-
Notifications
You must be signed in to change notification settings - Fork 0
/
example_language.py
74 lines (54 loc) · 2.28 KB
/
example_language.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
import os
import torch
import numpy as np
from perceiver_io.language_perceiver import LanguagePerceiver
from utils.bytes_tokenizer import BytesTokenizer
def pad(max_sequence_length: int, inputs, input_mask, pad_token):
input_len = inputs.shape[1]
assert input_len <= max_sequence_length
pad_len = max_sequence_length - input_len
padded_inputs = np.pad(
inputs,
pad_width=((0, 0), (0, pad_len)),
constant_values=pad_token)
padded_mask = np.pad(
input_mask,
pad_width=((0, 0), (0, pad_len)),
constant_values=0)
return padded_inputs, padded_mask
def language_example():
MAX_SEQ_LEN = 2048
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
tokenizer = BytesTokenizer()
perceiver = LanguagePerceiver(vocab_size=tokenizer.vocab_size)
perceiver.to(device)
perceiver.eval()
ckpt_file = "./pytorch_checkpoints/language_perceiver_io_bytes.pth"
# check if file exists
if not os.path.isfile(ckpt_file):
raise ValueError("Please download the model checkpoint and place it in /pytorch_checkpoints")
checkpoint = torch.load(ckpt_file, map_location=device)
perceiver.load_state_dict(checkpoint["model_state_dict"])
input_str = "This is an incomplete sentence where some words are missing."
input_tokens = tokenizer.to_int(input_str)
# Mask " missing.". Note that the model performs much better if the masked chunk
# starts with a space.
input_tokens[51:60] = tokenizer.mask_token
print("Tokenized string without masked bytes:")
print(tokenizer.to_string(input_tokens))
# Pad and reshape inputs
inputs = input_tokens[None]
input_mask = np.ones_like(inputs)
inputs, input_mask = pad(MAX_SEQ_LEN, inputs, input_mask, tokenizer.pad_token)
inputs = torch.from_numpy(inputs).to(device)
input_mask = torch.from_numpy(input_mask).bool().to(device)
# Predict
with torch.inference_mode():
out = perceiver(inputs, input_masks=input_mask)
masked_tokens_predictions = out[0, 51:60].argmax(axis=-1)
print("Greedy predictions:")
print(masked_tokens_predictions)
print("Predicted string:")
print(tokenizer.to_string(masked_tokens_predictions.cpu().numpy()))
if __name__ == "__main__":
language_example()