def test_rnn_deriv_ref(sequence_length, input_size, hidden_size, batch_size,
return_sequence, weight_initializer, bias_initializer,
transformer_factory):
assert batch_size == 1, "the recurrent reference implementation only support batch size 1"
assert return_sequence is True, "the reference rnn only supports sequences for deriv"
# Get input placeholder and numpy array
input_placeholder, input_value = make_placeholder(input_size, sequence_length, batch_size)
# Construct network weights and initial state, if desired
W_in, W_rec, b, init_state, init_state_value = make_weights(input_placeholder, hidden_size,
weight_initializer,
bias_initializer)
# Compute reference numpy RNN
rnn_ref = RefRecurrent(input_size, hidden_size, return_sequence=return_sequence)
rnn_ref.set_weights(W_in, W_rec, b.reshape(rnn_ref.bh.shape))
# Prepare deltas for gradient check
output_shape = (hidden_size, sequence_length, batch_size)
# generate random deltas tensor
deltas = np.random.randn(*output_shape)
# the reference code expects these shapes:
# input_shape: (seq_len, input_size, batch_size)
# output_shape: (seq_len, hidden_size, batch_size)
dW_in, dW_rec, db = rnn_ref.lossFun(input_value.transpose([1, 0, 2]),
deltas.copy().transpose([1, 0, 2]),
init_states=init_state_value)[:3]
# Generate ngraph RNN
rnn_ng = Recurrent(hidden_size, init=W_in, init_inner=W_rec, activation=Tanh(),
reset_cells=True, return_sequence=return_sequence)
# fprop ngraph RNN
out_ng = rnn_ng.train_outputs(input_placeholder)
deltas_constant = ng.constant(deltas, axes=out_ng.axes)
params = [(rnn_ng.W_input, W_in),
(rnn_ng.W_recur, W_rec),
(rnn_ng.b, b)]
with ExecutorFactory() as ex:
# Create derivative computations and execute
param_updates = list()
for px, _ in params:
update = ng.deriv(out_ng, px, error=deltas_constant)
param_updates.append(ex.executor(update, input_placeholder))
for update_fun, ref_val in zip(param_updates, [dW_in, dW_rec, db]):
ng.testing.assert_allclose(update_fun(input_value),
ref_val.squeeze(),
rtol=bprop_rtol, atol=bprop_atol)
def test_rnn_fprop(sequence_length, input_size, hidden_size, batch_size,
return_sequence, weight_initializer, bias_initializer,
init_state, extra_axes, backward, transformer_factory):
assert batch_size == 1, "the recurrent reference implementation only support batch size 1"
# Get input placeholder and numpy array
input_placeholder, input_value = make_placeholder(input_size, sequence_length, batch_size,
extra_axes=extra_axes)
# Construct network weights and initial state, if desired
W_in, W_rec, b, init_state, init_state_value = make_weights(input_placeholder, hidden_size,
weight_initializer,
bias_initializer,
init_state)
# Compute reference numpy RNN
rnn_ref = RefRecurrent(input_size, hidden_size, return_sequence=return_sequence)
rnn_ref.set_weights(W_in.reshape(rnn_ref.Wxh.shape), W_rec, b.reshape(rnn_ref.bh.shape))
# Compute reference numpy RNN
input_shape = (input_size, sequence_length, batch_size)
h_ref_list = rnn_ref.fprop_only(input_value.reshape(input_shape).transpose([1, 0, 2]),
init_states=init_state_value, backward=backward)
# Generate ngraph RNN
rnn_ng = Recurrent(hidden_size, init=W_in, init_inner=W_rec, activation=Tanh(),
reset_cells=True, return_sequence=return_sequence,
backward=backward)
# fprop ngraph RNN
out_ng = rnn_ng.train_outputs(input_placeholder, init_state=init_state)
with ExecutorFactory() as ex:
# Create computation and execute
if init_state is not None:
fprop_neon_fun = ex.executor(out_ng, input_placeholder, init_state)
fprop_neon = fprop_neon_fun(input_value, init_state_value)
else:
fprop_neon_fun = ex.executor(out_ng, input_placeholder)
fprop_neon = fprop_neon_fun(input_value)
# Compare output with reference implementation
if return_sequence is True:
fprop_neon = fprop_neon[:, :, 0]
ng.testing.assert_allclose(fprop_neon, h_ref_list, rtol=fprop_rtol, atol=fprop_atol)
def test_rnn_deriv_numerical(sequence_length, input_size, hidden_size, batch_size,
return_sequence, weight_initializer, bias_initializer,
backward, transformer_factory):
# Get input placeholder and numpy array
input_placeholder, input_value = make_placeholder(input_size, sequence_length, batch_size)
# Construct network weights and initial state, if desired
W_in, W_rec, b, init_state, init_state_value = make_weights(input_placeholder, hidden_size,
weight_initializer,
bias_initializer)
# Generate ngraph RNN
rnn_ng = Recurrent(hidden_size, init=W_in, init_inner=W_rec, activation=Tanh(),
reset_cells=True, return_sequence=return_sequence,
backward=backward)
# fprop ngraph RNN
out_ng = rnn_ng.train_outputs(input_placeholder)
params = [(rnn_ng.W_input, W_in),
(rnn_ng.W_recur, W_rec),
(rnn_ng.b, b)]
with ExecutorFactory() as ex:
# Create derivative computations and execute
param_updates = list()
for px, _ in params:
update = (ex.derivative(out_ng, px, input_placeholder),
ex.numeric_derivative(out_ng, px, delta, input_placeholder))
param_updates.append(update)
for (deriv_s, deriv_n), (_, val) in zip(param_updates, params):
ng.testing.assert_allclose(deriv_s(val, input_value),
deriv_n(val, input_value),
rtol=num_rtol, atol=num_atol)
def check_rnn(seq_len,
input_size,
hidden_size,
batch_size,
init_func,
return_seq=True):
# init_func is the initializer for the model params
assert batch_size == 1, "the recurrent reference implementation only support batch size 1"
# ========== neon model ==========
Cin = ng.make_axis(input_size)
REC = ng.make_axis(seq_len, recurrent=True)
N = ng.make_axis(batch_size, batch=True)
H = ng.make_axis(hidden_size)
ax_s = ng.make_axes([H, N])
ex = ExecutorFactory()
np.random.seed(0)
rnn_ng = Recurrent(hidden_size,
init_func,
activation=Tanh(),
reset_cells=True,
return_sequence=return_seq)
inp_ng = ng.placeholder([Cin, REC, N])
init_state_ng = ng.placeholder(ax_s)
# fprop graph
out_ng = rnn_ng.train_outputs(inp_ng, init_state=init_state_ng)
out_ng.input = True
rnn_W_input = rnn_ng.W_input
rnn_W_input.input = True
rnn_W_recur = rnn_ng.W_recur
rnn_W_recur.input = True
rnn_b = rnn_ng.b
rnn_b.input = True
fprop_neon_fun = ex.executor(out_ng, inp_ng, init_state_ng)
dWrecur_s_fun = ex.derivative(out_ng, rnn_W_recur, inp_ng, rnn_W_input,
rnn_b)
dWrecur_n_fun = ex.numeric_derivative(out_ng, rnn_W_recur, delta, inp_ng,
rnn_W_input, rnn_b)
dWinput_s_fun = ex.derivative(out_ng, rnn_W_input, inp_ng, rnn_W_recur,
rnn_b)
dWinput_n_fun = ex.numeric_derivative(out_ng, rnn_W_input, delta, inp_ng,
rnn_W_recur, rnn_b)
dWb_s_fun = ex.derivative(out_ng, rnn_b, inp_ng, rnn_W_input, rnn_W_recur)
dWb_n_fun = ex.numeric_derivative(out_ng, rnn_b, delta, inp_ng,
rnn_W_input, rnn_W_recur)
# fprop on random inputs
input_value = rng.uniform(-1, 1, inp_ng.axes)
init_state_value = rng.uniform(-1, 1, init_state_ng.axes)
fprop_neon = fprop_neon_fun(input_value, init_state_value).copy()
# after the rnn graph has been executed, can get the W values. Get copies so
# shared values don't confuse derivatives
Wxh_neon = rnn_ng.W_input.value.get(None).copy()
Whh_neon = rnn_ng.W_recur.value.get(None).copy()
bh_neon = rnn_ng.b.value.get(None).copy()
# bprop derivs
dWrecur_s = dWrecur_s_fun(Whh_neon, input_value, Wxh_neon, bh_neon)
dWrecur_n = dWrecur_n_fun(Whh_neon, input_value, Wxh_neon, bh_neon)
np.testing.assert_allclose(dWrecur_s, dWrecur_n, rtol=rtol, atol=atol)
dWb_s = dWb_s_fun(bh_neon, input_value, Wxh_neon, Whh_neon)
dWb_n = dWb_n_fun(bh_neon, input_value, Wxh_neon, Whh_neon)
np.testing.assert_allclose(dWb_s, dWb_n, rtol=rtol, atol=atol)
dWinput_s = dWinput_s_fun(Wxh_neon, input_value, Whh_neon, bh_neon)
dWinput_n = dWinput_n_fun(Wxh_neon, input_value, Whh_neon, bh_neon)
np.testing.assert_allclose(dWinput_s, dWinput_n, rtol=rtol, atol=atol)
# ========= reference model ==========
output_shape = (hidden_size, seq_len * batch_size)
# generate random deltas tensor
deltas = np.random.randn(*output_shape)
# the reference code expects these shapes:
# input_shape: (seq_len, input_size, batch_size)
# output_shape: (seq_len, hidden_size, batch_size)
deltas_ref = deltas.copy().T.reshape(seq_len, batch_size,
hidden_size).swapaxes(1, 2)
inp_ref = input_value.transpose([1, 0, 2])
# reference numpy RNN
rnn_ref = RefRecurrent(input_size, hidden_size)
rnn_ref.Wxh[:] = Wxh_neon
rnn_ref.Whh[:] = Whh_neon
rnn_ref.bh[:] = bh_neon.reshape(rnn_ref.bh.shape)
(dWxh_ref, dWhh_ref, db_ref, h_ref_list, dh_ref_list,
d_out_ref) = rnn_ref.lossFun(inp_ref,
deltas_ref,
init_states=init_state_value)
# comparing outputs
if return_seq is False:
h_ref_list = h_ref_list[:, -1].reshape(-1, 1)
else:
fprop_neon = fprop_neon[:, :, 0]
np.testing.assert_allclose(fprop_neon, h_ref_list, rtol=0.0, atol=1.0e-5)
return
dec = Recurrent(hidden_size,
init,
activation=Tanh(),
reset_cells=True,
return_sequence=True)
linear = Affine(init, activation=Softmax(), bias_init=init, axes=(ax.Y))
optimizer = RMSProp(decay_rate=0.95,
learning_rate=2e-3,
epsilon=1e-6,
gradient_clip_value=gradient_clip_value)
# build network graph
one_hot_enc_out = one_hot_enc.train_outputs(inputs['inp_txt'])
one_hot_dec_out = one_hot_dec.train_outputs(inputs['prev_tgt'])
enc_out = enc.train_outputs(one_hot_enc_out)
dec_out = dec.train_outputs(one_hot_dec_out, init_state=enc_out)
output_prob = linear.train_outputs(dec_out)
loss = ng.cross_entropy_multi(output_prob,
ng.one_hot(inputs['tgt_txt'], axis=ax.Y),
usebits=True)
mean_cost = ng.mean(loss, out_axes=[])
updates = optimizer(loss)
train_outputs = dict(batch_cost=mean_cost, updates=updates)
loss_outputs = dict(cross_ent_loss=loss)
######################
# Train Loop