VLM-Lens / src /models /base.py
marstin's picture
[martin-dev] add demo v1 test
d425e71
"""base.py.
Provides the common classes used such as the ModelSelection enum as well as the
abstract base class for models.
"""
import io
import logging
import os
import sqlite3
from abc import ABC, abstractmethod
from collections.abc import Iterator
from typing import Callable, List, Optional, TypedDict
import torch
import tqdm
from PIL import Image
from transformers import AutoProcessor
from transformers.feature_extraction_utils import BatchFeature
from .config import Config
class ModelInput(TypedDict):
"""Definition for the general model input dictionary."""
image: str | Image.Image
prompt: str
label: Optional[str]
data: BatchFeature
row_id: Optional[str]
class ModelBase(ABC):
"""Provides an abstract base class for everything to implement."""
def __init__(self, config: Config) -> None:
"""Initialization of the model base class.
Args:
config (Config): Parsed config.
"""
self.model_path = config.model_path
self.config = config
# log the modules -- note that this causes an exit
if self.config.log_named_modules:
self._log_named_modules()
exit(0)
# load the specific model
logging.debug(
f'Loading model {self.config.architecture.value}; {self.model_path}'
)
self._load_specific_model()
# load the processor
self._init_processor()
def _log_named_modules(self) -> None:
"""Logs the named modules based on the loaded model."""
file_path = 'logs/' + self.model_path + '.txt'
directory_path = os.path.dirname(file_path)
# if the path exists to the file, don't load the model again
if os.path.isfile(file_path):
logging.debug(f'Named modules are cached in {file_path}')
return
# in which case, we first load the model, then output its modules
self._load_specific_model()
# otherwise, we log the output to that file, and creating directories
# as needed
if not os.path.exists(directory_path):
os.makedirs(directory_path)
with open(file_path, 'w') as output_file:
output_file.writelines(
[f'{name}\n' for name, _ in self.model.named_modules()]
)
@abstractmethod
def _load_specific_model(self) -> None:
"""Abstract method that loads the specific model."""
pass
def _init_processor(self) -> None:
"""Initialize the self.processor by loading from the path."""
self.processor = AutoProcessor.from_pretrained(self.model_path)
def _generate_state_hook(self,
name: str,
model_input: ModelInput
) -> Callable[[torch.nn.Module, tuple, torch.Tensor], None]:
"""Generates the state hook depending on the embedding type.
Args:
name (str): The module name.
model_input (ModelInput): The input dictionary
containing the image path, prompt, label (if applicable) and
the data itself.
Returns:
hook function: The hook function to return.
"""
image_path, prompt = model_input['image'], model_input['prompt']
label = model_input.get('label', None)
row_id = model_input.get('row_id', None)
# Modify image path to be an absolute path if necessary
if isinstance(image_path, str) and image_path != self.config.NO_IMG_PROMPT:
image_path = os.path.abspath(image_path)
# this image path should already exist, error out if someone isn't
# properly providing an image path
assert os.path.exists(image_path)
def generate_states_hook(module: torch.nn.Module, input: tuple, output: torch.Tensor) -> None:
"""Hook handle function that saves the embedding output to a tensor.
This tensor will be saved within a SQL database, according to the
connection that was initialized previously.
Args:
module (torch.nn.Module): The module that save its hook on.
input (tuple): The input used.
output (torch.Tensor): The embeddings to save.
"""
if not isinstance(output, torch.Tensor):
logging.warning(f'Output type of {str(type(module))} is not a tensor, skipped.')
return
cursor = self.connection.cursor()
# Convert the tensor to a binary blob
tensor_blob = io.BytesIO()
# It currently averages the output across the sequence length dimension, i.e., mean pooling
# WARNING: When contributing new models, ensure that dim 1 is always the sequence length dimension
final_output = getattr(output, self.config.pooling_method)(dim=1) if hasattr(
self.config, 'pooling_method') and hasattr(output, self.config.pooling_method) else output
output_dim = final_output.shape[-1]
torch.save(final_output, tensor_blob)
# Insert the tensor into the table
cursor.execute(f"""
INSERT INTO {self.config.DB_TABLE_NAME}
(name, architecture, image_path, image_id, prompt, label, layer, pooling_method, tensor_dim, tensor)
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?);
""", (
self.model_path,
self.config.architecture.value,
image_path if isinstance(image_path, str) else None,
row_id,
prompt,
label,
name,
self.config.pooling_method if hasattr(self.config, 'pooling_method') else None,
output_dim,
tensor_blob.getvalue())
)
self.connection.commit()
logging.debug(
f'Ran hook and saved tensor for {image_path} using prompt '
f'{prompt} on layer {name}.'
)
return generate_states_hook
def _register_module_hooks(self,
model_input: ModelInput
) -> List[torch.utils.hooks.RemovableHandle]:
"""Register the generated hook function to the modules in the config.
At the same time, we need to add in the image path itself and the prompt
which will be used for the database input.
Args:
model_input (ModelInput): The input dictionary
containing the image path, prompt, label (if applicable) and
the data itself.
Raises:
RuntimeError: Calls a runtime error if no hooks were registered
Returns:
List[torch.utils.hooks.RemovableHandle]: A list of handles that one
can remove after the forward pass.
"""
logging.debug(
f'Registering module hook for {model_input["image"]} using prompt "{model_input["prompt"]}"'
)
# a list of hooks to remove after the forward pass
hooks = []
# for each module, register the state hook and save the output to database
for name, module in self.model.named_modules():
if self.config.matches_module(name):
hooks.append(module.register_forward_hook(
self._generate_state_hook(name, model_input)
))
logging.debug(f'Registered hook to {name}')
if len(hooks) == 0:
raise RuntimeError(
'No hooks were registered. Double-check the configured modules.'
)
return hooks
def _forward(self, data: BatchFeature) -> None:
"""Given some input data, performs a single forward pass.
This function itself can be overriden, while _hook_and_eval
should be left in tact.
Args:
data (BatchFeature): The given data tensor.
"""
data.to(self.config.device)
with torch.no_grad():
_ = self.model(**data)
logging.debug('Completed forward pass...')
def _hook_and_eval(self, model_input: ModelInput) -> None:
"""Given some input, performs a single forward pass.
Args:
model_input (ModelInput): The given input dictionary.
"""
logging.debug('Starting forward pass')
self.model.eval()
# now set up the modules to register the hook to
hooks = self._register_module_hooks(model_input)
# then ensure that the data is correct
self._forward(model_input['data'])
for hook in hooks:
hook.remove()
logging.debug('Unregistered all hooks..')
def _initialize_db(self) -> None:
"""Initializes a database based on config."""
# Connect to the database, creating it if it doesn't exist
self.connection = sqlite3.connect(self.config.output_db)
logging.debug(f'Database created at {self.config.output_db}')
cursor = self.connection.cursor()
# Create a table
cursor.execute(
f"""
CREATE TABLE IF NOT EXISTS {self.config.DB_TABLE_NAME} (
id INTEGER PRIMARY KEY,
name TEXT NOT NULL,
architecture TEXT NOT NULL,
timestamp DATETIME DEFAULT CURRENT_TIMESTAMP,
image_path TEXT NULL,
image_id INTEGER NULL,
prompt TEXT NOT NULL,
label TEXT NULL,
layer TEXT NOT NULL,
pooling_method TEXT NULL,
tensor_dim INTEGER NOT NULL,
tensor BLOB NOT NULL
);
"""
)
def _cleanup(self) -> None:
"""Cleanups the database by closing the connection."""
self.connection.close()
def _generate_processor_output(self, prompt: str, img_path: str | Image.Image) -> dict:
"""Generate the processor outputs from the prompt and image path.
Args:
prompt (str): The generated prompt string with the input text and
the image labels.
img_path (str | Image.Image): The specified input image path or image object.
Returns:
dict: The corresponding processor output per image and prompt.
"""
data = {
'text': prompt,
'return_tensors': 'pt'
}
if img_path:
img = Image.open(img_path) if isinstance(img_path, str) else img_path
data['images'] = [img.convert('RGB')]
return self.processor(**data)
def _generate_prompt(self, prompt: str, add_generation_prompt: bool = True, has_images: bool = False) -> str:
"""Generates the prompt string with the input messages.
TODO: move `add_generation_prompt` to the config.
[Note from Martin] I'd argue that we should keep it as a parameter here
since in gradio we want to hack these parameters a bit.
Args:
prompt (str): The input prompt string.
add_generation_prompt (bool): Whether to add a start token of a bot
response.
has_images (bool): Whether the model has images or not.
Returns:
str: The generated prompt with the input text and the image labels.
"""
logging.debug('Loading data...')
# build the input dict for the chat template
input_msgs_formatted = [{
'role': 'user',
'content': []
}]
# add the image if it exists
if self.config.has_images() or has_images:
input_msgs_formatted[0]['content'].append({
'type': 'image'
})
# add the prompt if it exists
if prompt:
input_msgs_formatted[0]['content'].append({
'type': 'text',
'text': prompt
})
# apply the chat template to get the prompt
return self.processor.apply_chat_template(
input_msgs_formatted,
add_generation_prompt=add_generation_prompt
)
def _load_input_data(self) -> Iterator[ModelInput]:
"""From a configuration, loads the input image and text data.
For each prompt and input image, create a separate batch feature that
will be ran separately and saved separately within the database.
Yields:
List[ModelInput]: List of input data, this input data is made of
a tuple of strings (first an image path, then a prompt) and
a batch feature which is either a torch.Tensor or a dictionary.
"""
# by default use the processor, which may not exist for each model
logging.debug('Generating embeddings through its processor...')
if self.config.dataset:
# Use the dataset to load input data, which includes (id, prompt, image_path)
for row in self.config.dataset:
prompt = self._generate_prompt(row['prompt'])
data = self._generate_processor_output(
prompt=prompt,
img_path=row['image']
)
yield {
'image': row['image'],
'prompt': row['prompt'],
'label': row['label'] if 'label' in self.config.dataset.column_names else None,
'data': data,
'row_id': row['id'],
}
else:
if not self.config.has_images():
yield {
'image': self.config.NO_IMG_PROMPT, # TODO: Check this?
'prompt': self.config.prompt,
'data': self._generate_processor_output(
prompt=self._generate_prompt(),
img_path=None
)
}
else:
prompt = self._generate_prompt(self.config.prompt)
for img_path in self.config.image_paths:
data = self._generate_processor_output(
prompt=prompt,
img_path=img_path
)
yield {
'image': img_path,
'prompt': self.config.prompt,
'data': data
}
@property
def _data_size(self) -> int:
"""Returns the total number of data points.
Returns:
int: The total number of data points.
"""
if self.config.dataset:
return len(self.config.dataset)
else:
if not self.config.has_images():
return 1
else:
return len(self.config.image_paths)
def run(self) -> None:
"""Get the hidden states from the model and saving them."""
# let's first initialize a database connection
self._initialize_db()
# then convert to gpu
self.model.to(self.config.device)
# then reset the starting point in tracking maximum GPU memory, if using cuda
if self.config.device.type == 'cuda':
torch.cuda.reset_peak_memory_stats(self.config.device)
# then run everything else
for item in tqdm.tqdm(self._load_input_data(), desc='Running forward hooks on data', total=self._data_size):
self._hook_and_eval(item)
# then output peak memory usage, if using cuda
if self.config.device.type == 'cuda':
logging.debug(f'Peak GPU memory allocated: {torch.cuda.max_memory_allocated(self.config.device) / 1e6:.2f} MB')
# finally clean up, closing database connection, etc.
self._cleanup()