def __init__(self, n_actions):
Serializable.__init__(self, n_actions)
cgt.set_precision('double')
n_in = 128
o_no = cgt.matrix("o_no",fixed_shape=(None,n_in))
a_n = cgt.vector("a_n",dtype='i8')
q_n = cgt.vector("q_n")
oldpdist_np = cgt.matrix("oldpdists")
h0 = (o_no - 128.0)/128.0
nhid = 64
h1 = cgt.tanh(nn.Affine(128,nhid,weight_init=nn.IIDGaussian(std=.1))(h0))
probs_na = nn.softmax(nn.Affine(nhid,n_actions,weight_init=nn.IIDGaussian(std=0.01))(h1))
logprobs_na = cgt.log(probs_na)
b = cgt.size(o_no, 0)
logps_n = logprobs_na[cgt.arange(b), a_n]
surr = (logps_n*q_n).mean()
kl = (oldpdist_np * cgt.log(oldpdist_np/probs_na)).sum(axis=1).mean()
params = nn.get_parameters(surr)
gradsurr = cgt.grad(surr, params)
flatgrad = cgt.concatenate([p.flatten() for p in gradsurr])
lam = cgt.scalar()
penobj = surr - lam * kl
self._f_grad_lagrangian = cgt.function([lam, oldpdist_np, o_no, a_n, q_n],
cgt.concatenate([p.flatten() for p in cgt.grad(penobj,params)]))
self.f_pdist = cgt.function([o_no], probs_na)
self.f_probs = cgt.function([o_no], probs_na)
self.f_surr_kl = cgt.function([oldpdist_np, o_no, a_n, q_n], [surr, kl])
self.f_gradlogp = cgt.function([oldpdist_np, o_no, a_n, q_n], flatgrad)
self.pc = ParamCollection(params)
def __init__(self, input_feature_size, input_time_size, num_units,
weight_init=HeUniform(),
activation=cgt.sigmoid,
cell_out_init=IIDUniform(-0.1, 0.1),
hid_out_init=IIDUniform(-0.1, 0.1),
#cell_out_init=Constant(0.0),
#hid_out_init=Constant(0.0),
backwards=False):
ingate = Gate(W_in=weight_init, W_hid=weight_init, W_cell=weight_init, nonlinearity=activation)
forgetgate = Gate(W_in=weight_init, W_hid=weight_init, W_cell=weight_init, nonlinearity=activation)
cell = Gate(W_cell=None, nonlinearity=cgt.tanh)
outgate = Gate(W_in=weight_init, W_hid=weight_init, W_cell=weight_init, nonlinearity=activation)
self.nonlinearity = activation
self.num_units = num_units
self.backwards = backwards
self.timesteps = input_time_size
def add_gate_params(gate, gate_name):
""" Convenience function for adding layer parameters from a Gate
instance. """
return (parameter(init_array(gate.W_in, (input_feature_size, num_units)), name=None),
parameter(init_array(gate.W_hid, (num_units, num_units)), name=None),
parameter(init_array(gate.b, (1, num_units)), name=None),
gate.nonlinearity)
# Add in parameters from the supplied Gate instances
(self.W_in_to_ingate, self.W_hid_to_ingate, self.b_ingate, self.nonlinearity_ingate) = add_gate_params(ingate, 'ingate')
(self.W_in_to_forgetgate, self.W_hid_to_forgetgate, self.b_forgetgate, self.nonlinearity_forgetgate) = add_gate_params(forgetgate, 'forgetgate')
(self.W_in_to_cell, self.W_hid_to_cell, self.b_cell, self.nonlinearity_cell) = add_gate_params(cell, 'cell')
(self.W_in_to_outgate, self.W_hid_to_outgate, self.b_outgate, self.nonlinearity_outgate) = add_gate_params(outgate, 'outgate')
self.hid_init = parameter(init_array(hid_out_init, (1, num_units)), name=None)
self.cell_init = parameter(init_array(cell_out_init, (1, num_units)), name=None)
# Stack input weight matrices into a (num_inputs, 4*num_units) #checks out
# matrix, which speeds up computation
self.W_in_stacked = cgt.concatenate(
[self.W_in_to_ingate, self.W_in_to_forgetgate,
self.W_in_to_cell, self.W_in_to_outgate], axis=1)
# Same for hidden weight matrices
self.W_hid_stacked = cgt.concatenate(
[self.W_hid_to_ingate, self.W_hid_to_forgetgate,
self.W_hid_to_cell, self.W_hid_to_outgate], axis=1)
# Stack biases into a (4*num_units) vector
self.b_stacked = cgt.concatenate(
[self.b_ingate, self.b_forgetgate,
self.b_cell, self.b_outgate], axis=1)
self.cell_prev = None
self.hid_prev = None
def circ_conv_1d(wg_bhn, s_bh3, axis=2):
"VERY inefficient way to implement circular convolution for the special case of filter size 3"
assert axis == 2
n = cgt.size(wg_bhn,2)
wback = cgt.concatenate([wg_bhn[:,:,n-1:n], wg_bhn[:,:,:n-1]], axis=2)
w = wg_bhn
wfwd = cgt.concatenate([wg_bhn[:,:,1:n], wg_bhn[:,:,0:1]], axis=2)
return cgt.broadcast("*", s_bh3[:,:,0:1] , wback, "xx1,xxx")\
+ cgt.broadcast("*", s_bh3[:,:,1:2] , w, "xx1,xxx")\
+ cgt.broadcast("*", s_bh3[:,:,2:3] , wfwd, "xx1,xxx")
def circ_conv_1d(wg_bhn, s_bh3, axis=2):
"VERY inefficient way to implement circular convolution for the special case of filter size 3"
assert axis == 2
n = cgt.size(wg_bhn,2)
wback = cgt.concatenate([wg_bhn[:,:,n-1:n], wg_bhn[:,:,:n-1]], axis=2)
w = wg_bhn
wfwd = cgt.concatenate([wg_bhn[:,:,1:n], wg_bhn[:,:,0:1]], axis=2)
return cgt.broadcast("*", s_bh3[:,:,0:1] , wback, "xx1,xxx")\
+ cgt.broadcast("*", s_bh3[:,:,1:2] , w, "xx1,xxx")\
+ cgt.broadcast("*", s_bh3[:,:,2:3] , wfwd, "xx1,xxx")
def __init__(self, obs_dim, ctrl_dim):
cgt.set_precision('double')
Serializable.__init__(self, obs_dim, ctrl_dim)
self.obs_dim = obs_dim
self.ctrl_dim = ctrl_dim
o_no = cgt.matrix("o_no",fixed_shape=(None,obs_dim))
a_na = cgt.matrix("a_na",fixed_shape = (None, ctrl_dim))
adv_n = cgt.vector("adv_n")
oldpdist_np = cgt.matrix("oldpdist", fixed_shape=(None, 2*ctrl_dim))
self.logstd = logstd_1a = nn.parameter(np.zeros((1, self.ctrl_dim)), name="std_1a")
std_1a = cgt.exp(logstd_1a)
# Here's where we apply the network
h0 = o_no
nhid = 32
h1 = cgt.tanh(nn.Affine(obs_dim,nhid,weight_init=nn.IIDGaussian(std=0.1))(h0))
h2 = cgt.tanh(nn.Affine(nhid,nhid,weight_init=nn.IIDGaussian(std=0.1))(h1))
mean_na = nn.Affine(nhid,ctrl_dim,weight_init=nn.IIDGaussian(std=0.01))(h2)
b = cgt.size(o_no, 0)
std_na = cgt.repeat(std_1a, b, axis=0)
oldmean_na = oldpdist_np[:, 0:self.ctrl_dim]
oldstd_na = oldpdist_np[:, self.ctrl_dim:2*self.ctrl_dim]
logp_n = ((-.5) * cgt.square( (a_na - mean_na) / std_na ).sum(axis=1)) - logstd_1a.sum()
oldlogp_n = ((-.5) * cgt.square( (a_na - oldmean_na) / oldstd_na ).sum(axis=1)) - cgt.log(oldstd_na).sum(axis=1)
ratio_n = cgt.exp(logp_n - oldlogp_n)
surr = (ratio_n*adv_n).mean()
pdists_np = cgt.concatenate([mean_na, std_na], axis=1)
# kl = cgt.log(sigafter/)
params = nn.get_parameters(surr)
oldvar_na = cgt.square(oldstd_na)
var_na = cgt.square(std_na)
kl = (cgt.log(std_na / oldstd_na) + (oldvar_na + cgt.square(oldmean_na - mean_na)) / (2 * var_na) - .5).sum(axis=1).mean()
lam = cgt.scalar()
penobj = surr - lam * kl
self._compute_surr_kl = cgt.function([oldpdist_np, o_no, a_na, adv_n], [surr, kl])
self._compute_grad_lagrangian = cgt.function([lam, oldpdist_np, o_no, a_na, adv_n],
cgt.concatenate([p.flatten() for p in cgt.grad(penobj,params)]))
self.f_pdist = cgt.function([o_no], pdists_np)
self.f_objs = cgt.function([oldpdist_np, o_no, a_na, adv_n], [surr, kl])
self.pc = ParamCollection(params)
def make_ff_controller(opt):
b, h, m, p, k = opt.b, opt.h, opt.m, opt.p, opt.k
H = 2*h
in_size = k + h*m
out_size = H*m + H + H + H*3 + H + h*m + h*m + p
# Previous reads
r_bhm = cgt.tensor3("r", fixed_shape = (b,h,m))
# External inputs
X_bk = cgt.matrix("x", fixed_shape = (b,k))
r_b_hm = r_bhm.reshape([r_bhm.shape[0], r_bhm.shape[1]*r_bhm.shape[2]])
# Input to controller
inp_bq = cgt.concatenate([X_bk, r_b_hm], axis=1)
hid_sizes = opt.ff_hid_sizes
activation = cgt.tanh
layer_out_sizes = [in_size] + hid_sizes + [out_size]
last_out = inp_bq
# feedforward part. we could simplify a bit by using nn.Affine
for i in xrange(len(layer_out_sizes)-1):
indim = layer_out_sizes[i]
outdim = layer_out_sizes[i+1]
W = cgt.shared(.02*nr.randn(indim, outdim), name="W%i"%i, fixed_shape_mask="all")
bias = cgt.shared(.02*nr.randn(1, outdim), name="b%i"%i, fixed_shape_mask="all")
last_out = cgt.broadcast("+",last_out.dot(W),bias,"xx,1x")
# Don't apply nonlinearity at the last layer
if i != len(layer_out_sizes)-2: last_out = activation(last_out)
idx = 0
k_bHm = last_out[:,idx:idx+H*m]; idx += H*m; k_bHm = k_bHm.reshape([b,H,m])
beta_bH = last_out[:,idx:idx+H]; idx += H
g_bH = last_out[:,idx:idx+H]; idx += H
s_bH3 = last_out[:,idx:idx+3*H]; idx += 3*H; s_bH3 = s_bH3.reshape([b,H,3])
gamma_bH = last_out[:,idx:idx+H]; idx += H
e_bhm = last_out[:,idx:idx+h*m]; idx += h*m; e_bhm = e_bhm.reshape([b,h,m])
a_bhm = last_out[:,idx:idx+h*m]; idx += h*m; a_bhm = a_bhm.reshape([b,h,m])
y_bp = last_out[:,idx:idx+p]; idx += p
k_bHm = cgt.tanh(k_bHm)
beta_bH = nn.softplus(beta_bH)
g_bH = cgt.sigmoid(g_bH)
s_bH3 = sum_normalize2(cgt.exp(s_bH3))
gamma_bH = cgt.sigmoid(gamma_bH)+1
e_bhm = cgt.sigmoid(e_bhm)
a_bhm = cgt.tanh(a_bhm)
# y_bp = y_bp
assert infer_shape(k_bHm) == (b,H,m)
assert infer_shape(beta_bH) == (b,H)
assert infer_shape(g_bH) == (b,H)
assert infer_shape(s_bH3) == (b,H,3)
assert infer_shape(gamma_bH) == (b,H)
assert infer_shape(e_bhm) == (b,h,m)
assert infer_shape(a_bhm) == (b,h,m)
assert infer_shape(y_bp) == (b,p)
return nn.Module([r_bhm, X_bk], [k_bHm, beta_bH, g_bH, s_bH3, gamma_bH, e_bhm, a_bhm, y_bp])
def make_funcs(config, dbg_out={}):
net_in, net_out = hybrid_network(config['num_inputs'], config['num_outputs'],
config['num_units'], config['num_sto'],
dbg_out=dbg_out)
if not config['dbg_out_full']: dbg_out = {}
# def f_sample(_inputs, num_samples=1, flatten=False):
# _mean, _var = f_step(_inputs)
# _samples = []
# for _m, _v in zip(_mean, _var):
# _s = np.random.multivariate_normal(_m, np.diag(np.sqrt(_v)), num_samples)
# if flatten: _samples.extend(_s)
# else: _samples.append(_s)
# return np.array(_samples)
Y_gt = cgt.matrix("Y")
Y_prec = cgt.tensor3('V', fixed_shape=(None, config['num_inputs'], config['num_inputs']))
params = nn.get_parameters(net_out)
size_batch, size_out = net_out.shape
inputs, outputs = [net_in], [net_out]
if config['no_bias']:
print "Excluding bias"
params = [p for p in params if not p.name.endswith(".b")]
loss_vec = dist.gaussian.logprob(Y_gt, net_out, Y_prec)
if config['weight_decay'] > 0.:
print "Applying penalty on parameter norm"
params_flat = cgt.concatenate([p.flatten() for p in params])
loss_param = config['weight_decay'] * cgt.sum(params_flat ** 2)
loss_vec -= loss_param # / size_batch
loss = cgt.sum(loss_vec) / size_batch
# TODO_TZ f_step seems not to fail if X has wrong dim
f_step = cgt.function(inputs, outputs)
f_surr = get_surrogate_func(inputs + [Y_prec, Y_gt], outputs,
[loss_vec], params, _dbg_out=dbg_out)
return params, f_step, None, None, None, f_surr
def make_ff_controller(opt):
b, h, m, p, k = opt.b, opt.h, opt.m, opt.p, opt.k
H = 2*h
in_size = k + h*m
out_size = H*m + H + H + H*3 + H + h*m + h*m + p
# Previous reads
r_bhm = cgt.tensor3("r", fixed_shape = (b,h,m))
# External inputs
X_bk = cgt.matrix("x", fixed_shape = (b,k))
r_b_hm = r_bhm.reshape([r_bhm.shape[0], r_bhm.shape[1]*r_bhm.shape[2]])
# Input to controller
inp_bq = cgt.concatenate([X_bk, r_b_hm], axis=1)
hid_sizes = opt.ff_hid_sizes
activation = cgt.tanh
layer_out_sizes = [in_size] + hid_sizes + [out_size]
last_out = inp_bq
# feedforward part. we could simplify a bit by using nn.Affine
for i in xrange(len(layer_out_sizes)-1):
indim = layer_out_sizes[i]
outdim = layer_out_sizes[i+1]
W = cgt.shared(.02*nr.randn(indim, outdim), name="W%i"%i, fixed_shape_mask="all")
bias = cgt.shared(.02*nr.randn(1, outdim), name="b%i"%i, fixed_shape_mask="all")
last_out = cgt.broadcast("+",last_out.dot(W),bias,"xx,1x")
# Don't apply nonlinearity at the last layer
if i != len(layer_out_sizes)-2: last_out = activation(last_out)
idx = 0
k_bHm = last_out[:,idx:idx+H*m]; idx += H*m; k_bHm = k_bHm.reshape([b,H,m])
beta_bH = last_out[:,idx:idx+H]; idx += H
g_bH = last_out[:,idx:idx+H]; idx += H
s_bH3 = last_out[:,idx:idx+3*H]; idx += 3*H; s_bH3 = s_bH3.reshape([b,H,3])
gamma_bH = last_out[:,idx:idx+H]; idx += H
e_bhm = last_out[:,idx:idx+h*m]; idx += h*m; e_bhm = e_bhm.reshape([b,h,m])
a_bhm = last_out[:,idx:idx+h*m]; idx += h*m; a_bhm = a_bhm.reshape([b,h,m])
y_bp = last_out[:,idx:idx+p]; idx += p
k_bHm = cgt.tanh(k_bHm)
beta_bH = nn.softplus(beta_bH)
g_bH = cgt.sigmoid(g_bH)
s_bH3 = sum_normalize2(cgt.exp(s_bH3))
gamma_bH = cgt.sigmoid(gamma_bH)+1
e_bhm = cgt.sigmoid(e_bhm)
a_bhm = cgt.tanh(a_bhm)
# y_bp = y_bp
assert infer_shape(k_bHm) == (b,H,m)
assert infer_shape(beta_bH) == (b,H)
assert infer_shape(g_bH) == (b,H)
assert infer_shape(s_bH3) == (b,H,3)
assert infer_shape(gamma_bH) == (b,H)
assert infer_shape(e_bhm) == (b,h,m)
assert infer_shape(a_bhm) == (b,h,m)
assert infer_shape(y_bp) == (b,p)
return nn.Module([r_bhm, X_bk], [k_bHm, beta_bH, g_bH, s_bH3, gamma_bH, e_bhm, a_bhm, y_bp])
def combo_layer(X, size_in, size_out, splits,
s_funcs=None, o_funcs=None,
name='?', dbg_out={}):
"""
Create a combination of specified sub-layers and non-linearity.
:param X: symbolic input
:param size_in: input size (except batch)
:param size_out: output size (except batch)
:param splits: split points for applying each sub-layer
:param s_funcs: list of functions to create each sub-layer
of signature (input, in_size, out_size, name) -> output
:param o_funcs: list of non-linearity functions
of signature (input) -> output
:param name: layer name
:type name: str
:return: symbolic output
Note for s_funcs and o_funcs:
- broadcasting enabled if not a list
- element with value None has no effect (skip)
"""
assert isinstance(splits, tuple) and len(splits) > 0
assert all(splits[i] < splits[i+1] for i in xrange(len(splits) - 1))
assert splits[0] >= 0 and splits[-1] <= size_out
splits = list(splits)
assert not isinstance(o_funcs, list) and not isinstance(s_funcs, list)
o_funcs = list(o_funcs) if isinstance(o_funcs, tuple) \
else [o_funcs] * (len(splits) + 1)
s_funcs = list(s_funcs) if isinstance(s_funcs, tuple) \
else [s_funcs] * (len(splits) + 1)
assert len(splits) + 1 == len(o_funcs) == len(s_funcs)
if splits[0] == 0:
splits.pop(0)
o_funcs.pop(0)
s_funcs.pop(0)
if len(splits) > 0 and splits[-1] == size_out:
splits.pop()
o_funcs.pop()
s_funcs.pop()
curr, names, ins, outs = 0, [], [], []
splits.append(size_out)
for _split, _f_s, _f_o in zip(splits, s_funcs, o_funcs):
_name = 'L%s[%d:%d]' % (name, curr, _split)
_s_out = _split - curr
_i = X if _f_s is None else _f_s(X, size_in, _s_out, name=_name)
_o = _i if _f_o is None else _f_o(_i)
curr = _split
ins.append(_i)
outs.append(_o)
names.append(_name)
out = cgt.concatenate(outs, axis=1) if len(outs) > 1 else outs[0]
dbg_out.update(dict(zip([_n + '~in' for _n in names], ins)))
dbg_out.update(dict(zip([_n + '~out' for _n in names], outs)))
return out
def initialize(self, loss, scale):
self._iter = 0
self.pc = Params(self.params)
cur_val = self.pc.get_value_flat()
idx = cur_val.nonzero()
new_val = np.random.uniform(-scale, scale, size=(self.pc.get_total_size(),))
new_val[idx] = cur_val[idx]
self.sync(new_val)
grad = cgt.concatenate([g.flatten() for g in cgt.grad(loss, self.params)])
return grad
def make_loss_and_grad(net):
X_b = inps[0] #cgt.matrix(dtype=cgt.floatX)
y_onehot = cgt.matrix(dtype='i4')
outputs = [logprobs]
loss = nn.crossent(outputs[0], y_onehot) / b_size
#gradloss = cgt.grad(loss, params)
gradloss = cgt.grad(loss, param_list)
# XXX use flatcat function
grad = cgt.concatenate([x.flatten() for x in gradloss])
#grad = gradloss
return cgt.make_function([X_b, y_onehot], [loss, grad, logprobs])
def __init__(self, n_actions):
Serializable.__init__(self, n_actions)
cgt.set_precision('double')
n_in = 128
o_no = cgt.matrix("o_no", fixed_shape=(None, n_in))
a_n = cgt.vector("a_n", dtype='i8')
q_n = cgt.vector("q_n")
oldpdist_np = cgt.matrix("oldpdists")
h0 = (o_no - 128.0) / 128.0
nhid = 64
h1 = cgt.tanh(
nn.Affine(128, nhid, weight_init=nn.IIDGaussian(std=.1))(h0))
probs_na = nn.softmax(
nn.Affine(nhid, n_actions,
weight_init=nn.IIDGaussian(std=0.01))(h1))
logprobs_na = cgt.log(probs_na)
b = cgt.size(o_no, 0)
logps_n = logprobs_na[cgt.arange(b), a_n]
surr = (logps_n * q_n).mean()
kl = (oldpdist_np * cgt.log(oldpdist_np / probs_na)).sum(axis=1).mean()
params = nn.get_parameters(surr)
gradsurr = cgt.grad(surr, params)
flatgrad = cgt.concatenate([p.flatten() for p in gradsurr])
lam = cgt.scalar()
penobj = surr - lam * kl
self._f_grad_lagrangian = cgt.function(
[lam, oldpdist_np, o_no, a_n, q_n],
cgt.concatenate([p.flatten() for p in cgt.grad(penobj, params)]))
self.f_pdist = cgt.function([o_no], probs_na)
self.f_probs = cgt.function([o_no], probs_na)
self.f_surr_kl = cgt.function([oldpdist_np, o_no, a_n, q_n],
[surr, kl])
self.f_gradlogp = cgt.function([oldpdist_np, o_no, a_n, q_n], flatgrad)
self.pc = ParamCollection(params)
def hybrid_layer(X, size_in, size_out, size_random, dbg_out=[]):
assert size_out >= size_random >= 0
out = cgt.sigmoid(nn.Affine(
size_in, size_out, name="InnerProd(%d->%d)" % (size_in, size_out)
)(X))
dbg_out.append(out)
if size_random == 0:
return out
if size_random == size_out:
out_s = cgt.bernoulli(out)
return out_s
out_s = cgt.bernoulli(out[:, :size_random])
out = cgt.concatenate([out_s, out[:, size_random:]], axis=1)
return out
def pyramidLayer(nn_input, temporal_resolution_decrease=2):
"""
Batch by time by features. Decreases temporal resolution and increases feature dimension by a resolution decrease factor.
"""
t_steps = cgt.infer_shape(nn_input)[1]
if t_steps % temporal_resolution_decrease != 0:
raise ValueError('number of timesteps is not divisable by resolution decrease!')
out_list = []
for iter_step in range(0, t_steps, temporal_resolution_decrease):
concentrate_list = []
for sub_iter_step in range(0, temporal_resolution_decrease):
concentrate_list.append(nn_input[:, iter_step + sub_iter_step, :])
out_list.append(cgt.concatenate(concentrate_list, axis=1))
return cgt.dimshuffle(cgt.stack(out_list), [1, 0, 2])
def hybrid_layer(X, size_in, size_out, size_random, dbg_out=[]):
assert size_out >= size_random >= 0
out = cgt.sigmoid(
nn.Affine(size_in,
size_out,
name="InnerProd(%d->%d)" % (size_in, size_out))(X))
dbg_out.append(out)
if size_random == 0:
return out
if size_random == size_out:
out_s = cgt.bernoulli(out)
return out_s
out_s = cgt.bernoulli(out[:, :size_random])
out = cgt.concatenate([out_s, out[:, size_random:]], axis=1)
return out
def mask_layer(func, X, size_in, i_start, i_end=None):
if i_end is None:
i_start, i_end = 0, i_start
assert isinstance(i_start, int) and isinstance(i_end, int)
assert -1 < i_start <= i_end <= size_in
if i_end == i_start:
return X
if i_end - i_start == size_in:
return func(X)
outs = []
if i_start > 0:
outs.append(X[:, :i_start])
outs.append(func(X[:, i_start:i_end]))
if i_end < size_in:
outs.append(X[:, i_end:])
out = cgt.concatenate(outs, axis=1)
return out
def get_features_bengio(nn_input, num_units=256, recurrent_layer=None):
if recurrent_layer is None:
recurrent_layer = recurrentLayer
w_init = IIDUniform(-0.1, 0.1)
activation = cgt.sigmoid
assert num_units % 2 == 0
num_units /= 2
l1_f = recurrent_layer(nn_input=nn_input, num_units=num_units, activation=activation, w_init=w_init)
l1_b = recurrent_layer(nn_input=nn_input, num_units=num_units, activation=activation, w_init=w_init)
l1_plus = cgt.concatenate([l1_f, l1_b], axis=2)
#l2_f = recurrent_layer(nn_input=l1_plus, num_units=num_units, activation=activation, w_init=w_init)
#l2_b = recurrent_layer(nn_input=l1_plus, num_units=num_units, activation=activation, w_init=w_init)
#l2_plus = cgt.concatenate([l2_f, l2_b], axis=2)
#l3_f = recurrent_layer(nn_input=l2_plus, num_units=num_units, activation=activation, w_init=w_init)
#l3_b = recurrent_layer(nn_input=l2_plus, num_units=num_units, activation=activation, w_init=w_init)
#l3_plus = cgt.concatenate([l3_f, l3_b], axis=2)
return l1_plus
def mask_layer(func, X, size_in, i_start, i_end=None):
if i_end is None:
i_start, i_end = 0, i_start
assert isinstance(i_start, int) and isinstance(i_end, int)
assert -1 < i_start <= i_end <= size_in
if i_end == i_start:
return X
if i_end - i_start == size_in:
return func(X)
outs = []
if i_start > 0:
outs.append(X[:, :i_start])
outs.append(func(X[:, i_start:i_end]))
if i_end < size_in:
outs.append(X[:, i_end:])
out = cgt.concatenate(outs, axis=1)
return out
def main(num_epochs=NUM_EPOCHS):
#cgt.set_precision('half')
print("Building network ...")
# Recurrent layers expect input of shape
# (batch size, max sequence length, number of features)
X = cgt.tensor3(name='X', fixed_shape=(N_BATCH, MAX_LENGTH, 2))
l_forward = nnbuilder.recurrentLayer(nn_input=X, num_units=N_HIDDEN)
l_backward = nnbuilder.recurrentLayer(nn_input=X, num_units=N_HIDDEN, backwards=True)
#l_forward = nnbuilder.LSTMLayer(nn_input=X, num_units=N_HIDDEN, activation=cgt.sigmoid)
#l_backward = nnbuilder.LSTMLayer(nn_input=X, num_units=N_HIDDEN, activation=cgt.sigmoid, backwards=True)
#l_forward = nnbuilder.GRULayer(nn_input=X, num_units=N_HIDDEN, activation=nn.rectify)
#l_backward = nnbuilder.GRULayer(nn_input=X, num_units=N_HIDDEN, activation=nn.rectify, backwards=True)
l_forward_slice = l_forward[:, MAX_LENGTH-1, :] # Take the last element in the forward slice time dimension
l_backward_slice = l_backward[:, 0, :] # And the first element in the backward slice time dimension
l_sum = cgt.concatenate([l_forward_slice, l_backward_slice], axis=1)
l_out = nnbuilder.denseLayer(l_sum, num_units=1, activation=cgt.tanh)
target_values = cgt.vector('target_output')
predicted_values = l_out[:, 0] # For this task we only need the last value
cost = cgt.mean((predicted_values - target_values)**2)
# Compute SGD updates for training
print("Computing updates ...")
updates = nn.rmsprop(cost, nn.get_parameters(l_out), LEARNING_RATE)
#updates = nn.nesterov_momentum(cost, nn.get_parameters(l_out), 0.05)
# cgt functions for training and computing cost
print("Compiling functions ...")
train = cgt.function([X, target_values], cost, updates=updates)
compute_cost = cgt.function([X, target_values], cost)
# We'll use this "validation set" to periodically check progress
X_val, y_val, mask_val = gen_data()
print("Training ...")
time_start = time.time()
try:
for epoch in range(num_epochs):
for _ in range(EPOCH_SIZE):
X, y, m = gen_data()
train(X, y)
cost_val = compute_cost(X_val, y_val)
print("Epoch {} validation cost = {}".format(epoch+1, cost_val))
print ('Epoch took ' + str(time.time() - time_start))
time_start = time.time()
except KeyboardInterrupt:
pass
def make_funcs(config, dbg_out=None):
params, Xs, Ys, C_0, H_0, C_T, H_T, C_1, H_1 = lstm_network(
config['rnn_steps'], config['num_inputs'], config['num_outputs'],
config['num_units'], config['num_mems'])
# basic
size_batch = Xs[0].shape[0]
dY = Ys[0].shape[-1]
Ys_gt = [
cgt.matrix(fixed_shape=(size_batch, dY), name='Y%d' % t)
for t in range(len(Ys))
]
Ys_var = [cgt.tensor3(fixed_shape=(size_batch, dY, dY)) for _ in Ys]
net_inputs, net_outputs = Xs + C_0 + H_0 + Ys_var, Ys + C_T + H_T
# calculate loss
loss_vec = []
for i in range(len(Ys)):
# if i == 0: continue
_l = dist.gaussian.logprob(Ys_gt[i], Ys[i], Ys_var[i])
loss_vec.append(_l)
loss_vec = cgt.add_multi(loss_vec)
if config['weight_decay'] > 0.:
params_flat = cgt.concatenate([p.flatten() for p in params])
loss_param = config['weight_decay'] * cgt.sum(params_flat**2)
loss_vec -= loss_param # / size_batch
loss = cgt.sum(loss_vec) / config['rnn_steps'] / size_batch
grad = cgt.grad(loss, params)
# functions
def f_init(size_batch):
c_0, h_0 = [], []
for _n_m in config['num_mems']:
if _n_m > 0:
c_0.append(np.zeros((size_batch, _n_m)))
h_0.append(np.zeros((size_batch, _n_m)))
return c_0, h_0
f_step = cgt.function([Xs[0]] + C_0 + H_0, [Ys[0]] + C_1 + H_1)
f_loss = cgt.function(net_inputs + Ys_gt, loss)
f_grad = cgt.function(net_inputs + Ys_gt, grad)
f_surr = cgt.function(net_inputs + Ys_gt, [loss] + net_outputs + grad)
return params, f_step, f_loss, f_grad, f_init, f_surr
def make_funcs(config, dbg_out=None):
params, Xs, Ys, C_0, H_0, C_T, H_T, C_1, H_1 = lstm_network(
config['rnn_steps'], config['num_inputs'], config['num_outputs'],
config['num_units'], config['num_mems']
)
# basic
size_batch = Xs[0].shape[0]
dY = Ys[0].shape[-1]
Ys_gt = [cgt.matrix(fixed_shape=(size_batch, dY), name='Y%d'%t)
for t in range(len(Ys))]
Ys_var = [cgt.tensor3(fixed_shape=(size_batch, dY, dY)) for _ in Ys]
net_inputs, net_outputs = Xs + C_0 + H_0 + Ys_var, Ys + C_T + H_T
# calculate loss
loss_vec = []
for i in range(len(Ys)):
# if i == 0: continue
_l = dist.gaussian.logprob(Ys_gt[i], Ys[i], Ys_var[i])
loss_vec.append(_l)
loss_vec = cgt.add_multi(loss_vec)
if config['weight_decay'] > 0.:
params_flat = cgt.concatenate([p.flatten() for p in params])
loss_param = config['weight_decay'] * cgt.sum(params_flat ** 2)
loss_vec -= loss_param # / size_batch
loss = cgt.sum(loss_vec) / config['rnn_steps'] / size_batch
grad = cgt.grad(loss, params)
# functions
def f_init(size_batch):
c_0, h_0 = [], []
for _n_m in config['num_mems']:
if _n_m > 0:
c_0.append(np.zeros((size_batch, _n_m)))
h_0.append(np.zeros((size_batch, _n_m)))
return c_0, h_0
f_step = cgt.function([Xs[0]] + C_0 + H_0, [Ys[0]] + C_1 + H_1)
f_loss = cgt.function(net_inputs + Ys_gt, loss)
f_grad = cgt.function(net_inputs + Ys_gt, grad)
f_surr = cgt.function(net_inputs + Ys_gt, [loss] + net_outputs + grad)
return params, f_step, f_loss, f_grad, f_init, f_surr
def make_funcs(config, dbg_out={}):
net_in, net_out = hybrid_network(config['num_inputs'],
config['num_outputs'],
config['num_units'],
config['num_sto'],
dbg_out=dbg_out)
if not config['dbg_out_full']: dbg_out = {}
# def f_sample(_inputs, num_samples=1, flatten=False):
# _mean, _var = f_step(_inputs)
# _samples = []
# for _m, _v in zip(_mean, _var):
# _s = np.random.multivariate_normal(_m, np.diag(np.sqrt(_v)), num_samples)
# if flatten: _samples.extend(_s)
# else: _samples.append(_s)
# return np.array(_samples)
Y_gt = cgt.matrix("Y")
Y_prec = cgt.tensor3('V',
fixed_shape=(None, config['num_inputs'],
config['num_inputs']))
params = nn.get_parameters(net_out)
size_batch, size_out = net_out.shape
inputs, outputs = [net_in], [net_out]
if config['no_bias']:
print "Excluding bias"
params = [p for p in params if not p.name.endswith(".b")]
loss_vec = dist.gaussian.logprob(Y_gt, net_out, Y_prec)
if config['weight_decay'] > 0.:
print "Applying penalty on parameter norm"
params_flat = cgt.concatenate([p.flatten() for p in params])
loss_param = config['weight_decay'] * cgt.sum(params_flat**2)
loss_vec -= loss_param # / size_batch
loss = cgt.sum(loss_vec) / size_batch
# TODO_TZ f_step seems not to fail if X has wrong dim
f_step = cgt.function(inputs, outputs)
f_surr = get_surrogate_func(inputs + [Y_prec, Y_gt],
outputs, [loss_vec],
params,
_dbg_out=dbg_out)
return params, f_step, None, None, None, f_surr
def make_funcs(net_in, net_out, config, dbg_out=None):
def f_grad (*x):
out = f_surr(*x)
return out['loss'], out['surr_loss'], out['surr_grad']
Y = cgt.matrix("Y")
params = nn.get_parameters(net_out)
if 'no_bias' in config and config['no_bias']:
print "Excluding bias"
params = [p for p in params if not p.name.endswith(".b")]
size_out, size_batch = Y.shape[1], net_in.shape[0]
f_step = cgt.function([net_in], [net_out])
# loss_raw of shape (size_batch, 1); loss should be a scalar
# sum-of-squares loss
sigma = 0.1
loss_raw = -cgt.sum((net_out - Y) ** 2, axis=1, keepdims=True) / sigma
# negative log-likelihood
# out_sigma = cgt.exp(net_out[:, size_out:]) + 1.e-6 # positive sigma
# loss_raw = -gaussian_diagonal.logprob(
# Y, net_out,
# out_sigma
# cgt.fill(.01, [size_batch, size_out])
# )
if 'param_penal_wt' in config:
print "Applying penalty on parameter norm"
assert config['param_penal_wt'] > 0
params_flat = cgt.concatenate([p.flatten() for p in params])
loss_param = cgt.fill(cgt.sum(params_flat ** 2), [size_batch, 1])
loss_param *= config['param_penal_wt']
loss_raw += loss_param
loss = cgt.sum(loss_raw) / size_batch
# end of loss definition
f_loss = cgt.function([net_in, Y], [net_out, loss])
f_surr = get_surrogate_func([net_in, Y],
[net_out] + dbg_out,
[loss_raw], params)
return params, f_step, f_loss, f_grad, f_surr
def make_funcs(net_in, net_out, config, dbg_out=None):
def f_grad(*x):
out = f_surr(*x)
return out['loss'], out['surr_loss'], out['surr_grad']
Y = cgt.matrix("Y")
params = nn.get_parameters(net_out)
if 'no_bias' in config and config['no_bias']:
print "Excluding bias"
params = [p for p in params if not p.name.endswith(".b")]
size_out, size_batch = Y.shape[1], net_in.shape[0]
f_step = cgt.function([net_in], [net_out])
# loss_raw of shape (size_batch, 1); loss should be a scalar
# sum-of-squares loss
sigma = 0.1
loss_raw = -cgt.sum((net_out - Y)**2, axis=1, keepdims=True) / sigma
# negative log-likelihood
# out_sigma = cgt.exp(net_out[:, size_out:]) + 1.e-6 # positive sigma
# loss_raw = -gaussian_diagonal.logprob(
# Y, net_out,
# out_sigma
# cgt.fill(.01, [size_batch, size_out])
# )
if 'param_penal_wt' in config:
print "Applying penalty on parameter norm"
assert config['param_penal_wt'] > 0
params_flat = cgt.concatenate([p.flatten() for p in params])
loss_param = cgt.fill(cgt.sum(params_flat**2), [size_batch, 1])
loss_param *= config['param_penal_wt']
loss_raw += loss_param
loss = cgt.sum(loss_raw) / size_batch
# end of loss definition
f_loss = cgt.function([net_in, Y], [net_out, loss])
f_surr = get_surrogate_func([net_in, Y], [net_out] + dbg_out, [loss_raw],
params)
return params, f_step, f_loss, f_grad, f_surr
def take_one_step(self, nn_input_bf, hid_out):
#PROBABLY BUGGED. SHOULD BE REWRITTEN.
self.num_batches = cgt.infer_shape(nn_input_bf)[0]
# (n_time_steps, n_batch, n_features)
#input_bf = cgt.dimshuffle(nn_input_bf, [1, 0, 2])
# Stack input weight matrices into a (num_inputs, 3*num_units)
# matrix, which speeds up computation
W_in_stacked = cgt.concatenate(
[self.W_in_to_resetgate, self.W_in_to_updategate,
self.W_in_to_hidden_update], axis=1)
# Same for hidden weight matrices
W_hid_stacked = cgt.concatenate(
[self.W_hid_to_resetgate, self.W_hid_to_updategate,
self.W_hid_to_hidden_update], axis=1)
# Stack gate biases into a (3*num_units) vector
b_stacked = cgt.concatenate(
[self.b_resetgate, self.b_updategate,
self.b_hidden_update], axis=1)
# At each loop, input_n will be (n_time_steps, 3*num_units).
# We define a slicing function that extract the input to each GRU gate
def slice_w(x, n):
return x[:, n*self.num_units:(n+1)*self.num_units]
# Create single recurrent computation step function
# input__n is the n'th vector of the input
def step(input_n, hid_previous, W_hid_stacked, W_in_stacked, b_stacked):
# Compute W_{hr} h_{t - 1}, W_{hu} h_{t - 1}, and W_{hc} h_{t - 1}
hid_input = cgt.dot(hid_previous, W_hid_stacked)
# Compute W_{xr}x_t + b_r, W_{xu}x_t + b_u, and W_{xc}x_t + b_c
input_n = cgt.broadcast("+", input_n.dot(W_in_stacked), b_stacked, "xx,1x")
# Reset and update gates
resetgate = slice_w(hid_input, 0) + slice_w(input_n, 0)
updategate = slice_w(hid_input, 1) + slice_w(input_n, 1)
resetgate = self.nonlinearity_resetgate(resetgate)
updategate = self.nonlinearity_updategate(updategate)
# Compute W_{xc}x_t + r_t \odot (W_{hc} h_{t - 1})
hidden_update_in = slice_w(input_n, 2)
hidden_update_hid = slice_w(hid_input, 2)
hidden_update = hidden_update_in + resetgate*hidden_update_hid
# Compute (1 - u_t)h_{t - 1} + u_t c_t
hid = (1 - updategate)*hid_previous + updategate*hidden_update
return self.nonlinearity_hid(hid) # adding this non-linearity seems to help stability.
#return hid
if hid_out is None:
if self.hid_out is None:
self.hid_out = cgt.dot(cgt.ones((self.num_batches, 1)), self.hid_init)
hid_out = self.hid_out
# Retrieve the dimensionality of the incoming layer
hid_out = step(nn_input_bf, hid_out, W_hid_stacked, W_in_stacked, b_stacked)
# dimshuffle back to (n_batch, n_time_steps, n_features))
# self.hid_out = cgt.dimshuffle(self.hid_out, [1, 0, 2])
# if scan is backward reverse the output
if self.backwards:
hid_out = cgt.flip(hid_out, [1])
self.hid_out = hid_out
return hid_out
def __call__(self, input_btf):
# (n_time_steps, n_batch, n_features)
input_tbf = cgt.dimshuffle(input_btf, [1, 0, 2])
self.num_batches = cgt.infer_shape(input_tbf)[1]
# Stack input weight matrices into a (num_inputs, 3*num_units)
# matrix, which speeds up computation
W_in_stacked = cgt.concatenate(
[self.W_in_to_resetgate, self.W_in_to_updategate,
self.W_in_to_hidden_update], axis=1)
# Same for hidden weight matrices
W_hid_stacked = cgt.concatenate(
[self.W_hid_to_resetgate, self.W_hid_to_updategate,
self.W_hid_to_hidden_update], axis=1)
# Stack gate biases into a (3*num_units) vector
b_stacked = cgt.concatenate(
[self.b_resetgate, self.b_updategate,
self.b_hidden_update], axis=1)
# At each loop, input_n will be (n_time_steps, 3*num_units).
# We define a slicing function that extract the input to each GRU gate
def slice_w(x, n):
return x[:, n*self.num_units:(n+1)*self.num_units]
# Create single recurrent computation step function
# input__n is the n'th vector of the input
def step(input_n, hid_previous, W_hid_stacked, W_in_stacked, b_stacked):
# Compute W_{hr} h_{t - 1}, W_{hu} h_{t - 1}, and W_{hc} h_{t - 1}
hid_input = cgt.dot(hid_previous, W_hid_stacked)
# Compute W_{xr}x_t + b_r, W_{xu}x_t + b_u, and W_{xc}x_t + b_c
input_n = cgt.broadcast("+", input_n.dot(W_in_stacked), b_stacked, "xx,1x")
# Reset and update gates
resetgate = slice_w(hid_input, 0) + slice_w(input_n, 0)
updategate = slice_w(hid_input, 1) + slice_w(input_n, 1)
resetgate = self.nonlinearity_resetgate(resetgate)
updategate = self.nonlinearity_updategate(updategate)
# Compute W_{xc}x_t + r_t \odot (W_{hc} h_{t - 1})
hidden_update_in = slice_w(input_n, 2)
hidden_update_hid = slice_w(hid_input, 2)
hidden_update = hidden_update_in + resetgate*hidden_update_hid
# Compute (1 - u_t)h_{t - 1} + u_t c_t
hid = (1 - updategate)*hid_previous + updategate*hidden_update
return hid
sequences = [input_tbf]
step_fun = step
hid_init = cgt.dot(cgt.ones((self.num_batches, 1)), self.hid_init)
# The hidden-to-hidden weight matrix is always used in step
non_seqs = [W_hid_stacked]
# When we aren't precomputing the input outside of scan, we need to
# provide the input weights and biases to the step function
non_seqs += [W_in_stacked, b_stacked]
# theano.scan only allows for positional arguments, so when
# self.precompute_input is True, we need to supply fake placeholder
# arguments for the input weights and biases.
# Retrieve the dimensionality of the incoming layer
hid_out = unroll_lstm(
fn=step_fun,
sequences=sequences,
outputs_info=[hid_init],
go_backwards=self.backwards,
non_sequences=non_seqs,
n_steps=self.timesteps)[0]
# dimshuffle back to (n_batch, n_time_steps, n_features))
hid_out = cgt.dimshuffle(hid_out, [1, 0, 2])
# if scan is backward reverse the output
if self.backwards:
hid_out = cgt.flip(hid_out, [1])
return hid_out
def concatenate(items, axis=0):
return cgt.concatenate(items, axis=axis)
def build_fcn_action_cond_encoder_net(input_shapes, levels=None):
x_shape, u_shape = input_shapes
x_c_dim = x_shape[0]
x1_c_dim = 16
levels = levels or [3]
levels = sorted(set(levels))
X = cgt.tensor4('X', fixed_shape=(None, ) + x_shape)
U = cgt.matrix('U', fixed_shape=(None, ) + u_shape)
# encoding
Xlevels = {}
for level in range(levels[-1] + 1):
if level == 0:
Xlevel = X
else:
if level == 1:
xlevelm1_c_dim = x_c_dim
xlevel_c_dim = x1_c_dim
else:
xlevelm1_c_dim = xlevel_c_dim
xlevel_c_dim = 2 * xlevel_c_dim
Xlevel_1 = nn.rectify(
nn.SpatialConvolution(xlevelm1_c_dim,
xlevel_c_dim,
kernelshape=(3, 3),
pad=(1, 1),
stride=(1, 1),
name='conv%d_1' % level,
weight_init=nn.IIDGaussian(std=0.01))(
Xlevels[level - 1]))
Xlevel_2 = nn.rectify(
nn.SpatialConvolution(
xlevel_c_dim,
xlevel_c_dim,
kernelshape=(3, 3),
pad=(1, 1),
stride=(1, 1),
name='conv%d_2' % level,
weight_init=nn.IIDGaussian(std=0.01))(Xlevel_1))
Xlevel = nn.max_pool_2d(Xlevel_2,
kernelshape=(2, 2),
pad=(0, 0),
stride=(2, 2))
Xlevels[level] = Xlevel
# bilinear
Xlevels_next_pred_0 = {}
Ylevels = OrderedDict()
Ylevels_diff_pred = OrderedDict()
for level in levels:
Xlevel = Xlevels[level]
Xlevel_diff_pred = Bilinear(input_shapes,
b=None,
axis=2,
name='bilinear%d' % level)(Xlevel, U)
Xlevels_next_pred_0[level] = Xlevel + Xlevel_diff_pred
Ylevels[level] = Xlevel.reshape(
(Xlevel.shape[0], cgt.mul_multi(Xlevel.shape[1:])))
Ylevels_diff_pred[level] = Xlevel_diff_pred.reshape(
(Xlevel_diff_pred.shape[0],
cgt.mul_multi(Xlevel_diff_pred.shape[1:])))
# decoding
Xlevels_next_pred = {}
for level in range(levels[-1] + 1)[::-1]:
if level == levels[-1]:
Xlevel_next_pred = Xlevels_next_pred_0[level]
else:
if level == 0:
xlevelm1_c_dim = x_c_dim
elif level < levels[-1] - 1:
xlevel_c_dim = xlevelm1_c_dim
xlevelm1_c_dim = xlevelm1_c_dim // 2
Xlevel_next_pred_2 = SpatialDeconvolution(
xlevel_c_dim,
xlevel_c_dim,
kernelshape=(2, 2),
pad=(0, 0),
stride=(2, 2),
name='upsample%d' % (level + 1),
weight_init=nn.IIDGaussian(std=0.01))(Xlevels_next_pred[
level +
1]) # TODO initialize with bilinear # TODO should rectify?
Xlevel_next_pred_1 = nn.rectify(
SpatialDeconvolution(
xlevel_c_dim,
xlevel_c_dim,
kernelshape=(3, 3),
pad=(1, 1),
stride=(1, 1),
name='deconv%d_2' % (level + 1),
weight_init=nn.IIDGaussian(std=0.01))(Xlevel_next_pred_2))
nonlinearity = nn.rectify if level > 0 else cgt.tanh
Xlevel_next_pred = nonlinearity(
SpatialDeconvolution(
xlevel_c_dim,
xlevelm1_c_dim,
kernelshape=(3, 3),
pad=(1, 1),
stride=(1, 1),
name='deconv%d_1' % (level + 1),
weight_init=nn.IIDGaussian(std=0.01))(Xlevel_next_pred_1))
if level in Xlevels_next_pred_0:
coefs = nn.parameter(nn.init_array(nn.Constant(0.5), (2, )),
name='sum%d.coef' % level)
Xlevel_next_pred = coefs[0] * Xlevel_next_pred + coefs[
1] * Xlevels_next_pred_0[level]
# TODO: tanh should be after sum
Xlevels_next_pred[level] = Xlevel_next_pred
X_next_pred = Xlevels_next_pred[0]
Y = cgt.concatenate(Ylevels.values(), axis=1)
Y_diff_pred = cgt.concatenate(Ylevels_diff_pred.values(), axis=1)
X_diff = cgt.tensor4('X_diff', fixed_shape=(None, ) + x_shape)
X_next = X + X_diff
loss = ((X_next - X_next_pred)**2).mean(axis=0).sum() / 2.
net_name = 'FcnActionCondEncoderNet_levels' + ''.join(
str(level) for level in levels)
input_vars = OrderedDict([(var.name, var) for var in [X, U, X_diff]])
pred_vars = OrderedDict([('Y_diff_pred', Y_diff_pred), ('Y', Y),
('X_next_pred', X_next_pred)])
return net_name, input_vars, pred_vars, loss
def stack(tensors, axis=0):
if axis is not 0:
raise ValueError('only axis=0 is supported under cgt')
return cgt.concatenate(map(lambda x: cgt.reshape(x, [1] + x.shape),
tensors),
axis=0)
def get_character_distribution(self, state_bf, context_bf):
total_state = cgt.concatenate([state_bf, context_bf], axis=1)
d1 = self.final_out_dense(total_state)
return softmax(d1, axis=1)
fixed_shape_mask="all")
yname = layer.top[0]
output = [cgt.broadcast("+", X.dot(W), b, "xx,1x")]
elif layer.type == "ReLU":
output = [nn.rectify(inputs[0])]
elif layer.type == "Softmax":
output = [nn.softmax(inputs[0])]
elif layer.type == "LRN":
# XXX needs params
param = layer.lrn_param
output = [
nn.lrn(inputs[0], param.alpha, param.beta, param.local_size)
]
elif layer.type == "Concat":
param = layer.concat_param
output = [cgt.concatenate(inputs, param.concat_dim)]
elif layer.type == "Dropout":
output = [nn.dropout(inputs[0])]
elif layer.type == "SoftmaxWithLoss":
output = [nn.loglik_softmax(inputs[0], inputs[1])]
elif layer.type == "Accuracy":
output = [nn.zero_one_loss(inputs[0], inputs[1])]
else:
cgt.error("unrecognized layer type %s" % layer.type)
assert output is not None
# assert isinstance(output, cgt.Node)
for i in xrange(len(layer.top)):
name2node[layer.top[i]] = output[i]
print "stored", layer.top[0]
def flatcat(xs):
return cgt.concatenate([x.flatten() for x in xs])
def stack(tensors, axis=0):
if axis is not 0:
raise ValueError('only axis=0 is supported under cgt')
return cgt.concatenate(map(lambda x: cgt.reshape(x, [1] + x.shape), tensors), axis=0)
bname = layer.param[1].name or layer.name+":b"
bval = np.empty(bshape, dtype=cgt.floatX)
b = name2node[bname] = cgt.shared(bval, name=bname, fixed_shape_mask="all")
yname = layer.top[0]
output = [cgt.broadcast("+",X.dot(W), b, "xx,1x") ]
elif layer.type == "ReLU":
output = [nn.rectify(inputs[0])]
elif layer.type == "Softmax":
output = [nn.softmax(inputs[0])]
elif layer.type == "LRN":
# XXX needs params
param = layer.lrn_param
output = [nn.lrn(inputs[0], param.alpha,param.beta, param.local_size)]
elif layer.type == "Concat":
param = layer.concat_param
output = [cgt.concatenate(inputs, param.concat_dim) ]
elif layer.type == "Dropout":
output = [nn.dropout(inputs[0])]
elif layer.type == "SoftmaxWithLoss":
output = [nn.loglik_softmax(inputs[0], inputs[1])]
elif layer.type == "Accuracy":
output = [nn.zero_one_loss(inputs[0], inputs[1])]
else:
cgt.error("unrecognized layer type %s"%layer.type)
assert output is not None
# assert isinstance(output, cgt.Node)
for i in xrange(len(layer.top)): name2node[layer.top[i]] = output[i]
print "stored", layer.top[0]
if layer.type != "Data":
def get_decoder_state(self, context_bf, prev_out_bc, prev_decoder_state):
input_bf = cgt.concatenate([context_bf, prev_out_bc], axis=1)
l1 = self.recurrent_decoder_one(input_bf, prev_decoder_state)
l2 = self.recurrent_decoder_two(l1)
return l2
def flatcat(xs):
return cgt.concatenate([x.flatten() for x in xs])
def concatenate(items, axis=0):
return cgt.concatenate(items, axis=axis)
def __call__(self, nn_input_btf):
# Because scan iterates over the first dimension we dimshuffle to
# (n_time_steps, n_batch, n_features)
nn_input_tbf = cgt.dimshuffle(nn_input_btf, [1, 0, 2])
seq_len, num_batch = nn_input_tbf.shape[0], nn_input_tbf.shape[1]
# Stack input weight matrices into a (num_inputs, 4*num_units) #checks out
# matrix, which speeds up computation
W_in_stacked = cgt.concatenate(
[self.W_in_to_ingate, self.W_in_to_forgetgate,
self.W_in_to_cell, self.W_in_to_outgate], axis=1)
# Same for hidden weight matrices
W_hid_stacked = cgt.concatenate(
[self.W_hid_to_ingate, self.W_hid_to_forgetgate,
self.W_hid_to_cell, self.W_hid_to_outgate], axis=1)
# Stack biases into a (4*num_units) vector
b_stacked = cgt.concatenate(
[self.b_ingate, self.b_forgetgate,
self.b_cell, self.b_outgate], axis=1)
def slice_w(x, n):
return x[:, n*self.num_units:(n+1)*self.num_units]
# Create single recurrent computation step function
# input_n is the n'th vector of the input
def step(input_n, cell_previous, hid_previous, W_hid_stacked, W_in_stacked, b_stacked):
input_n = cgt.broadcast("+", input_n.dot(W_in_stacked), b_stacked, "xx,1x")
# Calculate gates pre-activations and slice
gates = input_n + cgt.dot(hid_previous, W_hid_stacked)
# Extract the pre-activation gate values
ingate = slice_w(gates, 0)
forgetgate = slice_w(gates, 1)
cell_input = slice_w(gates, 2)
outgate = slice_w(gates, 3)
# Apply nonlinearities
ingate = self.nonlinearity_ingate(ingate)
forgetgate = self.nonlinearity_forgetgate(forgetgate)
cell_input = self.nonlinearity_cell(cell_input)
outgate = self.nonlinearity_outgate(outgate)
# Compute new cell value
cell = forgetgate*cell_previous + ingate*cell_input
# Compute new hidden unit activation
hid = outgate*self.nonlinearity(cell)
return [cell, hid]
sequences = nn_input_tbf
step_fun = step
ones = cgt.ones((num_batch, 1))
cell_init = cgt.dot(ones, self.cell_init)
hid_init = cgt.dot(ones, self.hid_init)
# The hidden-to-hidden weight matrix is always used in step
non_seqs = [W_hid_stacked]
non_seqs += [W_in_stacked, b_stacked]
cell_out, hid_out = unroll_lstm(
fn=step_fun,
sequences=sequences,
outputs_info=[cell_init, hid_init],
go_backwards=self.backwards,
non_sequences=non_seqs,
n_steps=self.timesteps)
# dimshuffle back to (n_batch, n_time_steps, n_features))
hid_out = cgt.dimshuffle(hid_out, [1, 0, 2])
# if scan is backward reverse the output
if self.backwards:
hid_out = cgt.flip(hid_out, [1])
return hid_out
def __init__(self, obs_dim, ctrl_dim):
cgt.set_precision('double')
Serializable.__init__(self, obs_dim, ctrl_dim)
self.obs_dim = obs_dim
self.ctrl_dim = ctrl_dim
o_no = cgt.matrix("o_no", fixed_shape=(None, obs_dim))
a_na = cgt.matrix("a_na", fixed_shape=(None, ctrl_dim))
adv_n = cgt.vector("adv_n")
oldpdist_np = cgt.matrix("oldpdist", fixed_shape=(None, 2 * ctrl_dim))
self.logstd = logstd_1a = nn.parameter(np.zeros((1, self.ctrl_dim)),
name="std_1a")
std_1a = cgt.exp(logstd_1a)
# Here's where we apply the network
h0 = o_no
nhid = 32
h1 = cgt.tanh(
nn.Affine(obs_dim, nhid, weight_init=nn.IIDGaussian(std=0.1))(h0))
h2 = cgt.tanh(
nn.Affine(nhid, nhid, weight_init=nn.IIDGaussian(std=0.1))(h1))
mean_na = nn.Affine(nhid,
ctrl_dim,
weight_init=nn.IIDGaussian(std=0.01))(h2)
b = cgt.size(o_no, 0)
std_na = cgt.repeat(std_1a, b, axis=0)
oldmean_na = oldpdist_np[:, 0:self.ctrl_dim]
oldstd_na = oldpdist_np[:, self.ctrl_dim:2 * self.ctrl_dim]
logp_n = ((-.5) * cgt.square(
(a_na - mean_na) / std_na).sum(axis=1)) - logstd_1a.sum()
oldlogp_n = ((-.5) * cgt.square(
(a_na - oldmean_na) / oldstd_na).sum(axis=1)
) - cgt.log(oldstd_na).sum(axis=1)
ratio_n = cgt.exp(logp_n - oldlogp_n)
surr = (ratio_n * adv_n).mean()
pdists_np = cgt.concatenate([mean_na, std_na], axis=1)
# kl = cgt.log(sigafter/)
params = nn.get_parameters(surr)
oldvar_na = cgt.square(oldstd_na)
var_na = cgt.square(std_na)
kl = (cgt.log(std_na / oldstd_na) +
(oldvar_na + cgt.square(oldmean_na - mean_na)) / (2 * var_na) -
.5).sum(axis=1).mean()
lam = cgt.scalar()
penobj = surr - lam * kl
self._compute_surr_kl = cgt.function([oldpdist_np, o_no, a_na, adv_n],
[surr, kl])
self._compute_grad_lagrangian = cgt.function(
[lam, oldpdist_np, o_no, a_na, adv_n],
cgt.concatenate([p.flatten() for p in cgt.grad(penobj, params)]))
self.f_pdist = cgt.function([o_no], pdists_np)
self.f_objs = cgt.function([oldpdist_np, o_no, a_na, adv_n],
[surr, kl])
self.pc = ParamCollection(params)
def combo_layer(X,
size_in,
size_out,
splits,
s_funcs=None,
o_funcs=None,
name='?',
dbg_out={}):
"""
Create a combination of specified sub-layers and non-linearity.
:param X: symbolic input
:param size_in: input size (except batch)
:param size_out: output size (except batch)
:param splits: split points for applying each sub-layer
:param s_funcs: list of functions to create each sub-layer
of signature (input, in_size, out_size, name) -> output
:param o_funcs: list of non-linearity functions
of signature (input) -> output
:param name: layer name
:type name: str
:return: symbolic output
Note for s_funcs and o_funcs:
- broadcasting enabled if not a list
- element with value None has no effect (skip)
"""
assert isinstance(splits, tuple) and len(splits) > 0
assert all(splits[i] < splits[i + 1] for i in xrange(len(splits) - 1))
assert splits[0] >= 0 and splits[-1] <= size_out
splits = list(splits)
assert not isinstance(o_funcs, list) and not isinstance(s_funcs, list)
o_funcs = list(o_funcs) if isinstance(o_funcs, tuple) \
else [o_funcs] * (len(splits) + 1)
s_funcs = list(s_funcs) if isinstance(s_funcs, tuple) \
else [s_funcs] * (len(splits) + 1)
assert len(splits) + 1 == len(o_funcs) == len(s_funcs)
if splits[0] == 0:
splits.pop(0)
o_funcs.pop(0)
s_funcs.pop(0)
if len(splits) > 0 and splits[-1] == size_out:
splits.pop()
o_funcs.pop()
s_funcs.pop()
curr, names, ins, outs = 0, [], [], []
splits.append(size_out)
for _split, _f_s, _f_o in zip(splits, s_funcs, o_funcs):
_name = 'L%s[%d:%d]' % (name, curr, _split)
_s_out = _split - curr
_i = X if _f_s is None else _f_s(X, size_in, _s_out, name=_name)
_o = _i if _f_o is None else _f_o(_i)
curr = _split
ins.append(_i)
outs.append(_o)
names.append(_name)
out = cgt.concatenate(outs, axis=1) if len(outs) > 1 else outs[0]
dbg_out.update(dict(zip([_n + '~in' for _n in names], ins)))
dbg_out.update(dict(zip([_n + '~out' for _n in names], outs)))
return out