def __getitem__(self, index, raw=False):
idx = index / 5
idx_cap = index % 5
path = self.content[int(idx)][0]
target = self.content[int(idx)][1][idx_cap]
if raw:
return path, target
img = Image.open(os.path.join(self.root, path)).convert('RGB')
if self.transform is not None:
img = self.transform(img)
#target = encode_sentence(target, self.params, self.dico)
if self.word2id is None:
if self.bpe:
target = bpe_encode(target, self.embed)
else:
target = encode_sentence_fasttext(target, self.embed)
else:
target = encode_sentence(target, self.embed, self.word2id)
return img, target
def __init__(
self,
root=path["COCO_ROOT"],
coco_json_file_path=path["COCO_RESTVAL_SPLIT"],
sset="train",
transform=None,
embed_type="bin",
embed_file="/data/m.portaz/wiki.en.bin",
embed_size=300,
):
self.root = os.path.join(root, "images/")
self.transform = transform
with open(coco_json_file_path, "r") as f:
datas = json.load(f)
if sset == "train":
self.content = [x for x in datas["images"] if x["split"] == "train"]
elif sset == "trainrv":
self.content = [
x
for x in datas["images"]
if x["split"] == "train" or x["split"] == "restval"
]
elif sset == "val":
self.content = [x for x in datas["images"] if x["split"] == "val"]
else:
self.content = [x for x in datas["images"] if x["split"] == "test"]
self.content = [
(
os.path.join(y["filepath"], y["filename"]),
[x["raw"] for x in y["sentences"]],
)
for y in self.content
]
self.word2id = None
self.bert = False
if embed_type == "bin":
print("Using binary file dictionary")
self.tokenizer = lambda x: encode_sentence_fasttext(x,fastText.load_model(embed_file))
elif embed_type == "multi" or embed_type == "bivec":
print("Using .vec file")
embed, _, word2id = load_vec(embed_file)
self.tokenizer = lambda x: encode_sentence(x, embed, word2id)
elif embed_type == "bert":
print("Using bert embeddings")
self.bertTokenizer = BertTokenizer.from_pretrained(
"bert-base-uncased"
)
self.tokenizer = lambda x: torch.LongTensor(self.bertTokenizer.encode(text=x))
elif embed_type == "xlm":
print("Using XLM embeddings")
self.xlmTokens = XLMTokenizer.from_pretrained("xlm-mlm-en-2048")
self.tokenizer = lambda x: torch.LongTensor(self.xlmTokens.encode(text=x))
else:
print("Unknown embed type :", embed_type)
print("Must be : bin, multi or xlm")
sys.exit(0)
def __getitem__(self, index):
#return self.sentences[index]
if self.fastText:
return encode_sentence_fasttext(
self.sentences[index][0], self.embed), self.sentences[index][1]
else:
return encode_sentence(self.sentences[index][0],
self.embed[0],
self.embed[2],
tokenize=False), self.sentences[index][1]
def __getitem__(self, index, raw=False):
path = self.imList[int(index)]
target = self.capList[int(index)]
img = Image.open(path).convert('RGB')
img = self.transform(img)
target = encode_sentence_fasttext(target, self.embed, False)
return img, target
def embedFile(args):
print("Compute embeddings for the file:", args.file)
embed = fastText.load_model(args.dict)
nl = file_len(args.file)
fout = open(args.output, 'w')
with open(args.file) as f:
for i, line in enumerate(f):
line = line.rstrip()
print("%2.2f" % (i / nl * 100.0), "\%", end='\r')
i, t = line.split("\t")
es = encode_sentence_fasttext(t.split(' '), embed, tokenize=False)
fout.write(i + '\t' + str(es) + '\n')
def __init__(self, text_path, word_dict="/data/m.portaz/wiki.multi.en.vec"):
lines = open(text_path).readlines()
self.sent_list = [line.rstrip("\n") for line in lines]
self.fastText = False
if word_dict[-4:] == ".vec":
emb, _, wordsID = load_vec("/data/m.portaz/wiki.multi.en.vec")
self.encoding = lambda x: encode_sentence(x, emb, wordsID, tokenize=False)
elif word_dict[-4:] == ".bin":
emb = fastText.load_model(word_dict)
self.encoding = lambda x: encode_sentence_fasttext(x, emb)
self.fastText = True
else:
print("Error : unknown dictionary type : ", word_dict)
sys.exit(1)