def calculate_loss(self, predictions, labels, weights=None, **unused_params):
with tf.name_scope("loss_xent"):
epsilon = 10e-6
if FLAGS.label_smoothing:
float_labels = smoothing(labels)
else:
float_labels = tf.cast(labels, tf.float32)
cross_entropy_loss = float_labels * tf.log(predictions + epsilon) + (
1 - float_labels) * tf.log(1 - predictions + epsilon)
cross_entropy_loss = tf.negative(cross_entropy_loss)
if weights is not None:
print cross_entropy_loss, weights
weighted_loss = tf.einsum("ij,i->ij", cross_entropy_loss, weights)
print "create weighted_loss", weighted_loss
return tf.reduce_mean(tf.reduce_sum(weighted_loss, 1))
else:
return tf.reduce_mean(tf.reduce_sum(cross_entropy_loss, 1))
python类einsum()的实例源码
def calculate_loss(self, predictions, labels, weights=None, **unused_params):
with tf.name_scope("loss_xent"):
epsilon = 10e-6
if FLAGS.label_smoothing:
float_labels = smoothing(labels)
else:
float_labels = tf.cast(labels, tf.float32)
cross_entropy_loss = float_labels * tf.log(predictions + epsilon) + (
1 - float_labels) * tf.log(1 - predictions + epsilon)
cross_entropy_loss = tf.negative(cross_entropy_loss)
if weights is not None:
print cross_entropy_loss, weights
weighted_loss = tf.einsum("ij,i->ij", cross_entropy_loss, weights)
print "create weighted_loss", weighted_loss
return tf.reduce_mean(tf.reduce_sum(weighted_loss, 1))
else:
return tf.reduce_mean(tf.reduce_sum(cross_entropy_loss, 1))
def create_model(self,
model_input,
vocab_size,
num_mixtures=None,
l2_penalty=1e-8,
sub_scope="",
original_input=None,
**unused_params):
num_methods = model_input.get_shape().as_list()[-1]
num_features = model_input.get_shape().as_list()[-2]
original_input = tf.nn.l2_normalize(original_input, dim=1)
gate_activations = slim.fully_connected(
original_input,
num_methods,
activation_fn=tf.nn.softmax,
weights_regularizer=slim.l2_regularizer(l2_penalty),
scope="gates"+sub_scope)
output = tf.einsum("ijk,ik->ij", model_input, gate_activations)
return {"predictions": output}
def create_model(self, model_input, vocab_size, l2_penalty=1e-8, original_input=None, **unused_params):
"""Creates a matrix regression model.
Args:
model_input: 'batch' x 'num_features' x 'num_methods' matrix of input features.
vocab_size: The number of classes in the dataset.
Returns:
A dictionary with a tensor containing the probability predictions of the
model in the 'predictions' key. The dimensions of the tensor are
batch_size x num_classes."""
num_features = model_input.get_shape().as_list()[-2]
num_methods = model_input.get_shape().as_list()[-1]
weight1d = tf.get_variable("ensemble_weight1d",
shape=[num_methods],
regularizer=slim.l2_regularizer(l2_penalty))
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_features, num_methods],
regularizer=slim.l2_regularizer(10 * l2_penalty))
weight = tf.nn.softmax(tf.einsum("ij,j->ij", weight2d, weight1d), dim=-1)
output = tf.einsum("ijk,jk->ij", model_input, weight)
return {"predictions": output}
def create_model(self, model_input, vocab_size, l2_penalty=1e-8, original_input=None, **unused_params):
"""Creates a linear regression model.
Args:
model_input: 'batch' x 'num_features' x 'num_methods' matrix of input features.
vocab_size: The number of classes in the dataset.
Returns:
A dictionary with a tensor containing the probability predictions of the
model in the 'predictions' key. The dimensions of the tensor are
batch_size x num_classes."""
num_methods = model_input.get_shape().as_list()[-1]
weight = tf.get_variable("ensemble_weight",
shape=[num_methods],
regularizer=slim.l2_regularizer(l2_penalty))
weight = tf.nn.softmax(weight)
output = tf.einsum("ijk,k->ij", model_input, weight)
return {"predictions": output}
def create_model(self, model_input, vocab_size, l2_penalty=1e-8, original_input=None, epsilon=1e-5, **unused_params):
"""Creates a non-unified matrix regression model.
Args:
model_input: 'batch' x 'num_features' x 'num_methods' matrix of input features.
vocab_size: The number of classes in the dataset.
Returns:
A dictionary with a tensor containing the probability predictions of the
model in the 'predictions' key. The dimensions of the tensor are
batch_size x num_classes."""
num_features = model_input.get_shape().as_list()[-2]
num_methods = model_input.get_shape().as_list()[-1]
log_model_input = tf.stop_gradient(tf.log((epsilon + model_input) / (1.0 + epsilon - model_input)))
weight = tf.get_variable("ensemble_weight",
shape=[num_features, num_methods],
regularizer=slim.l2_regularizer(l2_penalty))
weight = tf.nn.softmax(weight)
output = tf.nn.sigmoid(tf.einsum("ijk,jk->ij", log_model_input, weight))
return {"predictions": output}
def update_link_matrix(self, link_matrix_old, precedence_weighting_old, write_weighting):
"""
Updating the link matrix takes some effort (in order to vectorize the implementation)
Instead of the original index-by-index operation, it's all done at once.
:param link_matrix_old: from previous time step, shape [batch_size, memory_size, memory_size]
:param precedence_weighting_old: from previous time step, shape [batch_size, memory_size]
:param write_weighting: from current time step, shape [batch_size, memory_size]
:return: updated link matrix
"""
expanded = tf.expand_dims(write_weighting, axis=2)
# vectorizing the paper's original implementation
w = tf.tile(expanded, [1, 1, self.memory_size]) # shape [batch_size, memory_size, memory_size]
# shape of w_transpose is the same: [batch_size, memory_size, memory_size]
w_transp = tf.tile(tf.transpose(expanded, [0, 2, 1]), [1, self.memory_size, 1])
# in einsum, m and n are the same dimension because tensorflow doesn't support duplicated subscripts. Why?
lm = (1 - w - w_transp) * link_matrix_old + tf.einsum("bn,bm->bmn", precedence_weighting_old, write_weighting)
lm *= (1 - tf.eye(self.memory_size, batch_shape=[self.batch_size])) # making sure self links are off
return tf.identity(lm, name="Link_matrix")
def bilinear_answer_layer(size, encoded_question, question_length, encoded_support, support_length,
support2question, answer2support, is_eval, beam_size=1,
max_span_size=10000):
"""Answer layer for multiple paragraph QA."""
# computing single time attention over question
size = encoded_support.get_shape()[-1].value
question_state = compute_question_state(encoded_question, question_length)
# compute logits
hidden = tf.gather(tf.layers.dense(question_state, 2 * size, name="hidden"), support2question)
hidden_start, hidden_end = tf.split(hidden, 2, 1)
support_mask = misc.mask_for_lengths(support_length)
start_scores = tf.einsum('ik,ijk->ij', hidden_start, encoded_support)
start_scores = start_scores + support_mask
end_scores = tf.einsum('ik,ijk->ij', hidden_end, encoded_support)
end_scores = end_scores + support_mask
return compute_spans(start_scores, end_scores, answer2support, is_eval, support2question,
beam_size, max_span_size)
def compute_energy(hidden, state, attn_size, attn_keep_prob=None, pervasive_dropout=False, layer_norm=False,
mult_attn=False, **kwargs):
if attn_keep_prob is not None:
state_noise_shape = [1, tf.shape(state)[1]] if pervasive_dropout else None
state = tf.nn.dropout(state, keep_prob=attn_keep_prob, noise_shape=state_noise_shape)
hidden_noise_shape = [1, 1, tf.shape(hidden)[2]] if pervasive_dropout else None
hidden = tf.nn.dropout(hidden, keep_prob=attn_keep_prob, noise_shape=hidden_noise_shape)
if mult_attn:
state = dense(state, attn_size, use_bias=False, name='state')
hidden = dense(hidden, attn_size, use_bias=False, name='hidden')
return tf.einsum('ijk,ik->ij', hidden, state)
else:
y = dense(state, attn_size, use_bias=not layer_norm, name='W_a')
y = tf.expand_dims(y, axis=1)
if layer_norm:
y = tf.contrib.layers.layer_norm(y, scope='layer_norm_state')
hidden = tf.contrib.layers.layer_norm(hidden, center=False, scope='layer_norm_hidden')
f = dense(hidden, attn_size, use_bias=False, name='U_a')
v = get_variable('v_a', [attn_size])
s = f + y
return tf.reduce_sum(v * tf.tanh(s), axis=2)
def _linear(t_in, n_out):
v_w = tf.get_variable(
"w",
shape=(t_in.get_shape()[-1], n_out),
initializer=tf.uniform_unit_scaling_initializer(
factor=INIT_SCALE))
v_b = tf.get_variable(
"b",
shape=n_out,
initializer=tf.constant_initializer(0))
if len(t_in.get_shape()) == 2:
return tf.einsum("ij,jk->ik", t_in, v_w) + v_b
elif len(t_in.get_shape()) == 3:
return tf.einsum("ijk,kl->ijl", t_in, v_w) + v_b
else:
assert False
def apply(self, is_train, x, mask=None):
if self.key_mapper is not None:
with tf.variable_scope("map_keys"):
keys = self.key_mapper.apply(is_train, x, mask)
else:
keys = x
weights = tf.get_variable("weights", keys.shape.as_list()[-1], dtype=tf.float32,
initializer=get_keras_initialization(self.init))
dist = tf.tensordot(keys, weights, axes=[[2], [0]]) # (batch, x_words)
dist = exp_mask(dist, mask)
dist = tf.nn.softmax(dist)
out = tf.einsum("ajk,aj->ak", x, dist) # (batch, x_dim)
if self.post_process is not None:
with tf.variable_scope("post_process"):
out = self.post_process.apply(is_train, out)
return out
def apply(self, is_train, x, mask=None):
if self.key_mapper is not None:
with tf.variable_scope("map_keys"):
keys = self.key_mapper.apply(is_train, x, mask)
else:
keys = x
weights = tf.get_variable("weights", (keys.shape.as_list()[-1], self.n_encodings), dtype=tf.float32,
initializer=get_keras_initialization(self.init))
dist = tf.tensordot(keys, weights, axes=[[2], [0]]) # (batch, x_words, n_encoding)
if self.bias:
dist += tf.get_variable("bias", (1, 1, self.n_encodings),
dtype=tf.float32, initializer=tf.zeros_initializer())
if mask is not None:
bool_mask = tf.expand_dims(tf.cast(tf.sequence_mask(mask, tf.shape(x)[1]), tf.float32), 2)
dist = bool_mask * bool_mask + (1 - bool_mask) * VERY_NEGATIVE_NUMBER
dist = tf.nn.softmax(dist, dim=1)
out = tf.einsum("ajk,ajn->ank", x, dist) # (batch, n_encoding, feature)
if self.post_process is not None:
with tf.variable_scope("post_process"):
out = self.post_process.apply(is_train, out)
return out
def ntn(name, lhs, rhs, nr_output_channels,
use_bias=True, nonlin=__default_nonlin__,
W=None, b=None, param_dtype=__default_dtype__):
lhs, rhs= map(O.flatten2, [lhs, rhs])
assert lhs.static_shape[1] is not None and rhs.static_shape[1] is not None
W_shape = (lhs.static_shape[1], nr_output_channels, rhs.static_shape[1])
b_shape = (nr_output_channels, )
if W is None:
W = tf.contrib.layers.xavier_initializer()
W = O.ensure_variable('W', W, shape=W_shape, dtype=param_dtype)
if use_bias:
if b is None:
b = tf.constant_initializer()
b = O.ensure_variable('b', b, shape=b_shape, dtype=param_dtype)
out = tf.einsum('ia,abc,ic->ib', lhs.tft, W.tft, rhs.tft)
if use_bias:
out = tf.identity(out + b.add_axis(0), name='bias')
out = nonlin(out, name='nonlin')
return tf.identity(out, name='out')
def testFlatInnerTTTensbyTTTensBroadcasting(self):
# Inner product between two batch TT-tensors with broadcasting.
tt_1 = initializers.random_tensor_batch((2, 3, 4), batch_size=1)
tt_2 = initializers.random_tensor_batch((2, 3, 4), batch_size=3)
res_actual_1 = ops.flat_inner(tt_1, tt_2)
res_actual_2 = ops.flat_inner(tt_2, tt_1)
res_desired = tf.einsum('ijk,oijk->o', ops.full(tt_1[0]), ops.full(tt_2))
with self.test_session() as sess:
res = sess.run([res_actual_1, res_actual_2, res_desired])
res_actual_1_val, res_actual_2_val, res_desired_val = res
self.assertAllClose(res_actual_1_val, res_desired_val)
self.assertAllClose(res_actual_2_val, res_desired_val)
tt_1 = initializers.random_tensor_batch((2, 3, 4), batch_size=2)
with self.assertRaises(ValueError):
# The batch_sizes are different.
ops.flat_inner(tt_1, tt_2)
def testFullMatrix3d(self):
np.random.seed(1)
for rank in [1, 2]:
a = np.random.rand(3, 2, 3, rank).astype(np.float32)
b = np.random.rand(3, rank, 4, 5, rank).astype(np.float32)
c = np.random.rand(3, rank, 2, 2).astype(np.float32)
tt_cores = (a.reshape(3, 1, 2, 3, rank), b.reshape(3, rank, 4, 5, rank),
c.reshape(3, rank, 2, 2, 1))
# Basically do full by hand.
desired = np.einsum('oija,oaklb,obpq->oijklpq', a, b, c)
desired = desired.reshape((3, 2, 3, 4, 5, 2, 2))
desired = desired.transpose((0, 1, 3, 5, 2, 4, 6))
desired = desired.reshape((3, 2 * 4 * 2, 3 * 5 * 2))
with self.test_session():
tf_mat = TensorTrainBatch(tt_cores)
actual = ops.full(tf_mat)
self.assertAllClose(desired, actual.eval())
def create_model(self,
model_input,
vocab_size,
num_frames,
**unused_params):
shape = model_input.get_shape().as_list()
frames_sum = tf.reduce_sum(tf.abs(model_input),axis=2)
frames_true = tf.ones(tf.shape(frames_sum))
frames_false = tf.zeros(tf.shape(frames_sum))
frames_bool = tf.reshape(tf.where(tf.greater(frames_sum, frames_false), frames_true, frames_false),[-1,shape[1],1])
activation_1 = tf.reduce_max(model_input, axis=1)
activation_2 = tf.reduce_sum(model_input*frames_bool, axis=1)/(tf.reduce_sum(frames_bool, axis=1)+1e-6)
activation_3 = tf.reduce_min(model_input, axis=1)
model_input_1, final_probilities_1 = self.sub_moe(activation_1,vocab_size,scopename="_max")
model_input_2, final_probilities_2 = self.sub_moe(activation_2,vocab_size,scopename="_mean")
model_input_3, final_probilities_3 = self.sub_moe(activation_3,vocab_size,scopename="_min")
final_probilities = tf.stack((final_probilities_1,final_probilities_2,final_probilities_3),axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[shape[2], 3, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
activations = tf.stack((model_input_1, model_input_2, model_input_3), axis=2)
weight = tf.nn.softmax(tf.einsum("aij,ijk->ajk", activations, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1,vocab_size])
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters = [1024, 1024, 1024],
filter_sizes = [1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
cnn_output = slim.batch_norm(
cnn_output,
center=True,
scale=True,
is_training=FLAGS.train,
scope=sub_scope+"cluster_bn")
return cnn_output, max_frames
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
lstm_size = FLAGS.lstm_cells
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = self.rnn(cnn_output,lstm_size, num_frames,sub_scope="rnn%d"%(layer+1))
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale,vocab_size,scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
num_frames = tf.maximum(num_frames//pool_size,1)
final_probilities = tf.stack(final_probilities,axis=1)
moe_inputs = tf.stack(moe_inputs,axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1,vocab_size])
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters = [1024, 1024, 1024],
filter_sizes = [1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
cnn_output = slim.batch_norm(
cnn_output,
center=True,
scale=True,
is_training=FLAGS.train,
scope=sub_scope+"cluster_bn")
return cnn_output, max_frames
def create_model(self, model_input, vocab_size, num_frames, distill_labels=None, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
lstm_size = FLAGS.lstm_cells
pool_size = 2
cnn_input = model_input
cnn_size = FLAGS.cnn_cells
num_filters = [cnn_size, cnn_size, cnn_size*2]
filter_sizes = [1, 2, 3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = self.rnn(cnn_output,lstm_size, num_frames,sub_scope="rnn%d"%(layer+1))
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale,vocab_size,distill_labels=distill_labels, scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
num_frames = tf.maximum(num_frames//pool_size,1)
final_probilities = tf.stack(final_probilities,axis=1)
moe_inputs = tf.stack(moe_inputs,axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, lstm_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1,vocab_size])
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
lstm_size = FLAGS.lstm_cells
pool_size = 2
cnn_input = model_input
num_filters = [256, 256, 512]
filter_sizes = [1, 2, 3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = self.rnn_gate(cnn_output, lstm_size, num_frames, sub_scope="rnn%d"%(layer+1))
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale, vocab_size, scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
num_frames = tf.maximum(num_frames//pool_size,1)
final_probilities = tf.stack(final_probilities, axis=1)
moe_inputs = tf.stack(moe_inputs, axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1, vocab_size])
result["predictions"] = tf.reduce_mean(final_probilities, axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters=[1024,1024,1024],
filter_sizes=[1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
cnn_output = slim.batch_norm(
cnn_output,
center=True,
scale=True,
is_training=FLAGS.train,
scope=sub_scope+"cluster_bn")
return cnn_output, max_frames
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
lstm_size = FLAGS.lstm_cells
pool_size = 2
cnn_input = model_input
num_filters = [256, 256, 512]
filter_sizes = [1, 2, 3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = self.rnn_glu(cnn_output, lstm_size, num_frames, sub_scope="rnn%d"%(layer+1))
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale, vocab_size, scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
num_frames = tf.maximum(num_frames//pool_size,1)
final_probilities = tf.stack(final_probilities, axis=1)
moe_inputs = tf.stack(moe_inputs, axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1, vocab_size])
result["predictions"] = tf.reduce_mean(final_probilities, axis=1)
return result
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
lstm_size = FLAGS.lstm_cells
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = LstmMultiscaleModel().cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = LstmMultiscaleModel().rnn(cnn_output,lstm_size, num_frames,sub_scope="rnn%d"%(layer+1))
moe_inputs.append(cnn_multiscale)
final_probility = LstmMultiscaleModel().sub_moe(cnn_multiscale,vocab_size,scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
num_frames = tf.maximum(num_frames//pool_size,1)
final_probilities = tf.stack(final_probilities, axis=1)
moe_inputs = tf.stack(moe_inputs, axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", tf.stop_gradient(moe_inputs), weight2d), dim=1)
result = {}
result["predictions"] = tf.reduce_sum(tf.stop_gradient(final_probilities)*weight, axis=1)
return result
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = 10
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
if layer < 3:
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
else:
cnn_input = cnn_output
cnn_output, num_t = self.kmax(cnn_input, num_filters=features_size, filter_sizes=num_extend, sub_scope="kmax")
cnn_input = tf.reshape(cnn_output,[-1,features_size])
final_probilities = self.sub_moe(cnn_input,vocab_size)
final_probilities = tf.reshape(final_probilities,[-1,num_extend,vocab_size])
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", cnn_output, weight2d), dim=1)
result = {}
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters = [1024, 1024, 1024],
filter_sizes = [1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
cnn_output = slim.batch_norm(
cnn_output,
center=True,
scale=True,
is_training=FLAGS.train,
scope=sub_scope+"cluster_bn")
return cnn_output, max_frames
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = tf.reduce_max(cnn_output,axis=1)
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale,vocab_size,scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
final_probilities = tf.stack(final_probilities,axis=1)
moe_inputs = tf.stack(moe_inputs,axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1,vocab_size])
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = CnnKmaxModel().cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1), l2_penalty=0.0)
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = tf.reduce_max(cnn_output,axis=1)
moe_inputs.append(cnn_multiscale)
final_probility = CnnKmaxModel().sub_moe(cnn_multiscale,vocab_size,scopename="moe%d"%(layer+1), l2_penalty=0.0)
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
final_probilities = tf.stack(final_probilities,axis=1)
moe_inputs = tf.stack(moe_inputs,axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", tf.stop_gradient(moe_inputs), weight2d), dim=1)
result = {}
result["predictions"] = tf.reduce_sum(tf.stop_gradient(final_probilities)*weight, axis=1)
return result
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = tf.reduce_max(cnn_output,axis=1)
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale,vocab_size,scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
final_probilities = tf.stack(final_probilities,axis=1)
moe_inputs = tf.stack(moe_inputs,axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1,vocab_size])
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters=[1024, 1024, 1024],
filter_sizes=[1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
return cnn_output, max_frames
