def __init__(self,
env,
hidden_size,
entcoeff=0.001,
lr_rate=1e-3,
scope="adversary"):
self.scope = scope
self.observation_shape = env.observation_space.shape
self.actions_shape = env.action_space.shape
self.input_shape = tuple([
o + a for o, a in zip(self.observation_shape, self.actions_shape)
])
self.num_actions = env.action_space.shape[0]
self.hidden_size = hidden_size
self.build_ph()
# Build grpah
generator_logits = self.build_graph(self.generator_obs_ph,
self.generator_acs_ph,
reuse=False)
expert_logits = self.build_graph(self.expert_obs_ph,
self.expert_acs_ph,
reuse=True)
# Build accuracy
generator_acc = tf.reduce_mean(
tf.to_float(tf.nn.sigmoid(generator_logits) < 0.5))
expert_acc = tf.reduce_mean(
tf.to_float(tf.nn.sigmoid(expert_logits) > 0.5))
# Build regression loss
# let x = logits, z = targets.
# z * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x))
generator_loss = tf.nn.sigmoid_cross_entropy_with_logits(
logits=generator_logits, labels=tf.zeros_like(generator_logits))
generator_loss = tf.reduce_mean(generator_loss)
expert_loss = tf.nn.sigmoid_cross_entropy_with_logits(
logits=expert_logits, labels=tf.ones_like(expert_logits))
expert_loss = tf.reduce_mean(expert_loss)
# Build entropy loss
logits = tf.concat([generator_logits, expert_logits], 0)
entropy = tf.reduce_mean(logit_bernoulli_entropy(logits))
entropy_loss = -entcoeff * entropy
# Loss + Accuracy terms
self.losses = [
generator_loss, expert_loss, entropy, entropy_loss, generator_acc,
expert_acc
]
self.loss_name = [
"generator_loss", "expert_loss", "entropy", "entropy_loss",
"generator_acc", "expert_acc"
]
self.total_loss = generator_loss + expert_loss + entropy_loss
# Build Reward for policy
self.reward_op = -tf.log(1 - tf.nn.sigmoid(generator_logits) + 1e-8)
var_list = self.get_trainable_variables()
self.lossandgrad = U.function([
self.generator_obs_ph, self.generator_acs_ph, self.expert_obs_ph,
self.expert_acs_ph
], self.losses + [U.flatgrad(self.total_loss, var_list)])
def pdfromlatent(self, latent_vector, init_scale=1.0, init_bias=0.0):
mean = fc(latent_vector,
'pi',
self.size,
init_scale=init_scale,
init_bias=init_bias)
logstd = tf.get_variable(name='pi/logstd',
shape=[1, self.size],
initializer=tf.zeros_initializer())
pdparam = tf.concat([mean, mean * 0.0 + logstd], axis=1)
return self.pdfromflat(pdparam), mean
def __init__(self, inputs_tf, dimo, dimg, dimu, max_u, o_stats, g_stats,
hidden, layers, **kwargs):
"""The actor-critic network and related training pytorch.
Args:
inputs_tf (dict of tensors): all necessary inputs for the network: the
observation (o), the goal (g), and the action (u)
dimo (int): the dimension of the observations
dimg (int): the dimension of the goals
dimu (int): the dimension of the actions
max_u (float): the maximum magnitude of actions; action outputs will be scaled
accordingly
o_stats (baselines.her.Normalizer): normalizer for observations
g_stats (baselines.her.Normalizer): normalizer for goals
hidden (int): number of hidden units that should be used in hidden layers
layers (int): number of hidden layers
"""
self.o_tf = inputs_tf['o']
self.g_tf = inputs_tf['g']
self.u_tf = inputs_tf['u']
# Prepare inputs for actor and critic.
o = self.o_stats.normalize(self.o_tf)
g = self.g_stats.normalize(self.g_tf)
input_pi = tf.concat(axis=1, values=[o, g]) # for actor
# Networks.
with tf.variable_scope('pi'):
self.pi_tf = self.max_u * tf.tanh(
nn(input_pi, [self.hidden] * self.layers + [self.dimu]))
with tf.variable_scope('Q'):
# for policy training
input_Q = tf.concat(axis=1, values=[o, g, self.pi_tf / self.max_u])
self.Q_pi_tf = nn(input_Q, [self.hidden] * self.layers + [1])
# for critic training
input_Q = tf.concat(axis=1, values=[o, g, self.u_tf / self.max_u])
self._input_Q = input_Q # exposed for tests
self.Q_tf = nn(input_Q, [self.hidden] * self.layers + [1],
reuse=True)
def build_graph(self, obs_ph, acs_ph, reuse=False):
with tf.variable_scope(self.scope):
if reuse:
tf.get_variable_scope().reuse_variables()
with tf.variable_scope("obfilter"):
self.obs_rms = RunningMeanStd(shape=self.observation_shape)
obs = (obs_ph - self.obs_rms.mean / self.obs_rms.std)
_input = tf.concat(
[obs, acs_ph],
axis=1) # concatenate the two input -> form a transition
p_h1 = tf.contrib.layers.fully_connected(_input,
self.hidden_size,
activation_fn=tf.nn.tanh)
p_h2 = tf.contrib.layers.fully_connected(p_h1,
self.hidden_size,
activation_fn=tf.nn.tanh)
logits = tf.contrib.layers.fully_connected(
p_h2, 1, activation_fn=tf.identity)
return logits
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 getFactors(self, g, varlist):
graph = tf.get_default_graph()
factorTensors = {}
fpropTensors = []
bpropTensors = []
opTypes = []
fops = []
def searchFactors(gradient, graph):
# hard coded search stratergy
bpropOp = gradient.op
bpropOp_name = bpropOp.name
bTensors = []
fTensors = []
# combining additive gradient, assume they are the same op type and
# indepedent
if 'AddN' in bpropOp_name:
factors = []
for g in gradient.op.inputs:
factors.append(searchFactors(g, graph))
op_names = [item['opName'] for item in factors]
# TO-DO: need to check all the attribute of the ops as well
print(gradient.name)
print(op_names)
print(len(np.unique(op_names)))
assert len(np.unique(op_names)) == 1, gradient.name + \
' is shared among different computation OPs'
bTensors = reduce(lambda x, y: x + y,
[item['bpropFactors'] for item in factors])
if len(factors[0]['fpropFactors']) > 0:
fTensors = reduce(
lambda x, y: x + y,
[item['fpropFactors'] for item in factors])
fpropOp_name = op_names[0]
fpropOp = factors[0]['op']
else:
fpropOp_name = re.search(
'gradientsSampled(_[0-9]+|)/(.+?)_grad',
bpropOp_name).group(2)
fpropOp = graph.get_operation_by_name(fpropOp_name)
if fpropOp.op_def.name in KFAC_OPS:
# Known OPs
###
bTensor = [
i for i in bpropOp.inputs
if 'gradientsSampled' in i.name
][-1]
bTensorShape = fpropOp.outputs[0].get_shape()
if bTensor.get_shape()[0].value == None:
bTensor.set_shape(bTensorShape)
bTensors.append(bTensor)
###
if fpropOp.op_def.name == 'BiasAdd':
fTensors = []
else:
fTensors.append([
i for i in fpropOp.inputs
if param.op.name not in i.name
][0])
fpropOp_name = fpropOp.op_def.name
else:
# unknown OPs, block approximation used
bInputsList = [
i for i in bpropOp.inputs[0].op.inputs
if 'gradientsSampled' in i.name
if 'Shape' not in i.name
]
if len(bInputsList) > 0:
bTensor = bInputsList[0]
bTensorShape = fpropOp.outputs[0].get_shape()
if len(bTensor.get_shape()) > 0 and bTensor.get_shape(
)[0].value == None:
bTensor.set_shape(bTensorShape)
bTensors.append(bTensor)
fpropOp_name = opTypes.append('UNK-' + fpropOp.op_def.name)
return {
'opName': fpropOp_name,
'op': fpropOp,
'fpropFactors': fTensors,
'bpropFactors': bTensors
}
for t, param in zip(g, varlist):
if KFAC_DEBUG:
print(('get factor for ' + param.name))
factors = searchFactors(t, graph)
factorTensors[param] = factors
########
# check associated weights and bias for homogeneous coordinate representation
# and check redundent factors
# TO-DO: there may be a bug to detect associate bias and weights for
# forking layer, e.g. in inception models.
for param in varlist:
factorTensors[param]['assnWeights'] = None
factorTensors[param]['assnBias'] = None
for param in varlist:
if factorTensors[param]['opName'] == 'BiasAdd':
factorTensors[param]['assnWeights'] = None
for item in varlist:
if len(factorTensors[item]['bpropFactors']) > 0:
if (set(factorTensors[item]['bpropFactors']) == set(
factorTensors[param]['bpropFactors']
)) and (len(factorTensors[item]['fpropFactors']) > 0):
factorTensors[param]['assnWeights'] = item
factorTensors[item]['assnBias'] = param
factorTensors[param][
'bpropFactors'] = factorTensors[item][
'bpropFactors']
########
########
# concatenate the additive gradients along the batch dimension, i.e.
# assuming independence structure
for key in ['fpropFactors', 'bpropFactors']:
for i, param in enumerate(varlist):
if len(factorTensors[param][key]) > 0:
if (key + '_concat') not in factorTensors[param]:
name_scope = factorTensors[param][key][0].name.split(
':')[0]
with tf.name_scope(name_scope):
factorTensors[param][key + '_concat'] = tf.concat(
factorTensors[param][key], 0)
else:
factorTensors[param][key + '_concat'] = None
for j, param2 in enumerate(varlist[(i + 1):]):
if (len(factorTensors[param][key]) > 0) and (set(
factorTensors[param2][key]) == set(
factorTensors[param][key])):
factorTensors[param2][key] = factorTensors[param][key]
factorTensors[param2][
key + '_concat'] = factorTensors[param][key +
'_concat']
########
if KFAC_DEBUG:
for items in zip(varlist, fpropTensors, bpropTensors, opTypes):
print((items[0].name, factorTensors[item]))
self.factors = factorTensors
return factorTensors
def getKfacPrecondUpdates(self, gradlist, varlist):
updatelist = []
vg = 0.
assert len(self.stats) > 0
assert len(self.stats_eigen) > 0
assert len(self.factors) > 0
counter = 0
grad_dict = {var: grad for grad, var in zip(gradlist, varlist)}
for grad, var in zip(gradlist, varlist):
GRAD_RESHAPE = False
GRAD_TRANSPOSE = False
fpropFactoredFishers = self.stats[var]['fprop_concat_stats']
bpropFactoredFishers = self.stats[var]['bprop_concat_stats']
if (len(fpropFactoredFishers) + len(bpropFactoredFishers)) > 0:
counter += 1
GRAD_SHAPE = grad.get_shape()
if len(grad.get_shape()) > 2:
# reshape conv kernel parameters
KW = int(grad.get_shape()[0])
KH = int(grad.get_shape()[1])
C = int(grad.get_shape()[2])
D = int(grad.get_shape()[3])
if len(fpropFactoredFishers) > 1 and self._channel_fac:
# reshape conv kernel parameters into tensor
grad = tf.reshape(grad, [KW * KH, C, D])
else:
# reshape conv kernel parameters into 2D grad
grad = tf.reshape(grad, [-1, D])
GRAD_RESHAPE = True
elif len(grad.get_shape()) == 1:
# reshape bias or 1D parameters
D = int(grad.get_shape()[0])
grad = tf.expand_dims(grad, 0)
GRAD_RESHAPE = True
else:
# 2D parameters
C = int(grad.get_shape()[0])
D = int(grad.get_shape()[1])
if (self.stats[var]['assnBias']
is not None) and not self._blockdiag_bias:
# use homogeneous coordinates only works for 2D grad.
# TO-DO: figure out how to factorize bias grad
# stack bias grad
var_assnBias = self.stats[var]['assnBias']
grad = tf.concat(
[grad,
tf.expand_dims(grad_dict[var_assnBias], 0)], 0)
# project gradient to eigen space and reshape the eigenvalues
# for broadcasting
eigVals = []
for idx, stats in enumerate(
self.stats[var]['fprop_concat_stats']):
Q = self.stats_eigen[stats]['Q']
e = detectMinVal(self.stats_eigen[stats]['e'],
var,
name='act',
debug=KFAC_DEBUG)
Q, e = factorReshape(Q, e, grad, facIndx=idx, ftype='act')
eigVals.append(e)
grad = gmatmul(Q, grad, transpose_a=True, reduce_dim=idx)
for idx, stats in enumerate(
self.stats[var]['bprop_concat_stats']):
Q = self.stats_eigen[stats]['Q']
e = detectMinVal(self.stats_eigen[stats]['e'],
var,
name='grad',
debug=KFAC_DEBUG)
Q, e = factorReshape(Q, e, grad, facIndx=idx, ftype='grad')
eigVals.append(e)
grad = gmatmul(grad, Q, transpose_b=False, reduce_dim=idx)
##
#####
# whiten using eigenvalues
weightDecayCoeff = 0.
if var in self._weight_decay_dict:
weightDecayCoeff = self._weight_decay_dict[var]
if KFAC_DEBUG:
print(('weight decay coeff for %s is %f' %
(var.name, weightDecayCoeff)))
if self._factored_damping:
if KFAC_DEBUG:
print(('use factored damping for %s' % (var.name)))
coeffs = 1.
num_factors = len(eigVals)
# compute the ratio of two trace norm of the left and right
# KFac matrices, and their generalization
if len(eigVals) == 1:
damping = self._epsilon + weightDecayCoeff
else:
damping = tf.pow(self._epsilon + weightDecayCoeff,
1. / num_factors)
eigVals_tnorm_avg = [
tf.reduce_mean(tf.abs(e)) for e in eigVals
]
for e, e_tnorm in zip(eigVals, eigVals_tnorm_avg):
eig_tnorm_negList = [
item for item in eigVals_tnorm_avg
if item != e_tnorm
]
if len(eigVals) == 1:
adjustment = 1.
elif len(eigVals) == 2:
adjustment = tf.sqrt(e_tnorm /
eig_tnorm_negList[0])
else:
eig_tnorm_negList_prod = reduce(
lambda x, y: x * y, eig_tnorm_negList)
adjustment = tf.pow(
tf.pow(e_tnorm, num_factors - 1.) /
eig_tnorm_negList_prod, 1. / num_factors)
coeffs *= (e + adjustment * damping)
else:
coeffs = 1.
damping = (self._epsilon + weightDecayCoeff)
for e in eigVals:
coeffs *= e
coeffs += damping
#grad = tf.Print(grad, [tf.convert_to_tensor('1'), tf.convert_to_tensor(var.name), grad.get_shape()])
grad /= coeffs
#grad = tf.Print(grad, [tf.convert_to_tensor('2'), tf.convert_to_tensor(var.name), grad.get_shape()])
#####
# project gradient back to euclidean space
for idx, stats in enumerate(
self.stats[var]['fprop_concat_stats']):
Q = self.stats_eigen[stats]['Q']
grad = gmatmul(Q, grad, transpose_a=False, reduce_dim=idx)
for idx, stats in enumerate(
self.stats[var]['bprop_concat_stats']):
Q = self.stats_eigen[stats]['Q']
grad = gmatmul(grad, Q, transpose_b=True, reduce_dim=idx)
##
#grad = tf.Print(grad, [tf.convert_to_tensor('3'), tf.convert_to_tensor(var.name), grad.get_shape()])
if (self.stats[var]['assnBias']
is not None) and not self._blockdiag_bias:
# use homogeneous coordinates only works for 2D grad.
# TO-DO: figure out how to factorize bias grad
# un-stack bias grad
var_assnBias = self.stats[var]['assnBias']
C_plus_one = int(grad.get_shape()[0])
grad_assnBias = tf.reshape(
tf.slice(grad, begin=[C_plus_one - 1, 0], size=[1,
-1]),
var_assnBias.get_shape())
grad_assnWeights = tf.slice(grad,
begin=[0, 0],
size=[C_plus_one - 1, -1])
grad_dict[var_assnBias] = grad_assnBias
grad = grad_assnWeights
#grad = tf.Print(grad, [tf.convert_to_tensor('4'), tf.convert_to_tensor(var.name), grad.get_shape()])
if GRAD_RESHAPE:
grad = tf.reshape(grad, GRAD_SHAPE)
grad_dict[var] = grad
print(('projecting %d gradient matrices' % counter))
for g, var in zip(gradlist, varlist):
grad = grad_dict[var]
### clipping ###
if KFAC_DEBUG:
print(('apply clipping to %s' % (var.name)))
tf.Print(grad, [tf.sqrt(tf.reduce_sum(tf.pow(grad, 2)))],
"Euclidean norm of new grad")
local_vg = tf.reduce_sum(grad * g * (self._lr * self._lr))
vg += local_vg
# recale everything
if KFAC_DEBUG:
print('apply vFv clipping')
scaling = tf.minimum(1., tf.sqrt(self._clip_kl / vg))
if KFAC_DEBUG:
scaling = tf.Print(scaling, [
tf.convert_to_tensor('clip: '), scaling,
tf.convert_to_tensor(' vFv: '), vg
])
with tf.control_dependencies([tf.assign(self.vFv, vg)]):
updatelist = [grad_dict[var] for var in varlist]
for i, item in enumerate(updatelist):
updatelist[i] = scaling * item
return updatelist
def compute_stats(self, loss_sampled, var_list=None):
varlist = var_list
if varlist is None:
varlist = tf.trainable_variables()
gs = tf.gradients(loss_sampled, varlist, name='gradientsSampled')
self.gs = gs
factors = self.getFactors(gs, varlist)
stats = self.getStats(factors, varlist)
updateOps = []
statsUpdates = {}
statsUpdates_cache = {}
for var in varlist:
opType = factors[var]['opName']
fops = factors[var]['op']
fpropFactor = factors[var]['fpropFactors_concat']
fpropStats_vars = stats[var]['fprop_concat_stats']
bpropFactor = factors[var]['bpropFactors_concat']
bpropStats_vars = stats[var]['bprop_concat_stats']
SVD_factors = {}
for stats_var in fpropStats_vars:
stats_var_dim = int(stats_var.get_shape()[0])
if stats_var not in statsUpdates_cache:
old_fpropFactor = fpropFactor
B = (tf.shape(fpropFactor)[0]) # batch size
if opType == 'Conv2D':
strides = fops.get_attr("strides")
padding = fops.get_attr("padding")
convkernel_size = var.get_shape()[0:3]
KH = int(convkernel_size[0])
KW = int(convkernel_size[1])
C = int(convkernel_size[2])
flatten_size = int(KH * KW * C)
Oh = int(bpropFactor.get_shape()[1])
Ow = int(bpropFactor.get_shape()[2])
if Oh == 1 and Ow == 1 and self._channel_fac:
# factorization along the channels
# assume independence among input channels
# factor = B x 1 x 1 x (KH xKW x C)
# patches = B x Oh x Ow x (KH xKW x C)
if len(SVD_factors) == 0:
if KFAC_DEBUG:
print((
'approx %s act factor with rank-1 SVD factors'
% (var.name)))
# find closest rank-1 approx to the feature map
S, U, V = tf.batch_svd(
tf.reshape(fpropFactor, [-1, KH * KW, C]))
# get rank-1 approx slides
sqrtS1 = tf.expand_dims(tf.sqrt(S[:, 0, 0]), 1)
patches_k = U[:, :, 0] * sqrtS1 # B x KH*KW
full_factor_shape = fpropFactor.get_shape()
patches_k.set_shape(
[full_factor_shape[0], KH * KW])
patches_c = V[:, :, 0] * sqrtS1 # B x C
patches_c.set_shape([full_factor_shape[0], C])
SVD_factors[C] = patches_c
SVD_factors[KH * KW] = patches_k
fpropFactor = SVD_factors[stats_var_dim]
else:
# poor mem usage implementation
patches = tf.extract_image_patches(
fpropFactor,
ksizes=[
1, convkernel_size[0], convkernel_size[1],
1
],
strides=strides,
rates=[1, 1, 1, 1],
padding=padding)
if self._approxT2:
if KFAC_DEBUG:
print(('approxT2 act fisher for %s' %
(var.name)))
# T^2 terms * 1/T^2, size: B x C
fpropFactor = tf.reduce_mean(patches, [1, 2])
else:
# size: (B x Oh x Ow) x C
fpropFactor = tf.reshape(
patches, [-1, flatten_size]) / Oh / Ow
fpropFactor_size = int(fpropFactor.get_shape()[-1])
if stats_var_dim == (fpropFactor_size +
1) and not self._blockdiag_bias:
if opType == 'Conv2D' and not self._approxT2:
# correct padding for numerical stability (we
# divided out OhxOw from activations for T1 approx)
fpropFactor = tf.concat([
fpropFactor,
tf.ones([tf.shape(fpropFactor)[0], 1]) / Oh /
Ow
], 1)
else:
# use homogeneous coordinates
fpropFactor = tf.concat([
fpropFactor,
tf.ones([tf.shape(fpropFactor)[0], 1])
], 1)
# average over the number of data points in a batch
# divided by B
cov = tf.matmul(fpropFactor, fpropFactor,
transpose_a=True) / tf.cast(B, tf.float32)
updateOps.append(cov)
statsUpdates[stats_var] = cov
if opType != 'Conv2D':
# HACK: for convolution we recompute fprop stats for
# every layer including forking layers
statsUpdates_cache[stats_var] = cov
for stats_var in bpropStats_vars:
stats_var_dim = int(stats_var.get_shape()[0])
if stats_var not in statsUpdates_cache:
old_bpropFactor = bpropFactor
bpropFactor_shape = bpropFactor.get_shape()
B = tf.shape(bpropFactor)[0] # batch size
C = int(bpropFactor_shape[-1]) # num channels
if opType == 'Conv2D' or len(bpropFactor_shape) == 4:
if fpropFactor is not None:
if self._approxT2:
if KFAC_DEBUG:
print(('approxT2 grad fisher for %s' %
(var.name)))
bpropFactor = tf.reduce_sum(
bpropFactor, [1, 2]) # T^2 terms * 1/T^2
else:
bpropFactor = tf.reshape(
bpropFactor,
[-1, C]) * Oh * Ow # T * 1/T terms
else:
# just doing block diag approx. spatial independent
# structure does not apply here. summing over
# spatial locations
if KFAC_DEBUG:
print(('block diag approx fisher for %s' %
(var.name)))
bpropFactor = tf.reduce_sum(bpropFactor, [1, 2])
# assume sampled loss is averaged. TO-DO:figure out better
# way to handle this
bpropFactor *= tf.to_float(B)
##
cov_b = tf.matmul(bpropFactor,
bpropFactor,
transpose_a=True) / tf.to_float(
tf.shape(bpropFactor)[0])
updateOps.append(cov_b)
statsUpdates[stats_var] = cov_b
statsUpdates_cache[stats_var] = cov_b
if KFAC_DEBUG:
aKey = list(statsUpdates.keys())[0]
statsUpdates[aKey] = tf.Print(statsUpdates[aKey], [
tf.convert_to_tensor('step:'),
self.global_step,
tf.convert_to_tensor('computing stats'),
])
self.statsUpdates = statsUpdates
return statsUpdates
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])
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 _init(self,
ob_space,
ac_space,
hid_size,
num_hid_layers,
gaussian_fixed_var=True):
assert isinstance(ob_space, gym.spaces.Box)
self.pdtype = pdtype = make_pdtype(ac_space)
sequence_length = None
ob = U.get_placeholder(name="ob",
dtype=tf.float32,
shape=[sequence_length] + list(ob_space.shape))
with tf.variable_scope("obfilter"):
self.ob_rms = RunningMeanStd(shape=ob_space.shape)
obz = tf.clip_by_value((ob - self.ob_rms.mean) / self.ob_rms.std, -5.0,
5.0)
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(
dense(last_out,
hid_size,
"vffc%i" % (i + 1),
weight_init=U.normc_initializer(1.0)))
self.vpred = dense(last_out,
1,
"vffinal",
weight_init=U.normc_initializer(1.0))[:, 0]
last_out = obz
for i in range(num_hid_layers):
last_out = tf.nn.tanh(
dense(last_out,
hid_size,
"polfc%i" % (i + 1),
weight_init=U.normc_initializer(1.0)))
if gaussian_fixed_var and isinstance(ac_space, gym.spaces.Box):
mean = dense(last_out,
pdtype.param_shape()[0] // 2, "polfinal",
U.normc_initializer(0.01))
logstd = tf.get_variable(name="logstd",
shape=[1, pdtype.param_shape()[0] // 2],
initializer=tf.zeros_initializer())
pdparam = tf.concat([mean, mean * 0.0 + logstd], axis=1)
else:
pdparam = dense(last_out,
pdtype.param_shape()[0], "polfinal",
U.normc_initializer(0.01))
self.pd = pdtype.pdfromflat(pdparam)
self.state_in = []
self.state_out = []
# change for BC
stochastic = U.get_placeholder(name="stochastic",
dtype=tf.bool,
shape=())
ac = U.switch(stochastic, self.pd.sample(), self.pd.mode())
self.ac = ac
self._act = U.function([stochastic, ob], [ac, self.vpred])
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 __call__(self, obs, action, reuse=False):
with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
x = tf.concat([obs, action], axis=-1) # this assumes observation and action can be concatenated
x = self.network_builder(x)
x = tf.layers.dense(x, 1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
return x
