fp = conv_layer_2.get_shape()
num_rows, num_cols, num_fp = [int(x) for x in [num_rows, num_cols, num_fp]]
x_map = np.empty([num_rows, num_cols], np.float32)
y_map = np.empty([num_rows, num_cols], np.float32)
for i in range(num_rows):
for j in range(num_cols):
x_map[i, j] = (i - num_rows / 2.0) / num_rows
y_map[i, j] = (j - num_cols / 2.0) / num_cols
x_map = tf.convert_to_tensor(x_map)
y_map = tf.convert_to_tensor(y_map)
x_map = tf.reshape(x_map, [num_rows * num_cols])
y_map = tf.reshape(y_map, [num_rows * num_cols])
# rearrange features to be [batch_size, num_fp, num_rows, num_cols]
features = tf.reshape(tf.transpose(conv_layer_2, [0, 3, 1, 2]),
[-1, num_rows * num_cols])
softmax = tf.nn.softmax(features)
fp_x = tf.reduce_sum(tf.mul(x_map, softmax), [1], keep_dims=True)
fp_y = tf.reduce_sum(tf.mul(y_map, softmax), [1], keep_dims=True)
fp = tf.reshape(tf.concat(1, [fp_x, fp_y]), [-1, num_fp * 2])
fc_input = tf.concat(concat_dim=1, values=[fp, state_input])
fc_output, weights_FC, biases_FC = get_mlp_layers(fc_input, n_layers,
dim_hidden)
fc_vars = weights_FC + biases_FC
loss = euclidean_loss_layer(a=action,
b=fc_output,
precision=precision,
batch_size=batch_size)
nnet = TfMap.init_from_lists([nn_input, action, precision], [fc_output],
[loss],
fp=fp)
last_conv_vars = fc_input
return nnet, fc_vars, last_conv_vars
def multi_modal_network(dim_input=27,
dim_output=7,
batch_size=25,
network_config=None):
"""
An example a network in tf that has both state and image inputs.
Args:
dim_input: Dimensionality of input.
dim_output: Dimensionality of the output.
batch_size: Batch size.
network_config: dictionary of network structure parameters
Returns:
A tfMap object that stores inputs, outputs, and scalar loss.
"""
n_layers = 2
layer_size = 20
dim_hidden = (n_layers - 1) * [layer_size]
dim_hidden.append(dim_output)
pool_size = 2
filter_size = 3
# List of indices for state (vector) data and image (tensor) data in observation.
x_idx, img_idx, i = [], [], 0
for sensor in network_config['obs_include']:
dim = network_config['sensor_dims'][sensor]
if sensor in network_config['obs_image_data']:
img_idx = img_idx + list(range(i, i + dim))
else:
x_idx = x_idx + list(range(i, i + dim))
i += dim
nn_input, action, precision = get_input_layer(dim_input, dim_output)
state_input = nn_input[:, 0:x_idx[-1] + 1]
image_input = nn_input[:, x_idx[-1] + 1:img_idx[-1] + 1]
# image goes through 2 convnet layers
num_filters = network_config['num_filters']
im_height = network_config['image_height']
im_width = network_config['image_width']
num_channels = network_config['image_channels']
image_input = tf.reshape(image_input,
[-1, im_width, im_height, num_channels])
# we pool twice, each time reducing the image size by a factor of 2.
conv_out_size = int(im_width / (2.0 * pool_size) * im_height /
(2.0 * pool_size) * num_filters[1])
first_dense_size = conv_out_size + len(x_idx)
# Store layers weight & bias
weights = {
'wc1':
get_xavier_weights(
[filter_size, filter_size, num_channels, num_filters[0]],
(pool_size, pool_size)), # 5x5 conv, 1 input, 32 outputs
'wc2':
get_xavier_weights(
[filter_size, filter_size, num_filters[0], num_filters[1]],
(pool_size, pool_size)), # 5x5 conv, 32 inputs, 64 outputs
}
biases = {
'bc1': init_bias([num_filters[0]]),
'bc2': init_bias([num_filters[1]]),
}
conv_layer_0 = conv2d(img=image_input, w=weights['wc1'], b=biases['bc1'])
conv_layer_0 = max_pool(conv_layer_0, k=pool_size)
conv_layer_1 = conv2d(img=conv_layer_0, w=weights['wc2'], b=biases['bc2'])
conv_layer_1 = max_pool(conv_layer_1, k=pool_size)
conv_out_flat = tf.reshape(conv_layer_1, [-1, conv_out_size])
fc_input = tf.concat(concat_dim=1, values=[conv_out_flat, state_input])
fc_output, _, _ = get_mlp_layers(fc_input, n_layers, dim_hidden)
loss = euclidean_loss_layer(a=action,
b=fc_output,
precision=precision,
batch_size=batch_size)
return TfMap.init_from_lists([nn_input, action, precision], [fc_output],
[loss])