tools/llava.py

import torch
import torch.nn as nn
import sys

# Add the folder to sys.path
sys.path.append("/gpfs/home/ym621/UniPointMap")
import open_clip
from transformers import LlavaForConditionalGeneration, AutoProcessor, AutoTokenizer, AutoConfig


import torch.nn as nn
from transformers.activations import GELUActivation

class LlavaMultiModalProjector(nn.Module):
    def __init__(self, in_features: int, out_features: int):
        super().__init__()
        self.linear_1 = nn.Linear(in_features, out_features, bias=True)
        self.act = GELUActivation()
        self.linear_2 = nn.Linear(out_features, out_features, bias=True)

    def forward(self, x):
        x = self.linear_1(x)
        x = self.act(x)
        x = self.linear_2(x)
        return x
    
# ---------------------------------------------------------
# 1. Load OpenCLIP Vision Encoder (ViT-B/32)
# ---------------------------------------------------------
vision_model, _, preprocess = open_clip.create_model_and_transforms(
    "ViT-B-32", pretrained="openai"   # choices: openai, laion2b, laion400m, etc.
)

# Keep only the vision tower
custom_vision_tower = vision_model.visual
custom_hidden_size = custom_vision_tower.output_dim   # ViT-B/32 = 512

# ---------------------------------------------------------
# 2. Load Llava Model (base LLaVA checkpoint)
# ---------------------------------------------------------
model_id = "liuhaotian/llava-v1.5-7b"   # or 13b
model = LlavaForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True,
    device_map="auto"
)
breakpoint()

# ---------------------------------------------------------
# 3. Replace Vision Tower
# ---------------------------------------------------------
# Replace with OpenCLIP ViT-B/32
model.vision_tower = custom_vision_tower

# Fix mm_projector to map vision_dim -> LLM hidden size
llm_hidden_size = model.config.hidden_size   # usually 4096 for LLaVA-7B
model.multi_modal_projector = LlavaMultiModalProjector(
    in_features=512,
    out_features=4096   # match LLM hidden size
)


tokenizer = AutoTokenizer.from_pretrained(model_id)
config = AutoConfig.from_pretrained("liuhaotian/llava-v1.5-7b")

# ---------------------------------------------------------
# 4. Update Config
# ---------------------------------------------------------
model.config.vision_config.hidden_size = custom_hidden_size
model.config.vision_config.image_size = getattr(custom_vision_tower, "image_size", 224)  # ViT-B/32 default
model.config.vision_config.patch_size = getattr(custom_vision_tower, "patch_size", 32)

# ---------------------------------------------------------
# 5. Update Processor
# ---------------------------------------------------------
# processor = AutoProcessor.from_pretrained(model_id)
image_processor = preprocess   # swap OpenCLIP preprocessing


tokenizer = AutoTokenizer.from_pretrained("huggyllama/llama-7b",  use_fast=False)
processor = AutoProcessor.from_pretrained("liuhaotian/llava-v1.5-7b-pretrain")
# ---------------------------------------------------------
# 6. Test Inference (image + text prompt)
# ---------------------------------------------------------
from PIL import Image

# image = Image.open("test.jpg").convert("RGB")
prompt = "USER: Describe this image in detail. ASSISTANT:"

with torch.no_grad():
    output_ids = model.generate(**inputs, max_new_tokens=200)
print(processor.tokenizer.decode(output_ids[0], skip_special_tokens=True))