Emu2 / demo /meta.py
ryanzhangfan's picture
only resize large images
59624a6
raw
history blame
9.24 kB
# -*- coding: utf-8 -*-
# ===========================================================================================
#
# Copyright (c) Beijing Academy of Artificial Intelligence (BAAI). All rights reserved.
#
# Author : Fan Zhang
# Email : zhangfan@baai.ac.cn
# Institute : Beijing Academy of Artificial Intelligence (BAAI)
# Create On : 2023-12-12 02:54
# Last Modified : 2023-12-20 16:41
# File Name : meta.py
# Description :
#
# ===========================================================================================
import base64
from dataclasses import dataclass, field
import io
from enum import Enum
from PIL import Image
from typing import List, Tuple
import cv2
import numpy as np
from .constants import EVA_IMAGE_SIZE, GRD_SYMBOL, BOP_SYMBOL, EOP_SYMBOL, BOO_SYMBOL, EOO_SYMBOL
from .constants import DEFAULT_VIDEO_TOKEN, DEFAULT_EOS_TOKEN, USER_TOKEN, ASSISTANT_TOKEN, FAKE_VIDEO_END_TOKEN
from .utils import gen_id, frontend_logger as logging
class Role(Enum):
UNKNOWN = 0,
USER = 1,
ASSISTANT = 2,
class DataType(Enum):
UNKNOWN = 0,
TEXT = 1,
IMAGE = 2,
GROUNDING = 3,
VIDEO = 4,
ERROR = 5,
@dataclass
class DataMeta:
datatype: DataType = DataType.UNKNOWN
text: str = None
image: Image.Image = None
mask: Image.Image = None
coordinate: List[int] = None
frames: List[Image.Image] = None
stack_frame: Image.Image = None
@property
def grounding(self):
return self.coordinate is not None
@property
def text_str(self):
return self.text
@property
def image_str(self):
return self.image2str(self.image)
@property
def video_str(self):
ret = f'<div style="overflow:scroll"><b>[VIDEO]</b></div>{self.image2str(self.stack_frame)}'
return ret
@property
def grounding_str(self):
ret = ""
if self.text is not None:
ret += f'<div style="overflow:scroll"><b>[PHRASE]</b>{self.text}</div>'
ret += self.image2str(self.mask)
if self.image is not None:
ret += self.image2str(self.image)
return ret
def image2str(self, image):
buf = io.BytesIO()
image.save(buf, format="WEBP")
i_str = base64.b64encode(buf.getvalue()).decode()
return f'<div style="float:left"><img src="data:image/png;base64, {i_str}"></div>'
def format_chatbot(self):
match self.datatype:
case DataType.TEXT | DataType.ERROR:
return self.text_str
case DataType.IMAGE:
return self.image_str
case DataType.VIDEO:
return self.video_str
case DataType.GROUNDING:
return self.grounding_str
case _:
return ""
def format_prompt(self) -> List[str | Image.Image]:
match self.datatype:
case DataType.TEXT:
return [self.text]
case DataType.IMAGE:
return [self.image]
case DataType.VIDEO:
return [DEFAULT_VIDEO_TOKEN] + self.frames + [FAKE_VIDEO_END_TOKEN]
case DataType.GROUNDING:
ret = []
if self.text is not None:
ret.append(f"{BOP_SYMBOL}{self.text}{EOP_SYMBOL}")
ret += [BOO_SYMBOL, self.mask, EOO_SYMBOL]
if self.image is not None:
ret.append(self.image)
return ret
case _:
return []
def __str__(self):
s = ""
if self.text is not None:
s += f"T:{self.text}"
if self.image is not None:
w, h = self.image.size
s += f"[I:{h}x{w}]"
if self.coordinate is not None:
l, t, r, b = self.coordinate
s += f"[C:({l:03d},{t:03d}),({r:03d},{b:03d})]"
if self.frames is not None:
w, h = self.frames[0].size
s += f"[V:{len(self.frames)}x{h}x{w}]"
return s
@classmethod
def build(cls, text=None, image=None, coordinate=None, frames=None, is_error=False, *, resize: bool = True):
ins = cls()
ins.text = text if text != "" else None
# ins.image = cls.resize(image, force=resize)
ins.image = image
ins.coordinate = cls.fix(coordinate)
# ins.frames = cls.resize(frames, force=resize)
ins.frames = frames
if is_error:
ins.datatype = DataType.ERROR
elif coordinate is not None:
ins.datatype = DataType.GROUNDING
ins.draw_box()
elif image is not None:
ins.datatype = DataType.IMAGE
elif text is not None:
ins.datatype = DataType.TEXT
else:
ins.datatype = DataType.VIDEO
ins.stack()
return ins
@classmethod
def fix(cls, coordinate):
if coordinate is None:
return None
l, t, r, b = coordinate
l = min(EVA_IMAGE_SIZE, max(0, l))
t = min(EVA_IMAGE_SIZE, max(0, t))
r = min(EVA_IMAGE_SIZE, max(0, r))
b = min(EVA_IMAGE_SIZE, max(0, b))
return min(l, r), min(t, b), max(l, r), max(t, b)
@classmethod
def resize(cls, image: Image.Image | List[Image.Image] | None, *, force: bool = True):
if image is None:
return None
if not force:
return image
if isinstance(image, Image.Image):
image = [image]
for idx, im in enumerate(image):
w, h = im.size
if w < EVA_IMAGE_SIZE or h < EVA_IMAGE_SIZE:
continue
if w < h:
h = int(EVA_IMAGE_SIZE / w * h)
w = EVA_IMAGE_SIZE
else:
w = int(EVA_IMAGE_SIZE / h * w)
h = EVA_IMAGE_SIZE
image[idx] = im.resize((w, h))
return image if len(image) > 1 else image[0]
def draw_box(self):
left, top, right, bottom = self.coordinate
mask = np.zeros((EVA_IMAGE_SIZE, EVA_IMAGE_SIZE, 3), dtype=np.uint8)
mask = cv2.rectangle(mask, (left, top), (right, bottom), (255, 255, 255), 3)
self.mask = Image.fromarray(mask)
def stack(self):
w, h = self.frames[0].size
n = len(self.frames)
stack_frame = Image.new(mode="RGB", size=(w*n, h))
for idx, f in enumerate(self.frames):
stack_frame.paste(f, (idx*w, 0))
self.stack_frame = stack_frame
class ConvMeta:
def __init__(self):
self.system: str = "You are a helpful assistant, dedicated to delivering comprehensive and meticulous responses."
self.message: List[Tuple[Role, DataMeta]] = []
self.log_id: str = gen_id()
logging.info(f"{self.log_id}: create new round of chat")
def append(self, r: Role, p: DataMeta):
logging.info(f"{self.log_id}: APPEND [{r.name}] prompt element, type: {p.datatype.name}, message: {p}")
self.message.append((r, p))
def format_chatbot(self):
ret = []
for r, p in self.message:
cur_p = p.format_chatbot()
if r == Role.USER:
ret.append((cur_p, None))
else:
ret.append((None, cur_p))
return ret
def format_prompt(self):
ret = []
has_coor = False
for _, p in self.message:
has_coor |= (p.datatype == DataType.GROUNDING)
ret += p.format_prompt()
if has_coor:
ret.insert(0, GRD_SYMBOL)
logging.info(f"{self.log_id}: format generation prompt: {ret}")
return ret
def format_chat(self):
ret = [self.system]
prev_r = None
for r, p in self.message:
if prev_r != r:
if prev_r == Role.ASSISTANT:
ret.append(f"{DEFAULT_EOS_TOKEN}{USER_TOKEN}: ")
elif prev_r is None:
ret.append(f" {USER_TOKEN}: ")
else:
ret.append(f" {ASSISTANT_TOKEN}: ")
ret += p.format_prompt()
prev_r = r
else:
ret += p.format_prompt()
ret.append(f" {ASSISTANT_TOKEN}:")
logging.info(f"{self.log_id}: format chat prompt: {ret}")
return ret
def clear(self):
logging.info(f"{self.log_id}: clear chat history, end current chat round.")
del self.message
self.message = []
self.log_id = gen_id()
def pop(self):
if self.has_gen:
logging.info(f"{self.log_id}: pop out previous generation / chat result")
self.message.pop()
def pop_error(self):
new_message = []
for r, p in self.message:
if p.datatype == DataType.ERROR:
logging.info(f"{self.log_id}: pop error message: {p.text_str}")
else:
new_message.append((r, p))
del self.message
self.message = new_message
@property
def has_gen(self):
if len(self.message) == 0:
return False
if self.message[-1][0] == Role.USER:
return False
return True