def __init__(self, config, *args, **kwargs):
if not hasattr(config, "vocab"):
raise AttributeError(
"config passed to the processor has no attribute vocab")
self.vocab = Vocab(*args, **config.vocab, **kwargs)
self._init_extras(config)
def __call__(self, item):
if not self._already_downloaded:
self.vocab = Vocab(*self._args, **self.config.vocab,
**self._kwargs)
self._already_downloaded = True
indices = super().__call__(item)["text"]
embeddings = torch.zeros(
(len(indices), self.vocab.get_embedding_dim()), dtype=torch.float)
for idx, index in enumerate(indices):
embeddings[idx] = self.vocab.vectors[index]
return {"text": embeddings}
class CaptionProcessor(BaseProcessor):
"""Processes a caption with start, end and pad tokens and returns raw string.
Args:
config (DictConfig): Configuration for caption processor.
"""
def __init__(self, config, *args, **kwargs):
if not hasattr(config, "vocab"):
raise AttributeError("config passed to the processor has no "
"attribute vocab")
self.vocab = Vocab(*args, **config.vocab, **kwargs)
def __call__(self, item):
for idx, v in enumerate(item):
if v == self.vocab.EOS_INDEX:
item = item[:idx]
break
tokens = [
self.vocab.get_itos()[w] for w in item if w not in
{self.vocab.SOS_INDEX, self.vocab.EOS_INDEX, self.vocab.PAD_INDEX}
]
caption = " ".join(tokens)
return {"tokens": tokens, "caption": caption}
class VocabEmbedding(nn.Module):
def __init__(self, embedding_dim, **vocab_params):
super().__init__()
self.vocab = Vocab(**vocab_params)
self.module = self.vocab.get_embedding(nn.Embedding,
embedding_dim=embedding_dim)
def forward(self, x):
return self.module(x)
class GloVeProcessor(VocabProcessor):
"""Inherits VocabProcessor, and returns GloVe vectors for each of the
words. Maps them to index using vocab processor, and then gets GloVe vectors
corresponding to those indices.
Args:
config (DictConfig): Configuration parameters for GloVe same as
:func:`~VocabProcessor`.
"""
def __init__(self, config, *args, **kwargs):
if not hasattr(config, "vocab"):
raise AttributeError(
"Config passed to the processor has no attribute vocab")
vocab_processor_config = copy.deepcopy(config)
# GloVeProcessor needs vocab type to be "intersected"
vocab_processor_config.vocab.type = "intersected"
if "vocab_file" not in vocab_processor_config.vocab:
warnings.warn("'vocab_file' key is not present in the config."
" Switching to pretrained vocab.")
vocab_processor_config.vocab.type = "pretrained"
self._init_extras(vocab_processor_config)
self.config = vocab_processor_config
self._already_downloaded = False
self._args = args
self._kwargs = kwargs
def __call__(self, item):
if not self._already_downloaded:
self.vocab = Vocab(*self._args, **self.config.vocab,
**self._kwargs)
self._already_downloaded = True
indices = super().__call__(item)["text"]
embeddings = torch.zeros(
(len(indices), self.vocab.get_embedding_dim()), dtype=torch.float)
for idx, index in enumerate(indices):
embeddings[idx] = self.vocab.vectors[index]
return {"text": embeddings}
class VocabProcessor(BaseProcessor):
"""Use VocabProcessor when you have vocab file and you want to process
words to indices. Expects UNK token as "<unk>" and pads sentences using
"<pad>" token. Config parameters can have ``preprocessor`` property which
is used to preprocess the item passed and ``max_length`` property which
points to maximum length of the sentence/tokens which can be convert to
indices. If the length is smaller, the sentence will be padded. Parameters
for "vocab" are necessary to be passed.
**Key**: vocab
Example Config::
dataset_config:
vqa2:
data_dir: ${env.data_dir}
processors:
text_processor:
type: vocab
params:
max_length: 14
vocab:
type: intersected
embedding_name: glove.6B.300d
vocab_file: vqa2/defaults/extras/vocabs/vocabulary_100k.txt
Args:
config (DictConfig): node containing configuration parameters of
the processor
Attributes:
vocab (Vocab): Vocab class object which is abstraction over the vocab
file passed.
"""
MAX_LENGTH_DEFAULT = 50
PAD_TOKEN = "<pad>"
PAD_INDEX = 0
def __init__(self, config, *args, **kwargs):
if not hasattr(config, "vocab"):
raise AttributeError(
"config passed to the processor has no attribute vocab")
self.vocab = Vocab(*args, **config.vocab, **kwargs)
self._init_extras(config)
def _init_extras(self, config, *args, **kwargs):
self.preprocessor = None
if hasattr(config, "max_length"):
self.max_length = config.max_length
else:
warnings.warn("No 'max_length' parameter in Processor's "
"configuration. Setting to {}.".format(
self.MAX_LENGTH_DEFAULT))
self.max_length = self.MAX_LENGTH_DEFAULT
if "preprocessor" in config:
self.preprocessor = Processor(config.preprocessor, *args, **kwargs)
if self.preprocessor is None:
raise ValueError(
f"No text processor named {config.preprocessor} is defined."
)
def __call__(self, item):
"""Call requires item to have either "tokens" attribute or either
"text" attribute. If "text" is present, it will tokenized using
the preprocessor.
Args:
item (Dict): Dict containing the "text" or "tokens".
Returns:
Dict: Dict containing indices in "text" key, "tokens" in "tokens"
key and "length" of the string in "length" key.
"""
indices = None
if not isinstance(item, dict):
raise TypeError("Argument passed to the processor must be "
"a dict with either 'text' or 'tokens' as "
"keys")
if "tokens" in item:
tokens = item["tokens"]
indices = self._map_strings_to_indices(item["tokens"])
elif "text" in item:
if self.preprocessor is None:
raise AssertionError("If tokens are not provided, a text "
"processor must be defined in the config")
tokens = self.preprocessor({"text": item["text"]})["text"]
indices = self._map_strings_to_indices(tokens)
else:
raise AssertionError("A dict with either 'text' or 'tokens' keys "
"must be passed to the processor")
tokens, length = self._pad_tokens(tokens)
return {"text": indices, "tokens": tokens, "length": length}
def _pad_tokens(self, tokens):
padded_tokens = [self.PAD_TOKEN] * self.max_length
token_length = min(len(tokens), self.max_length)
padded_tokens[:token_length] = tokens[:token_length]
token_length = torch.tensor(token_length, dtype=torch.long)
return padded_tokens, token_length
def get_pad_index(self):
"""Get index of padding <pad> token in vocabulary.
Returns:
int: index of the padding token.
"""
return self.vocab.get_pad_index()
def get_vocab_size(self):
"""Get size of the vocabulary.
Returns:
int: size of the vocabulary.
"""
return self.vocab.get_size()
def _map_strings_to_indices(self, tokens):
length = min(len(tokens), self.max_length)
tokens = tokens[:length]
output = torch.zeros(self.max_length, dtype=torch.long)
output.fill_(self.vocab.get_pad_index())
for idx, token in enumerate(tokens):
output[idx] = self.vocab.stoi[token]
return output
def __init__(self, embedding_dim, **vocab_params):
super().__init__()
self.vocab = Vocab(**vocab_params)
self.module = self.vocab.get_embedding(nn.Embedding,
embedding_dim=embedding_dim)