def test_save_cudnn_rnn(self):
np.random.seed(5218)
X = K.variable(np.random.rand(25, 12, 8))
num_layers = 2
num_gates = 'lstm'
skip_input = False
is_bidirectional = False
path = '/tmp/rnn'
weights, biases = K.init_rnn(input_dim=8, hidden_dim=18,
b_init=init_ops.random_normal_initializer(),
num_layers=num_layers, num_gates=num_gates,
skip_input=skip_input,
is_bidirectional=is_bidirectional)
rnn = N.CudnnRNN(num_units=18,
W_init=weights, b_init=biases,
rnn_mode=num_gates, num_layers=num_layers,
skip_input=skip_input, is_bidirectional=is_bidirectional,
return_states=False,
dropout=0., name="CudnnRNNTest")
y = rnn(X)
K.initialize_all_variables()
y = K.eval(y)
N.serialize(nnops=rnn, path=path, binary_output=True, override=True)
test_script = r"""
from __future__ import print_function, division, absolute_import
import os
os.environ['ODIN'] = 'gpu,float32,seed=5218'
import pickle
import numpy as np
import tensorflow as tf
from tensorflow.python.ops import init_ops
from odin.config import randint
from odin import backend as K, nnet as N
np.random.seed(5218)
X = K.variable(np.random.rand(25, 12, 8))
rnn = N.deserialize("%s", force_restore_vars=True)
y = rnn(X)
K.initialize_all_variables()
y = K.eval(y)
print(len(rnn.variables),
sum(np.sum(K.eval(i)) for i in rnn.variables
if K.role.has_roles(i, K.role.Weight)),
sum(np.sum(K.eval(i)) for i in rnn.variables
if K.role.has_roles(i, K.role.Bias)),
y.sum(),
(y**2).sum())
""" % path
outputs = run_script(test_script)[1]
num_variables, w, b, s1, s2 = outputs.split(' ')
assert int(num_variables) == len(rnn.variables)
assert np.allclose(float(w),
sum(np.sum(K.eval(i)) for i in rnn.variables
if K.role.has_roles(i, K.role.Weight)))
assert np.allclose(float(b),
sum(np.sum(K.eval(i)) for i in rnn.variables
if K.role.has_roles(i, K.role.Bias)))
assert np.allclose(float(s1), y.sum())
assert np.allclose(float(s2), (y**2).sum())
def test_cudnn_rnn_nnet(self):
if get_device() == 'cpu':
return
print()
np.random.seed(1208)
batch_size = 6
hidden_size = 4
X_linear = K.placeholder(shape=(None, 3, 8), name='X_linear')
X_skip = K.placeholder(shape=(None, 3, hidden_size), name='X_skip')
for direction_mode in ['bidirectional', 'unidirectional']:
is_bidirectional = direction_mode == 'bidirectional'
for nb_layers in [2]:
real_layers = nb_layers * 2 if is_bidirectional else nb_layers
for rnn_mode in ['gru', 'lstm', 'rnn_relu', 'rnn_tanh']:
for init_state, init_state_name in zip(
[
None, # None init
K.init.uniform, # function init
K.variable(
np.random.rand(real_layers, 1,
hidden_size)), # variable
K.variable(
np.random.rand(real_layers, batch_size,
hidden_size)), # variable
K.zeros(shape=(real_layers, 1, hidden_size)),
K.ones(shape=(real_layers, batch_size,
hidden_size))
],
[
'None', 'Function', 'Var1', 'VarB', 'Tensor1',
'TensorB'
]):
for input_mode in ['linear', 'skip']:
if input_mode == 'linear':
X = X_linear
x = np.random.rand(batch_size, 3, 8)
else:
X = X_skip
x = np.random.rand(batch_size, 3, hidden_size)
start = timeit.default_timer()
f = N.CudnnRNN(num_units=hidden_size,
rnn_mode=rnn_mode,
input_mode=input_mode,
num_layers=nb_layers,
direction_mode=direction_mode,
params_split=False,
return_states=True)
# perform function
y = f(X, h0=init_state, c0=init_state)
f = K.function(X, y)
output = f(x)
benchmark = timeit.default_timer() - start
self.assertTrue([list(i.shape)
for i in output] == [[
batch_size if j is None else j
for j in K.get_shape(i)
] for i in y])
print(
"*PASSED* [Layers]%s [Mode]%-8s [Input]%-6s [Direction]%-12s [State]%s [Benchmark]%.4f"
% (nb_layers, rnn_mode, input_mode,
direction_mode, init_state_name, benchmark))
name='input%d' % i) for i, shape in enumerate(train.shape)
]
print("Inputs:", ctext(inputs, 'cyan'))
# ====== create the network ====== #
f_encoder = N.Sequence([
N.Dimshuffle(pattern=(0, 1, 2, 'x')),
N.Conv(
num_filters=32, filter_size=(7, 7), b_init=None, activation=K.linear),
N.BatchNorm(),
N.Pool(pool_size=(3, 2), strides=2),
],
debug=True,
name='Encoder')
f_latent = N.Sequence([
N.Flatten(outdim=3),
N.CudnnRNN(
num_units=128, num_layers=1, is_bidirectional=False, rnn_mode='lstm'),
],
debug=True,
name='Latent')
f_decoder = N.Sequence([
N.Flatten(outdim=2),
N.Dense(num_units=1024, b_init=None, activation=K.linear),
N.BatchNorm(axes=0, activation=K.relu)
],
debug=True,
name='Decoder')
f_output = N.Sequence([N.Dense(len(digits), activation=K.linear)],
debug=True,
name='Output')
# ====== applying ====== #
E = f_encoder(inputs[0])
# ===========================================================================
# Build network
# ===========================================================================
ops = N.Sequence([
N.Dimshuffle((0, 1, 2, 'x')) if USE_MNIST_DATA else N.Dimshuffle(
(0, 2, 3, 1)),
N.Conv(32, filter_size=3, strides=1, pad='same', activation=K.linear),
N.BatchNorm(axes='auto', activation=K.relu),
N.Pool(pool_size=2, strides=None),
N.Dimshuffle(pattern=(0, 3, 1, 2)),
N.Flatten(outdim=3),
N.CudnnRNN(18,
initial_states=None,
rnn_mode='lstm',
num_layers=2,
input_mode='linear',
direction_mode='unidirectional',
params_split=False),
N.Flatten(outdim=2),
N.Dense(128, activation=K.relu),
N.Dense(10, activation=K.softmax)
],
debug=True)
ops = cPickle.loads(cPickle.dumps(ops)) # test if the ops is pickle-able
K.set_training(True)
y_pred_train = ops(X)
K.set_training(False)
y_pred_score = ops(X)
def test_cudnn_rnn(self):
if get_ngpu() == 0:
return
print()
batch_size = 2
time_steps = 5
input_dim = 12
hidden_dim = 8
X = K.variable(value=np.random.rand(batch_size, time_steps, input_dim),
dtype='float32',
name='X')
for rnn_mode in ('lstm', 'rnn_relu', 'gru'):
for num_layers in [1, 2]:
for W_init in [
init_ops.glorot_uniform_initializer(seed=1234),
init_ops.random_normal_initializer(seed=1234)
]:
for b_init in [0, 1]:
for bidirectional in (True, False):
for skip_input in (False, ):
print('RNNmode:%s' % rnn_mode,
"#Layers:%d" % num_layers,
'Bidirectional:%s' % bidirectional,
'SkipInput:%s' % skip_input)
weights, biases = K.init_rnn(
input_dim=input_dim,
hidden_dim=hidden_dim,
num_gates=rnn_mode,
num_layers=num_layers,
W_init=W_init,
b_init=b_init,
skip_input=skip_input,
cudnn_vector=False,
is_bidirectional=bidirectional,
name=None)
# ====== check number of params ====== #
params1 = K.params_to_cudnn(weights, biases)
n = params1.shape[0].value
nb_params = cudnn_rnn_ops.cudnn_rnn_opaque_params_size(
rnn_mode=rnn_mode,
num_layers=num_layers,
num_units=hidden_dim,
input_size=input_dim,
input_mode='skip_input'
if skip_input else 'linear_input',
direction='bidirectional'
if bidirectional else 'unidirectional')
nb_params = K.eval(nb_params)
assert n == nb_params
# ====== check cannonical shape match ====== #
kwargs = {
'num_layers':
num_layers,
'num_units':
hidden_dim,
'input_mode':
'skip_input'
if skip_input else 'linear_input',
'direction':
'bidirectional'
if bidirectional else 'unidirectional'
}
if rnn_mode == 'lstm':
rnn = cudnn_rnn.CudnnLSTM(**kwargs)
elif rnn_mode == 'gru':
rnn = cudnn_rnn.CudnnGRU(**kwargs)
if rnn_mode == 'rnn_relu':
rnn = cudnn_rnn.CudnnRNNRelu(**kwargs)
if rnn_mode == 'rnn_tanh':
rnn = cudnn_rnn.CudnnRNNTanh(**kwargs)
rnn.build(input_shape=(None, None, input_dim))
assert len(weights) == len(
rnn.canonical_weight_shapes)
assert len(biases) == len(
rnn.canonical_bias_shapes)
for w, s in zip(weights,
rnn.canonical_weight_shapes):
assert tuple(w.shape.as_list()) == s
# ====== check params conversion ====== #
K.initialize_all_variables()
params2 = cudnn_rnn_ops.cudnn_rnn_canonical_to_opaque_params(
rnn_mode=rnn_mode,
num_layers=num_layers,
num_units=hidden_dim,
input_size=input_dim,
input_mode='skip_input'
if skip_input else 'linear_input',
direction='bidirectional'
if bidirectional else 'unidirectional',
weights=weights,
biases=biases)
assert np.all(
K.eval(params1) == K.eval(params2))
# ====== odin cudnn implementation ====== #
name = 'TEST' + uuid(length=25)
outputs = K.cudnn_rnn(
X=X,
num_units=hidden_dim,
rnn_mode=rnn_mode,
num_layers=num_layers,
parameters=None,
skip_input=skip_input,
is_bidirectional=bidirectional,
dropout=0.1,
name=name)
K.initialize_all_variables()
s0 = K.eval(outputs[0]).sum()
s1 = K.eval(outputs[1]).sum()
all_variables = K.get_all_variables(scope=name)
new_weights = [
i for i in all_variables
if K.role.has_roles(i, roles=K.role.Weight)
]
new_biases = [
i for i in all_variables
if K.role.has_roles(i, roles=K.role.Bias)
]
new_weights, new_biases = K.sort_cudnn_params(
new_weights, new_biases, rnn_mode=rnn_mode)
assert len(weights) == len(weights)
assert len(biases) == len(biases)
for i, j in zip(weights + biases,
new_weights + new_biases):
assert i.name.split(
'/')[-1] == j.name.split('/')[-1]
# ====== CudnnRNN wrapper ====== #
rnn = N.CudnnRNN(
num_units=hidden_dim,
W_init=new_weights,
b_init=new_biases,
rnn_mode=rnn_mode,
num_layers=num_layers,
skip_input=skip_input,
is_bidirectional=bidirectional,
return_states=True,
dropout=0.)
outputs = rnn(X)
K.initialize_all_variables()
y0 = K.eval(outputs[0]).sum()
y1 = K.eval(outputs[1]).sum()
assert y0 == s0
assert y1 == s1