def test_MpiAdam():
np.random.seed(0)
tf.set_random_seed(0)
a = tf.Variable(np.random.randn(3).astype('float32'))
b = tf.Variable(np.random.randn(2, 5).astype('float32'))
loss = tf.reduce_sum(tf.square(a)) + tf.reduce_sum(tf.sin(b))
stepsize = 1e-2
update_op = tf.train.AdamOptimizer(stepsize).minimize(loss)
do_update = U.function([], loss, updates=[update_op])
tf.get_default_session().run(tf.global_variables_initializer())
for i in range(10):
print(i, do_update())
tf.set_random_seed(0)
tf.get_default_session().run(tf.global_variables_initializer())
var_list = [a, b]
lossandgrad = U.function([], [loss, U.flatgrad(loss, var_list)],
updates=[update_op])
adam = MpiAdam(var_list)
for i in range(10):
l, g = lossandgrad()
adam.update(g, stepsize)
print(i, l)
def kl(self, other):
return tf.reduce_sum(
tf.nn.sigmoid_cross_entropy_with_logits(logits=other.logits,
labels=self.ps),
axis=-1) - tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(
logits=self.logits, labels=self.ps),
axis=-1)
def kl(self, other):
a0 = self.logits - tf.reduce_max(self.logits, axis=-1, keepdims=True)
a1 = other.logits - tf.reduce_max(other.logits, axis=-1, keepdims=True)
ea0 = tf.exp(a0)
ea1 = tf.exp(a1)
z0 = tf.reduce_sum(ea0, axis=-1, keepdims=True)
z1 = tf.reduce_sum(ea1, axis=-1, keepdims=True)
p0 = ea0 / z0
return tf.reduce_sum(p0 * (a0 - tf.log(z0) - a1 + tf.log(z1)), axis=-1)
def euclidean_loss_layer(a, b, precision, batch_size):
""" Math: out = (action - mlp_out)'*precision*(action-mlp_out)
= (u-uhat)'*A*(u-uhat)"""
scale_factor = tf.constant(2*batch_size, dtype='float')
uP = batched_matrix_vector_multiply(a-b, precision)
uPu = tf.reduce_sum(uP*(a-b)) # this last dot product is then summed, so we just the sum all at once.
return uPu/scale_factor
def kl(self, other):
assert isinstance(other, DiagGaussianPd)
return tf.reduce_sum(
other.logstd - self.logstd +
(tf.square(self.std) + tf.square(self.mean - other.mean)) /
(2.0 * tf.square(other.std)) - 0.5,
axis=-1)
def kl_div(action_dist1, action_dist2, action_size):
mean1, std1 = action_dist1[:, :action_size], action_dist1[:, action_size:]
mean2, std2 = action_dist2[:, :action_size], action_dist2[:, action_size:]
numerator = tf.square(mean1 - mean2) + tf.square(std1) - tf.square(std2)
denominator = 2 * tf.square(std2) + 1e-8
return tf.reduce_sum(
numerator/denominator + tf.log(std2) - tf.log(std1),reduction_indices=-1)
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_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 learn(env,
policy_func,
reward_giver,
expert_dataset,
rank,
pretrained,
pretrained_weight,
*,
g_step,
d_step,
entcoeff,
save_per_iter,
ckpt_dir,
log_dir,
timesteps_per_batch,
task_name,
gamma,
lam,
max_kl,
cg_iters,
cg_damping=1e-2,
vf_stepsize=3e-4,
d_stepsize=3e-4,
vf_iters=3,
max_timesteps=0,
max_episodes=0,
max_iters=0,
callback=None):
nworkers = MPI.COMM_WORLD.Get_size()
rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
# Setup losses and stuff
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_func("pi",
ob_space,
ac_space,
reuse=(pretrained_weight != None))
oldpi = policy_func("oldpi", ob_space, ac_space)
atarg = tf.placeholder(
dtype=tf.float32,
shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
ob = U.get_placeholder_cached(name="ob")
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 = entcoeff * meanent
vferr = tf.reduce_mean(tf.square(pi.vpred - 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 = pi.get_trainable_variables()
var_list = [
v for v in all_var_list
if v.name.startswith("pi/pol") or v.name.startswith("pi/logstd")
]
vf_var_list = [v for v in all_var_list if v.name.startswith("pi/vff")]
assert len(var_list) == len(vf_var_list) + 1
d_adam = MpiAdam(reward_giver.get_trainable_variables())
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(oldpi.get_variables(), pi.get_variables())
])
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()
th_init = get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
set_from_flat(th_init)
d_adam.sync()
vfadam.sync()
if rank == 0:
print("Init param sum", th_init.sum(), flush=True)
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi,
env,
reward_giver,
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
true_rewbuffer = deque(maxlen=40)
assert sum([max_iters > 0, max_timesteps > 0, max_episodes > 0]) == 1
g_loss_stats = stats(loss_names)
d_loss_stats = stats(reward_giver.loss_name)
ep_stats = stats(["True_rewards", "Rewards", "Episode_length"])
# if provide pretrained weight
if pretrained_weight is not None:
U.load_state(pretrained_weight, var_list=pi.get_variables())
while True:
if callback: callback(locals(), globals())
if max_timesteps and timesteps_so_far >= max_timesteps:
break
elif max_episodes and episodes_so_far >= max_episodes:
break
elif max_iters and iters_so_far >= max_iters:
break
# Save model
if rank == 0 and iters_so_far % save_per_iter == 0 and ckpt_dir is not None:
fname = os.path.join(ckpt_dir, task_name)
os.makedirs(os.path.dirname(fname), exist_ok=True)
saver = tf.train.Saver()
saver.save(tf.get_default_session(), fname)
logger.log("********** Iteration %i ************" % iters_so_far)
def fisher_vector_product(p):
return allmean(compute_fvp(p, *fvpargs)) + cg_damping * p
# ------------------ Update G ------------------
logger.log("Optimizing Policy...")
for _ in range(g_step):
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, "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]
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:])
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=128):
if hasattr(pi, "ob_rms"):
pi.ob_rms.update(
mbob) # update running mean/std for policy
g = allmean(compute_vflossandgrad(mbob, mbret))
vfadam.update(g, vf_stepsize)
g_losses = meanlosses
for (lossname, lossval) in zip(loss_names, meanlosses):
logger.record_tabular(lossname, lossval)
logger.record_tabular("ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
# ------------------ Update D ------------------
logger.log("Optimizing Discriminator...")
logger.log(fmt_row(13, reward_giver.loss_name))
ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob))
batch_size = len(ob) // d_step
d_losses = [
] # list of tuples, each of which gives the loss for a minibatch
for ob_batch, ac_batch in dataset.iterbatches(
(ob, ac), include_final_partial_batch=False,
batch_size=batch_size):
ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob_batch))
# update running mean/std for reward_giver
if hasattr(reward_giver, "obs_rms"):
reward_giver.obs_rms.update(
np.concatenate((ob_batch, ob_expert), 0))
*newlosses, g = reward_giver.lossandgrad(ob_batch, ac_batch,
ob_expert, ac_expert)
d_adam.update(allmean(g), d_stepsize)
d_losses.append(newlosses)
logger.log(fmt_row(13, np.mean(d_losses, axis=0)))
lrlocal = (seg["ep_lens"], seg["ep_rets"], seg["ep_true_rets"]
) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples
lens, rews, true_rets = map(flatten_lists, zip(*listoflrpairs))
true_rewbuffer.extend(true_rets)
lenbuffer.extend(lens)
rewbuffer.extend(rews)
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
logger.record_tabular("EpTrueRewMean", np.mean(true_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()
def neglogp(self, x):
return tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(
logits=self.logits, labels=tf.to_float(x)),
axis=-1)
def entropy(self):
return tf.reduce_sum(self.logstd + .5 * np.log(2.0 * np.pi * np.e),
axis=-1)
def neglogp(self, x):
return 0.5 * tf.reduce_sum(tf.square((x - self.mean) / self.std), axis=-1) \
+ 0.5 * np.log(2.0 * np.pi) * tf.to_float(tf.shape(x)[-1]) \
+ tf.reduce_sum(self.logstd, axis=-1)
def entropy(self):
a0 = self.logits - tf.reduce_max(self.logits, axis=-1, keepdims=True)
ea0 = tf.exp(a0)
z0 = tf.reduce_sum(ea0, axis=-1, keepdims=True)
p0 = ea0 / z0
return tf.reduce_sum(p0 * (tf.log(z0) - a0), axis=-1)
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 entropy(self):
return tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(
logits=self.logits, labels=self.ps),
axis=-1)
def __init__(self, policy, ob_space, ac_space, nenvs, nsteps, nstack, num_procs,
ent_coef, q_coef, gamma, max_grad_norm, lr,
rprop_alpha, rprop_epsilon, total_timesteps, lrschedule,
c, trust_region, alpha, delta):
sess = get_session()
nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch]) # actions
D = tf.placeholder(tf.float32, [nbatch]) # dones
R = tf.placeholder(tf.float32, [nbatch]) # rewards, not returns
MU = tf.placeholder(tf.float32, [nbatch, nact]) # mu's
LR = tf.placeholder(tf.float32, [])
eps = 1e-6
step_ob_placeholder = tf.placeholder(dtype=ob_space.dtype, shape=(nenvs,) + ob_space.shape[:-1] + (ob_space.shape[-1] * nstack,))
train_ob_placeholder = tf.placeholder(dtype=ob_space.dtype, shape=(nenvs*(nsteps+1),) + ob_space.shape[:-1] + (ob_space.shape[-1] * nstack,))
with tf.variable_scope('acer_model', reuse=tf.AUTO_REUSE):
step_model = policy(observ_placeholder=step_ob_placeholder, sess=sess)
train_model = policy(observ_placeholder=train_ob_placeholder, sess=sess)
params = find_trainable_variables("acer_model")
print("Params {}".format(len(params)))
for var in params:
print(var)
# create polyak averaged model
ema = tf.train.ExponentialMovingAverage(alpha)
ema_apply_op = ema.apply(params)
def custom_getter(getter, *args, **kwargs):
v = ema.average(getter(*args, **kwargs))
print(v.name)
return v
with tf.variable_scope("acer_model", custom_getter=custom_getter, reuse=True):
polyak_model = policy(observ_placeholder=train_ob_placeholder, sess=sess)
# Notation: (var) = batch variable, (var)s = seqeuence variable, (var)_i = variable index by action at step i
# action probability distributions according to train_model, polyak_model and step_model
# poilcy.pi is probability distribution parameters; to obtain distribution that sums to 1 need to take softmax
train_model_p = tf.nn.softmax(train_model.pi)
polyak_model_p = tf.nn.softmax(polyak_model.pi)
step_model_p = tf.nn.softmax(step_model.pi)
v = tf.reduce_sum(train_model_p * train_model.q, axis = -1) # shape is [nenvs * (nsteps + 1)]
# strip off last step
f, f_pol, q = map(lambda var: strip(var, nenvs, nsteps), [train_model_p, polyak_model_p, train_model.q])
# Get pi and q values for actions taken
f_i = get_by_index(f, A)
q_i = get_by_index(q, A)
# Compute ratios for importance truncation
rho = f / (MU + eps)
rho_i = get_by_index(rho, A)
# Calculate Q_retrace targets
qret = q_retrace(R, D, q_i, v, rho_i, nenvs, nsteps, gamma)
# Calculate losses
# Entropy
# entropy = tf.reduce_mean(strip(train_model.pd.entropy(), nenvs, nsteps))
entropy = tf.reduce_mean(cat_entropy_softmax(f))
# Policy Graident loss, with truncated importance sampling & bias correction
v = strip(v, nenvs, nsteps, True)
check_shape([qret, v, rho_i, f_i], [[nenvs * nsteps]] * 4)
check_shape([rho, f, q], [[nenvs * nsteps, nact]] * 2)
# Truncated importance sampling
adv = qret - v
logf = tf.log(f_i + eps)
gain_f = logf * tf.stop_gradient(adv * tf.minimum(c, rho_i)) # [nenvs * nsteps]
loss_f = -tf.reduce_mean(gain_f)
# Bias correction for the truncation
adv_bc = (q - tf.reshape(v, [nenvs * nsteps, 1])) # [nenvs * nsteps, nact]
logf_bc = tf.log(f + eps) # / (f_old + eps)
check_shape([adv_bc, logf_bc], [[nenvs * nsteps, nact]]*2)
gain_bc = tf.reduce_sum(logf_bc * tf.stop_gradient(adv_bc * tf.nn.relu(1.0 - (c / (rho + eps))) * f), axis = 1) #IMP: This is sum, as expectation wrt f
loss_bc= -tf.reduce_mean(gain_bc)
loss_policy = loss_f + loss_bc
# Value/Q function loss, and explained variance
check_shape([qret, q_i], [[nenvs * nsteps]]*2)
ev = q_explained_variance(tf.reshape(q_i, [nenvs, nsteps]), tf.reshape(qret, [nenvs, nsteps]))
loss_q = tf.reduce_mean(tf.square(tf.stop_gradient(qret) - q_i)*0.5)
# Net loss
check_shape([loss_policy, loss_q, entropy], [[]] * 3)
loss = loss_policy + q_coef * loss_q - ent_coef * entropy
if trust_region:
g = tf.gradients(- (loss_policy - ent_coef * entropy) * nsteps * nenvs, f) #[nenvs * nsteps, nact]
# k = tf.gradients(KL(f_pol || f), f)
k = - f_pol / (f + eps) #[nenvs * nsteps, nact] # Directly computed gradient of KL divergence wrt f
k_dot_g = tf.reduce_sum(k * g, axis=-1)
adj = tf.maximum(0.0, (tf.reduce_sum(k * g, axis=-1) - delta) / (tf.reduce_sum(tf.square(k), axis=-1) + eps)) #[nenvs * nsteps]
# Calculate stats (before doing adjustment) for logging.
avg_norm_k = avg_norm(k)
avg_norm_g = avg_norm(g)
avg_norm_k_dot_g = tf.reduce_mean(tf.abs(k_dot_g))
avg_norm_adj = tf.reduce_mean(tf.abs(adj))
g = g - tf.reshape(adj, [nenvs * nsteps, 1]) * k
grads_f = -g/(nenvs*nsteps) # These are turst region adjusted gradients wrt f ie statistics of policy pi
grads_policy = tf.gradients(f, params, grads_f)
grads_q = tf.gradients(loss_q * q_coef, params)
grads = [gradient_add(g1, g2, param) for (g1, g2, param) in zip(grads_policy, grads_q, params)]
avg_norm_grads_f = avg_norm(grads_f) * (nsteps * nenvs)
norm_grads_q = tf.global_norm(grads_q)
norm_grads_policy = tf.global_norm(grads_policy)
else:
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, norm_grads = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=rprop_alpha, epsilon=rprop_epsilon)
_opt_op = trainer.apply_gradients(grads)
# so when you call _train, you first do the gradient step, then you apply ema
with tf.control_dependencies([_opt_op]):
_train = tf.group(ema_apply_op)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
# Ops/Summaries to run, and their names for logging
run_ops = [_train, loss, loss_q, entropy, loss_policy, loss_f, loss_bc, ev, norm_grads]
names_ops = ['loss', 'loss_q', 'entropy', 'loss_policy', 'loss_f', 'loss_bc', 'explained_variance',
'norm_grads']
if trust_region:
run_ops = run_ops + [norm_grads_q, norm_grads_policy, avg_norm_grads_f, avg_norm_k, avg_norm_g, avg_norm_k_dot_g,
avg_norm_adj]
names_ops = names_ops + ['norm_grads_q', 'norm_grads_policy', 'avg_norm_grads_f', 'avg_norm_k', 'avg_norm_g',
'avg_norm_k_dot_g', 'avg_norm_adj']
def train(obs, actions, rewards, dones, mus, states, masks, steps):
cur_lr = lr.value_steps(steps)
td_map = {train_model.X: obs, polyak_model.X: obs, A: actions, R: rewards, D: dones, MU: mus, LR: cur_lr}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
td_map[polyak_model.S] = states
td_map[polyak_model.M] = masks
return names_ops, sess.run(run_ops, td_map)[1:] # strip off _train
def _step(observation, **kwargs):
return step_model._evaluate([step_model.action, step_model_p, step_model.state], observation, **kwargs)
self.train = train
self.save = functools.partial(save_variables, sess=sess, variables=params)
self.train_model = train_model
self.step_model = step_model
self._step = _step
self.step = self.step_model.step
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
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
