def gradient_calc(seq_len, input_size, hidden_size, batch_size,
epsilon=None, rand_scale=None, inp_bl=None):
NervanaObject.be.bsz = NervanaObject.be.batch_size = batch_size
input_shape = (input_size, seq_len * batch_size)
# generate input if one is not given
if inp_bl is None:
inp_bl = np.random.randn(*input_shape)
# neon lstm instance
lstm = LSTM(hidden_size, Gaussian(), Tanh(), Logistic())
inpa = lstm.be.array(np.copy(inp_bl))
# run fprop on the baseline input
out_bl = lstm.fprop(inpa).get()
# random scaling/hash to generate fake loss
if rand_scale is None:
rand_scale = np.random.random(out_bl.shape) * 2.0 - 1.0
# loss function would be:
# loss_bl = np.sum(rand_scale * out_bl)
# run back prop with rand_scale as the errors
# use copy to avoid any interactions
deltas_neon = lstm.bprop(lstm.be.array(np.copy(rand_scale))).get()
# add a perturbation to each input element
grads_est = np.zeros(inpa.shape)
inp_pert = inp_bl.copy()
for pert_ind in range(inpa.size):
save_val = inp_pert.flat[pert_ind]
inp_pert.flat[pert_ind] = save_val + epsilon
reset_lstm(lstm)
out_pos = lstm.fprop(lstm.be.array(inp_pert)).get()
inp_pert.flat[pert_ind] = save_val - epsilon
reset_lstm(lstm)
out_neg = lstm.fprop(lstm.be.array(inp_pert)).get()
# calculate the loss with perturbations
loss_pos = np.sum(rand_scale*out_pos)
loss_neg = np.sum(rand_scale*out_neg)
# compute the gradient estimate
grad = 0.5*(loss_pos-loss_neg)/epsilon
grads_est.flat[pert_ind] = grad
# reset the perturbed input element
inp_pert.flat[pert_ind] = save_val
del lstm
return (grads_est, deltas_neon)
def gradient_calc(seq_len,
input_size,
hidden_size,
batch_size,
epsilon=None,
rand_scale=None,
inp_bl=None):
NervanaObject.be.bsz = NervanaObject.be.batch_size = batch_size
input_shape = (input_size, seq_len * batch_size)
# generate input if one is not given
if inp_bl is None:
inp_bl = np.random.randn(*input_shape)
# neon lstm instance
lstm = LSTM(hidden_size, Gaussian(), Tanh(), Logistic())
inpa = lstm.be.array(np.copy(inp_bl))
# run fprop on the baseline input
out_bl = lstm.fprop(inpa).get()
# random scaling/hash to generate fake loss
if rand_scale is None:
rand_scale = np.random.random(out_bl.shape) * 2.0 - 1.0
# loss function would be:
# loss_bl = np.sum(rand_scale * out_bl)
# run back prop with rand_scale as the errors
# use copy to avoid any interactions
deltas_neon = lstm.bprop(lstm.be.array(np.copy(rand_scale))).get()
# add a perturbation to each input element
grads_est = np.zeros(inpa.shape)
inp_pert = inp_bl.copy()
for pert_ind in range(inpa.size):
save_val = inp_pert.flat[pert_ind]
inp_pert.flat[pert_ind] = save_val + epsilon
reset_lstm(lstm)
out_pos = lstm.fprop(lstm.be.array(inp_pert)).get()
inp_pert.flat[pert_ind] = save_val - epsilon
reset_lstm(lstm)
out_neg = lstm.fprop(lstm.be.array(inp_pert)).get()
# calculate the loss with perturbations
loss_pos = np.sum(rand_scale * out_pos)
loss_neg = np.sum(rand_scale * out_neg)
# compute the gradient estimate
grad = 0.5 * (loss_pos - loss_neg) / epsilon
grads_est.flat[pert_ind] = grad
# reset the perturbed input element
inp_pert.flat[pert_ind] = save_val
del lstm
return (grads_est, deltas_neon)
def test_biLSTM_fprop_rnn(backend_default, fargs):
# basic sanity check with 0 weights random inputs
seq_len, input_size, hidden_size, batch_size = fargs
in_shape = (input_size, seq_len)
out_shape = (hidden_size, seq_len)
NervanaObject.be.bsz = batch_size
# setup the bi-directional rnn
init_glorot = GlorotUniform()
bilstm = BiLSTM(hidden_size, gate_activation=Logistic(),
activation=Tanh(), init=init_glorot, reset_cells=True)
bilstm.configure(in_shape)
bilstm.prev_layer = True
bilstm.allocate()
# setup the bi-directional rnn
init_glorot = GlorotUniform()
rnn = LSTM(hidden_size, gate_activation=Logistic(),
activation=Tanh(), init=init_glorot, reset_cells=True)
rnn.configure(in_shape)
rnn.prev_layer = True
rnn.allocate()
# same weight for bi-rnn backward and rnn weights
nout = hidden_size
bilstm.W_input_b[:] = bilstm.W_input_f
bilstm.W_recur_b[:] = bilstm.W_recur_f
bilstm.b_b[:] = bilstm.b_f
bilstm.dW[:] = 0
rnn.W_input[:] = bilstm.W_input_f
rnn.W_recur[:] = bilstm.W_recur_f
rnn.b[:] = bilstm.b_f
rnn.dW[:] = 0
# inputs - random and flipped left-to-right inputs
lr = np.random.random((input_size, seq_len * batch_size))
lr_rev = list(reversed(get_steps(lr.copy(), in_shape)))
rl = con(lr_rev, axis=1)
inp_lr = bilstm.be.array(lr)
inp_rl = bilstm.be.array(rl)
inp_rnn = rnn.be.array(lr)
# outputs
out_lr = bilstm.fprop(inp_lr).get().copy()
bilstm.h_buffer[:] = 0
out_rl = bilstm.fprop(inp_rl).get()
out_rnn = rnn.fprop(inp_rnn).get().copy()
# views
out_lr_f_s = get_steps(out_lr[:nout], out_shape)
out_lr_b_s = get_steps(out_lr[nout:], out_shape)
out_rl_f_s = get_steps(out_rl[:nout], out_shape)
out_rl_b_s = get_steps(out_rl[nout:], out_shape)
out_rnn_s = get_steps(out_rnn, out_shape)
# asserts for fprop
for x_rnn, x_f, x_b, y_f, y_b in zip(out_rnn_s, out_lr_f_s, out_lr_b_s,
reversed(out_rl_f_s), reversed(out_rl_b_s)):
assert allclose_with_out(x_f, y_b, rtol=0.0, atol=1.0e-5)
assert allclose_with_out(x_b, y_f, rtol=0.0, atol=1.0e-5)
assert allclose_with_out(x_rnn, x_f, rtol=0.0, atol=1.0e-5)
assert allclose_with_out(x_rnn, y_b, rtol=0.0, atol=1.0e-5)
def test_biLSTM_fprop_rnn(backend_default, fargs):
# basic sanity check with 0 weights random inputs
seq_len, input_size, hidden_size, batch_size = fargs
in_shape = (input_size, seq_len)
out_shape = (hidden_size, seq_len)
NervanaObject.be.bsz = batch_size
# setup the bi-directional rnn
init_glorot = GlorotUniform()
bilstm = BiLSTM(hidden_size,
gate_activation=Logistic(),
activation=Tanh(),
init=init_glorot,
reset_cells=True)
bilstm.configure(in_shape)
bilstm.prev_layer = True
bilstm.allocate()
# setup the bi-directional rnn
init_glorot = GlorotUniform()
rnn = LSTM(hidden_size,
gate_activation=Logistic(),
activation=Tanh(),
init=init_glorot,
reset_cells=True)
rnn.configure(in_shape)
rnn.prev_layer = True
rnn.allocate()
# same weight for bi-rnn backward and rnn weights
nout = hidden_size
bilstm.W_input_b[:] = bilstm.W_input_f
bilstm.W_recur_b[:] = bilstm.W_recur_f
bilstm.b_b[:] = bilstm.b_f
bilstm.dW[:] = 0
rnn.W_input[:] = bilstm.W_input_f
rnn.W_recur[:] = bilstm.W_recur_f
rnn.b[:] = bilstm.b_f
rnn.dW[:] = 0
# inputs - random and flipped left-to-right inputs
lr = np.random.random((input_size, seq_len * batch_size))
lr_rev = list(reversed(get_steps(lr.copy(), in_shape)))
rl = con(lr_rev, axis=1)
inp_lr = bilstm.be.array(lr)
inp_rl = bilstm.be.array(rl)
inp_rnn = rnn.be.array(lr)
# outputs
out_lr = bilstm.fprop(inp_lr).get().copy()
bilstm.h_buffer[:] = 0
out_rl = bilstm.fprop(inp_rl).get()
out_rnn = rnn.fprop(inp_rnn).get().copy()
# views
out_lr_f_s = get_steps(out_lr[:nout], out_shape)
out_lr_b_s = get_steps(out_lr[nout:], out_shape)
out_rl_f_s = get_steps(out_rl[:nout], out_shape)
out_rl_b_s = get_steps(out_rl[nout:], out_shape)
out_rnn_s = get_steps(out_rnn, out_shape)
# asserts for fprop
for x_rnn, x_f, x_b, y_f, y_b in zip(out_rnn_s, out_lr_f_s, out_lr_b_s,
reversed(out_rl_f_s),
reversed(out_rl_b_s)):
assert allclose_with_out(x_f, y_b, rtol=0.0, atol=1.0e-5)
assert allclose_with_out(x_b, y_f, rtol=0.0, atol=1.0e-5)
assert allclose_with_out(x_rnn, x_f, rtol=0.0, atol=1.0e-5)
assert allclose_with_out(x_rnn, y_b, rtol=0.0, atol=1.0e-5)
def test_beamsearch(backend_default):
"""
Simlulated beam search on a minibatch of 2, for 4 time steps. The
LSTM states are real but the "softmax outputs" z are hardcoded and
not taken from the network.
There are 6 tokens the network outputs, and they have probabilities
like exp(1), exp(5), exp(7)
The test asserts that the score_lists assigned by _beamsearch_step(z_list)
are equal to the probabilities computed manually adding probabilities
to z_list.
"""
be = backend_default
batch_size = 2
be.bsz = batch_size
time_steps = 4
nout = 6
num_beams = 3
# create unused layers
activation = Tanh()
gate_activation = Logistic()
init_ary = np.eye(nout)
init = Array(init_ary)
encoder = LSTM(nout, init,
activation=activation, gate_activation=gate_activation,
name="Enc")
decoder = LSTM(nout, init,
activation=activation, gate_activation=gate_activation,
name="Dec")
class DummyFProp():
"""
Constructs an artificial beam search example with known correct outputs.
This is called inside a nested loop over steps, num_life. In the first
time step there is one life beam, after that, 3 life beams per step.
There are 4 time steps total. Each beamsearch_step builds one list over
num_life beams.
At t=0, the winners for ex0 are 1, 4, 5 (indexed by their position) and
winners for ex1 are 2,4,5. From there we continue the beam for ex0:
12, 13, 14 6+2=8 6+3=9 6+2=8
40, 43, 45 with scores 5+4=9 5+3=8 5+7=12 three new winners 45, 52, 55
50, 52, 55 5+4=9 5+6=11 5+5=10
for ex2
1 4 5 with scores 5 4 7
we get the three winners 1, 4, 5 and continue (just taking the
3 in order, no sorting)
10 12 13 14 (not unique!) 5+2=7 5+2=7 5+3=8
41 42 43 with scores 4+6=10 4+5=9 4+7=11 winners 43 51 52
51 52 53 7+4=11 7+6=13 7+3=10 scores 11 11 13
continue from the three winners 43 51 52
431 433 434 11+10=21 11+3=14 11+9=20
511 512 513 with scores 11+6=17 11+5=16 11+7=18 winners 431 434 520
520 521 522 13+8=21 13+4=17 13+6=19 scores 21 20 21
continue from three winners 431 511 513 (going along beams, the matches
in a beam)
4310 4312 4313 4314 21+2=23 21+2=23 21+3=24 21+10=31 (not unique!)
4341 4342 4343 with scores 20+10=30 20+5=25 20+7=27 winners 4314 4341 5204
5200 5202 5204 21+8=29 21+6=27 21+10=31 scores 31 30 31
overall winners are 4314 4341 5204
"""
def __init__(self):
self.i = -1
# t=0
# X x x <-- winners: 1, 4, 5 (for example 0)
z = be.array(np.exp(np.array([[1, 6, 2, 1, 5, 5],
[1, 5, 2, 2, 4, 7]]))).T
# t=1
# x x x <-- give we picked 4: new winners 2,3,4
z1 = be.array(np.exp(np.array([[1, 1, 2, 3, 2, 1],
[2, 1, 2, 3, 2, 1]]))).T
# x x x <-- give we picked 5:
# new winners 0,3,[5]
# score 12
z2 = be.array(np.exp(np.array([[4, 1, 2, 3, 1, 7],
[2, 6, 5, 7, 2, 4]]))).T
# x X X <-- give we picked 1:
# new winners 0,[2],[5]
# scores 12, 11
z3 = be.array(np.exp(np.array([[4, 1, 6, 3, 1, 5],
[1, 4, 6, 3, 2, 1]]))).T
# t=2
# example 0: given constructed (1, 5), score 11: 3, 4; scores 21, 20
z4 = be.array(np.exp(np.array([[1, 1, 2, 10, 9, 1],
[2, 10, 2, 3, 9, 1]]))).T
# example 0: given constructed (5, 5), score 12: none selected from this beam
z5 = be.array(np.exp(np.array([[4, 1, 2, 3, 1, 7],
[2, 6, 5, 7, 2, 4]]))).T
# example 0: given constructed (1, 2), score 12: 1; score 20
z6 = be.array(np.exp(np.array([[4, 8, 6, 3, 1, 5],
[8, 4, 6, 3, 1, 1]]))).T
# t=3
# example 0: given constructed (1, 5, 4), score 20: 1, score 30
z7 = be.array(np.exp(np.array([[1, 10, 2, 1, 1, 1],
[2, 1, 2, 3, 10, 1]]))).T
# example 0: given constructed (1, 2, 1), score 20: 5, score 30
z8 = be.array(np.exp(np.array([[4, 1, 2, 3, 1, 10],
[2, 10, 5, 7, 2, 4]]))).T
# example 0: given constructed (1, 5, 3), score 21: 4, score 31
z9 = be.array(np.exp(np.array([[4, 8, 6, 3, 10, 5],
[8, 4, 6, 3, 10, 1]]))).T
self.z_list = [z, z1, z2, z3, z4, z5, z6, z7, z8, z9]
def fprop(self, z, inference=True, init_state=None):
self.i += 1
return self.z_list[self.i]
def final_state():
return be.zeros_like(decoder.h[-1])
class InObj(NervanaObject):
def __init__(self):
self.shape = (nout, time_steps)
self.decoder_shape = (nout, time_steps)
decoder.fprop = DummyFProp().fprop
layers = Seq2Seq([encoder, decoder], decoder_connections=[0])
layers.decoder._recurrent[0].final_state = final_state
in_obj = InObj()
layers.configure(in_obj) # made zeros because zeros have shape
layers.allocate()
layers.allocate_deltas(None)
beamsearch = BeamSearch(layers)
inputs = be.iobuf(in_obj.shape)
beamsearch.beamsearch(inputs, num_beams=num_beams)
ex0 = np.array([[1, 5, 4, 1],
[1, 2, 1, 5],
[1, 5, 3, 4]])
ex1 = np.array([[5, 1, 4, 4],
[5, 1, 1, 1],
[5, 2, 0, 4]])
# extract all candidates
examples = reformat_samples(beamsearch, num_beams, batch_size)
assert allclose_with_out(examples[0], ex0)
assert allclose_with_out(examples[1], ex1)
def check_lstm(seq_len,
input_size,
hidden_size,
batch_size,
init_func,
inp_moms=[0.0, 1.0]):
# init_func is the initializer for the model params
# inp_moms is the [ mean, std dev] of the random input
input_shape = (input_size, seq_len * batch_size)
hidden_shape = (hidden_size, seq_len * batch_size)
NervanaObject.be.bsz = NervanaObject.be.batch_size = batch_size
# neon LSTM
lstm = LSTM(hidden_size,
init_func,
activation=Tanh(),
gate_activation=Logistic())
inp = np.random.rand(*input_shape) * inp_moms[1] + inp_moms[0]
inpa = lstm.be.array(inp)
# import pdb; pdb.set_trace()
# run neon fprop
lstm.fprop(inpa)
# reference numpy LSTM
lstm_ref = RefLSTM()
WLSTM = lstm_ref.init(input_size, hidden_size)
# make ref weights and biases with neon model
WLSTM[0, :] = lstm.b.get().T
WLSTM[1:input_size + 1, :] = lstm.W_input.get().T
WLSTM[input_size + 1:] = lstm.W_recur.get().T
# transpose input X and do fprop
inp_ref = inp.copy().T.reshape(seq_len, batch_size, input_size)
(Hout_ref, cprev, hprev, batch_cache) = lstm_ref.forward(inp_ref, WLSTM)
# the output needs transpose as well
Hout_ref = Hout_ref.reshape(seq_len * batch_size, hidden_size).T
IFOGf_ref = batch_cache['IFOGf'].reshape(seq_len * batch_size,
hidden_size * 4).T
Ct_ref = batch_cache['Ct'].reshape(seq_len * batch_size, hidden_size).T
# compare results
print '====Verifying IFOG===='
allclose_with_out(lstm.ifog_buffer.get(), IFOGf_ref, rtol=0.0, atol=1.0e-5)
print '====Verifying cell states===='
allclose_with_out(lstm.c_act_buffer.get(), Ct_ref, rtol=0.0, atol=1.0e-5)
print '====Verifying hidden states===='
allclose_with_out(lstm.h_buffer.get(), Hout_ref, rtol=0.0, atol=1.0e-5)
print 'fprop is verified'
# now test the bprop
# generate random deltas tensor
deltas = np.random.randn(*hidden_shape)
lstm.bprop(lstm.be.array(deltas))
# grab the delta W from gradient buffer
dWinput_neon = lstm.dW_input.get()
dWrecur_neon = lstm.dW_recur.get()
db_neon = lstm.db.get()
# import pdb; pdb.set_trace()
deltas_ref = deltas.copy().T.reshape(seq_len, batch_size, hidden_size)
(dX_ref, dWLSTM_ref, dc0_ref,
dh0_ref) = lstm_ref.backward(deltas_ref, batch_cache)
dWrecur_ref = dWLSTM_ref[-hidden_size:, :]
dWinput_ref = dWLSTM_ref[1:input_size + 1, :]
db_ref = dWLSTM_ref[0, :]
dX_ref = dX_ref.reshape(seq_len * batch_size, input_size).T
# compare results
print 'Making sure neon LSTM match numpy LSTM in bprop'
print '====Verifying update on W_recur===='
assert allclose_with_out(dWrecur_neon,
dWrecur_ref.T,
rtol=0.0,
atol=1.0e-5)
print '====Verifying update on W_input===='
assert allclose_with_out(dWinput_neon,
dWinput_ref.T,
rtol=0.0,
atol=1.0e-5)
print '====Verifying update on bias===='
assert allclose_with_out(db_neon.flatten(), db_ref, rtol=0.0, atol=1.0e-5)
print '====Verifying output delta===='
assert allclose_with_out(lstm.out_deltas_buffer.get(),
dX_ref,
rtol=0.0,
atol=1.0e-5)
print 'bprop is verified'
return
def check_lstm(seq_len, input_size, hidden_size,
batch_size, init_func, inp_moms=[0.0, 1.0]):
# init_func is the initializer for the model params
# inp_moms is the [ mean, std dev] of the random input
input_shape = (input_size, seq_len * batch_size)
hidden_shape = (hidden_size, seq_len * batch_size)
NervanaObject.be.bsz = NervanaObject.be.batch_size = batch_size
# neon LSTM
lstm = LSTM(hidden_size,
init_func,
activation=Tanh(),
gate_activation=Logistic())
inp = np.random.rand(*input_shape)*inp_moms[1] + inp_moms[0]
inpa = lstm.be.array(inp)
# run neon fprop
lstm.configure((input_size, seq_len))
lstm.prev_layer = True # Hack to force allocating a delta buffer
lstm.allocate()
lstm.set_deltas([lstm.be.iobuf(lstm.in_shape)])
lstm.fprop(inpa)
# reference numpy LSTM
lstm_ref = RefLSTM()
WLSTM = lstm_ref.init(input_size, hidden_size)
# make ref weights and biases with neon model
WLSTM[0, :] = lstm.b.get().T
WLSTM[1:input_size+1, :] = lstm.W_input.get().T
WLSTM[input_size+1:] = lstm.W_recur.get().T
# transpose input X and do fprop
inp_ref = inp.copy().T.reshape(seq_len, batch_size, input_size)
(Hout_ref, cprev, hprev, batch_cache) = lstm_ref.forward(inp_ref,
WLSTM)
# the output needs transpose as well
Hout_ref = Hout_ref.reshape(seq_len * batch_size, hidden_size).T
IFOGf_ref = batch_cache['IFOGf'].reshape(seq_len * batch_size, hidden_size * 4).T
Ct_ref = batch_cache['Ct'].reshape(seq_len * batch_size, hidden_size).T
# compare results
print '====Verifying IFOG===='
allclose_with_out(lstm.ifog_buffer.get(),
IFOGf_ref,
rtol=0.0,
atol=1.0e-5)
print '====Verifying cell states===='
allclose_with_out(lstm.c_act_buffer.get(),
Ct_ref,
rtol=0.0,
atol=1.0e-5)
print '====Verifying hidden states===='
allclose_with_out(lstm.outputs.get(),
Hout_ref,
rtol=0.0,
atol=1.0e-5)
print 'fprop is verified'
# now test the bprop
# generate random deltas tensor
deltas = np.random.randn(*hidden_shape)
lstm.bprop(lstm.be.array(deltas))
# grab the delta W from gradient buffer
dWinput_neon = lstm.dW_input.get()
dWrecur_neon = lstm.dW_recur.get()
db_neon = lstm.db.get()
deltas_ref = deltas.copy().T.reshape(seq_len, batch_size, hidden_size)
(dX_ref, dWLSTM_ref, dc0_ref, dh0_ref) = lstm_ref.backward(deltas_ref,
batch_cache)
dWrecur_ref = dWLSTM_ref[-hidden_size:, :]
dWinput_ref = dWLSTM_ref[1:input_size+1, :]
db_ref = dWLSTM_ref[0, :]
dX_ref = dX_ref.reshape(seq_len * batch_size, input_size).T
# compare results
print 'Making sure neon LSTM match numpy LSTM in bprop'
print '====Verifying update on W_recur===='
assert allclose_with_out(dWrecur_neon,
dWrecur_ref.T,
rtol=0.0,
atol=1.0e-5)
print '====Verifying update on W_input===='
assert allclose_with_out(dWinput_neon,
dWinput_ref.T,
rtol=0.0,
atol=1.0e-5)
print '====Verifying update on bias===='
assert allclose_with_out(db_neon.flatten(),
db_ref,
rtol=0.0,
atol=1.0e-5)
print '====Verifying output delta===='
assert allclose_with_out(lstm.out_deltas_buffer.get(),
dX_ref,
rtol=0.0,
atol=1.0e-5)
print 'bprop is verified'
return
def test_beamsearch(backend_default):
"""
Simlulated beam search on a minibatch of 2, for 4 time steps. The
LSTM states are real but the "softmax outputs" z are hardcoded and
not taken from the network.
There are 6 tokens the network outputs, and they have probabilities
like exp(1), exp(5), exp(7)
The test asserts that the score_lists assigned by _beamsearch_step(z_list)
are equal to the probabilities computed manually adding probabilities
to z_list.
"""
be = backend_default
batch_size = 2
be.bsz = batch_size
time_steps = 4
nout = 6
num_beams = 3
# create unused layers
activation = Tanh()
gate_activation = Logistic()
init_ary = np.eye(nout)
init = Array(init_ary)
encoder = LSTM(nout,
init,
activation=activation,
gate_activation=gate_activation,
name="Enc")
decoder = LSTM(nout,
init,
activation=activation,
gate_activation=gate_activation,
name="Dec")
class DummyFProp():
"""
Constructs an artificial beam search example with known correct outputs.
This is called inside a nested loop over steps, num_life. In the first
time step there is one life beam, after that, 3 life beams per step.
There are 4 time steps total. Each beamsearch_step builds one list over
num_life beams.
At t=0, the winners for ex0 are 1, 4, 5 (indexed by their position) and
winners for ex1 are 2,4,5. From there we continue the beam for ex0:
12, 13, 14 6+2=8 6+3=9 6+2=8
40, 43, 45 with scores 5+4=9 5+3=8 5+7=12 three new winners 45, 52, 55
50, 52, 55 5+4=9 5+6=11 5+5=10
for ex2
1 4 5 with scores 5 4 7
we get the three winners 1, 4, 5 and continue (just taking the
3 in order, no sorting)
10 12 13 14 (not unique!) 5+2=7 5+2=7 5+3=8
41 42 43 with scores 4+6=10 4+5=9 4+7=11 winners 43 51 52
51 52 53 7+4=11 7+6=13 7+3=10 scores 11 11 13
continue from the three winners 43 51 52
431 433 434 11+10=21 11+3=14 11+9=20
511 512 513 with scores 11+6=17 11+5=16 11+7=18 winners 431 434 520
520 521 522 13+8=21 13+4=17 13+6=19 scores 21 20 21
continue from three winners 431 511 513 (going along beams, the matches
in a beam)
4310 4312 4313 4314 21+2=23 21+2=23 21+3=24 21+10=31 (not unique!)
4341 4342 4343 with scores 20+10=30 20+5=25 20+7=27 winners 4314 4341 5204
5200 5202 5204 21+8=29 21+6=27 21+10=31 scores 31 30 31
overall winners are 4314 4341 5204
"""
def __init__(self):
self.i = -1
# t=0
# X x x <-- winners: 1, 4, 5 (for example 0)
z = be.array(
np.exp(np.array([[1, 6, 2, 1, 5, 5], [1, 5, 2, 2, 4, 7]]))).T
# t=1
# x x x <-- give we picked 4: new winners 2,3,4
z1 = be.array(
np.exp(np.array([[1, 1, 2, 3, 2, 1], [2, 1, 2, 3, 2, 1]]))).T
# x x x <-- give we picked 5:
# new winners 0,3,[5]
# score 12
z2 = be.array(
np.exp(np.array([[4, 1, 2, 3, 1, 7], [2, 6, 5, 7, 2, 4]]))).T
# x X X <-- give we picked 1:
# new winners 0,[2],[5]
# scores 12, 11
z3 = be.array(
np.exp(np.array([[4, 1, 6, 3, 1, 5], [1, 4, 6, 3, 2, 1]]))).T
# t=2
# example 0: given constructed (1, 5), score 11: 3, 4; scores 21, 20
z4 = be.array(
np.exp(np.array([[1, 1, 2, 10, 9, 1], [2, 10, 2, 3, 9, 1]]))).T
# example 0: given constructed (5, 5), score 12: none selected from this beam
z5 = be.array(
np.exp(np.array([[4, 1, 2, 3, 1, 7], [2, 6, 5, 7, 2, 4]]))).T
# example 0: given constructed (1, 2), score 12: 1; score 20
z6 = be.array(
np.exp(np.array([[4, 8, 6, 3, 1, 5], [8, 4, 6, 3, 1, 1]]))).T
# t=3
# example 0: given constructed (1, 5, 4), score 20: 1, score 30
z7 = be.array(
np.exp(np.array([[1, 10, 2, 1, 1, 1], [2, 1, 2, 3, 10, 1]]))).T
# example 0: given constructed (1, 2, 1), score 20: 5, score 30
z8 = be.array(
np.exp(np.array([[4, 1, 2, 3, 1, 10], [2, 10, 5, 7, 2, 4]]))).T
# example 0: given constructed (1, 5, 3), score 21: 4, score 31
z9 = be.array(
np.exp(np.array([[4, 8, 6, 3, 10, 5], [8, 4, 6, 3, 10, 1]]))).T
self.z_list = [z, z1, z2, z3, z4, z5, z6, z7, z8, z9]
def fprop(self, z, inference=True, init_state=None):
self.i += 1
return self.z_list[self.i]
def final_state():
return be.zeros_like(decoder.h[-1])
class InObj(NervanaObject):
def __init__(self):
self.shape = (nout, time_steps)
self.decoder_shape = (nout, time_steps)
decoder.fprop = DummyFProp().fprop
layers = Seq2Seq([encoder, decoder], decoder_connections=[0])
layers.decoder._recurrent[0].final_state = final_state
in_obj = InObj()
layers.configure(in_obj) # made zeros because zeros have shape
layers.allocate()
layers.allocate_deltas(None)
beamsearch = BeamSearch(layers)
inputs = be.iobuf(in_obj.shape)
beamsearch.beamsearch(inputs, num_beams=num_beams)
ex0 = np.array([[1, 5, 4, 1], [1, 2, 1, 5], [1, 5, 3, 4]])
ex1 = np.array([[5, 1, 4, 4], [5, 1, 1, 1], [5, 2, 0, 4]])
# extract all candidates
examples = reformat_samples(beamsearch, num_beams, batch_size)
assert allclose_with_out(examples[0], ex0)
assert allclose_with_out(examples[1], ex1)
