def def_model():
qnd.add_flag("hidden_layer_size",
type=int,
default=64,
help="Hidden layer size")
def model(image, number=None, mode=None):
h = tf.contrib.layers.fully_connected(image,
qnd.FLAGS.hidden_layer_size)
h = tf.contrib.layers.fully_connected(h, 10, activation_fn=None)
predictions = tf.argmax(h, axis=1)
if mode == tf.contrib.learn.ModeKeys.INFER:
return predictions
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=number,
logits=h))
return predictions, loss, minimize(loss), {
"accuracy":
tf.contrib.metrics.streaming_accuracy(predictions, number)[1],
}
return model
def add_flags():
qnd.add_flag("regularization_scale", type=float, default=1e-8)
adder = qnd.FlagAdder()
adder.add_flag("word_embedding_size", type=int, default=100)
adder.add_flag("sentence_embedding_size", type=int, default=100)
adder.add_flag("document_embedding_size", type=int, default=100)
adder.add_flag("context_vector_size", type=int, default=100)
return adder
def add_flags():
qnd.add_flag("regularization_scale", type=float, default=1e-8)
adder = qnd.FlagAdder()
adder.add_flag("word_embedding_size", type=int, default=100)
adder.add_flag("sentence_embedding_size", type=int, default=100)
adder.add_flag("document_embedding_size", type=int, default=100)
adder.add_flag("context_vector_size", type=int, default=100)
return adder
def def_ar_lm():
qnd.add_flag('cell_size', type=int, default=128)
qnd.add_flag('num_unroll', type=int, default=16)
qnd.add_flag('batch_size', type=int, default=64)
qnd.add_flag('num_batch_threads', type=int, default=os.cpu_count())
qnd.add_flag('batch_queue_capacity', type=int, default=1024)
def ar_lm(key, sentence, labels, *, char_embeddings):
cell = tf.contrib.rnn.LayerNormBasicLSTMCell(qnd.FLAGS.cell_size)
batch = tf.contrib.training.batch_sequences_with_states(
key,
input_sequences={
'sentence': tf.gather(char_embeddings, sentence),
'labels': labels,
},
input_context={},
input_length=None,
initial_states={
'c': tf.zeros([cell.state_size.c], tf.float32),
'h': tf.zeros([cell.state_size.h], tf.float32),
},
num_unroll=qnd.FLAGS.num_unroll,
batch_size=qnd.FLAGS.batch_size,
num_threads=qnd.FLAGS.num_batch_threads,
capacity=qnd.FLAGS.batch_queue_capacity)
outputs, _ = tf.nn.state_saving_rnn(cell,
tf.unstack(
batch.sequences['sentence'],
axis=1),
sequence_length=batch.length,
state_saver=batch,
state_name=('c', 'h'))
logits = batch_linear(outputs, ex.static_shape(char_embeddings)[0])
labels = batch.sequences['labels']
loss = sequence_labeling_loss(logits, labels, batch.length)
return (
{
'key':
key,
'labels': (tf.argmax(logits, axis=2) *
tf.sequence_mask(batch.length, dtype=tf.int64)),
},
loss,
ex.minimize(loss),
)
return ar_lm
def def_char_lm():
get_chars = qndex.nlp.def_chars()
qnd.add_flag('char_embedding_size', type=int, default=100)
ar_lm = def_ar_lm()
def char_lm(key, sentence, labels):
return ar_lm(key,
sentence,
labels,
char_embeddings=ex.embeddings(
id_space_size=len(get_chars()),
embedding_size=qnd.FLAGS.char_embedding_size,
name='char_embeddings'))
return char_lm
def add_flags():
adder = add_child_flags()
adder.add_flag("dropout_keep_prob", type=float, default=0.5)
adder.add_flag("nums_of_cnn_channels",
type=argtyp.int_list,
default=[32] * 4)
adder.add_flag("nums_of_attention_cnn_channels",
type=argtyp.int_list,
default=[32] * 3)
qnd.add_required_flag("font_file")
qnd.add_flag("font_size", type=int, default=32)
qnd.add_flag("save_font_array_file")
return adder
def add_flags():
adder = add_child_flags()
adder.add_flag("dropout_keep_prob", type=float, default=0.5)
adder.add_flag("nums_of_cnn_channels",
type=argtyp.int_list,
default=[32] * 4)
adder.add_flag("nums_of_attention_cnn_channels",
type=argtyp.int_list,
default=[32] * 3)
qnd.add_required_flag("font_file")
qnd.add_flag("font_size", type=int, default=32)
qnd.add_flag("save_font_array_file")
return adder
def add_flags():
adder = add_child_flags()
adder.add_flag("dropout_keep_prob", type=float, default=0.5)
qnd.add_required_flag("font_file")
qnd.add_flag("font_size", type=int, default=32)
qnd.add_flag("save_font_array_file")
qnd.add_flag("regularization_scale", type=float, default=1e-8)
return adder