Upload infer.py
Browse files
infer.py
ADDED
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import GPT2LMHeadModel, AutoTokenizer
|
2 |
+
import demo_util
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
from PIL import Image
|
6 |
+
import os
|
7 |
+
|
8 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
9 |
+
torch.manual_seed(0)
|
10 |
+
|
11 |
+
device = "cuda:1"
|
12 |
+
dtype = torch.float16
|
13 |
+
|
14 |
+
config = demo_util.get_config("configs/titok_l32.yaml")
|
15 |
+
titok_tokenizer = demo_util.get_titok_tokenizer(config)
|
16 |
+
titok_tokenizer = titok_tokenizer.to(device)
|
17 |
+
tokenizer = AutoTokenizer.from_pretrained("./image_tokenizer")
|
18 |
+
model = GPT2LMHeadModel.from_pretrained("./checkpoint-20000").to(device).to(dtype).eval()
|
19 |
+
|
20 |
+
def detokenize(tokens):
|
21 |
+
encoded_tokens = torch.from_numpy(np.array(tokens).astype(np.int64)).view(1, 1, -1).to(device)
|
22 |
+
reconstructed_image = titok_tokenizer.decode_tokens(encoded_tokens)
|
23 |
+
reconstructed_image = torch.clamp(reconstructed_image, 0.0, 1.0)
|
24 |
+
reconstructed_image = (reconstructed_image * 255.0).permute(0, 2, 3, 1).to("cpu", dtype=torch.uint8).numpy()[0]
|
25 |
+
return Image.fromarray(reconstructed_image)
|
26 |
+
|
27 |
+
prompt = ""
|
28 |
+
|
29 |
+
inputs = tokenizer(f"{text}<|startofimage|>", return_tensors="pt").to(device)
|
30 |
+
input_ids = inputs["input_ids"]
|
31 |
+
init = input_ids.shape[-1]
|
32 |
+
soi_token = tokenizer.encode("<|image:0|>")[0]
|
33 |
+
|
34 |
+
for _ in range(33):
|
35 |
+
logits = model(input_ids).logits[:, -1, :]
|
36 |
+
probas = torch.nn.functional.softmax(logits, dim=-1)
|
37 |
+
pred_idx = torch.argmax(probas, dim=-1, keepdim=True)
|
38 |
+
input_ids = torch.cat((input_ids, pred_idx), dim=-1)
|
39 |
+
tokenizer.decode(input_ids[0])
|
40 |
+
tokens = input_ids[:, init:-1].detach().cpu().squeeze().numpy() - soi_token
|
41 |
+
|
42 |
+
if np.any(tokens < 0) or np.any(tokens >= 4096):
|
43 |
+
print("Illegal Image Tokens")
|
44 |
+
|
45 |
+
else:
|
46 |
+
img = detokenize(tokens)
|
47 |
+
img.save(f"./out.png")
|