csarron commited on
Commit
b27ebb2
1 Parent(s): 085a591
Files changed (1) hide show
  1. README.md +56 -0
README.md ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ see [read_pyarrow.py](https://gist.github.com/csarron/df712e53c9e0dcaad4eb6843e7a3d51c#file-read_pyarrow-py) for how to read one pyarrow file.
2
+
3
+ example PyTorch dataset:
4
+
5
+ ```python
6
+ from torch.utils.data import Dataset
7
+
8
+ class ImageCaptionArrowDataset(Dataset):
9
+ def __init__(
10
+ self,
11
+ dataset_file,
12
+ tokenizer,
13
+ ):
14
+
15
+ import pyarrow as pa
16
+
17
+ data = [pa.ipc.open_file(pa.memory_map(f, "rb")).read_all() for f in glob.glob(dataset_file)]
18
+ self.data = pa.concat_tables(data)
19
+ # do other initialization, like init image preprocessing fn,
20
+
21
+ def __getitem__(self, index):
22
+ # item_id = self.data["id"][index].as_py()
23
+ text = self.data["text"][index].as_py() # get text
24
+ if isinstance(text, list):
25
+ text = random.choice(text)
26
+
27
+ img_bytes = self.data["image"][index].as_py() # get image bytes
28
+
29
+ # do some processing with image and text, return the features
30
+
31
+
32
+ # img_feat = self.image_bytes_to_tensor(img_bytes)
33
+ # inputs = self.tokenizer(
34
+ # text,
35
+ # padding="max_length",
36
+ # max_length=self.max_text_len,
37
+ # truncation=True,
38
+ # return_token_type_ids=True,
39
+ # return_attention_mask=True,
40
+ # add_special_tokens=True,
41
+ # return_tensors="pt",
42
+ # )
43
+ # input_ids = inputs.input_ids.squeeze(0)
44
+ # attention_mask = inputs.attention_mask.squeeze(0)
45
+ # return {
46
+ # # "item_ids": item_id,
47
+ # "text_ids": input_ids,
48
+ # "input_ids": input_ids,
49
+ # "text_masks": attention_mask,
50
+ # "pixel_values": img_feat,
51
+ # }
52
+ def __len__(self):
53
+ return len(self.data)
54
+
55
+
56
+ ```