def test_einsum():
x = cgt.tensor3()
y = cgt.tensor3()
sizes = {'i': 2, 'j': 3, 'k': 5, 'l': 7}
xaxes = 'ijk'
yaxes = 'ikl'
zaxes = 'ijl'
for i in xrange(10):
xperm = xaxes
(yperm,
zperm) = permaxes = [[chars[i] for i in np.random.permutation(3)]
for chars in [yaxes, zaxes]]
desc = "%s,%s->%s" % tuple("".join(chars)
for chars in [xperm] + permaxes)
z = cgt.einsum(desc, x, y)
xval = nr.randn(*(sizes[c] for c in xperm))
yval = nr.randn(*(sizes[c] for c in yperm))
np.testing.assert_allclose(cgt.numeric_eval(z, {
x: xval,
y: yval
}),
np.einsum(desc, xval, yval),
atol={
"single": 1e-3,
"double": 1e-6
}[cgt.get_precision()])
def make_ntm(opt):
Mprev_bnm = cgt.tensor3("M", fixed_shape=(opt.b, opt.n, opt.m))
X_bk = cgt.matrix("X", fixed_shape=(opt.b, opt.k))
wprev_bHn = cgt.tensor3("w", fixed_shape=(opt.b, opt.h*2, opt.n))
rprev_bhm = cgt.tensor3("r", fixed_shape=(opt.b, opt.h, opt.m))
controller = make_ff_controller(opt)
M_bnm, w_bHn, r_bhm, y_bp = ntm_step(opt, Mprev_bnm, X_bk, wprev_bHn, rprev_bhm, controller)
# in this form it looks like a standard seq-to-seq model
# external input and output are first elements
ntm = nn.Module([X_bk, Mprev_bnm, wprev_bHn, rprev_bhm], [y_bp, M_bnm, w_bHn, r_bhm])
return ntm
def make_ntm(opt):
Mprev_bnm = cgt.tensor3("M", fixed_shape=(opt.b, opt.n, opt.m))
X_bk = cgt.matrix("X", fixed_shape=(opt.b, opt.k))
wprev_bHn = cgt.tensor3("w", fixed_shape=(opt.b, opt.h*2, opt.n))
rprev_bhm = cgt.tensor3("r", fixed_shape=(opt.b, opt.h, opt.m))
controller = make_ff_controller(opt)
M_bnm, w_bHn, r_bhm, y_bp = ntm_step(opt, Mprev_bnm, X_bk, wprev_bHn, rprev_bhm, controller)
# in this form it looks like a standard seq-to-seq model
# external input and output are first elements
ntm = nn.Module([X_bk, Mprev_bnm, wprev_bHn, rprev_bhm], [y_bp, M_bnm, w_bHn, r_bhm])
return ntm
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 test_get_character_distribution():
batch_size = 32
feat_t_steps = 20
feat_num_features = 42
num_out_classes = 28 # This is the index of the start token.
num_out_classes_true = num_out_classes + 2 # Add start and end tokens automatically.
decoder_size = 50
tau = np.reshape(np.random.normal(0.1, 0.2, batch_size*feat_t_steps*feat_num_features), (batch_size, feat_t_steps, feat_num_features))
tau2 = np.reshape(np.random.normal(0.1, 0.2, batch_size*feat_num_features), (batch_size, feat_num_features))
tau3 = np.reshape(np.random.normal(0.1, 0.2, batch_size*decoder_size), (batch_size, decoder_size))
feats = cgt.tensor3(fixed_shape=(batch_size, feat_t_steps, feat_num_features))
s = nnbuilder.Seq2Seq(nn_input_btf=feats, num_out_classes=num_out_classes,
decoder_size=decoder_size, feature_size=feat_num_features)
context_bf = cgt.matrix(fixed_shape=(batch_size, feat_num_features))
state_bf = cgt.matrix(fixed_shape=(batch_size, decoder_size))
m_out = s.get_character_distribution(state_bf, context_bf)
out_fun = cgt.function([feats, context_bf, state_bf], [m_out])
m = out_fun(tau, tau2, tau3)[0]
assert m.shape == (batch_size, num_out_classes_true)
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_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 test_get_decoder_state():
batch_size = 32
feat_t_steps = 20
feat_num_features = 42
num_out_classes = 28
num_out_classes_true = num_out_classes + 2 # Start, end, are added
decoder_size = 50
tau = np.reshape(np.random.normal(0.1, 0.2, batch_size*feat_t_steps*feat_num_features), (batch_size, feat_t_steps, feat_num_features))
tau2 = np.reshape(np.random.normal(0.1, 0.2, batch_size*feat_num_features), (batch_size, feat_num_features))
tau3 = np.reshape(np.random.normal(0.1, 0.2, batch_size*num_out_classes_true), (batch_size, num_out_classes_true))
feats = cgt.tensor3(fixed_shape=(batch_size, feat_t_steps, feat_num_features))
s = nnbuilder.Seq2Seq(nn_input_btf=feats, num_out_classes=num_out_classes,
decoder_size=decoder_size, feature_size=feat_num_features)
context_bf = cgt.matrix(fixed_shape=(batch_size, feat_num_features))
prev_out_bc = cgt.matrix(fixed_shape=(batch_size, num_out_classes_true))
state_i_bf = nn.parameter(nn.init_array(nn.IIDGaussian(0.1), (batch_size, decoder_size)), name="decoder_init")
decoder_out = s.get_decoder_state(context_bf, prev_out_bc, state_i_bf)
decode_fun = cgt.function([feats, context_bf, prev_out_bc], [decoder_out])
m = decode_fun(tau, tau2, tau3)[0]
assert m.shape == (batch_size, decoder_size)
assert np.mean(m) < 1.0
def make_loss_and_grad_and_step(arch, size_input, size_output, size_mem,
size_batch, n_layers, n_unroll, k_in, k_h):
# symbolic variables
x_tnk = cgt.tensor3()
targ_tnk = cgt.tensor3()
#make_network = make_deep_lstm if arch=="lstm" else make_deep_gru
make_network = make_deep_rrnn_rot_relu
network = make_network(size_input, size_mem, n_layers, size_output,
size_batch, k_in, k_h)
init_hiddens = [
cgt.matrix() for _ in xrange(get_num_hiddens(arch, n_layers))
]
# TODO fixed sizes
cur_hiddens = init_hiddens
loss = 0
for t in xrange(n_unroll):
outputs = network([x_tnk[t]] + cur_hiddens)
cur_hiddens, prediction_logprobs = outputs[:-1], outputs[-1]
# loss = loss + nn.categorical_negloglik(prediction_probs, targ_tnk[t]).sum()
loss = loss - (prediction_logprobs * targ_tnk[t]).sum()
cur_hiddens = outputs[:-1]
final_hiddens = cur_hiddens
loss = loss / (n_unroll * size_batch)
params = network.get_parameters()
gradloss = cgt.grad(loss, params)
flatgrad = flatcat(gradloss)
with utils.Message("compiling loss+grad"):
f_loss_and_grad = cgt.function([x_tnk, targ_tnk] + init_hiddens,
[loss, flatgrad] + final_hiddens)
f_loss = cgt.function([x_tnk, targ_tnk] + init_hiddens, loss)
assert len(init_hiddens) == len(final_hiddens)
x_nk = cgt.matrix('x')
outputs = network([x_nk] + init_hiddens)
f_step = cgt.function([x_nk] + init_hiddens, outputs)
# print "node count", cgt.count_nodes(flatgrad)
return network, f_loss, f_loss_and_grad, f_step
def test_seq2seq_init():
batch_size = 32
feat_t_steps = 5
feat_num_features = 256
feats = cgt.tensor3(fixed_shape=(batch_size, feat_t_steps, feat_num_features))
s = nnbuilder.Seq2Seq(nn_input_btf=feats, num_out_classes=28)
#assert type(s.recurrent_decoder_one) == nnbuilder.GRULayer
assert s.get_features_fun == s.get_features_bengio
def test_make_prediction():
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 = 27 + 2
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 prediction objective'
pred = seq2seq.make_prediction(ground_labels_basis_btc=ground_labels_basis, max_label_length=feat_t_steps)
print 'that took ' + str(time.time() - last_time) + ' seconds'
last_time = time.time()
print 'compiling pred function'
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.mean(test_data)
print 'now predicting'
last_time = time.time()
batch_iter = 0
batch = test_data[batch_iter, :, 0:feat_t_steps, :]
batch = batch - data_mean
batch = batch / data_sd
labels = test_labels[batch_iter, :, 0:feat_t_steps]
labels_basis = ind_to_basis(num_out_classes_true, labels)
prediction_final = pred_fun(batch, labels_basis)[0]
one_pred = prediction_final[0]
print 'that took ' + str(time.time() - last_time) + ' seconds'
def make_loss_and_grad_and_step(arch, size_input, size_output, size_mem, size_batch, n_layers, n_unroll, k_in, k_h):
# symbolic variables
x_tnk = cgt.tensor3()
targ_tnk = cgt.tensor3()
#make_network = make_deep_lstm if arch=="lstm" else make_deep_gru
make_network = make_deep_rrnn_rot_relu
network = make_network(size_input, size_mem, n_layers, size_output, size_batch, k_in, k_h)
init_hiddens = [cgt.matrix() for _ in xrange(get_num_hiddens(arch, n_layers))]
# TODO fixed sizes
cur_hiddens = init_hiddens
loss = 0
for t in xrange(n_unroll):
outputs = network([x_tnk[t]] + cur_hiddens)
cur_hiddens, prediction_logprobs = outputs[:-1], outputs[-1]
# loss = loss + nn.categorical_negloglik(prediction_probs, targ_tnk[t]).sum()
loss = loss - (prediction_logprobs*targ_tnk[t]).sum()
cur_hiddens = outputs[:-1]
final_hiddens = cur_hiddens
loss = loss / (n_unroll * size_batch)
params = network.get_parameters()
gradloss = cgt.grad(loss, params)
flatgrad = flatcat(gradloss)
with utils.Message("compiling loss+grad"):
f_loss_and_grad = cgt.function([x_tnk, targ_tnk] + init_hiddens, [loss, flatgrad] + final_hiddens)
f_loss = cgt.function([x_tnk, targ_tnk] + init_hiddens, loss)
assert len(init_hiddens) == len(final_hiddens)
x_nk = cgt.matrix('x')
outputs = network([x_nk] + init_hiddens)
f_step = cgt.function([x_nk]+init_hiddens, outputs)
# print "node count", cgt.count_nodes(flatgrad)
return network, f_loss, f_loss_and_grad, f_step
def test_einsum():
x = cgt.tensor3()
y = cgt.tensor3()
sizes = {"i": 2, "j": 3, "k": 5, "l": 7}
xaxes = "ijk"
yaxes = "ikl"
zaxes = "ijl"
for i in xrange(10):
xperm = xaxes
(yperm, zperm) = permaxes = [[chars[i] for i in np.random.permutation(3)] for chars in [yaxes, zaxes]]
desc = "%s,%s->%s" % tuple("".join(chars) for chars in [xperm] + permaxes)
z = cgt.einsum(desc, x, y)
xval = nr.randn(*(sizes[c] for c in xperm))
yval = nr.randn(*(sizes[c] for c in yperm))
np.testing.assert_allclose(
cgt.numeric_eval(z, {x: xval, y: yval}),
np.einsum(desc, xval, yval),
atol={"single": 1e-3, "double": 1e-6}[cgt.get_precision()],
)
def test_einsum():
cgt.reset_config()
cgt.set_precision("double")
x = cgt.tensor3()
y = cgt.tensor3()
sizes = {'i':2,'j':3,'k':5,'l':7}
xaxes = 'ijk'
yaxes = 'ikl'
zaxes = 'ijl'
for i in xrange(10):
xperm = xaxes
(yperm,zperm) = permaxes = [[chars[i] for i in np.random.permutation(3)] for chars in [yaxes,zaxes]]
desc = "%s,%s->%s"%tuple("".join(chars) for chars in [xperm] + permaxes)
z = cgt.einsum(desc, x, y)
xval = nr.randn(*(sizes[c] for c in xperm))
yval = nr.randn(*(sizes[c] for c in yperm))
np.testing.assert_allclose(
cgt.numeric_eval(z, {x : xval, y : yval}),
np.einsum(desc, xval, yval))
def make_funcs(opt, ntm, total_time, loss_timesteps):
x_tbk = cgt.tensor3("x", fixed_shape=(total_time, opt.b, opt.k))
y_tbp = cgt.tensor3("y", fixed_shape=(total_time, opt.b, opt.p))
loss_timesteps = set(loss_timesteps)
initial_states = make_ntm_initial_states(opt)
params = ntm.get_parameters() + get_parameters(initial_states)
# params = ntm.get_parameters()
lossCE = 0
loss01 = 0
state_arrs = initial_states
for t in xrange(total_time):
tmp = ntm([x_tbk[t]] + state_arrs)
raw_pred = tmp[0]
state_arrs = tmp[1:4]
if t in loss_timesteps:
p_pred = cgt.sigmoid(raw_pred)
ce = bernoulli_crossentropy(
y_tbp[t],
p_pred).sum() # cross-entropy of bernoulli distribution
lossCE = lossCE + ce
loss01 = loss01 + cgt.cast(cgt.equal(y_tbp[t], round01(p_pred)),
cgt.floatX).sum()
lossCE = lossCE / (len(loss_timesteps) * opt.p * opt.b) / np.log(2)
loss01 = loss01 / (len(loss_timesteps) * opt.p * opt.b)
gradloss = cgt.grad(lossCE, params)
flatgrad = flatcat(gradloss)
f_loss = cgt.function([x_tbk, y_tbp], lossCE)
f_loss_and_grad = cgt.function([x_tbk, y_tbp], [lossCE, loss01, flatgrad])
print "number of nodes in computation graph:", core.count_nodes(
[lossCE, loss01, flatgrad])
return f_loss, f_loss_and_grad, params
def test_get_features_default():
batch_size = 32
feat_t_steps = 20
feat_num_features = 42
num_out_classes = 28
num_units = 20
feats = cgt.tensor3(fixed_shape=(batch_size, feat_t_steps, feat_num_features))
s = nnbuilder.Seq2Seq(nn_input_btf=feats, num_out_classes=num_out_classes)
out = s.get_features_fun(feats, num_units=num_units)
out_fun = cgt.function([feats], [out])
tau = np.reshape(np.random.normal(0.1, 0.2, batch_size*feat_t_steps*feat_num_features), (batch_size, feat_t_steps, feat_num_features))
m = out_fun(tau)[0]
assert m.shape == (batch_size, feat_t_steps, num_units)
assert np.mean(m) < 1
def make_funcs(opt, ntm, total_time, loss_timesteps):
x_tbk = cgt.tensor3("x", fixed_shape=(total_time, opt.b, opt.k))
y_tbp = cgt.tensor3("y", fixed_shape=(total_time, opt.b, opt.p))
loss_timesteps = set(loss_timesteps)
initial_states = make_ntm_initial_states(opt)
params = ntm.get_parameters() + get_parameters(initial_states)
# params = ntm.get_parameters()
lossCE = 0
loss01 = 0
state_arrs = initial_states
for t in xrange(total_time):
tmp = ntm([x_tbk[t]] + state_arrs)
raw_pred = tmp[0]
state_arrs = tmp[1:4]
if t in loss_timesteps:
p_pred = cgt.sigmoid(raw_pred)
ce = bernoulli_crossentropy(y_tbp[t] , p_pred).sum() # cross-entropy of bernoulli distribution
lossCE = lossCE + ce
loss01 = loss01 + cgt.cast(cgt.equal(y_tbp[t], round01(p_pred)),cgt.floatX).sum()
lossCE = lossCE / (len(loss_timesteps) * opt.p * opt.b) / np.log(2)
loss01 = loss01 / (len(loss_timesteps) * opt.p * opt.b)
gradloss = cgt.grad(lossCE, params)
flatgrad = flatcat(gradloss)
f_loss = cgt.function([x_tbk, y_tbp], lossCE)
f_loss_and_grad = cgt.function([x_tbk, y_tbp], [lossCE, loss01, flatgrad])
print "number of nodes in computation graph:", core.count_nodes([lossCE, loss01, flatgrad])
return f_loss, f_loss_and_grad, params
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 test_get_context():
batch_size = 32
feat_t_steps = 3
feat_num_features = 30
state_num_features = 20
num_out_classes = 28
feats = cgt.tensor3(fixed_shape=(batch_size, feat_t_steps, feat_num_features))
prev_out = cgt.matrix(fixed_shape=(batch_size, state_num_features))
sigmoided = cgt.sigmoid(prev_out)
s = nnbuilder.Seq2Seq(nn_input_btf=feats, num_out_classes=num_out_classes, feature_size=feat_num_features, decoder_size=state_num_features)
mm = cgt.infer_shape(s.features_post_mlp_btf)
assert mm == (batch_size, feat_t_steps, feat_num_features)
context_out = s.get_context(sigmoided)
out_fun = cgt.function([feats, prev_out], [context_out])
tau = np.reshape(np.random.normal(0.1, 0.2, batch_size*feat_t_steps*feat_num_features), (batch_size, feat_t_steps, feat_num_features))
tau2 = np.reshape(np.random.normal(0.1, 0.2, batch_size*state_num_features), (batch_size, state_num_features))
m = out_fun(tau, tau2)[0]
assert m.shape == (batch_size, feat_num_features)
assert np.mean(m) < 1
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
from gru import GRUCell
import time
from cgt.utils import Message
import numpy as np
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--horizon", type=int)
args = parser.parse_args()
horizon = args.horizon
assert horizon is not None
size = 128
batchsize = 64
cell = GRUCell([size], size)
X = cgt.tensor3()
init = cgt.matrix()
prev_h = init
for i in xrange(horizon):
prev_h = cell(X[i], prev_h)
loss = prev_h.sum()
with Message("compiling"):
f = cgt.function([X, init], cgt.grad(loss, cell.params()))
with Message("running"):
xval = np.zeros((horizon, batchsize, size), cgt.floatX)
initval = np.zeros((batchsize, size), cgt.floatX)
for i in xrange(100):
f(xval, initval)
import gru, cgt, numpy as np
import sys
from time import time
elapsed = []
horizons = 2**np.arange(2, 10)
for horizon in horizons:
print "HORIZON", horizon
tstart = time()
batch_size = 6
dim_x = 16
mem_size = 10
X_tnk = cgt.tensor3("X")
cell = gru.GRUCell([dim_x], mem_size)
Minit_nk = cgt.zeros((X_tnk.shape[0], X_tnk.shape[1]), cgt.floatX)
M = Minit_nk
for t in xrange(horizon):
M = cell(M, X_tnk[t])
# cgt.print_tree(M)
print "simplifying..."
M_simp = cgt.simplify([M])
print "done"
# cgt.print_tree(M_simp)
print "fn before:", cgt.count_nodes(M)
import gru,cgt, numpy as np
import sys
from time import time
elapsed = []
horizons = 2**np.arange(2, 10)
for horizon in horizons:
print "HORIZON",horizon
tstart = time()
batch_size = 6
dim_x = 16
mem_size = 10
X_tnk = cgt.tensor3("X")
cell = gru.GRUCell([dim_x], mem_size)
Minit_nk = cgt.zeros((X_tnk.shape[0], X_tnk.shape[1]),cgt.floatX)
M = Minit_nk
for t in xrange(horizon):
M = cell(M, X_tnk[t])
# cgt.print_tree(M)
print "simplifying..."
M_simp = cgt.simplify([M])
print "done"
# cgt.print_tree(M_simp)
print "fn before:",cgt.count_nodes(M)
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
from gru import GRUCell
import time
from cgt.utils import Message
import numpy as np
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--horizon",type=int)
args = parser.parse_args()
horizon = args.horizon
assert horizon is not None
size=128
batchsize=64
cell = GRUCell([size],size)
X = cgt.tensor3()
init = cgt.matrix()
prev_h = init
for i in xrange(horizon):
prev_h = cell(X[i], prev_h)
loss = prev_h.sum()
with Message("compiling"):
f = cgt.function([X, init],cgt.grad(loss, cell.params()))
with Message("running"):
xval = np.zeros((horizon,batchsize,size),cgt.floatX)
initval = np.zeros((batchsize, size), cgt.floatX)
for i in xrange(100):
f(xval, initval)
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
