3D-GRAND / model.py
jedyang97's picture
fix zero gpu
7cb0f3c
raw
history blame
2.46 kB
# model.py
import spaces
import json
import torch
from torch.utils.data import DataLoader
from llava.dataset.obj_identifier_dataset import (
ObjIdentifierDataset,
DataCollatorForBatchDecodingObjIdentifierDataset,
)
from llava.mm_utils import get_model_name_from_path
from llava.model.builder import load_pretrained_model
from llava.utils import disable_torch_init
def load_model_and_dataloader(model_path, model_base, scene_to_obj_mapping, obj_context_feature_type="text", load_8bit=False, load_4bit=False, load_bf16=False, device_map='auto'):
model_name = get_model_name_from_path(model_path)
tokenizer, model, image_processor, context_len = load_pretrained_model(
model_path=model_path,
model_base=model_base,
model_name=model_name,
load_8bit=load_8bit,
load_4bit=load_4bit,
load_bf16=load_bf16,
device_map=device_map,
)
dataset = ObjIdentifierDataset(
tokenizer=tokenizer,
data_path=[],
scene_to_obj_mapping=scene_to_obj_mapping,
obj_context_feature_type=obj_context_feature_type,
mode="generate",
)
collator = DataCollatorForBatchDecodingObjIdentifierDataset(tokenizer=tokenizer)
data_loader = DataLoader(
dataset, collate_fn=collator, batch_size=1, num_workers=0, shuffle=False
)
return tokenizer, model, data_loader
def get_model_response(model, tokenizer, data_loader, scene_id, user_input, max_new_tokens=50, temperature=0.2, top_p=0.9):
input_data = [
{
"id": f"interactive@{scene_id}@input",
"scene_id": scene_id,
"conversations": [{"from": "human", "value": f"Ground these sentences: <refer_expression>{user_input}<refer_expression>\n<image>"}],
"ground_truth": "User provided description for interactive session.",
}
]
data_loader.dataset.update_data(input_data)
for batch in data_loader:
input_ids = batch["input_ids"].squeeze(dim=1).cuda()
with torch.inference_mode():
output_ids = model.generate(
input_ids,
max_new_tokens=max_new_tokens,
temperature=temperature,
top_p=top_p,
use_cache=True,
)
prompt = tokenizer.batch_decode(input_ids)[0]
outputs = tokenizer.batch_decode(output_ids[:, input_ids.shape[-1]:], skip_special_tokens=True)
return prompt, outputs[0]