model-editing / app.py
Charles Lin
All algs except KE working.
8335d0c
import streamlit as st
import pandas as pd
import time
import copy
import importlib
from torch.cuda import is_available as use_cuda
import algs
import config
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import utils
EDIT_ALGS = [
"MEND: Model editor networks using gradient decomposition",
"SERAC: Semi-parametric editing with a retrieval-augmented counterfactual model",
"ENN: Editable neural networks",
"KE: KnowledgeEditor",
"FT: Fine-tuning",
"LU: Lookup Cache",
]
def get_alg_class(alg_abbrv):
alg_module = importlib.import_module(f"algs.{alg_abbrv.lower()}")
alg_class = getattr(alg_module, alg_abbrv.upper())
return alg_class
def load_editable_model(alg_abbrv):
alg_module = importlib.import_module(f"algs.{alg_abbrv.lower()}")
alg_class = getattr(alg_module, alg_abbrv.upper())
st.session_state.config = getattr(config, f"{alg_abbrv.lower()}_config")
with st.spinner('Loading model...'):
st.session_state.editable_model = alg_class(
st.session_state.model,
st.session_state.config,
lambda: copy.deepcopy(st.session_state.model),
).eval()
if "archive" in st.session_state.config:
archive, st.session_state.config.archive = utils.load_archive(str(st.session_state.config.archive))
print(f"Loading archive from {st.session_state.config.archive}")
st.session_state.editable_model.load_state_dict(archive["model"])
def generate(ids):
output_ids = st.session_state.editable_model.generate(input_ids=ids, max_new_tokens=20, min_length=1,
num_return_sequences=1, num_beams=3)
return st.session_state.tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]
def reset():
st.session_state.edits.drop(st.session_state.edits.index, inplace=True)
st.session_state.model_outputs.drop(st.session_state.edits.index, inplace=True)
selected_alg = st.session_state.alg_selector
alg_abbrv = selected_alg[:selected_alg.index(":")]
load_editable_model(alg_abbrv)
def apply_edit():
st.session_state.edits.loc[len(st.session_state.edits)] = [str(edit_input), str(edit_label)]
with st.spinner("Editing model..."):
input_ids = st.session_state.tokenizer(str(edit_input), return_tensors="pt")["input_ids"].to(st.session_state.device)
label_ids = st.session_state.tokenizer(str(edit_label), return_tensors="pt")["input_ids"].to(st.session_state.device)
edit_sample = {"input_ids": input_ids, "labels": label_ids}
st.session_state.editable_model, _ = st.session_state.editable_model.edit(edit_sample, detach_history=True)
def sample_model():
input_str = str(test_input)
with st.spinner('Generating completion...'):
encoding = st.session_state.tokenizer(input_str, return_tensors="pt")
ids = encoding["input_ids"].to(st.session_state.device)
model_output = generate(ids)
n_edits = len(st.session_state.edits)
alg_name = st.session_state.alg_selector
alg_abbrv = alg_name[:alg_name.index(":")]
st.session_state.model_outputs.loc[len(st.session_state.model_outputs)] = [input_str, model_output, n_edits, alg_abbrv]
################################
#### Backend initialization ####
################################
if "init" not in st.session_state:
st.session_state.edits = pd.DataFrame([], columns=["Edit input", "Edit label"])
st.session_state.model_outputs = pd.DataFrame([], columns=["Input", "Output", "N edits", "Alg"])
st.session_state.init = True
st.session_state.device = "cpu" # "cuda" if use_cuda() else "cpu"
with st.spinner('Loading model...'):
st.session_state.tokenizer = AutoTokenizer.from_pretrained("google/t5-large-ssm-nq")
st.session_state.model = AutoModelForSeq2SeqLM.from_pretrained("google/t5-large-ssm-nq").to(st.session_state.device).eval()
# There is a "Loading model..." spinner in load_editable_model
alg_abbrv = "MEND" # Default initial alg of dropdown selector
load_editable_model(alg_abbrv)
########################
#### Interface code ####
########################
st.title("Language Model Editing")
st.markdown("**Note: this HF space is currently under development and doesn't actually work yet!**")
st.markdown("The goal of this demo is to give you a sense of the *abilities* and *limitations* of existing methods for **editing** pre-trained language models. **Model editing** algorithms use a single input-output pair to update a pre-trained model's behavior for that input (and ideally, related inputs).")
st.markdown("This demo uses a [T5-large](https://huggingface.co/google/t5-large-ssm-nq) model fine-tuned on [Natural Questions](https://arxiv.org/pdf/2002.08910.pdf) as the base pre-trained model.")
st.write("You can choose from a variety of algorithms for model editing in the dropdown below. At the bottom of the page, you can query the model for whatever input you want before/after editing.")
st.markdown("***")
col1, col2 = st.columns([5,1])
with col1:
alg_selector = st.selectbox("Editing algorithm:", EDIT_ALGS, key="alg_selector", on_change=reset)
with col2:
st.text("ㅤ")
st.button("Clear edits", on_click=reset)
st.write("Edits applied so far:")
st.table(st.session_state.edits)
col1, col2, col3 = st.columns([3, 2, 1])
with col1:
edit_input = st.text_input("Edit input:", placeholder="e.g., 'What is the tallest mountain on Earth?'")
with col2:
edit_label = st.text_input("Edit target:", placeholder="e.g., 'Denali'")
with col3:
st.text("ㅤ")
edit_button = st.button("Apply edit", on_click=apply_edit)
st.markdown("***")
if len(st.session_state.edits) == 0:
title = "Input to sample from *unedited* model:"
else:
title = f"Input to sample from *edited* model:"
col1, col2 = st.columns([5, 1])
with col1:
test_input = st.text_input(title, placeholder="e.g., 'What is the earth's tallest mountain?'")
with col2:
st.text("ㅤ")
generate_button = st.button("Generate", on_click=sample_model)
st.write("Model generation history:")
st.table(st.session_state.model_outputs)