def network(insize=1, size=256, winlen=19, stride=2, alphabet_info=None):
nbase = 4 if alphabet_info is None else alphabet_info.nbase
return Serial([
Convolution(insize, size, winlen, stride=stride, fun=tanh),
Reverse(GruMod(size, size)),
GruMod(size, size),
Reverse(GruMod(size, size)),
GruMod(size, size),
Reverse(GruMod(size, size)),
GlobalNormFlipFlop(size, nbase),
])
def network(insize=1, size=512, winlen=19, stride=2, outsize=40):
nbase = nbase_flipflop(outsize)
return Serial([
Convolution(insize, size, winlen, stride=stride, fun=tanh),
Reverse(GruMod(size, size)),
GruMod(size, size),
Reverse(GruMod(size, size)),
GruMod(size, size),
Reverse(GruMod(size, size)),
GlobalNormFlipFlop(size, nbase),
])
def network(insize=1, size=256, winlen=19, stride=2, outsize=40):
nbase = nbase_flipflop(
outsize) #2 * nbase * (nbase + 1) -> 40 if nbase=4, but why??
return Serial([
Convolution(insize, size, winlen, stride=stride, fun=tanh),
Reverse(GruMod(size, size)),
GruMod(size, size),
Reverse(GruMod(size, size)),
GruMod(size, size),
Reverse(GruMod(size, size)),
GlobalNormFlipFlop(size, nbase),
])
def network(insize=1, size=256, winlen=19, stride=2, outsize=40):
nbase = int(np.sqrt(outsize / 2))
assert 2 * nbase * (nbase + 1) == outsize,\
"Invalid size for a flipflop model: nbase = {}, size = {}".format(nbase, outsize)
return Serial([
Convolution(insize, size, winlen, stride=stride, fun=tanh),
Reverse(GruMod(size, size)),
GruMod(size, size),
Reverse(GruMod(size, size)),
GruMod(size, size),
Reverse(GruMod(size, size)),
GlobalNormFlipFlop(size, nbase),
])
def network(insize=1, size=256, winlen=19, stride=5, alphabet_info=None):
nbase = 4 if alphabet_info is None else alphabet_info.nbase
winlen2 = 5
return Serial([
Convolution(insize, 4, winlen2, stride=1, fun=swish),
Convolution(4, 16, winlen2, stride=1, fun=swish),
Convolution(16, size, winlen, stride=stride, fun=swish),
Reverse(Lstm(size, size)),
Lstm(size, size),
Reverse(Lstm(size, size)),
Lstm(size, size),
Reverse(Lstm(size, size)),
GlobalNormFlipFlop(size, nbase),
])
def parse_sublayer(sublayer):
# TODO apply additional attributes (e.g. has_bias, convolutional padding)
if sublayer['type'] == 'convolution':
if sublayer['activation'] != 'tanh':
sys.stderr.write((
'Incompatible convolutional layer activation fucntion ' +
'({}) encountered.\n').format(sublayer['type']))
sys.exit(1)
sys.stderr.write((
'Loading convolutional layer with attributes:\n\tin size: {}\n' +
'\tout size: {}\n\twinlen: {}\n\tstride: {}\n').format(
sublayer['insize'], sublayer['size'], sublayer['winlen'],
sublayer['stride']))
layer = Convolution(
sublayer['insize'], sublayer['size'], sublayer['winlen'],
stride=sublayer['stride'], fun=tanh)
elif sublayer['type'] == 'LSTM':
sys.stderr.write((
'Loading LSTM layer with attributes:\n\tin size: {}\n' +
'\tout size: {}\n').format(
sublayer['insize'], sublayer['size']))
layer = Lstm(sublayer['insize'], sublayer['size'])
elif sublayer['type'] == 'GruMod':
sys.stderr.write((
'Loading GRU layer with attributes:\n\tin size: {}\n' +
'\tout size: {}\n').format(
sublayer['insize'], sublayer['size']))
layer = GruMod(sublayer['insize'], sublayer['size'])
elif sublayer['type'] == 'reverse':
sublayer = sublayer['sublayers']
if sublayer['type'] == 'GruMod':
sys.stderr.write((
'Loading Reverse GRU layer with attributes:\n\tin size: {}\n' +
'\tout size: {}\n').format(
sublayer['insize'], sublayer['size']))
layer = Reverse(GruMod(sublayer['insize'], sublayer['size']))
elif sublayer['type'] == 'LSTM':
sys.stderr.write((
'Loading Reverse LSTM layer with attributes:\n' +
'\tin size: {}\n\tout size: {}\n').format(
sublayer['insize'], sublayer['size']))
layer = Reverse(Lstm(sublayer['insize'], sublayer['size']))
else:
sys.stderr.write((
'Invalid reversed-time layer type ({})\n').format(
sublayer['type']))
sys.exit(1)
elif sublayer['type'] == 'GlobalNormTwoState':
nbase = nbase_flipflop(sublayer['size'])
sys.stderr.write((
'Loading flip-flop layer with attributes:\n\tin size: {}\n' +
'\tnbases: {}\n').format(sublayer['insize'], nbase))
layer = GlobalNormFlipFlop(sublayer['insize'], nbase)
elif sublayer['type'] == 'GlobalNormTwoStateCatMod':
output_alphabet = sublayer['output_alphabet']
curr_can_base = 0
collapse_alphabet = ''
for can_i_nmod in sublayer['can_nmods']:
collapse_alphabet += output_alphabet[curr_can_base] * (
can_i_nmod + 1)
curr_can_base += can_i_nmod + 1
alphabet_info = alphabet.AlphabetInfo(
output_alphabet, collapse_alphabet,
sublayer['modified_base_long_names'], do_reorder=False)
sys.stderr.write((
'Loading modified bases flip-flop layer with attributes:\n' +
'\tin size: {}\n\tmod bases: {}\n').format(
sublayer['insize'], alphabet_info.mod_long_names))
layer = GlobalNormFlipFlopCatMod(sublayer['insize'], alphabet_info)
else:
sys.stderr.write('Encountered invalid layer type ({}).\n'.format(
sublayer['type']))
sys.exit(1)
layer = set_params(layer, sublayer['params'], sublayer['type'])
return layer