def conv2d(x,
kernel,
strides=(1, 1),
padding='valid',
data_format='channels_first',
image_shape=None,
filter_shape=None):
"""2D convolution.
# Arguments
x: Input tensor
kernel: kernel tensor.
strides: strides tuple.
padding: string, "same" or "valid".
data_format: 'channels_first' or 'channels_last'.
Whether to use Theano or TensorFlow dimension
ordering in inputs/kernels/ouputs.
image_shape: Optional, the input tensor shape
filter_shape: Optional, the kernel shape.
# Returns
x convolved with the kernel.
# Raises
Exception: In case of invalid border mode or data format.
"""
if padding == 'same':
padding = 'SAME'
elif padding == 'valid':
padding = 'VALID'
else:
raise Exception('Invalid border mode: ' + str(padding))
strides = (1, ) + strides + (1, )
if floatx() == 'float64':
# tf conv2d only supports float32
x = tf.cast(x, 'float32')
kernel = tf.cast(kernel, 'float32')
if data_format == 'channels_first':
# TF uses the last dimension as channel dimension,
# instead of the 2nd one.
# TH input shape: (samples, input_depth, rows, cols)
# TF input shape: (samples, rows, cols, input_depth)
# TH kernel shape: (depth, input_depth, rows, cols)
# TF kernel shape: (rows, cols, input_depth, depth)
x = tf.transpose(x, (0, 2, 3, 1))
kernel = tf.transpose(kernel, (2, 3, 1, 0))
x = tf.nn.conv2d(x, kernel, strides, padding=padding)
x = tf.transpose(x, (0, 3, 1, 2))
elif data_format == 'channels_last':
x = tf.nn.conv2d(x, kernel, strides, padding=padding)
else:
raise Exception('Unknown data_format: ' + str(data_format))
if floatx() == 'float64':
x = tf.cast(x, 'float64')
return x
def testMicroAddElementsFail(self):
data = [m(3, 3), m(3, 3)]
test_func = self.testAddElements
args = list()
###############
x = [
pkb.placeholder(shape=t.shape) for t in data
if isinstance(t, np.ndarray)
]
xv = [
pkb.variable(t, dtype=floatx()) for t in data
if isinstance(t, np.ndarray)
]
par = [t for t in data if not isinstance(t, np.ndarray)]
grad_funcs = test_func(pkb, *(x + par + list(args)))
funcs = test_func(pkb, *(xv + par + list(args)))
#for gf, f in zip(grad_funcs, funcs):
gf = grad_funcs[0]
f = funcs[0]
df = pkb.gradients(pkb.mean(gf), x)
gfn = pkb.function(x, df, updates=[])
fr = f.eval()
gr = gfn([t for t in data if isinstance(t, np.ndarray)])
if args.verbose:
print(pkb, fr, gr)
results.append((fr, gr))
return results
def m(*args, **kwargs):
dtype = kwargs.get('dtype', floatx())
"""Makes a test matrix whose dimensions are the supplied arguments."""
total = functools.reduce(operator.mul, args, 1)
arr = np.array(range(-2, total - 2), dtype=dtype)
arr = np.reshape(arr, args)
return arr
def make_sparse(value, dtype=None):
if dtype is None:
dtype = floatx()
assert hasattr(value, 'tocoo')
_assert_sparse_module()
var = th_sparse_module.as_sparse_variable(value)
return var
def run_one_backend(self, data, test_func, b, *args):
tf_session = tensorflow.Session()
tf.set_session(tf_session)
results = []
with tf_session.as_default():
x = [
b.placeholder(shape=t.shape) for t in data
if hasattr(t, 'shape')
]
xv = [
b.variable(t, dtype=floatx()) for t in data
if hasattr(t, 'shape')
]
ps = [t for t in data if not hasattr(t, 'shape')]
grad_funcs = test_func(self, b, *(x + ps + list(args)))
funcs = test_func(self, b, *(xv + ps + list(args)))
tf_session.run(tensorflow.global_variables_initializer())
for gf, f in zip(grad_funcs, funcs):
df = b.gradients(b.mean(gf), x)
gfn = b.function(x, df, updates=[])
fr = f.eval()
gr = gfn([t for t in data if hasattr(t, 'shape')])
if verbose:
print(b, fr, gr)
results.append((fr, gr))
tf_session.close()
return results
def placeholder(shape=None, ndim=None, dtype=None, sparse=False, name=None):
dtype = plaidml.DType.from_numpy(dtype or floatx())
# TODO: Need to support empty shapes; once supported, convert below to `if _ is not None`
if shape:
return _KerasNode('placeholder', shape=edsl.LogicalShape(dtype, shape), name=name)
if ndim:
return _KerasNode('placeholder', shape=edsl.LogicalShape(dtype, [0] * ndim), name=name)
raise ValueError()
def constant(value, dtype=None, shape=None, name=None):
if dtype is None:
dtype = floatx()
if shape is None:
shape = ()
np_value = value * np.ones(shape)
np_value.astype(dtype)
return np_value
def random_uniform(shape, minval=0.0, maxval=1.0, dtype=None, seed=None):
rng_state = _make_rng_state(seed)
R = edsl.prng(rng_state.tensor, shape)
dtype = dtype or floatx()
if dtype != 'float32':
R = edsl.cast(R, plaidml.DType.from_numpy(dtype))
O = (maxval - minval) * R + minval
return _KerasNode('random_uniform', tensor=O)
def truncated_normal(shape, mean=0.0, stddev=1.0, dtype=None, seed=None):
logger.debug('truncated_normal(shape: {}, mean: {}, stddev: {}, dtype: {}, seed: {})'.format(
shape, mean, stddev, dtype, seed))
if dtype is None:
dtype = floatx()
if seed:
np.random.seed(seed)
return variable(stddev * scipy.stats.truncnorm.rvs(-2.0, 2.0, size=shape) + mean, dtype)
def random_normal_variable(shape, mean, scale, dtype=None, name=None, seed=None):
if dtype is None:
dtype = floatx()
elif isinstance(dtype, plaidml.DType):
dtype = ptile.convert_pml_dtype_to_np(dtype)
if seed:
np.random.seed(seed)
data = np.random.normal(mean, scale, shape).astype(dtype)
return variable(data, dtype=dtype, name=name)
def constant(value, dtype=None, shape=None, name=None):
if shape is None:
if isinstance(value, np.ndarray):
shape = value.shape
elif isinstance(value, list) or isinstance(value, tuple):
shape = (len(value),)
else:
shape = (1,)
np_value = np.full(shape, value, dtype=dtype or floatx())
return _KerasNode('constant', name=name, value=np_value)
def zeros_like(x, dtype=None, name=None):
value = np.full((1), 0, dtype=dtype or floatx())
zero = _create_var('a_zero', value)
I = x.tensor
ndim = I.shape.ndims
dims = edsl.TensorDims(ndim)
idxs = edsl.TensorIndexes(ndim)
I.bind_dims(*dims)
O = edsl.TensorOutput(*dims)
O[idxs] = zero[0]
return _KerasNode('zeros_like', name=name, tensor=O)
def make_sparse(value, dtype=None):
if dtype is None:
dtype = floatx()
assert hasattr(value, 'tocoo')
sparse_coo = value.tocoo()
indices = np.concatenate(
(np.expand_dims(sparse_coo.row, 1), np.expand_dims(sparse_coo.col, 1)),
1)
v = tf.SparseTensor(indices=indices,
values=sparse_coo.data,
dense_shape=sparse_coo.shape)
return v
def n(*args):
"""Makes a test matrix whose dimensions are the supplied arguments.
Differs from m only in what values it has."""
total = functools.reduce(operator.mul, args, 1)
arr = np.array(range(-11, total - 11), dtype=floatx())
arr = np.reshape(arr, args)
for i in range(5):
if len(args) > i + 1:
np.swapaxes(arr, 0, i + 1)
arr = np.reshape(arr, args)
return arr
def random_normal(shape, mean=0.0, stddev=1.0, dtype=None, seed=None):
if dtype is None:
dtype = floatx()
if seed:
np.random.seed(seed)
# TODO: We only use half of the Box-Muller here
u1 = random_uniform(shape, dtype='float32')
u2 = random_uniform(shape, dtype='float32')
z0 = sqrt(-2.0 * log(u1 + (1.0 / (2**33)))) * cos(2.0 * math.pi * u2)
z0 = stddev * z0
z0 = z0 + mean
if dtype != 'float32':
z0 = cast(z0, dtype)
return z0
def __init__(self, extend_labels=False, multilabel=False):
self._multilabel = multilabel
self._extend_labels = extend_labels
self._labels = np.loadtxt(label_file(), dtype=floatx())
self._class_index = open(class_index_file(), 'r').read().split('\n')
if extend_labels:
self._word_tree = Tree()
self._extra_labels = {
"layer_1":
np.zeros((self._labels.shape[0],
len(self._word_tree.extra_activations["layer_1"]))),
"layer_2":
np.zeros((self._labels.shape[0],
len(self._word_tree.extra_activations["layer_2"]))),
"layer_3":
np.zeros((self._labels.shape[0],
len(self._word_tree.extra_activations["layer_3"])))
}
for sample in range(self._labels.shape[0]):
#Removing Multilabel 1
if not multilabel:
search = np.where(self._labels[sample] == 1)
if search[0].shape[0] > 1:
self._labels[sample] = np.zeros(
(self._labels.shape[1]), dtype=floatx())
self._labels[sample, search[0][0]] = 1.
for o_layer in list(
self._word_tree.coocurrence_indexes.keys()):
activations = np.where(self._labels[sample] == 1.)[0]
groups = self._word_tree.coocurrence_indexes[o_layer][
activations]
try:
self._extra_labels[o_layer][sample, groups] = 1.
except Exception:
print(Exception, o_layer, activations)
exit()
def _postprocess_conv2d_output(x, data_format):
"""Transpose and cast the output from conv2d if needed.
# Arguments
x: A tensor.
data_format: string, `"channels_last"` or `"channels_first"`.
# Returns
A tensor.
"""
if data_format == 'channels_first':
x = tf.transpose(x, (0, 3, 1, 2))
if floatx() == 'float64':
x = tf.cast(x, 'float64')
return x
def sobel(img, border_mode='zero'):
filter = np.array([
[1, 0, -1],
[2, 0, -2],
[1, 0, -1],
], dtype=floatx())
img, conv_border = add_border(img, border=1, mode=border_mode)
kernel_x = theano.shared(filter[np.newaxis, np.newaxis])
kernel_y = theano.shared(np.transpose(filter)[np.newaxis, np.newaxis])
conv_x = T.nnet.conv2d(img, kernel_x, border_mode=conv_border,
input_shape=(None, 1, None, None),
filter_shape=(1, 1, 3, 3))
conv_y = T.nnet.conv2d(img, kernel_y, border_mode=conv_border,
input_shape=(None, 1, None, None),
filter_shape=(1, 1, 3, 3))
return conv_x, conv_y
def variable(value, dtype=None, name=None, constraint=None):
if name is None:
name = 'anon'
dtype = dtype or floatx()
if isinstance(value, _KerasNode):
value = value.eval()
if isinstance(value, float) or isinstance(value, six.integer_types):
value = np.array(value, dtype=dtype)
if isinstance(value, list) or isinstance(value, tuple):
value = np.array(value, dtype=dtype)
if isinstance(value, np.ndarray):
if dtype != value.dtype:
logger.debug(
'Casting to requested dtype in variable, received {} and requested {}'.format(
value.dtype, dtype))
value = value.astype(dtype)
return _KerasNode('variable', name=name, value=value)
raise TypeError('Unknown type for variable: {}'.format(type(value)))
def sobel(img, border_mode='zero'):
filter = np.array([
[1, 0, -1],
[2, 0, -2],
[1, 0, -1],
], dtype=floatx())
img, conv_border = add_border(img, border=1, mode=border_mode)
kernel_x = theano.shared(filter[np.newaxis, np.newaxis])
kernel_y = theano.shared(np.transpose(filter)[np.newaxis, np.newaxis])
conv_x = T.nnet.conv2d(img,
kernel_x,
border_mode=conv_border,
input_shape=(None, 1, None, None),
filter_shape=(1, 1, 3, 3))
conv_y = T.nnet.conv2d(img,
kernel_y,
border_mode=conv_border,
input_shape=(None, 1, None, None),
filter_shape=(1, 1, 3, 3))
return conv_x, conv_y
def sparse_mean(x, axis=None):
"""Mean of a tensor, alongside the specified axis.
"""
# bool is available since theano v0.9dev
if 'int' in x.dtype or x.dtype == 'bool':
dtype = floatx()
else:
dtype = x.dtype
if isinstance(axis, (integer_types, np.integer)):
if axis == -1:
axis = max(x.ndim - 1, 0)
s = th_sparse_module.sp_sum(x, axis, True)
shp = shape(x)
if s.dtype in ('float16', 'float32', 'complex64'):
shp = cast(shp, 'float32')
else:
shp = cast(shp, 'float64')
if axis is None:
axis = list(range(len(x.data.shape)))
elif isinstance(axis, (integer_types, np.integer)):
axis = [axis]
elif isinstance(axis, np.ndarray) and axis.ndim == 0:
axis = [int(axis)]
else:
axis = [int(a) for a in axis]
for i in axis:
s = true_div(s, shp[i])
if s.dtype != shp.dtype and s.dtype in discrete_dtypes:
s = cast(s, shp.dtype)
if dtype == 'float16' or (dtype is None and x.dtype == 'float16'):
s = cast(s, 'float16')
s.name = 'mean'
return s
def truncated_normal(shape, mean=0.0, stddev=1.0, dtype=None, seed=None):
if dtype is None:
dtype = floatx()
if seed:
np.random.seed(seed)
return variable(stddev * scipy.stats.truncnorm.rvs(-2.0, 2.0, size=shape) + mean, dtype)
def zeros(shape, dtype=None, name=None):
logger.debug('zeros(shape: {}, dtype: {}, name: {})'.format(shape, dtype, name))
value = np.full(shape, 0, dtype=dtype or floatx())
return _KerasNode('zeros', name=name, value=value)
def ones_like(x, dtype=floatx(), name=None):
return np.ones_like(x, dtype=dtype)
def train(db, entity_db, vocab, word2vec, out_file, mode, text_len, dim_size,
negative, epoch, batch_size, word_static, entity_static,
include_title, optimizer, dev_size, patience, num_links,
random_seed):
np.random.seed(random_seed)
click.echo('Initializing weights...')
word_embedding = np.random.uniform(low=-0.05,
high=0.05,
size=(vocab.word_size, dim_size))
word_embedding = np.vstack([np.zeros(dim_size), word_embedding])
word_embedding = word_embedding.astype(floatx())
entity_embedding = np.random.uniform(low=-0.05,
high=0.05,
size=(vocab.entity_size, dim_size))
entity_embedding = entity_embedding.astype(floatx())
if word2vec:
for word in vocab.words():
try:
vec = word2vec.get_word_vector(word)
except KeyError:
continue
if vec is not None:
word_embedding[vocab.get_word_index(word) + 1] = vec
for entity in vocab.entities():
try:
vec = word2vec.get_entity_vector(entity)
except KeyError:
continue
if vec is not None:
entity_embedding[vocab.get_entity_index(
entity)] = vec / np.linalg.norm(vec, 2)
tokenizer = RegexpTokenizer()
if mode == 'sentence':
sentence_detector = OpenNLPSentenceDetector()
def generate_data(keys, count_links=False, shuffle=True, loop=True):
num_entities = entity_embedding.shape[0]
labels = np.zeros((batch_size, negative + 1), dtype=np.int)
labels[:, 0] = 1
while True:
word_batch = []
entity_batch = []
if shuffle:
keys = np.random.permutation(keys)
for key in keys:
value = db[key]
text = value['text']
links = value['links']
target_data = []
if mode == 'paragraph':
target_data = [(text, links)]
if include_title:
target_data[0][1].append(
(None, key.decode('utf-8'), None))
elif mode == 'sentence':
for (start,
end) in sentence_detector.sent_pos_detect(text):
target_data.append((text[start:end], [
(l[0], l[1], (l[2][0] - start, l[2][1] - start))
for l in links if start <= l[2][0] < end
]))
if include_title:
target_data[-1][1].append(
(None, key.decode('utf-8'), None))
else:
raise NotImplementedError()
for (target_text, target_links) in target_data:
word_indices = []
word_offsets = []
for token in tokenizer.tokenize(target_text):
word_index = vocab.get_word_index(token.text.lower())
if word_index is not None:
word_indices.append(word_index + 1)
word_offsets.append(token.span[0])
positive_ids = [
vocab.get_entity_index(
entity_db.resolve_redirect(title))
for (_, title, _) in target_links
]
positive_id_set = frozenset(
[o for o in positive_ids if o is not None])
for (positive_id,
(_, title, span)) in zip(positive_ids, target_links):
if positive_id is None:
continue
if not word_indices:
continue
if count_links:
yield 1
continue
negatives = []
while True:
negative_id = np.random.randint(0, num_entities)
if negative_id not in positive_id_set:
negatives.append(negative_id)
if len(negatives) == negative:
break
word_batch.append(word_indices)
entity_indices = [positive_id] + negatives
entity_batch.append(entity_indices)
if len(word_batch) == batch_size:
yield ([
pad_sequences(word_batch, maxlen=text_len),
np.vstack(entity_batch)
], labels)
word_batch = []
entity_batch = []
if word_batch:
yield ([
pad_sequences(word_batch, maxlen=text_len),
np.vstack(entity_batch)
], labels[:len(word_batch)])
if not loop or count_links:
break
(train_keys, dev_keys) = train_test_split(db.keys(),
test_size=dev_size,
random_state=random_seed)
if num_links is None:
click.echo('Counting links...')
with click.progressbar(train_keys) as bar:
num_links = sum(
list(generate_data(bar, count_links=True, shuffle=False)))
click.echo('The number of links: %d' % num_links)
dev_data = list(generate_data(dev_keys, loop=False))
dev_data = ([
np.vstack([d[0][0] for d in dev_data]),
np.vstack([d[0][1] for d in dev_data])
], np.vstack([d[1] for d in dev_data]))
callbacks = []
callbacks.append(
ModelCheckpoint(out_file + '_checkpoint.h5',
monitor='val_acc',
save_best_only=True))
if patience:
callbacks.append(EarlyStopping(monitor='val_acc', patience=patience))
model = build_model(
text_len=text_len,
negative_size=negative,
optimizer=optimizer,
word_size=word_embedding.shape[0],
entity_size=entity_embedding.shape[0],
dim_size=word_embedding.shape[1],
word_static=word_static,
entity_static=entity_static,
word_embedding=word_embedding,
entity_embedding=entity_embedding,
)
model.fit_generator(generate_data(train_keys),
samples_per_epoch=num_links,
nb_epoch=epoch,
validation_data=dev_data,
max_q_size=1000,
callbacks=callbacks)
db.close()
word_embedding = model.get_layer('word_embedding').get_weights()[0][1:]
entity_embedding = model.get_layer('entity_embedding').get_weights()[0]
ret = dict(
word_embedding=word_embedding,
entity_embedding=entity_embedding,
vocab=vocab,
)
ret['W'] = model.get_layer('text_layer').get_weights()[0]
ret['b'] = model.get_layer('text_layer').get_weights()[1]
joblib.dump(ret, out_file + '.joblib', protocol=-1)
def eye(size, dtype=None, name=None):
if dtype is None:
dtype = floatx()
elif isinstance(dtype, plaidml.DType):
dtype = dtype.into_numpy()
return variable(np.eye(size, dtype=dtype), name=name, dtype=dtype)
# -*- coding: utf-8 -*-
import numpy as np
import theano
from keras import backend as K
from keras.backend.common import floatx
from keras.engine.topology import Layer
from theano import tensor as T
FLOATX = floatx()
FLOAT_MIN = np.finfo('float32').min + K.epsilon()
FLOAT_MAX = np.finfo('float32').max - K.epsilon()
class TextRepresentationLayer(Layer):
def __init__(self, W=None, b=None, *args, **kwargs):
super(TextRepresentationLayer, self).__init__(*args, **kwargs)
self.W = W
self.b = b
def build(self, input_shape):
if self.W is None:
self.W = K.variable(np.identity(input_shape[0][2]))
elif isinstance(self.W, np.ndarray):
self.W = K.variable(self.W)
else:
raise RuntimeError()
if self.b is None:
self.b = K.random_uniform_variable((input_shape[0][2],), -0.05, 0.05)
def zeros(shape, dtype=None, name=None):
value = np.full(shape, 0, dtype=dtype or floatx())
return _KerasNode('zeros', name=name, value=value)
def build_model():
import keras.applications as kapp
from keras.layers import Input
from keras.backend.common import floatx
inputLayer = Input(shape=(224, 224, 3), dtype=floatx())
return kapp.VGG19(input_tensor=inputLayer)
def zeros_like(x, dtype=floatx(), name=None):
return np.zeros_like(x, dtype=dtype)
def ones(shape, dtype=None, name=None):
value = np.full(shape, 1, dtype=dtype or floatx())
return _KerasNode('ones', name=name, value=value)
def ones(shape, dtype=floatx(), name=None):
return np.ones(shape, dtype=dtype)
