ViLaSR / eval /infer_multi.py
cornuHGF's picture
Upload folder using huggingface_hub
1efcb3c verified
import argparse
import traceback
import random
import re
import copy
import torch
import torch.distributed as dist
import os
import json, jsonlines
from tqdm import tqdm
import pdb
import numpy as np
import socket
from vllm import LLM, SamplingParams
from transformers import Qwen2_5_VLProcessor, Qwen2_5_VLForConditionalGeneration
from transformers import AutoProcessor, AutoTokenizer
from qwen_vl_utils import process_vision_info, fetch_image
from torch.utils.data import Dataset, DataLoader
from PIL import Image
from datasets import load_dataset
import sys
sys.path.append('/mnt/beegfs/dzhu6/ViLaSR')
from utils.edit_image import merge_bbox_movement, parse_bbox_and_movement, plot_movement, plot_bounding_boxes
from typing import List, Dict, Any
from concurrent.futures import ThreadPoolExecutor
from datetime import datetime
import time, csv
csv.field_size_limit(sys.maxsize)
SYSTEM_PROMPT = """### Guidance:
You are a spatial reasoning assistant with access to two powerful visualization tools.
Your task is to break down complex spatial problems and iteratively refine your solution through visualization feedback.
### Available tools:
You can use the following two tools to visualize. After each tool usage, you must wait for and analyze the visualization feedback before proceeding.
1. **Object Mapper**
- Purpose: Identifies and maps key items in the space
- Input format: JSON
```json
[{{
"index": i, # Image index
"bbox_2d": [x1, y1, x2, y2],
"label": "object name/description"
}}]
```
- Output: Generates bounding boxes for visual inspection of the i-th image
2. **Path Tracer**
- Purpose: Plots movement or connections between points
- Input format: JSON
```json
[{{
"index": i, # Image index
"start_point_2d": [x1, y1],
"end_point_2d": [x2, y2],
"label": "trace_description"
}}]
```
- Output: Generates visual paths for verification of the i-th image
### Required Output Format:
For each reasoning step, you must structure your response as follows:
<think> [Your detailed reasoning process] </think> Action: [Object Mapper/Path Tracer]
```json
[JSON format coordinates]
```
After your reasoning and iteratively refine your solution through visualization feedback, you should arrive at a final answer and structure your response as follows:
<think> [Your detailed reasoning process] </think> Action: Answer
<answer> [Your final answer] </answer>
### Please NOTE the following reasoning techniques:
1. Initial Analysis
- Break down the spatial problem
- Plan your approach
2. Iterative Reasoning for Each Step
- Choose appropriate tool
- Provide absolute coordinates in JSON format (The top-left corner of the image is (0, 0) and the bottom-right corner is ({width}, {height}))
- Observe the visualization output
- Reflect on the visualization:
* Is the placement/path accurate?
* Does it align with your reasoning?
* What adjustments are needed?
- Backtrack and Adjust:
* If errors found, backtrack to previous step to modify actions or decisions as needed"""
# (You can only edit ONE image at each step)
PROMPT_TEMPLATE = """
### Question:
{question}
Begin your reasoning. After each tool use, critically evaluate the visualization and adjust if needed:
"""
BSZ=1 # 50 reduce it if GPU OOM
MAX_IMAGES=45
SUBIMAGE_PATTERN = r".*\#\#\#\[([\d\.]+),\s*([\d\.]+),\s*([\d\.]+),\s*([\d\.]+)\]"
TYPE_TEMPLATE = {
# ----------- same as cold start -----------
"multiple choice": '\nAnswer with the option\'s letter from the given choices directly.', # " Please provide only the single option letter (e.g., A, B, C, D, etc.) within the <answer> </answer> tags.",
"free-form": '', # " Please provide your text answer within the <answer> </answer> tags.",
"regression": '\nPlease answer the question using a single word or phrase (e.g., 42 or 3.14).', # " Please provide the numerical value (e.g., 42 or 3.14) within the <answer> </answer> tags."
"numerical": '\nPlease answer the question using a single word or phrase (e.g., 42 or 3.14).', # same as regression
"vci": "", # for spar-bench
}
from dataclasses import dataclass
@dataclass
class ProcessData:
index: int
response: str
mm_data: Dict
bbox_list_origin: Dict
movement_list_origin: Dict
finish_reason: str
is_finished: bool
grid_size: int
def calculate_grid_centers(image_size=616, grid_size=5):
"""for maze data"""
# matplotlib default margins
margin_left = int(image_size * 0.125)
margin_right = int(image_size * 0.1)
margin_bottom = int(image_size * 0.11)
margin_top = int(image_size * 0.12)
usable_width = image_size - (margin_left + margin_right)
usable_height = image_size - (margin_top + margin_bottom)
cell_width = usable_width / grid_size
cell_height = usable_height / grid_size
centers = []
for i in range(grid_size):
for j in range(grid_size):
# 计算中心坐标,考虑不同的边距
center_x = margin_left + cell_width/2 + j * cell_width
center_y = margin_top + cell_height/2 + i * cell_height
centers.append((center_x, center_y))
# print("calculate_grid_centers:", image_size, margin_left, margin_right, cell_width, cell_height)
return centers, (cell_width+cell_height)/2 # 返回x和y方向的格子大小
def check_path_tracer(movement_list, centers, cell_size):
for movement in movement_list:
for key in ['start_point_2d', 'end_point_2d']:
x, y = int(movement[key][0]), int(movement[key][1])
min_distance = min([np.sqrt((x-c[0])**2 + (y-c[1])**2) for c in centers])
if min_distance > cell_size/2:
# print(key, movement[key], min_distance, cell_size/2, centers)
return False
return True
def check_repetition(allindex, bbox_list_origin, movement_list_origin):
for cnt, tmp_index in enumerate(allindex):
for bbox_list in list(bbox_list_origin.values()):
for bbox in bbox_list:
if bbox in allindex[tmp_index]["bbox_list"]:
return True
for movement_list in list(movement_list_origin.values()):
for movement in movement_list:
if movement in allindex[tmp_index]["movement_list"]:
return True
return False
def process_single_response(data: ProcessData):
"""处理单个响应的函数"""
if data.is_finished is True:
return {
'index': data.index,
'response': data.response,
'finish_reason': data.finish_reason,
'is_finished': data.is_finished,
'processed_image_idx': [None],
}
try:
# 解析和绘图
bbox_list_new, movement_list_new = parse_bbox_and_movement(data.response)
current_image_index = len(data.mm_data['image'])
image_index_list, image_list = [], []
bbox_list, movement_list = data.bbox_list_origin, data.movement_list_origin
finish_reason = None
try:
allindex = {}
for tmp_bbox_list in bbox_list_new:
tmp_bbox_list = copy.deepcopy(tmp_bbox_list)
if tmp_bbox_list["index"] in allindex:
if "bbox_list" in allindex[tmp_bbox_list["index"]]:
allindex[tmp_bbox_list["index"]]["bbox_list"].append(tmp_bbox_list)
else:
allindex[tmp_bbox_list["index"]]["bbox_list"] = [tmp_bbox_list]
else:
allindex[tmp_bbox_list["index"]] = {'bbox_list': [tmp_bbox_list], 'movement_list': []}
for tmp_movement_list in movement_list_new:
tmp_movement_list = copy.deepcopy(tmp_movement_list)
if tmp_movement_list["index"] in allindex:
if "movement_list" in allindex[tmp_movement_list["index"]]:
allindex[tmp_movement_list["index"]]["movement_list"].append(tmp_movement_list)
else:
allindex[tmp_movement_list["index"]]["movement_list"] = [tmp_movement_list]
else:
allindex[tmp_movement_list["index"]] = {'bbox_list': [], 'movement_list': [tmp_movement_list]}
except Exception as e:
traceback.print_exc()
print("bbox_list_new, movement_list_new: ", bbox_list_new, movement_list_new)
finish_reason = "ToolGenError"
if len(allindex) == 0:
finish_reason = "ToolError"
elif len(data.mm_data['image']) >= MAX_IMAGES+1:
finish_reason = "TooManyImages"
if finish_reason is not None:
return {
'index': data.index,
'processed_image_idx': [None],
'image': [data.mm_data['image'][0].copy()],
'response': data.response,
'finish_reason': finish_reason,
'bbox_list': bbox_list,
'movement_list': movement_list,
'is_finished': True,
}
for cnt, tmp_index in enumerate(allindex):
bbox_list_new, movement_list_new = allindex[tmp_index]["bbox_list"], allindex[tmp_index]["movement_list"]
image_index_new = current_image_index + cnt
image_index, bbox_list, movement_list = merge_bbox_movement(
bbox_list_origin=data.bbox_list_origin,
movement_list_origin=data.movement_list_origin,
bbox_list_new=bbox_list_new,
movement_list_new=movement_list_new,
image_index_new=image_index_new,
)
image_index_list.append(image_index)
if image_index == -1:
return {
'index': data.index,
'processed_image_idx': [None],
'image': [data.mm_data['image'][0].copy()],
'response': data.response,
'finish_reason': "ToolError",
'bbox_list': bbox_list,
'movement_list': movement_list,
'is_finished': True,
}
image = data.mm_data['image'][image_index].copy()
assert isinstance(image, Image.Image)
input_width, input_height = image.size
# draw bbox and lines
plot_bounding_boxes(image, bbox_list[image_index_new], input_height=input_height, input_width=input_width)
plot_movement(image, movement_list[image_index_new], input_height=input_height, input_width=input_width)
image_list.append(image)
return {
'index': data.index,
'processed_image_idx': image_index_list,
'image': image_list,
'response': data.response,
'finish_reason': data.finish_reason,
'bbox_list': bbox_list,
'movement_list': movement_list,
'is_finished': data.is_finished
}
except Exception as e:
print(f"Error processing response {data.index}: {str(e)}")
traceback.print_exc()
return None
def save_samples_info(samples_info, save_dir):
def get_unique_dir(base_path, prefix='generation'):
"""generate unique dirctory"""
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
counter = 0
while True:
if counter == 0:
dir_name = f"{prefix}_{timestamp}"
else:
dir_name = f"{prefix}_{timestamp}_{counter}"
full_path = os.path.join(base_path, dir_name)
if not os.path.exists(full_path):
return full_path
counter += 1
all_sample_dir = []
for idx, sample in enumerate(samples_info):
if 'qid' in sample:
sample_dir = os.path.join(save_dir, sample['qid'])
elif 'id' in sample:
sample_dir = os.path.join(save_dir, sample['id'])
elif 'index' in sample:
sample_dir = os.path.join(save_dir, sample['index'])
else:
sample_dir = get_unique_dir(save_dir, f'sample')
os.makedirs(sample_dir, exist_ok=True)
all_sample_dir.append(sample_dir)
text_data = {
'prompt': sample['prompt'],
'sequence': sample['sequence'],
'response': sample['response'],
'finish_reason': sample['finish_reason'],
'execution_pass': sample['execution_pass']
}
with open(os.path.join(sample_dir, 'text_data.json'), 'w', encoding='utf-8') as f:
json.dump(text_data, f, indent=2, ensure_ascii=False)
# save images
if 'multi_modal_data' in sample and 'image' in sample['multi_modal_data']:
images_dir = os.path.join(sample_dir, 'images')
os.makedirs(images_dir, exist_ok=True)
for img_idx, img in enumerate(sample['multi_modal_data']['image']):
if isinstance(img, Image.Image):
img_path = os.path.join(images_dir, f'image_{img_idx}.png')
img.save(img_path)
return all_sample_dir
def get_qwen_chat(model, processor, texts, images, sampling_params={}):
responses = []
for text, image in zip(texts, images):
inputs = processor(text=[text], images=[image], padding=True, return_tensors='pt').to(model.device)
with torch.inference_mode():
generated_ids = model.generate(
**inputs,
pad_token_id=processor.tokenizer.eos_token_id,
temperature=sampling_params.temperature,
top_p=sampling_params.top_p,
max_new_tokens=sampling_params.max_tokens,
)
generated_ids_trimmed = [
out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
response = processor.batch_decode(
generated_ids_trimmed,
skip_special_tokens=True,
clean_up_tokenization_spaces=False,
)[0]
responses.append(response)
return responses
def multi_turn_generate(inference_engine, processor, tokenizer, vllm_inputs=None, sampling_params=None, prompt_token_ids=None, use_tqdm=False, save_dir=None, max_num_steps=10):
def _get_prompts_and_indices(samples_info):
prompts, multi_modal_data, indices=[], [], []
for index, info in enumerate(samples_info):
if not info['stop'] and len(info['multi_modal_data']['image']) <= MAX_IMAGES:
prompts.append(info['sequence'])
multi_modal_data.append(info['multi_modal_data'])
indices.append(info['index'])
return prompts, multi_modal_data, indices
sampling_params=copy.deepcopy(sampling_params)
new_vllm_inputs = []
for single_vllm_input in vllm_inputs:
prompt = tokenizer.decode(single_vllm_input['prompt_token_ids'], skip_special_tokens=False)
new_vllm_inputs.extend([{
"id": single_vllm_input['id'],
"prompt": prompt,
"multi_modal_data": single_vllm_input['multi_modal_data'],
"grid_size": single_vllm_input['grid_size'],
} for _ in range(sampling_params.n)])
sampling_params.n=1
sampling_params.detokenize=True # True convert ids to text
samples_info = []
for index, item in enumerate(new_vllm_inputs):
processed_image = [fetch_image({'image': origin_image}) for origin_image in item['multi_modal_data']['image']]
sample_info = {
"id": item["id"],
"prompt": item["prompt"],
"sequence": item["prompt"],
"multi_modal_data": {"image": processed_image},
"response": "",
"stop": False,
"finish_reason": None,
"processed_image_idx": [],
"index": index,
"mask_info": [],
"execution_pass": 0,
"bbox_list": {img_idx: [] for img_idx in range(len(processed_image))}, # save bbox_list for each image index
"movement_list": {img_idx: [] for img_idx in range(len(processed_image))}, # save movement_list for each image index
"grid_size": item['grid_size'],
}
samples_info.append(sample_info)
intermediate_prompt = 'The index of the given image is {current_image_idx} (width: {width}, height: {height}). Continue your reasoning. After each tool use, critically evaluate the visualization and adjust if needed:'
final_prompt = 'The index of the given image is {current_image_idx} (width: {width}, height: {height}). Then, you can not invoke the Object Mapper or Path Tracer tool. Please answer the initial question and structure your response as required:'
intermediate_template = """<|im_end|>
<|im_start|>user
{pad}
{prompt}
<|im_end|>
<|im_start|>assistant
""" # connect multi-turn conversations
num_llm_calls_available = max_num_steps - 1
while num_llm_calls_available >= 0:
num_llm_calls_available-=1
input_prompts, multi_modal_data, indices=_get_prompts_and_indices(samples_info) # _get_prompts_and_indices
# print("input_prompts:", len(input_prompts), input_prompts[0])
# print("multi_modal_data:", len(multi_modal_data), multi_modal_data[0])
# print("indices:", len(indices), indices[0])
# print("num_llm_calls_available:", num_llm_calls_available)
if type(inference_engine) == LLM:
input_prompts = [{
'prompt_token_ids': tokenizer.encode(prompt, add_special_tokens=False)[:],
'multi_modal_data': mm_data # {'image', list()}
} for prompt, mm_data in zip(input_prompts, multi_modal_data)]
outputs = inference_engine.generate(prompts=input_prompts, sampling_params=sampling_params, use_tqdm=use_tqdm)
else:
responses = get_qwen_chat(inference_engine, processor, input_prompts, [m['image'] for m in multi_modal_data], sampling_params)
class A:
pass
outputs = []
for request_id, response in enumerate(responses):
a = A()
a.outputs = [A()]
a.outputs[0].text = response
a.outputs[0].finish_reason = "stop"
a.outputs[0].stop_reason = None
a.request_id = request_id
outputs.append(a)
sorted_outputs = sorted(outputs, key=lambda output: int(output.request_id))
responses=[x.outputs[0].text for x in sorted_outputs]
finish_reason=[x.outputs[0].finish_reason for x in sorted_outputs] # "stop", "length"
stop_reason=[x.outputs[0].stop_reason for x in sorted_outputs] # None: have EOS
if num_llm_calls_available==-1:
for i ,index in enumerate(indices):
samples_info[index]['response']+=responses[i]
samples_info[index]['sequence']+=responses[i]
samples_info[index]['stop']=True
samples_info[index]['finish_reason']=finish_reason[i]
break
def _is_finished(finish_reason, stop_reason, response):
if finish_reason=='stop' and stop_reason==None and "<answer>" in response and "</answer>" in response:
return True
if finish_reason=='length':
return True
if finish_reason=='rule':
return True
return False
# breakpoint()
is_finished=[_is_finished(finish_reason[i], stop_reason[i], responses[i]) for i in range(len(finish_reason))]
# check if all samples are finished
if all([x for x in is_finished]):
for i ,index in enumerate(indices):
samples_info[index]['response']+=responses[i]
samples_info[index]['sequence']+=responses[i]
samples_info[index]['stop']=True
samples_info[index]['finish_reason']=finish_reason[i]
break
# ----------- Parallel Process -----------
# Prepare Data
process_data_list = [
ProcessData(
index=index,
response=responses[i],
mm_data=samples_info[index]['multi_modal_data'],
bbox_list_origin=samples_info[index]["bbox_list"],
movement_list_origin=samples_info[index]["movement_list"],
finish_reason=finish_reason[i],
is_finished=is_finished[i], # if is_finished == True, stop reasoning
grid_size=samples_info[index]['grid_size'],
) for i, index in enumerate(indices)]
with ThreadPoolExecutor(max_workers=max(min(len(indices), os.cpu_count()//2, 64), 1) ) as executor:
results = list(executor.map(process_single_response, process_data_list))
# update samples_info
for result in results:
if result is not None:
index = result['index']
samples_info[index]['response'] += result['response']
samples_info[index]['stop'] = result['is_finished']
samples_info[index]['finish_reason'] = result['finish_reason']
samples_info[index]['processed_image_idx'].extend(result['processed_image_idx'])
if result['is_finished'] is False:
current_image_count = len(samples_info[index]['multi_modal_data']['image'])
if len(result["image"]) > 1: # 处理多图片情况
current_image_idx = current_image_count + 1
pad_prompt = ""
for tmp_image_idx, tmp_image in enumerate(result["image"]):
width, height = fetch_image({"image": tmp_image}).size
if current_image_count + tmp_image_idx + 1>=MAX_IMAGES:
pad_prompt += f"<|vision_start|><|image_pad|><|vision_end|>" + final_prompt.format(
current_image_idx=current_image_idx + tmp_image_idx,
width=width,
height=height,
)
samples_info[index]['multi_modal_data']['image'].append(tmp_image)
break
else:
if tmp_image_idx <= len(result["image"]) - 2:
pad_prompt += f"<|vision_start|><|image_pad|><|vision_end|>The index of the given image is {current_image_idx+tmp_image_idx} (width: {width}, height: {height}).\n"
else:
if num_llm_calls_available > 0:
pad_prompt += f"<|vision_start|><|image_pad|><|vision_end|>" + intermediate_prompt.format(
current_image_idx=current_image_idx + tmp_image_idx,
width=width,
height=height,
)
else:
pad_prompt += f"<|vision_start|><|image_pad|><|vision_end|>" + final_prompt.format(
current_image_idx=current_image_idx + tmp_image_idx,
width=width,
height=height,
)
samples_info[index]['multi_modal_data']['image'].append(tmp_image)
samples_info[index]['sequence'] += result['response'] + intermediate_template.format(prompt="", pad=pad_prompt)
else:
current_image_idx = current_image_count + 1
width, height = fetch_image({"image": result["image"][0]}).size
if current_image_count + 1>= MAX_IMAGES:
# Maximum limit reached, switching to final_prompt
prompt = final_prompt.format(
current_image_idx=current_image_idx,
width=width,
height=height,
)
else:
prompt = (intermediate_prompt if num_llm_calls_available > 0 else final_prompt).format(
current_image_idx=current_image_idx,
width=width,
height=height,
)
samples_info[index]['sequence'] += result['response'] + intermediate_template.format(
prompt=prompt,
pad="<|vision_start|><|image_pad|><|vision_end|>"
)
samples_info[index]['multi_modal_data']['image'].append(result['image'][0])
# 更新其他信息
samples_info[index]['bbox_list'] = result['bbox_list']
samples_info[index]["movement_list"] = result['movement_list']
else:
samples_info[index]['sequence'] += result['response']
for i, line in enumerate(samples_info):
if samples_info[i]['finish_reason']!='length':
samples_info[i]['sequence']+=tokenizer.eos_token # add end of sentence
batch_sequences = [sample['sequence'] for sample in samples_info]
if save_dir:
all_sample_dir = save_samples_info(samples_info, save_dir)
return batch_sequences, all_sample_dir
return batch_sequences
def parse_dialog(serialized_content):
# segement dialogue
segments = re.split(r'<\|im_start\|>|<\|im_end\|>', serialized_content)
segments = [s for s in segments if s]
conversations = []
current_role = None
current_content = []
system_content = None
if segments[0].startswith('system'):
system_content = segments[0].replace('system\n\n', '', 1) # only replace the first time
segments = segments[1:]
if system_content:
conversations.append({
"role": "system",
"content": system_content
})
for segment in segments:
if segment.startswith('user'):
has_vision = '<|vision_start|><|image_pad|><|vision_end|>' in segment
text = segment.replace('user\n', '', 1) # only replace the first time
# text = text.replace('<|vision_start|><|image_pad|><|vision_end|>\n', '', 1) # keep <|vision_start|><|image_pad|><|vision_end|>
content = []
if has_vision:
content.append({
"type": "image",
"image": "image_path",
"nframes": "args.max_frames",
"max_pixels": args.max_pixels
})
content.append({
"type": "text",
"text": text
})
conversations.append({
"role": "user",
"content": content
})
elif segment.startswith('assistant'):
text = segment.replace('assistant\n', '', 1) # only replace the first time
conversations.append({
"role": "assistant",
"content": text
})
return conversations
def setup_distributed():
"""Setup distributed training environment for SLURM"""
if "SLURM_PROCID" in os.environ:
rank = int(os.environ["SLURM_PROCID"])
world_size = int(os.environ["SLURM_NTASKS"])
local_rank = int(os.environ["SLURM_LOCALID"])
master_addr = os.environ.get("MASTER_ADDR", "127.0.0.1")
master_port = os.environ.get("MASTER_PORT", "29500")
else:
rank = int(os.environ.get("RANK", 0))
world_size = int(os.environ.get("WORLD_SIZE", 1))
local_rank = int(os.environ.get("LOCAL_RANK", 0))
master_addr = os.environ.get("MASTER_ADDR", "127.0.0.1")
master_port = os.environ.get("MASTER_PORT", "29500")
os.environ["MASTER_ADDR"] = master_addr
os.environ["MASTER_PORT"] = master_port
dist.init_process_group("nccl", rank=rank, world_size=world_size)
torch.cuda.set_device(local_rank)
return rank, world_size, local_rank
def eval_model(args):
# rank, world_size, local_rank = setup_distributed()
# print(f"RANK and WORLD_SIZE: {rank}/{world_size}, local rank: {local_rank}")
# hostname = socket.gethostname()
# print(f"Process {rank} is running on node {hostname}")
# Model
model_path = args.model_path
model_name = args.model_name
print(f"prune: {args.prune}")
print(f"resize: {args.resize}")
print(f"Loading from {model_path}")
print("torch.cuda.device_count():", torch.cuda.device_count())
if args.prune and 'vscan' in args.prune.lower():
from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
Qwen2_5_VisionTransformerPretrainedModel,
Qwen2_5_VLVisionBlock,
Qwen2_5_VLVisionSdpaAttention,
Qwen2_5_VLVisionFlashAttention2,
Qwen2_5_VisionPatchEmbed,
Qwen2_5_VLModel
)
sys.path.append('/mnt/beegfs/dzhu6/SelfEvolvingAgent/VScan/')
from qwen.model.qwen2_5_vl_custom import (
Qwen2_5_VLForConditionalGeneration_X,
Qwen2_5_VisionTransformerPretrainedModel_X,
Qwen2_5_VLVisionBlock_X,
Qwen2_5_VLVisionSdpaAttention_X,
Qwen2_5_VLVisionFlashAttention2_X,
Qwen2_5_VisionPatchEmbed_X,
Qwen2_5_VLModel_X
)
Qwen2_5_VLForConditionalGeneration.forward = Qwen2_5_VLForConditionalGeneration_X.forward
Qwen2_5_VisionTransformerPretrainedModel.forward = Qwen2_5_VisionTransformerPretrainedModel_X.forward
Qwen2_5_VLVisionBlock.forward = Qwen2_5_VLVisionBlock_X.forward
Qwen2_5_VLVisionSdpaAttention.forward = Qwen2_5_VLVisionSdpaAttention_X.forward
Qwen2_5_VLVisionFlashAttention2.forward = Qwen2_5_VLVisionFlashAttention2_X.forward
Qwen2_5_VisionPatchEmbed.forward = Qwen2_5_VisionPatchEmbed_X.forward
Qwen2_5_VLModel.forward = Qwen2_5_VLModel_X.forward
Qwen2_5_VLModel.layer_prune = Qwen2_5_VLModel_X.layer_prune
print("Qwen2_5_VLForConditionalGeneration.forward:", Qwen2_5_VLForConditionalGeneration.forward)
llm = Qwen2_5_VLForConditionalGeneration.from_pretrained(
args.model_path,
torch_dtype=torch.bfloat16,
device_map=f"auto",
attn_implementation="flash_attention_2",
trust_remote_code=True
).eval()
llm.model.layer_list = [14]
llm.model.image_token_ratio_list = [0.333]
llm.image_token_ratio = 0.167
min_pixels = 4*28*28
max_pixels = 1280*28*28
else:
# llm = Qwen2_5_VLForConditionalGeneration.from_pretrained(
# args.model_path,
# torch_dtype=torch.bfloat16,
# device_map=f"auto",
# attn_implementation="flash_attention_2",
# trust_remote_code=True
# ).eval()
tensor_parallel_size = int(os.environ.get("SLURM_NTASKS", "1"))
pipeline_parallel_size = int(os.environ.get("VLLM_PIPELINE_PARALLEL_SIZE", "1"))
print(f"Loading from {model_path}, tensor_parallel_size: {tensor_parallel_size}, pipeline_parallel_size: {pipeline_parallel_size}")
llm = LLM(
trust_remote_code=True,
model=model_path,
dtype="bfloat16",
tensor_parallel_size=tensor_parallel_size,
pipeline_parallel_size=pipeline_parallel_size,
limit_mm_per_prompt={"image": 62, "video": 10},
gpu_memory_utilization=0.85, # default 0.9
enable_prefix_caching=True # cache
)
processor = AutoProcessor.from_pretrained(model_path)
sampling_params = SamplingParams(
temperature=args.temperature,
top_p=args.top_p,
max_tokens=16384,
stop_token_ids=[],
)
processor = AutoProcessor.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
tokenizer.padding_side = "left"
processor.tokenizer = tokenizer
'''
file_path = args.input_file
if file_path.endswith('.jsonl'):
with open(file_path, 'r', encoding='utf-8') as f:
data = [json.loads(line) for line in f]
else:
with open(file_path, 'r', encoding='utf-8') as f:
data = json.load(f)
'''
if 'blink' in args.dataset.lower():
sys.path.append('/mnt/beegfs/dzhu6/VisualSketchpad/')
from infer2 import get_blink_dataset
data = get_blink_dataset()
elif 'spatialeval' in args.input_file.lower():
data = load_dataset("MilaWang/SpatialEval", "vqa", split="test")
elif '3dsr' in args.dataset.lower():
dp = "/mnt/beegfs/dzhu6/correlation/3dsrbench_v1_vlmevalkit_circular.tsv"
with open(dp, 'r') as f:
reader = csv.DictReader(f, delimiter='\t')
data = list(reader)
st, ed = (len(data)*args.split)//args.all, (len(data)*(args.split+1))//args.all
print(f"{len(data)} lines found, generating from {st} to {ed}")
print("Data: ", len(data))
if type(data) is list:
data = data[st:ed]
else:
data = data.select(range(st, ed))
print("Data: ", len(data))
batches = []
messages = []
ids = []
def get_key(x):
return x['id'] if 'id' in x else x['qid'] if 'qid' in x else x['index']
start_idx = 0
old = set()
if args.all > 1:
# 分split
output_dir = os.path.join(args.output_dir, f"split_{args.split}_all_{args.all}") # args.output_dir
else:
output_dir = args.output_dir
save_dir = output_dir
if args.over_write:
os.system(f"rm -rf {output_dir} && mkdir {output_dir}")
else:
if not os.path.exists(output_dir):
os.system(f"mkdir {output_dir}")
print("Output Dir: ", output_dir)
output_file_path = f"{output_dir}/results.jsonl"
if os.path.exists(output_file_path):
mode = "a"
with jsonlines.open(output_file_path) as fin:
for line in fin:
old.add(get_key(line))
# start_idx += 1
# start_idx = len(set([x['id'] if 'id' in x else x['qid'] for x in jsonlines.open(output_file_path)]))
else:
mode = "w"
print("loaded ", len(old), " lines from ", output_file_path)
for xidx, x in enumerate(data):
if get_key(x) in old:
continue
# if 'spatialeval' in args.dataset.lower() and 'real' not in get_key(x):
# continue
ids.append(get_key(x))
batches.append(x.copy())
batches[-1].pop("image", None)
# if xidx == 0:
# print(batches[-1])
'''
{
"dataset": "SpatialEval",
"task": "spatialreal",
"split": "test",
"question_id": "spatialreal.vqa.sa_1543979.0",
"question": "Please answer the following question based on the image. How many individual cartons of strawberries are there in the wooden basket? Available options:\nA. eight\nB. ten\nC. four\nD. six\n",
"answer": "D",
"image_path": [
"SpatialEval/spatialreal.vqa.sa_1543979.0.png"
],
"data_type": "image",
"problem_type": "multiple choice"
},
'''
if args.dataset in ["vsi_bench"]:
prompt = f"These are frames from a video, numbered from 1 to {args.max_frames} in sequence. That is, the index of each image is 1, 2, 3, ..., {args.max_frames}.\n\nAnswer the quesntion with appropriate tools:\n" + x['question'] # + '\n\nThe final answer should be a single word or phrase.'
if x['problem_type'] == 'multiple choice' and 'options' in x:
prompt = prompt + '\n' + '\n'.join(x['options'])
prompt = prompt + TYPE_TEMPLATE[x['problem_type'].lower()]
width, height = fetch_image({"image": os.path.join(args.image_folder, x["image_path"][0]), "max_pixels": args.max_pixels}).size
image_messages = []
for image_idx, image_path in enumerate(x["image_path"]):
image_messages.extend([
{
"type": "image",
"image": os.path.join(args.image_folder, image_path),
"nframes": args.max_frames,
"max_pixels": args.max_pixels
},
{
"type": "text",
"text": f"The index of the given image is {image_idx+1} (width: {width}, height: {height}).\n",
}
])
image_messages.append({
"type": "text",
"text": PROMPT_TEMPLATE.format(question=prompt)
})
msg = [
{
"role": "system",
"content": SYSTEM_PROMPT.format(width=width, height=height)
},
{
"role": "user",
"content": image_messages,
}
]
elif "3dsrbench" in args.dataset.lower():
x['index'] = x['qid']
question = x['question']
question += f"\nA. {x['A']}"
question += f"\nB. {x['B']}"
if "C" in x and x['C'] is not None and x['C'].strip() != "":
question += f"\nC. {x['C']}"
if "D" in x and x['D'] is not None and x['D'].strip() != "":
question += f"\nD. {x['D']}"
question += "\n"
prompt = question
if "flip" in x['index'].lower():
if "flip" in x['index'].lower():
ext = x['image_url'].split('.')[-1]
flip = x['image_url'].replace("http://images.cocodataset.org/", "/mnt/beegfs/dzhu6/coco_images/").replace("." + ext, "_flip." + ext)
if not os.path.exists(flip):
image = Image.open(x['image_url'].replace("http://images.cocodataset.org/", "/mnt/beegfs/dzhu6/coco_images/")).convert("RGB")
image = image.transpose(method=Image.FLIP_LEFT_RIGHT)
image.save(flip)
x['image_url'] = x['image_url'].replace('.' + ext, "_flip." + ext)
x["problem_type"] = "multiple choice"
x['image_path'] = [x['image_url'].replace("http://images.cocodataset.org/", "")]
prompt = prompt + '\nThe index of the given image is 1.' + TYPE_TEMPLATE[x['problem_type'].lower()]
width, height = fetch_image({"image": os.path.join(args.image_folder, x["image_path"][0]), "max_pixels": args.max_pixels}).size
if args.resize:
width, height = width//2, height//2
msg = [
{
"role": "system",
"content": SYSTEM_PROMPT.format(width=width, height=height)
},
{
"role": "user",
"content": [
{
"type": "image",
"image": os.path.join(args.image_folder, x["image_path"][0]),
"nframes": args.max_frames,
"grid_size": x["grid_size"] if "grid_size" in x else None,
"max_pixels": args.max_pixels
},
{
"type": "text",
"text": PROMPT_TEMPLATE.format(question=prompt)
}
]
}]
elif "blink" in args.dataset.lower():
image_inputs, _ = process_vision_info(x['messages'])
if image_inputs:
width, height = image_inputs[0].size
if args.resize:
width, height = width//2, height//2
msg = [
{
"role": "system",
"content": SYSTEM_PROMPT.format(width=width, height=height)
},
]
else:
msg = [
{
"role": "system",
"content": SYSTEM_PROMPT.format(width=0, height=0)
},
]
msg += x['messages']
elif 'spatialeval' in args.dataset.lower(): # 3dsrbench_v1
x["problem_type"] = "multiple choice"
x['image_path'] = [x['id'] + '.png']
prompt = x["text"]
prompt = prompt + '\nThe index of the given image is 1.' + TYPE_TEMPLATE[x['problem_type'].lower()]
width, height = fetch_image({"image": os.path.join(args.image_folder, x["image_path"][0]), "max_pixels": args.max_pixels}).size
if args.resize:
width, height = width//2, height//2
msg = [
{
"role": "system",
"content": SYSTEM_PROMPT.format(width=width, height=height)
},
{
"role": "user",
"content": [
{
"type": "image",
"image": os.path.join(args.image_folder, x["image_path"][0]),
"nframes": args.max_frames,
"grid_size": x["grid_size"] if "grid_size" in x else None,
"max_pixels": args.max_pixels
},
{
"type": "text",
"text": PROMPT_TEMPLATE.format(question=prompt)
}
]
}]
elif args.dataset in ["maze", "SpatialEval_spatialreal",]:
prompt = x["question"]
if x['problem_type'] == 'multiple choice' and 'options' in x:
prompt = prompt + '\n' + '\n'.join(x['options'])
prompt = prompt + '\nThe index of the given image is 1.' + TYPE_TEMPLATE[x['problem_type'].lower()]
width, height = fetch_image({"image": os.path.join(args.image_folder, x["image_path"][0]), "max_pixels": args.max_pixels}).size
msg = [
{
"role": "system",
"content": SYSTEM_PROMPT.format(width=width, height=height)
},
{
"role": "user",
"content": [
{
"type": "image",
"image": os.path.join(args.image_folder, x["image_path"][0]),
"nframes": args.max_frames,
"grid_size": x["grid_size"] if "grid_size" in x else None,
"max_pixels": args.max_pixels
},
{
"type": "text",
"text": PROMPT_TEMPLATE.format(question=prompt)
}
]
}]
elif args.dataset in ["spar_bench", "spar_bench_tiny", "mmsi_bench"]:
prompt = x["question"]
if x['problem_type'] == 'multiple choice' and x.get('options', None) is not None:
prompt = prompt + '\n' + '\n'.join(x['options'])
prompt = prompt.replace("Your answer can only include one of options A, B, C or D.", "")
prompt = prompt.replace("Answer using a single number and nothing else.", "")
post_prompt = ""
if x.get('original_question_type', None) in ['position_matching', "camera_motion_infer"]:
post_prompt = "The values represent the bounding box coordinates normalized to a 0-1000 scale, with the top-left corner as the origin of the image."
prompt = prompt + "\n" + post_prompt
if x['data_type'] == 'single_view':
prompt = prompt + '\nThe index of the given image is 1.' + TYPE_TEMPLATE[x['problem_type'].lower()]
width, height = fetch_image({"image": os.path.join(args.image_folder, x["image_path"][0]), "max_pixels": args.max_pixels}).size
msg = [
{
"role": "system",
"content": SYSTEM_PROMPT.format(width=width, height=height)
},
{
"role": "user",
"content": [
{
"type": "image",
"image": os.path.join(args.image_folder, x["image_path"][0]),
"max_pixels": args.max_pixels
},
{
"type": "text",
"text": PROMPT_TEMPLATE.format(question=prompt)
}
]
}
]
elif x['data_type'] == 'multi_view': # multi_view
n_frames = len(x["image_path"])
width, height = fetch_image({"image": os.path.join(args.image_folder, x["image_path"][0]), "max_pixels": args.max_pixels}).size
image_messages = []
for image_idx, image_path in enumerate(x["image_path"]):
image_messages.extend([
{
"type": "image",
"image": os.path.join(args.image_folder, image_path),
"max_pixels": args.max_pixels
},
{
"type": "text",
"text": f"The index of the given image is {image_idx+1} (width: {width}, height: {height}).\n"
}
])
prompt = prompt + TYPE_TEMPLATE[x['problem_type'].lower()]
image_messages.append({
"type": "text",
"text": PROMPT_TEMPLATE.format(question=prompt)
})
msg = [
{
"role": "system",
"content": SYSTEM_PROMPT.format(width=width, height=height)
},
{
"role": "user",
"content": image_messages
}
]
else:
raise Exception(f"UNKNON args.dataset: {args.dataset}")
messages.append(msg)
print("messages: ", len(messages))
print("batches: ", len(batches))
with open(output_file_path, mode, encoding="utf-8") as fout:
print("Message Example:", messages[0])
print(f"Start from the {start_idx} example")
for i in tqdm(range(start_idx, len(messages), BSZ), desc="Processing batches"):
batch_messages = messages[i:i + BSZ]
prompts = [processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True) for msg in batch_messages]
image_num = []
for msg in batch_messages:
current_image_num = 0
for turn in msg:
if isinstance(turn["content"], list):
for turn_content in turn["content"]:
if turn_content["type"] == "image":
current_image_num += 1
if args.dataset in ["vsi_bench"]:
assert current_image_num == args.max_frames, f"wrong image number: {current_image_num} != {args.max_frames}"
elif args.dataset in ["maze", "SpatialEval_spatialreal", "SpatialEval"]:
assert current_image_num == 1, f"wrong image number: {current_image_num}"
image_num.append(current_image_num)
image_inputs, video_inputs, video_kwargs = process_vision_info(batch_messages, return_video_kwargs=True)
image_idx = 0
video_idx = 0
llm_inputs = []
for idx, (prompt, msg) in enumerate(zip(prompts, batch_messages)):
mm_type = batch_messages[idx][1]['content'][0]['type']
sample_mm_data = {}
sample_video_kw = {}
if mm_type == 'image':
sample_mm_data["image"] = []
for current_idx in range(image_num[idx]):
width, height = image_inputs[image_idx].size
image = image_inputs[image_idx]
if args.resize:
image = image.resize((width//2, height//2), resample=Image.Resampling.LANCZOS)
if args.dataset in ["video", "vsi_bench"]:
sample_mm_data["image"].append(image)
else:
sample_mm_data["image"].append(image)
image_idx += 1
elif mm_type == 'video':
sample_mm_data["video"] = [video_inputs[video_idx]]
for key, value in video_kwargs.items():
sample_video_kw[key] = value[video_idx]
video_idx += 1
llm_inputs.append({
'id': ids[i + idx],
"prompt": prompt,
"prompt_token_ids": tokenizer.encode(prompt, add_special_tokens=False),
"multi_modal_data": sample_mm_data,
"mm_processor_kwargs": sample_video_kw,
"grid_size": msg[1]["content"][0]["grid_size"] if args.dataset == 'maze' else None
})
if image_inputs is not None:
assert image_idx == len(image_inputs), f"Image index mismatch: {image_idx} != {len(image_inputs)}"
if video_inputs is not None:
assert video_idx == len(video_inputs), f"Video index mismatch: {video_idx} != {len(video_inputs)}"
try:
if i < 1e9:
batch_sequences = multi_turn_generate(llm, processor, tokenizer, vllm_inputs=llm_inputs, sampling_params=sampling_params, save_dir=save_dir,
max_num_steps=20 if args.dataset=="maze" else 10)
batch_sequences, all_sample_dir = batch_sequences
else:
batch_sequences = multi_turn_generate(llm, processor, tokenizer, vllm_inputs=llm_inputs, sampling_params=sampling_params, save_dir=None,
max_num_steps=20 if args.dataset=="maze" else 10)
all_sample_dir = [None] * len(batch_sequences)
batch_conversations = [parse_dialog(sequence) for sequence in batch_sequences]
print(f"Processed batch {(i)//BSZ + 1}/{(len(messages) + BSZ - 1)//BSZ}. ")
except Exception as e:
print(f"Error processing batch starting at index {i}: {e}")
continue
for input_example, model_output, sample_dir in zip(batches[i:i + BSZ], batch_conversations, all_sample_dir):
result = input_example.copy()
result['conversations'] = model_output
result['model_output'] = model_output[-1]['content']
result['model_id'] = model_name
result['sample_dir'] = sample_dir
fout.write(
json.dumps(result)
+ "\n"
)
fout.flush()
if 'spatialeval' in args.dataset.lower():
with jsonlines.open(f"/mnt/beegfs/dzhu6/SpatialEval/outputs/vqa/spatialreal/m-{model_name}_bare_prune_{args.prune}{'_resize_True' if args.resize else ''}_split_{args.split}_all_{args.all}.jsonl", "a") as writer:
result = input_example.copy()
result['answer'] = model_output[-1]['content']
writer.write(result)
elif '3dsr' in args.dataset.lower():
with jsonlines.open(f"/mnt/beegfs/dzhu6/correlation/results/{model_name}_prune_{args.prune}{'_resize_True' if args.resize else ''}_split_{args.split}_all_{args.all}.jsonl", "a") as writer:
result = input_example.copy()
result['response'] = model_output[-1]['content']
writer.write(result)
elif 'blink' in args.dataset.lower():
with jsonlines.open(f"/mnt/beegfs/dzhu6/VisualSketchpad/{model_name}_traj_False_prune_{args.prune}{'_resize_True' if args.resize else ''}_maxnewtokens_16384_split_{args.split}_all_{args.all}.jsonl", "a") as writer:
result = input_example.copy()
result.pop('messages', None)
result['response'] = model_output[-1]['content']
writer.write(result)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--model-path", type=str, required=True)
parser.add_argument("--model-name", type=str, required=True)
parser.add_argument("--model-base", type=str, default=None)
parser.add_argument("--dataset", type=str, required=True, help="")
parser.add_argument("--image-folder", type=str, default="")
parser.add_argument("--input-file", type=str, required=True, help="Path to the question file")
parser.add_argument("--output-dir", type=str, default="./result")
parser.add_argument("--temperature", type=float, default=0.75)
parser.add_argument("--top_p", type=float, default=0.9)
parser.add_argument("--num_beams", type=int, default=1)
parser.add_argument("--max-frames", type=int, default=32)
parser.add_argument("--max-pixels", type=int, default=256*28*28)
parser.add_argument("--over_write", type=int, default=0, help="Whether to overwrite the output directory")
parser.add_argument("--split", type=int, default=1)
parser.add_argument("--all", type=int, default=1)
parser.add_argument("--prune", type=str, default=None)
parser.add_argument("--resize", action="store_true", default=False)
args = parser.parse_args()
if args.image_folder == "None":
args.image_folder = ""
eval_model(args)