def add_lstm_cell(network, name, size, xin, cin=None, hin=None):
'''
Add a LSTM cell named `name` to the `network`, with the dimension `size`.
`xin`, `cin`, `hin` are the layers' names whose outputs are x_t, C_{t-1},
h_{t-1}, respectively. Return the layers' names whose outputs are C_t, h_t.
'''
from nn_dataflow.core import Network
from nn_dataflow.core import InputLayer, FCLayer, EltwiseLayer
if not isinstance(network, Network):
raise TypeError('add_lstm_cell: network must be a Network instance.')
if cin is None:
cin = '{}_cinit'.format(name)
network.add_ext(cin, InputLayer(size, 1))
if hin is None:
hin = '{}_hinit'.format(name)
network.add_ext(hin, InputLayer(size, 1))
if (cin not in network) or (hin not in network) or (xin not in network):
raise ValueError('add_lstm_cell: cin {}, hin {}, xin {} must all be '
'in the network.'.format(cin, hin, xin))
def gate_name(gate):
''' Name of a gate. '''
return '{}_{}gate'.format(name, gate)
# Candidate.
cand_name = '{}_cand'.format(name)
prevs = (hin, xin) if hin else (xin, )
network.add(cand_name, FCLayer(len(prevs) * size, size), prevs=prevs)
# Three gates.
prevs = (hin, xin) if hin else (xin, )
for g in ['i', 'f', 'o']:
network.add(gate_name(g),
FCLayer(len(prevs) * size, size),
prevs=prevs)
# C_t.
cout_name = '{}_cout'.format(name)
cout_f_name = cout_name + '_f'
prevs = (cin, gate_name('f')) if cin else (gate_name('f'), )
network.add(cout_f_name, EltwiseLayer(size, 1, len(prevs)), prevs=prevs)
cout_i_name = cout_name + '_i'
prevs = (cand_name, gate_name('i'))
network.add(cout_i_name, EltwiseLayer(size, 1, 2), prevs=prevs)
prevs = (cout_i_name, cout_f_name)
network.add(cout_name, EltwiseLayer(size, 1, 2), prevs=prevs)
# h_t.
hout_name = '{}_hout'.format(name)
prevs = (cout_name, gate_name('o'))
network.add(hout_name, EltwiseLayer(size, 1, 2), prevs=prevs)
return cout_name, hout_name
def test_add(self):
''' Modifier add. '''
self.assertEqual(len(self.network), 3)
self.network.add('f2', FCLayer(64, 2000, 7), prevs='p1')
self.network.add('f3', FCLayer(3000, 1000), prevs=('f1', 'f2'))
self.network.add('e4', EltwiseLayer(1000, 1, 2), prevs=('f1', 'f3'))
self.network.add('f4', FCLayer(1000, 1000), prevs='e4')
self.assertEqual(len(self.network), 7)
def test_repr(self):
''' __repr__. '''
# pylint: disable=eval-used
for l in [
LocalRegionLayer(64, 28, 2, 1),
LocalRegionLayer(64, [28, 14], 1, [2, 4]),
LocalRegionLayer(64, [28, 14], 1, [2, 4], 7),
LocalRegionLayer(64, 28, 1, 4, 7)
]:
self.assertIn('LocalRegionLayer', repr(l))
self.assertEqual(eval(repr(l)), l)
for l in [
PoolingLayer(64, 28, 2),
PoolingLayer(64, 28, 3, strd=2),
PoolingLayer(64, [28, 14], [3, 4], strd=[2, 3])
]:
self.assertIn('PoolingLayer', repr(l))
self.assertEqual(eval(repr(l)), l)
for l in [EltwiseLayer(64, 32, 3), EltwiseLayer(64, 28, 4)]:
self.assertIn('EltwiseLayer', repr(l))
self.assertEqual(eval(repr(l)), l)
def test_nexts(self):
''' Get nexts. '''
self.network.add('f2', FCLayer(64, 2000, 7), prevs='p1')
self.network.add('f3', FCLayer(3000, 1000), prevs=('f1', 'f2'))
self.network.add('e4', EltwiseLayer(1000, 1, 2), prevs=('f1', 'f3'))
self.network.add('f4', FCLayer(1000, 1000), prevs='e4')
nexts = self.network.nexts('p1')
self.assertTupleEqual(nexts, ('f1', 'f2'))
nexts = self.network.nexts('f1')
self.assertTupleEqual(nexts, ('f3', 'e4'))
nexts = self.network.nexts('f2')
self.assertTupleEqual(nexts, ('f3',))
nexts = self.network.nexts('f3')
self.assertTupleEqual(nexts, ('e4',))
def test_len(self):
''' Accessor len. '''
self.assertEqual(len(self.network), 3)
network = Network('test_net')
self.assertEqual(len(network), 0)
network.set_input_layer(InputLayer(3, 224))
self.assertEqual(len(network), 0)
network.add('c1', ConvLayer(3, 4, 224, 1))
self.assertEqual(len(network), 1)
self.network.add('f2', FCLayer(64, 2000, 7), prevs='p1')
self.assertEqual(len(self.network), 4)
self.network.add('f3', FCLayer(3000, 1000), prevs=('f1', 'f2'))
self.assertEqual(len(self.network), 5)
self.network.add('e4', EltwiseLayer(1000, 1, 2), prevs=('f1', 'f3'))
self.assertEqual(len(self.network), 6)
self.network.add('f4', FCLayer(1000, 1000), prevs='e4')
self.assertEqual(len(self.network), 7)
NN.add('conv1', ConvLayer(3, 64, 112, 7, 2))
NN.add('pool1', PoolingLayer(64, 56, 3, 2))
RES_PREV = 'pool1'
for i in range(3):
NN.add('conv2_{}_a'.format(i), ConvLayer(64 if i == 0 else 256, 64, 56, 1))
NN.add('conv2_{}_b'.format(i), ConvLayer(64, 64, 56, 3))
NN.add('conv2_{}_c'.format(i), ConvLayer(64, 256, 56, 1))
# With residual shortcut.
if i == 0:
NN.add('conv2_br', ConvLayer(64, 256, 56, 1), prevs=(RES_PREV, ))
RES_PREV = 'conv2_br'
NN.add('conv2_{}_res'.format(i),
EltwiseLayer(256, 56, 2),
prevs=(RES_PREV, 'conv2_{}_c'.format(i)))
RES_PREV = 'conv2_{}_res'.format(i)
for i in range(8):
NN.add(
'conv3_{}_a'.format(i),
ConvLayer(256, 128, 28, 1, 2) if i == 0 else ConvLayer(
512, 128, 28, 1))
NN.add('conv3_{}_b'.format(i), ConvLayer(128, 128, 28, 3))
NN.add('conv3_{}_c'.format(i), ConvLayer(128, 512, 28, 1))
# With residual shortcut.
if i == 0:
NN.add('conv3_br', ConvLayer(256, 512, 28, 1, 2), prevs=(RES_PREV, ))
RES_PREV = 'conv3_br'
from nn_dataflow.core import Network
from nn_dataflow.core import InputLayer, EltwiseLayer
from nn_dataflow.nns import add_lstm_cell
'''
LSTM from GNMT.
Sutskever, Vinyals, Le, Google, NIPS 2014
'''
NN = Network('GNMT')
NN.set_input_layer(InputLayer(1000, 1))
NL = 4
# Word embedding is a simple lookup.
# Exclude or ignore embedding processing.
WE = NN.INPUT_LAYER_KEY
# layered LSTM.
X = WE
for l in range(NL):
cell = 'cell_l{}'.format(l)
C, H = add_lstm_cell(NN, cell, 1000, X)
X = H
# log(p), softmax.
NN.add('Wd', EltwiseLayer(1000, 1, 1), prevs=(X, ))
NN.add('conv1', ConvLayer(3, 64, 112, 7, 2))
NN.add('pool1', PoolingLayer(64, 56, 3, 2))
RES_PREV = 'pool1'
for i in range(3):
NN.add('conv2_{}_a'.format(i), ConvLayer(64 if i == 0 else 256, 64, 56, 1))
NN.add('conv2_{}_b'.format(i), ConvLayer(64, 64, 56, 3))
NN.add('conv2_{}_c'.format(i), ConvLayer(64, 256, 56, 1))
# With residual shortcut.
if i == 0:
NN.add('conv2_br', ConvLayer(64, 256, 56, 1), prevs=(RES_PREV,))
RES_PREV = 'conv2_br'
NN.add('conv2_{}_res'.format(i), EltwiseLayer(256, 56, 2),
prevs=(RES_PREV, 'conv2_{}_c'.format(i)))
RES_PREV = 'conv2_{}_res'.format(i)
for i in range(4):
NN.add('conv3_{}_a'.format(i),
ConvLayer(256, 128, 28, 1, 2) if i == 0
else ConvLayer(512, 128, 28, 1))
NN.add('conv3_{}_b'.format(i), ConvLayer(128, 128, 28, 3))
NN.add('conv3_{}_c'.format(i), ConvLayer(128, 512, 28, 1))
# With residual shortcut.
if i == 0:
NN.add('conv3_br', ConvLayer(256, 512, 28, 1, 2), prevs=(RES_PREV,))
RES_PREV = 'conv3_br'
NN.add('conv3_{}_res'.format(i), EltwiseLayer(512, 28, 2),
def test_eltwiselayer(self):
''' EltwiseLayer init. '''
elayer = EltwiseLayer(64, 28, 3)
self.assertEqual(elayer.ops_per_neuron(), 3)
self.assertEqual(elayer.nifm, 3 * elayer.nofm)
self.assertEqual(elayer.ifmap_size(), elayer.ofmap_size())
WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
PARTICULAR PURPOSE. See the BSD-3 License for more details.
You should have received a copy of the Modified BSD-3 License along with this
program. If not, see <https://opensource.org/licenses/BSD-3-Clause>.
"""
from nn_dataflow.core import Network
from nn_dataflow.core import InputLayer, EltwiseLayer
from nn_dataflow.nns import add_lstm_cell
'''
LSTM from Show and Tell.
Vinyals et al., Google, CVPR 2015
'''
NN = Network('ShowTell')
NN.set_input_layer(InputLayer(512, 1))
# Word embedding is a simple lookup.
# Exclude or ignore embedding processing.
WE = NN.INPUT_LAYER_KEY
# LSTM.
C, H = add_lstm_cell(NN, 'cell', 512, WE)
# log(p), softmax.
NN.add('Wd', EltwiseLayer(512, 1, 1), prevs=(H, ))