Spaces:
Runtime error
Runtime error
Create process.py
Browse files- process.py +111 -0
process.py
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import io
|
| 2 |
+
import spaces
|
| 3 |
+
import argparse
|
| 4 |
+
import numpy as np
|
| 5 |
+
import torch
|
| 6 |
+
from decord import cpu, VideoReader, bridge
|
| 7 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 8 |
+
from transformers import BitsAndBytesConfig
|
| 9 |
+
|
| 10 |
+
MODEL_PATH = "THUDM/cogvlm2-llama3-caption"
|
| 11 |
+
DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
|
| 12 |
+
TORCH_TYPE = torch.bfloat16 if torch.cuda.is_available() and torch.cuda.get_device_capability()[
|
| 13 |
+
0] >= 8 else torch.float16
|
| 14 |
+
|
| 15 |
+
parser = argparse.ArgumentParser(description="Apollo tyre delay reasoning")
|
| 16 |
+
parser.add_argument('--quant', type=int, choices=[4, 8], help='Enable 4-bit or 8-bit precision loading', default=4)
|
| 17 |
+
args = parser.parse_args([])
|
| 18 |
+
|
| 19 |
+
def load_video(video_data, strategy='chat'):
|
| 20 |
+
bridge.set_bridge('torch')
|
| 21 |
+
mp4_stream = video_data
|
| 22 |
+
num_frames = 24
|
| 23 |
+
decord_vr = VideoReader(io.BytesIO(mp4_stream), ctx=cpu(0))
|
| 24 |
+
frame_id_list = None
|
| 25 |
+
total_frames = len(decord_vr)
|
| 26 |
+
|
| 27 |
+
if strategy == 'base':
|
| 28 |
+
clip_end_sec = 60
|
| 29 |
+
clip_start_sec = 0
|
| 30 |
+
start_frame = int(clip_start_sec * decord_vr.get_avg_fps())
|
| 31 |
+
end_frame = min(total_frames,
|
| 32 |
+
int(clip_end_sec * decord_vr.get_avg_fps())) if clip_end_sec is not None else total_frames
|
| 33 |
+
frame_id_list = np.linspace(start_frame, end_frame - 1, num_frames, dtype=int)
|
| 34 |
+
elif strategy == 'chat':
|
| 35 |
+
timestamps = decord_vr.get_frame_timestamp(np.arange(total_frames))
|
| 36 |
+
timestamps = [i[0] for i in timestamps]
|
| 37 |
+
max_second = round(max(timestamps)) + 1
|
| 38 |
+
frame_id_list = []
|
| 39 |
+
for second in range(max_second):
|
| 40 |
+
closest_num = min(timestamps, key=lambda x: abs(x - second))
|
| 41 |
+
index = timestamps.index(closest_num)
|
| 42 |
+
frame_id_list.append(index)
|
| 43 |
+
if len(frame_id_list) >= num_frames:
|
| 44 |
+
break
|
| 45 |
+
|
| 46 |
+
video_data = decord_vr.get_batch(frame_id_list)
|
| 47 |
+
video_data = video_data.permute(3, 0, 1, 2)
|
| 48 |
+
return video_data
|
| 49 |
+
|
| 50 |
+
# Configure quantization
|
| 51 |
+
quantization_config = BitsAndBytesConfig(
|
| 52 |
+
load_in_4bit=True,
|
| 53 |
+
bnb_4bit_compute_dtype=TORCH_TYPE,
|
| 54 |
+
bnb_4bit_use_double_quant=True,
|
| 55 |
+
bnb_4bit_quant_type="nf4"
|
| 56 |
+
)
|
| 57 |
+
|
| 58 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
| 59 |
+
MODEL_PATH,
|
| 60 |
+
trust_remote_code=True,
|
| 61 |
+
)
|
| 62 |
+
|
| 63 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 64 |
+
MODEL_PATH,
|
| 65 |
+
torch_dtype=TORCH_TYPE,
|
| 66 |
+
trust_remote_code=True,
|
| 67 |
+
quantization_config=quantization_config,
|
| 68 |
+
device_map="auto"
|
| 69 |
+
).eval()
|
| 70 |
+
|
| 71 |
+
@spaces.GPU
|
| 72 |
+
def predict(prompt, video_data, temperature):
|
| 73 |
+
strategy = 'chat'
|
| 74 |
+
video = load_video(video_data, strategy=strategy)
|
| 75 |
+
history = []
|
| 76 |
+
query = prompt
|
| 77 |
+
inputs = model.build_conversation_input_ids(
|
| 78 |
+
tokenizer=tokenizer,
|
| 79 |
+
query=query,
|
| 80 |
+
images=[video],
|
| 81 |
+
history=history,
|
| 82 |
+
template_version=strategy
|
| 83 |
+
)
|
| 84 |
+
|
| 85 |
+
inputs = {
|
| 86 |
+
'input_ids': inputs['input_ids'].unsqueeze(0).to(DEVICE),
|
| 87 |
+
'token_type_ids': inputs['token_type_ids'].unsqueeze(0).to(DEVICE),
|
| 88 |
+
'attention_mask': inputs['attention_mask'].unsqueeze(0).to(DEVICE),
|
| 89 |
+
'images': [[inputs['images'][0].to(DEVICE).to(TORCH_TYPE)]],
|
| 90 |
+
}
|
| 91 |
+
|
| 92 |
+
gen_kwargs = {
|
| 93 |
+
"max_new_tokens": 2048,
|
| 94 |
+
"pad_token_id": 128002,
|
| 95 |
+
"top_k": 1,
|
| 96 |
+
"do_sample": False,
|
| 97 |
+
"top_p": 0.1,
|
| 98 |
+
"temperature": temperature,
|
| 99 |
+
}
|
| 100 |
+
|
| 101 |
+
with torch.no_grad():
|
| 102 |
+
outputs = model.generate(**inputs, **gen_kwargs)
|
| 103 |
+
outputs = outputs[:, inputs['input_ids'].shape[1]:]
|
| 104 |
+
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 105 |
+
return response
|
| 106 |
+
|
| 107 |
+
def inference(video, prompt):
|
| 108 |
+
temperature = 0.8
|
| 109 |
+
video_data = open(video, 'rb').read()
|
| 110 |
+
response = predict(prompt, video_data, temperature)
|
| 111 |
+
return response
|