def make_deep_lstm(size_input, size_mem, n_layers, size_output, size_batch):
inputs = [cgt.matrix(fixed_shape=(size_batch, size_input))]
for _ in xrange(2 * n_layers):
inputs.append(cgt.matrix(fixed_shape=(size_batch, size_mem)))
outputs = []
for i_layer in xrange(n_layers):
prev_h = inputs[i_layer * 2]
prev_c = inputs[i_layer * 2 + 1]
if i_layer == 0:
x = inputs[0]
size_x = size_input
else:
x = outputs[(i_layer - 1) * 2]
size_x = size_mem
input_sums = nn.Affine(size_x, 4 * size_mem)(x) + nn.Affine(
size_x, 4 * size_mem)(prev_h)
sigmoid_chunk = cgt.sigmoid(input_sums[:, 0:3 * size_mem])
in_gate = sigmoid_chunk[:, 0:size_mem]
forget_gate = sigmoid_chunk[:, size_mem:2 * size_mem]
out_gate = sigmoid_chunk[:, 2 * size_mem:3 * size_mem]
in_transform = cgt.tanh(input_sums[:, 3 * size_mem:4 * size_mem])
next_c = forget_gate * prev_c + in_gate * in_transform
next_h = out_gate * cgt.tanh(next_c)
outputs.append(next_c)
outputs.append(next_h)
category_activations = nn.Affine(size_mem, size_output)(outputs[-1])
logprobs = nn.logsoftmax(category_activations)
outputs.append(logprobs)
return nn.Module(inputs, outputs)
def make_deep_rrnn_rot_relu(size_input, size_mem, n_layers, size_output,
size_batch_in, k_in, k_h):
inputs = [cgt.matrix() for i_layer in xrange(n_layers + 1)]
outputs = []
print 'input_size: ', size_input
for i_layer in xrange(n_layers):
prev_h = inputs[
i_layer +
1] # note that inputs[0] is the external input, so we add 1
x = inputs[0] if i_layer == 0 else outputs[i_layer - 1]
size_x = size_input if i_layer == 0 else size_mem
size_batch = prev_h.shape[0]
xform_h_param = nn.TensorParam((2 * k_h, size_mem), name="rotxform")
xform_h_non = xform_h_param.weight
xform_h_non.props["is_rotation"] = True
xform_h_norm = cgt.norm(xform_h_non, axis=1, keepdims=True)
xform_h = cgt.broadcast('/', xform_h_non, xform_h_norm, "xx,x1")
add_in_lin = nn.Affine(size_x, size_mem)(x)
add_in_relu = nn.rectify(add_in_lin)
prev_h_scaled = nn.scale_mag(prev_h)
h_in_added = prev_h_scaled + add_in_relu
inters_h = [h_in_added]
colon = slice(None, None, None)
for i in xrange(2 * k_h):
inter_in = inters_h[-1]
r_cur = xform_h[i, :]
#r_cur = cgt.subtensor(xform_h, [i, colon])
r_cur_2_transpose = cgt.reshape(r_cur, (size_mem, 1))
r_cur_2 = cgt.reshape(r_cur, (1, size_mem))
ref_cur = cgt.dot(cgt.dot(inter_in, r_cur_2_transpose), r_cur_2)
inter_out = inter_in - 2 * ref_cur
inters_h.append(inter_out)
next_h = inters_h[-1]
outputs.append(next_h)
category_activations = nn.Affine(size_mem, size_output,
name="pred")(outputs[-1])
logprobs = nn.logsoftmax(category_activations)
outputs.append(logprobs)
#print 'len outputs:', len(outputs)
#print 'len inputs:', len(inputs)
return nn.Module(inputs, outputs)
def build_fc_return_loss(X, y):
"""
Build fully connected network and return loss
"""
np.random.seed(0)
h1 = nn.rectify(
nn.Affine(28 * 28, 256, weight_init=nn.IIDGaussian(std=.1))(X))
h2 = nn.rectify(
nn.Affine(256, 256, weight_init=nn.IIDGaussian(std=.1))(h1))
logprobs = nn.logsoftmax(
nn.Affine(256, 10, weight_init=nn.IIDGaussian(std=.1))(h2))
neglogliks = -logprobs[cgt.arange(X.shape[0]), y]
loss = neglogliks.mean()
return loss
def hybrid_network(size_in, size_out, num_units, num_stos, dbg_out={}):
assert len(num_units) == len(num_stos)
net_in = cgt.matrix("X", fixed_shape=(None, size_in))
prev_num_units, prev_out = size_in, net_in
dbg_out['NET~in'] = net_in
curr_layer = 1
for (curr_num_units, curr_num_sto) in zip(num_units, num_stos):
assert curr_num_units >= curr_num_sto >= 0
prev_out = combo_layer(
prev_out,
prev_num_units,
curr_num_units, (curr_num_sto, ),
s_funcs=s_func_ip,
o_funcs=(lambda x: cgt.bernoulli(cgt.sigmoid(x)), cgt.nn.rectify),
name=str(curr_layer),
dbg_out=dbg_out)
dbg_out['L%d~out' % curr_layer] = prev_out
prev_num_units = curr_num_units
curr_layer += 1
net_out = nn.Affine(prev_num_units,
size_out,
name="InnerProd(%d->%d)" %
(prev_num_units, size_out))(prev_out)
dbg_out['NET~out'] = net_out
return net_in, net_out
def lstm_network_t(size_in, size_out, num_units, num_mems, dbg_out={}):
def s_func_lstm(_in, _s_in, _s_out, name=''):
c_prev = cgt.matrix(fixed_shape=(None, _s_out))
h_prev = cgt.matrix(fixed_shape=(None, _s_out))
c_cur, h_cur = lstm_block(h_prev, c_prev, _in, _s_in, _s_out, name)
net_c_prev.append(c_prev)
net_h_prev.append(h_prev)
net_c_curr.append(c_cur)
net_h_curr.append(h_cur)
return h_cur
assert len(num_units) == len(num_mems)
net_c_prev, net_h_prev, net_c_curr, net_h_curr = [], [], [], []
net_in = cgt.matrix(fixed_shape=(None, size_in))
prev_num_units, prev_out = size_in, net_in
curr_layer = 1
for curr_num_units, curr_num_mem in zip(num_units, num_mems):
assert curr_num_units >= curr_num_mem >= 0
prev_out = combo_layer(prev_out,
prev_num_units,
curr_num_units, (curr_num_mem, ),
s_funcs=(s_func_lstm, s_func_ip),
o_funcs=(None, cgt.sigmoid),
name=str(curr_layer),
dbg_out=dbg_out)
dbg_out['L%d~out' % curr_layer] = prev_out
prev_num_units = curr_num_units
curr_layer += 1
net_out = nn.Affine(prev_num_units, size_out, name="Out")(prev_out)
dbg_out['NET~out'] = net_out
return net_in, net_out, net_c_prev, net_h_prev, net_c_curr, net_h_curr
def hybrid_network(size_in, size_out, num_units, num_stos, dbg_out=[]):
assert len(num_units) == len(num_stos)
X = cgt.matrix("X", fixed_shape=(None, size_in))
prev_num_units, prev_out = size_in, X
dbg_out.append(X)
for (curr_num_units, curr_num_sto) in zip(num_units, num_stos):
_layer_dbg_out = []
prev_out = hybrid_layer(prev_out,
prev_num_units,
curr_num_units,
curr_num_sto,
dbg_out=_layer_dbg_out)
prev_num_units = curr_num_units
dbg_out.extend(_layer_dbg_out)
dbg_out.append(prev_out)
# TODO_TZ bigger problem! param cannot deterministically influence cost
# otherwise the surrogate cost is not complete log likelihood
net_out = nn.Affine(prev_num_units,
size_out,
name="InnerProd(%d->%d)" %
(prev_num_units, size_out))(prev_out)
dbg_out.append(net_out)
# assert prev_num_units == size_out
# net_out = prev_out
return X, net_out
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--profile",action="store_true")
parser.add_argument("--unittest",action="store_true")
parser.add_argument("--epochs",type=int,default=10)
args = parser.parse_args()
batchsize = 64
Xshape = (batchsize, 3, 32, 32)
X = cgt.tensor4("X", fixed_shape = Xshape)
y = cgt.vector("y", fixed_shape = (batchsize,), dtype='i4')
conv1 = nn.SpatialConvolution(3, 32, kernelshape=(5,5), pad=(2,2),
weight_init=nn.IIDGaussian(std=1e-4))(X)
relu1 = nn.rectify(conv1)
pool1 = nn.max_pool_2d(relu1, kernelshape=(3,3), stride=(2,2))
conv2 = nn.SpatialConvolution(32, 32, kernelshape=(5,5), pad=(2,2),
weight_init=nn.IIDGaussian(std=0.01))(relu1)
relu2 = nn.rectify(conv2)
pool2 = nn.max_pool_2d(relu2, kernelshape=(3,3), stride=(2,2))
conv3 = nn.SpatialConvolution(32, 64, kernelshape=(5,5), pad=(2,2),
weight_init=nn.IIDGaussian(std=0.01))(pool2)
pool3 = nn.max_pool_2d(conv3, kernelshape=(3,3), stride=(2,2))
relu3 = nn.rectify(pool3)
d0,d1,d2,d3 = relu3.shape
flatlayer = relu3.reshape([d0,d1*d2*d3])
nfeats = cgt.infer_shape(flatlayer)[1]
ip1 = nn.Affine(nfeats, 10)(flatlayer)
logprobs = nn.logsoftmax(ip1)
loss = -logprobs[cgt.arange(batchsize), y].mean()
params = nn.get_parameters(loss)
updates = rmsprop_updates(loss, params, stepsize=1e-3)
train = cgt.function(inputs=[X, y], outputs=[loss], updates=updates)
if args.profile: cgt.profiler.start()
data = np.load("/Users/joschu/Data/cifar-10-batches-py/cifar10.npz")
Xtrain = data["X_train"]
ytrain = data["y_train"]
print fmt_row(10, ["Epoch","Train NLL","Train Err","Test NLL","Test Err","Epoch Time"])
for i_epoch in xrange(args.epochs):
for start in xrange(0, Xtrain.shape[0], batchsize):
tstart = time.time()
end = start+batchsize
print train(Xtrain[start:end], ytrain[start:end]), time.time()-tstart
if start > batchsize*5: break
# elapsed = time.time() - tstart
# trainerr, trainloss = computeloss(Xtrain[:len(Xtest)], ytrain[:len(Xtest)])
# testerr, testloss = computeloss(Xtest, ytest)
# print fmt_row(10, [i_epoch, trainloss, trainerr, testloss, testerr, elapsed])
if args.profile:
cgt.profiler.print_stats()
return
if args.unittest:
break
def make_deep_gru(size_input, size_mem, n_layers, size_output, size_batch):
inputs = [cgt.matrix() for i_layer in xrange(n_layers + 1)]
outputs = []
for i_layer in xrange(n_layers):
prev_h = inputs[
i_layer +
1] # note that inputs[0] is the external input, so we add 1
x = inputs[0] if i_layer == 0 else outputs[i_layer - 1]
size_x = size_input if i_layer == 0 else size_mem
update_gate = cgt.sigmoid(
nn.Affine(size_x, size_mem, name="i2u")(x) +
nn.Affine(size_mem, size_mem, name="h2u")(prev_h))
reset_gate = cgt.sigmoid(
nn.Affine(size_x, size_mem, name="i2r")(x) +
nn.Affine(size_mem, size_mem, name="h2r")(prev_h))
gated_hidden = reset_gate * prev_h
p2 = nn.Affine(size_mem, size_mem)(gated_hidden)
p1 = nn.Affine(size_x, size_mem)(x)
hidden_target = cgt.tanh(p1 + p2)
next_h = (1.0 - update_gate) * prev_h + update_gate * hidden_target
outputs.append(next_h)
category_activations = nn.Affine(size_mem, size_output,
name="pred")(outputs[-1])
logprobs = nn.logsoftmax(category_activations)
outputs.append(logprobs)
return nn.Module(inputs, outputs)
def lstm_block(h_prev, c_prev, x_curr, size_x, size_c, name=''):
"""
Construct a LSTM cell block of specified number of cells
:param h_prev: self activations at previous time step
:param c_prev: self memory state at previous time step
:param x_curr: inputs from previous layer at current time step
:param size_x: size of inputs
:param size_c: size of both c and h
:return: c and h at current time step
:rtype:
"""
input_sums = nn.Affine(size_x, 4 * size_c, name=name+'*x')(x_curr) + \
nn.Affine(size_c, 4 * size_c, name=name+'*h')(h_prev)
c_new = cgt.tanh(input_sums[:, 3 * size_c:])
sigmoid_chunk = cgt.sigmoid(input_sums[:, :3 * size_c])
in_gate = sigmoid_chunk[:, :size_c]
forget_gate = sigmoid_chunk[:, size_c:2 * size_c]
out_gate = sigmoid_chunk[:, 2 * size_c:3 * size_c]
c_curr = forget_gate * c_prev + in_gate * c_new
h_curr = out_gate * cgt.tanh(c_curr)
return c_curr, h_curr
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, num_features=None, num_hidden=100):
stepsize = 0.01
# with shape (batchsize, ncols)
X = cgt.matrix("X", fixed_shape=(1, num_features))
# y: a symbolic variable representing the rewards, which are integers
y = cgt.scalar("y", dtype='float64')
hid1 = nn.rectify(
nn.Affine(num_features,
num_hidden,
weight_init=nn.IIDGaussian(std=.1),
bias_init=nn.Constant(1))(X))
# One final fully-connected layer, and then a linear activation output for reward
output = nn.Affine(num_hidden,
1,
weight_init=nn.IIDGaussian(std=.1),
bias_init=nn.Constant(1))(hid1)
abs_deviation = cgt.abs(output - y).mean()
params = nn.get_parameters(abs_deviation)
gparams = cgt.grad(abs_deviation, params)
updates = [(p, p - stepsize * gp) for (p, gp) in zip(params, gparams)]
self.predictor = cgt.function([X], output)
self.updater = cgt.function([X, y], abs_deviation, updates=updates)
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 build_convnet_return_loss(X, y):
np.random.seed(0)
conv1 = nn.rectify(
nn.SpatialConvolution(1,
32,
kernelshape=(3, 3),
pad=(0, 0),
weight_init=nn.IIDGaussian(std=.1))(X))
pool1 = nn.max_pool_2d(conv1, kernelshape=(3, 3), stride=(2, 2))
conv2 = nn.rectify(
nn.SpatialConvolution(32,
32,
kernelshape=(3, 3),
pad=(0, 0),
weight_init=nn.IIDGaussian(std=.1))(pool1))
pool2 = nn.max_pool_2d(conv2, kernelshape=(3, 3), stride=(2, 2))
d0, d1, d2, d3 = pool2.shape
flatlayer = pool2.reshape([d0, d1 * d2 * d3])
nfeats = cgt.infer_shape(flatlayer)[1]
logprobs = nn.logsoftmax(nn.Affine(nfeats, 10)(flatlayer))
loss = -logprobs[cgt.arange(X.shape[0]), y].mean()
return loss
import cgt
from cgt import nn, utils
import numpy as np, numpy.random as nr
from numpy.linalg import norm
from param_collection import ParamCollection
k_in = 1
size_x = 3
size_mem = 4
size_batch = 4
x = cgt.matrix(fixed_shape=(size_batch, size_x))
prev_h = cgt.matrix(fixed_shape=(size_batch, size_mem))
r_vec = nn.Affine(size_x, 2 * k_in * size_mem)(x)
r_non = cgt.reshape(r_vec, (size_batch, 2 * k_in, size_mem))
r_norm = cgt.norm(r_non, axis=2, keepdims=True)
r = cgt.broadcast('/', r_non, r_norm, "xxx,xx1")
prev_h_3 = cgt.reshape(prev_h, (size_batch, size_mem, 1))
inters = [prev_h_3]
for i in xrange(k_in * 2):
inter_in = inters[-1]
r_cur = r[:, i, :]
r_cur_3_transpose = cgt.reshape(r_cur, (size_batch, 1, size_mem))
r_cur_3 = cgt.reshape(r_cur, (size_batch, size_mem, 1))
ref_cur = cgt.batched_matmul(
r_cur_3, cgt.batched_matmul(r_cur_3_transpose, inter_in))
inter_out = inter_in - ref_cur
inters.append(inter_out)
h = inters[-1]
def s_func_ip(X, size_in, size_out, name):
return nn.Affine(size_in, size_out, name=name)(X)
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)