def test_get_train_objective():
batch_size = 32
feat_t_steps = 5
feat_num_features = 256
max_label_length = 5
num_out_classes = 27
feats = cgt.tensor3(fixed_shape=(batch_size, feat_t_steps, feat_num_features))
ground_labels_basis = cgt.tensor3(fixed_shape=(batch_size, max_label_length, num_out_classes))
seq2seq = nnbuilder.Seq2Seq(nn_input_btf=feats, num_out_classes=num_out_classes)
train_objective = seq2seq.get_train_objective(max_label_length=max_label_length,
ground_labels_basis_btc=ground_labels_basis)
train_shape = cgt.infer_shape(train_objective)
assert train_shape == ()
nn.get_parameters(train_objective)
def save_weights(network_out_layer, file_name_for_saving):
all_params = get_parameters(network_out_layer)
param_values = []
for param in all_params:
param_values.append(param.op.get_value())
pickle.dump(param_values, open(file_name_for_saving+".p", "wb"))
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 __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 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 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))(pool1)
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 = fetch_dataset("http://rll.berkeley.edu/cgt-data/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_updater_fc():
X = cgt.matrix("X", fixed_shape=(None, 28 * 28))
y = cgt.vector("y", dtype="i8")
stepsize = cgt.scalar("stepsize")
loss = build_fc_return_loss(X, y)
params = nn.get_parameters(loss)
gparams = cgt.grad(loss, params)
updates = [(p, p - stepsize * gp) for (p, gp) in zip(params, gparams)]
return cgt.function([X, y, stepsize], loss, updates=updates)
def make_updater_fc():
X = cgt.matrix("X", fixed_shape=(None, 28 * 28))
y = cgt.vector("y", dtype='i8')
stepsize = cgt.scalar("stepsize")
loss = build_fc_return_loss(X, y)
params = nn.get_parameters(loss)
gparams = cgt.grad(loss, params)
updates = [(p, p - stepsize * gp) for (p, gp) in zip(params, gparams)]
return cgt.function([X, y, stepsize], loss, updates=updates)
def make_updater_convnet():
X = cgt.tensor4("X", fixed_shape=(None, 1, 28, 28)) # so shapes can be inferred
y = cgt.vector("y", dtype="i8")
stepsize = cgt.scalar("stepsize")
loss = build_convnet_return_loss(X, y)
params = nn.get_parameters(loss)
gparams = cgt.grad(loss, params)
updates = [(p, p - stepsize * gp) for (p, gp) in zip(params, gparams)]
return cgt.function([X, y, stepsize], loss, updates=updates)
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_updater_convnet():
X = cgt.tensor4("X", fixed_shape=(None, 1, 28,
28)) # so shapes can be inferred
y = cgt.vector("y", dtype='i8')
stepsize = cgt.scalar("stepsize")
loss = build_convnet_return_loss(X, y)
params = nn.get_parameters(loss)
gparams = cgt.grad(loss, params)
updates = [(p, p - stepsize * gp) for (p, gp) in zip(params, gparams)]
return cgt.function([X, y, stepsize], loss, updates=updates)
def make_updater_convnet_parallel():
X = cgt.tensor4("X", fixed_shape=(None, 1, 28, 28)) # so shapes can be inferred
y = cgt.vector("y", dtype="i8")
stepsize = cgt.scalar("stepsize")
loss = build_convnet_return_loss(X, y)
m = nn.Module([X, y], [loss])
split_loss = 0
for start in xrange(0, batch_size, batch_size // 4):
sli = slice(start, start + batch_size // 4)
split_loss += m([X[sli], y[sli]])[0]
split_loss /= 4
params = nn.get_parameters(loss)
gparams = cgt.grad(split_loss, params)
updates2 = [(p, p - stepsize * gp) for (p, gp) in zip(params, gparams)]
return cgt.function([X, y, stepsize], split_loss, updates=updates2)
def make_updater_fc_parallel():
X = cgt.matrix("X", fixed_shape=(None,28*28))
y = cgt.vector("y",dtype='i8')
stepsize = cgt.scalar("stepsize")
loss = build_fc_return_loss(X,y)
params = nn.get_parameters(loss)
m = nn.Module([X,y], [loss])
split_loss = 0
for start in xrange(0, batch_size, batch_size//4):
sli = slice(start, start+batch_size//4)
split_loss += m([X[sli], y[sli]])[0]
split_loss /= 4
gparams = cgt.grad(split_loss, params)
updates2 = [(p, p-stepsize*gp) for (p, gp) in zip(params, gparams)]
return cgt.function([X,y, stepsize], split_loss, updates=updates2)
def make_updater_fc_parallel():
X = cgt.matrix("X", fixed_shape=(None, 28 * 28))
y = cgt.vector("y", dtype='i8')
stepsize = cgt.scalar("stepsize")
loss = build_fc_return_loss(X, y)
params = nn.get_parameters(loss)
m = nn.Module([X, y], [loss])
split_loss = 0
for start in xrange(0, batch_size, batch_size // 4):
sli = slice(start, start + batch_size // 4)
split_loss += m([X[sli], y[sli]])[0]
split_loss /= 4
gparams = cgt.grad(split_loss, params)
updates2 = [(p, p - stepsize * gp) for (p, gp) in zip(params, gparams)]
return cgt.function([X, y, stepsize], split_loss, updates=updates2)
def main():
print("Loading data...")
X = cgt.matrix("X", fixed_shape=(None, 28*28))
y = cgt.vector("y", dtype='i8')
model = build_model(X, 0.0)
loss = -cgt.mean(categorical.loglik(y, model))
updates = nn.rmsprop(loss, nn.get_parameters(loss), 0.01)
train = cgt.function(inputs=[X, y], outputs=[], updates=updates)
y_nodrop = cgt.argmax(model, axis=1)
cost_nodrop = -cgt.mean(categorical.loglik(y, model))
err_nodrop = cgt.cast(cgt.not_equal(y_nodrop, y), cgt.floatX).mean()
computeloss = cgt.function(inputs=[X, y], outputs=[err_nodrop, cost_nodrop])
batch_size=128
Xdata, ydata = load_data()
Xtrain = Xdata[0:60000]
ytrain = ydata[0:60000]
Xtest = Xdata[60000:70000]
ytest = ydata[60000:70000]
sortinds = np.random.permutation(60000)
Xtrain = Xtrain[sortinds]
ytrain = ytrain[sortinds]
print fmt_row(10, ["Epoch","Train NLL","Train Err","Test NLL","Test Err","Epoch Time"])
for i_epoch in xrange(3):
tstart = time.time()
for start in xrange(0, Xtrain.shape[0], batch_size):
end = start+batch_size
train(Xtrain[start:end], ytrain[start:end])
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])
nnbuilder.save_weights(model, 'mnist')
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 __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 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 __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 main():
X = cgt.matrix(name='data', dtype=cgt.floatX, fixed_shape=(None, 2212))
y = cgt.vector("y", dtype='i8')
model = build_nn(X)
loss = -cgt.mean(categorical.loglik(y, model))
updates = nn.adagrad(loss, nn.get_parameters(loss), 0.01)
y_nodrop = cgt.argmax(model, axis=1)
cost_nodrop = -cgt.mean(categorical.loglik(y, model))
err_nodrop = cgt.cast(cgt.not_equal(y_nodrop, y), cgt.floatX).mean()
train = cgt.function(inputs=[X, y], outputs=[], updates=updates)
computeloss = cgt.function(inputs=[X, y], outputs=[err_nodrop, cost_nodrop])
batch_size = 20
Xdata, ydata = load_data()
Xtrain = Xdata[0:5200]
ytrain = ydata[0:5200]
Xtest = Xdata[5200:5573]
ytest = ydata[5200:5573]
sortinds = np.random.permutation(5200)
Xtrain = Xtrain[sortinds]
ytrain = ytrain[sortinds]
print fmt_row(10, ["Epoch","Train NLL","Train Err","Test NLL","Test Err","Epoch Time"])
for i_epoch in xrange(20):
tstart = time.time()
for start in xrange(0, Xtrain.shape[0], batch_size):
end = start+batch_size
train(Xtrain[start:end], ytrain[start:end])
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])
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 __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 __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 test_seq_2_seq():
batch_size = 32 # How many samples do you want to batch.
feat_t_steps = 3 # How many 10ms sound clips.
feat_num_features = 10 # The dimension of the 10ms clips.
max_label_length = feat_t_steps # The maximal label length of the transcription.
num_out_classes = 27 # 26 letters and space.
num_out_classes_true = 27 + 2 # Start and end tokens are added.
num_batches = 512 # 1032
num_epochs = 40
feats = cgt.tensor3(fixed_shape=(batch_size, feat_t_steps, feat_num_features))
ground_labels_basis = cgt.tensor3(fixed_shape=(batch_size, max_label_length, num_out_classes_true))
last_time = time.time()
print 'initializing seq2seq'
seq2seq = nnbuilder.Seq2Seq(nn_input_btf=feats, num_out_classes=num_out_classes)
print 'that took ' + str(time.time() - last_time) + ' seconds'
last_time = time.time()
print 'making train objective'
train_objective = seq2seq.get_train_objective(max_label_length=max_label_length,
ground_labels_basis_btc=ground_labels_basis)
print 'that took ' + str(time.time() - last_time) + ' seconds'
last_time = time.time()
print 'making updates'
updates = nn.rmsprop(train_objective, nn.get_parameters(train_objective), learning_rate=0.0001)
#updates = nn.nesterov_momentum(train_objective, nn.get_parameters(train_objective), learning_rate=0.0001, mu=0.4)
#updates = nn.momentum(train_objective, nn.get_parameters(train_objective), learning_rate=0.00001, mu=0.4)
#updates = nn.adadelta(train_objective, nn.get_parameters(train_objective), learning_rate=0.0001, rho=0.95)
print 'that took ' + str(time.time() - last_time) + ' seconds'
last_time = time.time()
print 'compiling train function, test function, and prediction output function'
train_function = cgt.function([feats, ground_labels_basis], [], updates=updates)
test_function = cgt.function([feats, ground_labels_basis], [train_objective])
pred = seq2seq.make_prediction(ground_labels_basis_btc=ground_labels_basis, max_label_length=feat_t_steps)
pred_fun = cgt.function([feats, ground_labels_basis], [pred])
print 'that took ' + str(time.time() - last_time) + ' seconds'
test_data = np.load('test_data.npy')
test_labels = np.load('test_labels.npy')
data_mean = np.mean(test_data)
data_sd = np.std(test_data)
print 'now training'
last_time = time.time()
for one_epoch in range(0, num_epochs):
tested = 0
print 'starting epoch ' + str(one_epoch)
for batch_iter in range(0, num_batches):
batch, labels_basis = normalize_batch_and_labels(test_data, batch_iter, feat_t_steps, data_mean, data_sd,
test_labels, num_out_classes_true)
train_function(batch, labels_basis)
for batch_iter in range(0, num_batches):
batch, labels_basis = normalize_batch_and_labels(test_data, batch_iter, feat_t_steps, data_mean, data_sd,
test_labels, num_out_classes_true)
tested += test_function(batch, labels_basis)[0]
tested = tested / batch_iter
print 'train loss for batch ' + str(batch_iter) + ' is ' + str(tested)
print 'an actual prediction is '
print pred_fun(batch, labels_basis)[0]
print 'the truth is'
print test_labels[batch_iter, :, 0:feat_t_steps]
print 'that took ' + str(time.time() - last_time) + ' seconds'
last_time = time.time()
prediction_final = pred_fun(batch, labels_basis)[0]
print prediction_final
def test_the_test_problem():
#Works
batch_size = 32 # How many samples do you want to batch.
feat_t_steps = 20 # How many 10ms sound clips.
feat_num_features = 10 # The dimension of the 10ms clips.
max_label_length = feat_t_steps # The maximal label length of the transcription. includes start character.
num_out_classes = 27
num_out_classes_true = num_out_classes + 2
num_batches = 756
num_epochs = 30
feats = cgt.tensor3(fixed_shape=(batch_size, feat_t_steps, feat_num_features))
ground_labels_basis = cgt.tensor3(fixed_shape=(batch_size, max_label_length, num_out_classes_true))
last_time = time.time()
print 'initializing temporal dense layer'
d1 = nnbuilder.temporalDenseLayer(feats, num_units=128, activation=cgt.sigmoid)
#d2 = nnbuilder.temporalDenseLayer(d1, num_units=128, activation=cgt.sigmoid)
d3 = nnbuilder.temporalDenseLayer(d1, num_units=num_out_classes_true, activation=nnbuilder.linear)
out = nn.three_d_softmax(d3, axis=2)
log_probs = None
for iter_step in range(0, max_label_length):
this_character_dist_bc = out[:, iter_step, :]
prev_out_bc = ground_labels_basis[:, iter_step, :]
log_probs_pre = prev_out_bc * this_character_dist_bc
log_probs_pre = cgt.log(cgt.sum(log_probs_pre, axis=1))
if log_probs is None:
log_probs = cgt.sum(log_probs_pre)
else:
log_probs += cgt.sum(log_probs_pre)
log_probs = -log_probs
print 'that took ' + str(time.time() - last_time) + ' seconds'
last_time = time.time()
print 'compiling objective function'
updates = nn.rmsprop(log_probs, nn.get_parameters(log_probs), learning_rate=0.01)
pred_train = cgt.function([feats, ground_labels_basis], [], updates=updates)
pred_fun = cgt.function([feats, ground_labels_basis], [log_probs])
most_likely_chars = cgt.argmax(out, axis=1)
actual_predictions = cgt.function([feats, ground_labels_basis], [most_likely_chars])
print 'that took ' + str(time.time() - last_time) + ' seconds'
test_data = np.load('test_data.npy')
test_labels = np.load('test_labels.npy')
data_mean = np.mean(test_data)
data_sd = np.mean(test_data)
print 'now training'
for one_epoch in range(0, num_epochs):
trained = 0
last_time = time.time()
print 'starting epoch ' + str(one_epoch)
for batch_iter in range(0, num_batches):
batch, labels_basis = normalize_batch_and_labels(test_data, batch_iter, feat_t_steps, data_mean, data_sd,
test_labels, num_out_classes_true)
pred_train(batch, labels_basis)
for batch_iter in range(0, num_batches):
batch, labels_basis = normalize_batch_and_labels(test_data, batch_iter, feat_t_steps, data_mean, data_sd,
test_labels, num_out_classes_true)
trained += pred_fun(batch, labels_basis)[0]
trained = trained/batch_iter
print 'train loss is ' + str(trained)
print 'that took ' + str(time.time() - last_time) + ' seconds'
act_pred = actual_predictions(batch, labels_basis)[0]
print 'an actual prediction is '
print act_pred
conv2 = nn.rectify(
nn.SpatialConvolution(32, 32, kernelshape=(3,3), stride=(1,1), pad=(1,1), weight_init=nn.IIDGaussian(std=.1))(pool1)
)
pool2 = nn.max_pool_2d(conv2, kernelshape=(2,2), stride=(2,2))
d0, d1, d2, d3 = pool2.shape
flat = pool2.reshape([d0, d1*d2*d3])
nfeats = cgt.infer_shape(flat)[1]
probs = nn.softmax(nn.Affine(nfeats, 10)(flat))
cost = -categorical.loglik(y, probs).mean()
y_preds = cgt.argmax(probs, axis=1)
err = cgt.cast(cgt.not_equal(y, y_preds), cgt.floatX).mean()
params = nn.get_parameters(cost)
updates = nn.sgd(cost, params, 1e-3)
# training function
f = cgt.function(inputs=[X, y], outputs=[], updates=updates)
# compute the cost and error
cost_and_err = cgt.function(inputs=[X, y], outputs=[cost, err])
for i in xrange(epochs):
t0 = time.time()
for start in xrange(0, Xtrain.shape[0], batch_size):
end = batch_size + start
f(Xtrainimg[start:end], ytrain[start:end])
elapsed = time.time() - t0
costval, errval = cost_and_err(Xtestimg, ytest)
print("Epoch {} took {}, test cost = {}, test error = {}".format(i, elapsed, costval, errval))
def set_all_weights_helper(network_out_layer, param_values):
all_params = get_parameters(network_out_layer)
for param, param_value in zip(all_params, param_values):
param.op.set_value(param_value)
def get_all_weights(network_out_layer):
all_params = get_parameters(network_out_layer)
param_values = []
for param in all_params:
param_values.append(param.op.get_value())
return param_values
def set_all_weights(network_out_layer, pickled_list_of_weights):
all_params = get_parameters(network_out_layer)
param_values = pickle.load(open(pickled_list_of_weights, "rb"))
for param, param_value in zip(all_params, param_values):
param.op.set_value(param_value)
