def __init__(self, mode, config):
self._mode = mode
self.config = config
if self._mode != Mode.Predict:
self.input = tf.placeholder(name='inputs', dtype=tf.int32, shape=[None, None])
self.pos_input = tf.placeholder(name='pos_input', dtype=tf.int32, shape=[None])
self.morph_input = tf.placeholder(name='morph_input', dtype=tf.int32, shape=[None, config.max_morph_tags])
else:
k, v = list(zip(*config.dec_char_numberer.value2num.items()))
i = KeyValueTensorInitializer(v, k,key_dtype=tf.int64,value_dtype=tf.string)
self.dec2char = HashTable(i, default_value="<UNK>")
self.enc_chars = self.lookup(config.enc_char_numberer)
self.morph_tags = self.lookup(config.morph_numberer)
self.pos_tags = self.lookup(config.pos_numberer)
self.input_plc =tf.placeholder(name='inputs', dtype=tf.string, shape=[None, None])
self.pos_plc = tf.placeholder(name='pos_input', dtype=tf.string, shape=[None])
self.morph_plc = tf.placeholder(name='morph_input', dtype=tf.string, shape=[None, config.max_morph_tags])
self.input = self.enc_chars.lookup(self.input_plc)
self.pos_input = self.pos_tags.lookup(self.pos_plc)
self.morph_input = self.morph_tags.lookup(self.morph_plc)
self.input_lengths = tf.placeholder(name='input_lengths', dtype=tf.int32, shape=[None])
self.enc_character_embeddings = tf.get_variable('enc_character_embeddings',
shape=[self.config.enc_vocab_size,
self.config.char_embedding_size])
self.dec_character_embeddings = tf.get_variable('dec_character_embeddings',
shape=[self.config.dec_vocab_size,
self.config.char_embedding_size])
self.pos_embeddings = tf.get_variable('pos_embeddings',
shape=[self.config.pos_vocab_size,
self.config.pos_embedding_size])
self.morph_embeddings = tf.get_variable('morph_embeddings',
shape=[self.config.morph_vocab_size,
self.config.morph_embedding_size])
if self._mode != Mode.Train:
self.start_tokens = tf.tile([config.start_idx], [tf.shape(self.input)[0]])
if self._mode != Mode.Predict:
# length +2 since lengths are <bow> + word + <eow>
self.dec_input = tf.placeholder(name='dec_in', dtype=tf.int32, shape=[None, None])
# exclude start token from targets for loss-computations since we feed the start token and don't want to
# predict it
self.decoder_targets = self.dec_input[:, 1:]
self.dec_lengths = tf.placeholder(name='dec_lengths', dtype=tf.int32, shape=[None])
def __init__(self, returns_dict=False):
embeddings = [
("", [0, 0, 0, 0]), # OOV items are mapped to this embedding.
("hello world", [1, 2, 3, 4]),
("pair-programming", [5, 5, 5, 5]),
]
keys = tf.constant([item[0] for item in embeddings], dtype=tf.string)
indices = tf.constant(list(range(len(embeddings))), dtype=tf.int64)
tbl_init = KeyValueTensorInitializer(keys, indices)
self.table = HashTable(tbl_init, 0)
self.weights = tf.Variable(
list([item[1] for item in embeddings]), dtype=tf.float32)
self.variables = [self.weights]
self.trainable_variables = self.variables
self._returns_dict = returns_dict
def text_module_fn():
embeddings = [
("", [0, 0, 0, 0]), # OOV items are mapped to this embedding.
("hello world", [1, 2, 3, 4]),
("pair-programming", [5, 5, 5, 5]),
]
keys = tf.constant([item[0] for item in embeddings], dtype=tf.string)
indices = tf.constant(list(range(len(embeddings))), dtype=tf.int64)
tbl_init = KeyValueTensorInitializer(keys, indices)
table = HashTable(tbl_init, 0)
weights_initializer = tf.cast(
tf.constant(list([item[1] for item in embeddings])), tf.float32)
weights = tf_v1.get_variable(
"weights", dtype=tf.float32, initializer=weights_initializer)
text_tensor = tf_v1.placeholder(dtype=tf.string, name="text", shape=[None])
indices_tensor = table.lookup(text_tensor)
embedding_tensor = tf.gather(weights, indices_tensor)
hub.add_signature(inputs=text_tensor, outputs=embedding_tensor)
def input_fn():
with tf.variable_scope("input_fn"), tf.device("/cpu:0"):
caption_dataset = Dataset.from_tensor_slices(list(captions))
filename_dataset = Dataset.from_tensor_slices(list(filenames))
table_init = KeyValueTensorInitializer(
list(self.word_to_idx.keys()),
list(self.word_to_idx.values()),
key_dtype=tf.string,
value_dtype=tf.int32)
table = HashTable(table_init, default_value=0)
def split_sentence(sentence):
words = tf.string_split(tf.reshape(sentence, (1, ))).values
words = tf.concat([
tf.constant(["<START>"]), words,
tf.constant(["<END>"])
],
axis=0)
return table.lookup(words)
index_dataset = caption_dataset.map(split_sentence,
num_threads=8)
def decode_image(filename):
image = tf.image.decode_jpeg(tf.read_file(filename),
channels=3)
# image = tf.image.resize_images(image, [224, 224])
image = tf.to_float(image)
return image
image_dataset = filename_dataset.map(decode_image,
num_threads=8)
caption_structure = {
"raw": caption_dataset,
"index": index_dataset
}
return image_dataset, caption_structure
def lookup(self, numberer):
k, v = list(zip(*numberer.value2num.items()))
i = KeyValueTensorInitializer(k, v)
return HashTable(i, default_value=numberer.unknown_idx)