wissamantoun's picture
Update app.py
3b165bd verified
raw
history blame contribute delete
No virus
22.9 kB
import json
from typing import Tuple
import numpy as np
import pandas as pd
import plotly.express as px
import plotly.figure_factory as ff
import plotly.graph_objects as go
import streamlit as st
from plotly.subplots import make_subplots
from exp_utils import MODELS
from visualize_utils import viridis_rgb
st.set_page_config(
page_title="Results Viewer",
page_icon="📊",
initial_sidebar_state="expanded",
layout="wide",
)
MODELS_SIZE_MAPPING = {k: v["model_size"] for k, v in MODELS.items()}
MODELS_FAMILY_MAPPING = {k: v["model_family"] for k, v in MODELS.items()}
MODEL_FAMILES = set([model["model_family"] for model in MODELS.values()])
Q_W_MODELS = [
"llama-7b",
"llama-2-7b",
"llama-13b",
"llama-2-13b",
"llama-30b",
"llama-65b",
"llama-2-70b",
]
Q_W_MODELS = [f"{model}_quantized" for model in Q_W_MODELS] + [
f"{model}_watermarked" for model in Q_W_MODELS
]
MODEL_NAMES = list(MODELS.keys()) + Q_W_MODELS
MODEL_NAMES_SORTED_BY_NAME_AND_SIZE = sorted(
MODEL_NAMES,
key=lambda x: (
MODELS[x.replace("_quantized", "").replace("_watermarked", "")]["model_family"],
MODELS[x.replace("_quantized", "").replace("_watermarked", "")]["model_size"],
),
)
MODEL_NAMES_SORTED_BY_SIZE = sorted(
MODEL_NAMES,
key=lambda x: (
MODELS[x.replace("_quantized", "").replace("_watermarked", "")]["model_size"],
MODELS[x.replace("_quantized", "").replace("_watermarked", "")]["model_family"],
),
)
# sort MODELS_SIZE_MAPPING by value then by key
MODELS_SIZE_MAPPING = {
k: v
for k, v in sorted(MODELS_SIZE_MAPPING.items(), key=lambda item: (item[1], item[0]))
}
MODELS_SIZE_MAPPING_LIST = list(MODELS_SIZE_MAPPING.keys())
CHAT_MODELS = [
x
for x in MODEL_NAMES_SORTED_BY_NAME_AND_SIZE
if MODELS[x.replace("_quantized", "").replace("_watermarked", "")]["is_chat"]
]
def clean_dataframe(df: pd.DataFrame) -> pd.DataFrame:
# remove all columns that have "_loss" and "_runtime" in them
words_to_remove = [
"epoch",
"loss",
"runtime",
"samples_per_second",
"steps_per_second",
"samples",
"results_dir",
]
df = df.loc[
:,
~df.columns.str.contains("|".join(words_to_remove), case=False, regex=True),
]
# rename the rest of the columns by replacing "_roc_auc" with ""
df.columns = df.columns.str.replace("_roc_auc", "")
df.columns = df.columns.str.replace("eval_", "")
df["model_family"] = df["model_name"].apply(
lambda x: MODELS_FAMILY_MAPPING[
x.replace("_quantized", "").replace("_watermarked", "")
]
)
# create a dict with the model_name and the model_family
model_family_dict = {
k: v
for k, v in zip(
df["model_name"].values.tolist(), df["model_family"].values.tolist()
)
}
# average the results over the 5 seeds for each model (seed column is exp_seed)
df_avg = df.groupby(["model_name"]).mean()
df_std = df.groupby(["model_name"]).std()
# remove the exp_seed column
df_avg = df_avg.drop(columns=["exp_seed"])
df_std = df_std.drop(columns=["exp_seed"])
df_avg["model_family"] = df_avg.index.map(model_family_dict)
df_std["model_family"] = df_std.index.map(model_family_dict)
df_avg["model_size"] = df_avg.index.map(
lambda x: MODELS_SIZE_MAPPING[
x.replace("_quantized", "").replace("_watermarked", "")
]
)
df_std["model_size"] = df_std.index.map(
lambda x: MODELS_SIZE_MAPPING[
x.replace("_quantized", "").replace("_watermarked", "")
]
)
# sort rows by model family then model size
df_avg = df_avg.sort_values(
by=["model_family", "model_size"], ascending=[True, True]
)
df_std = df_std.sort_values(
by=["model_family", "model_size"], ascending=[True, True]
)
availables_rows = [x for x in df_avg.columns if x in df_avg.index]
df_avg = df_avg.reindex(availables_rows)
availables_rows = [x for x in df_std.columns if x in df_std.index]
df_std = df_std.reindex(availables_rows)
df_avg["is_quantized"] = df_avg.index.str.contains("quantized")
df_avg["is_watermarked"] = df_avg.index.str.contains("watermarked")
df_std["is_quantized"] = df_std.index.str.contains("quantized")
df_std["is_watermarked"] = df_std.index.str.contains("watermarked")
return df_avg, df_std
def get_data(path) -> Tuple[pd.DataFrame, pd.DataFrame]:
df, df_std = clean_dataframe(pd.read_csv(path, index_col=0))
return df, df_std
def filter_df(
df: pd.DataFrame,
model_family_train: list,
model_family_test: list,
model_size_train: tuple,
model_size_test: tuple,
is_chat_train: bool,
is_chat_test: bool,
is_quantized_train: bool,
is_quantized_test: bool,
is_watermarked_train: bool,
is_watermarked_test: bool,
sort_by_size: bool,
split_chat_models: bool,
split_quantized_models: bool,
split_watermarked_models: bool,
filter_empty_col_row: bool,
is_debug: bool,
) -> pd.DataFrame:
# remove all columns and rows that have "pythia-70m" in the name
# filter rows
if is_debug:
st.write("No filters")
st.write(df)
df = df.loc[
(df["model_size"] >= model_size_train[0] * 1e9)
& (df["model_size"] <= model_size_train[1] * 1e9)
]
if is_debug:
st.write("Filter model size train")
st.write(df)
df = df.loc[df["model_family"].isin(model_family_train)]
if is_debug:
st.write("Filter model family train")
st.write(df)
if is_chat_train != "Both":
df = df.loc[df["is_chat"] == is_chat_train]
if is_debug:
st.write("Filter is chat train")
st.write(df)
if is_quantized_train != "Both":
df = df.loc[df["is_quantized"] == is_quantized_train]
if is_debug:
st.write("Filter is quantized train")
st.write(df)
if is_watermarked_train != "Both":
df = df.loc[df["is_watermarked"] == is_watermarked_train]
if is_debug:
st.write("Filter is watermark train")
st.write(df)
# filter columns
if is_debug:
st.write("No filters")
st.write(df)
columns_to_keep = []
for column in df.columns:
if (
column.replace("_quantized", "").replace("_watermarked", "")
in MODELS.keys()
):
model_size = MODELS[
column.replace("_quantized", "").replace("_watermarked", "")
]["model_size"]
if (
model_size >= model_size_test[0] * 1e9
and model_size <= model_size_test[1] * 1e9
):
columns_to_keep.append(column)
df = df[list(sorted(list(set(columns_to_keep))))]
if is_debug:
st.write("Filter model size test")
st.write(df)
# filter columns
columns_to_keep = []
for column in df.columns:
for model_family in model_family_test:
if (
model_family
== MODELS[column.replace("_quantized", "").replace("_watermarked", "")][
"model_family"
]
):
columns_to_keep.append(column)
df = df[list(sorted(list(set(columns_to_keep))))]
if is_debug:
st.write("Filter model family test")
st.write(df)
if is_chat_test != "Both":
# filter columns
columns_to_keep = []
for column in df.columns:
if (
MODELS[column.replace("_quantized", "").replace("_watermarked", "")][
"is_chat"
]
== is_chat_test
):
columns_to_keep.append(column)
df = df[list(sorted(list(set(columns_to_keep))))]
if is_debug:
st.write("Filter is chat test")
st.write(df)
if is_quantized_test != "Both":
# filter columns
columns_to_keep = []
for column in df.columns:
if "quantized" in column and is_quantized_test:
columns_to_keep.append(column)
elif "quantized" not in column and not is_quantized_test:
columns_to_keep.append(column)
df = df[list(sorted(list(set(columns_to_keep))))]
if is_debug:
st.write("Filter is quantized test")
st.write(df)
if is_watermarked_test != "Both":
# filter columns
columns_to_keep = []
for column in df.columns:
if "watermark" in column and is_watermarked_test:
columns_to_keep.append(column)
elif "watermark" not in column and not is_watermarked_test:
columns_to_keep.append(column)
df = df[list(sorted(list(set(columns_to_keep))))]
if is_debug:
st.write("Filter is watermark test")
st.write(df)
df = df.select_dtypes(include="number")
if is_debug:
st.write("Select dtypes to be only numbers")
st.write(df)
if sort_by_size:
columns_in = [x for x in MODEL_NAMES_SORTED_BY_SIZE if x in df.columns]
else:
columns_in = [x for x in MODEL_NAMES_SORTED_BY_NAME_AND_SIZE if x in df.columns]
df = df[columns_in]
if is_debug:
st.write("Sort columns")
st.write(df)
# sort rows by size according the MODELS_SIZE_MAPPING_LIST
if sort_by_size:
availables_rows = [x for x in MODEL_NAMES_SORTED_BY_SIZE if x in df.index]
df = df.reindex(availables_rows)
else:
availables_rows = [
x for x in MODEL_NAMES_SORTED_BY_NAME_AND_SIZE if x in df.index
]
df = df.reindex(availables_rows)
if is_debug:
st.write("Sort rows")
st.write(df)
if split_chat_models:
# put chat models at the end of the columns
chat_models = [x for x in CHAT_MODELS if x in df.columns]
# sort chat models by size
chat_models = sorted(chat_models, key=lambda x: MODELS[x]["model_size"])
df = df[[x for x in df.columns if x not in chat_models] + chat_models]
# put chat models at the end of the rows
chat_models = [x for x in CHAT_MODELS if x in df.index]
# sort chat models by size
chat_models = sorted(chat_models, key=lambda x: MODELS[x]["model_size"])
df = df.reindex([x for x in df.index if x not in chat_models] + chat_models)
if is_debug:
st.write("Split chat models")
st.write(df)
if split_quantized_models:
# put chat models at the end of the columns
quantized_models = [
x for x in Q_W_MODELS if x in df.columns and "quantized" in x
]
# sort chat models by size
quantized_models = sorted(
quantized_models,
key=lambda x: MODELS[
x.replace("_quantized", "").replace("_watermarked", "")
]["model_size"],
)
df = df[[x for x in df.columns if x not in quantized_models] + quantized_models]
# put chat models at the end of the rows
quantized_models = [x for x in Q_W_MODELS if x in df.index and "quantized" in x]
# sort chat models by size
quantized_models = sorted(
quantized_models,
key=lambda x: MODELS[
x.replace("_quantized", "").replace("_watermarked", "")
]["model_size"],
)
df = df.reindex(
[x for x in df.index if x not in quantized_models] + quantized_models
)
if split_watermarked_models:
# put chat models at the end of the columns
watermarked_models = [
x for x in Q_W_MODELS if x in df.columns and "watermarked" in x
]
# sort chat models by size
watermarked_models = sorted(
watermarked_models,
key=lambda x: MODELS[
x.replace("_quantized", "").replace("_watermarked", "")
]["model_size"],
)
df = df[
[x for x in df.columns if x not in watermarked_models] + watermarked_models
]
# put chat models at the end of the rows
watermarked_models = [
x for x in Q_W_MODELS if x in df.index and "watermarked" in x
]
# sort chat models by size
watermarked_models = sorted(
watermarked_models,
key=lambda x: MODELS[
x.replace("_quantized", "").replace("_watermarked", "")
]["model_size"],
)
df = df.reindex(
[x for x in df.index if x not in watermarked_models] + watermarked_models
)
if is_debug:
st.write("Split chat models")
st.write(df)
if filter_empty_col_row:
# remove all for which the row and column are Nan
df = df.dropna(axis=0, how="all")
df = df.dropna(axis=1, how="all")
return df
df, df_std = get_data("./deberta_results.csv")
df_q_w, df_std_q_w = get_data("./results_qantized_watermarked.csv")
df = df.merge(
df_q_w[
df_q_w.columns[
df_q_w.columns.str.contains("quantized|watermarked", case=False, regex=True)
]
],
how="outer",
left_index=True,
right_index=True,
)
df_std = df_std.merge(
df_std_q_w[
df_std_q_w.columns[
df_std_q_w.columns.str.contains(
"quantized|watermarked", case=False, regex=True
)
]
],
how="outer",
left_index=True,
right_index=True,
)
df.columns = df.columns.str.replace("_y", "", regex=True)
df_std.columns = df_std.columns.str.replace("_y", "", regex=True)
df = df.drop(columns=["is_quantized_x", "is_watermarked_x"])
df.update(df_q_w)
df_std.update(df_std_q_w)
df["is_chat"].fillna(False, inplace=True)
df_std["is_chat"].fillna(False, inplace=True)
df["is_watermarked"].fillna(False, inplace=True)
df_std["is_watermarked"].fillna(False, inplace=True)
df["is_quantized"].fillna(False, inplace=True)
df_std["is_quantized"].fillna(False, inplace=True)
with open("./ood_results.json", "r") as f:
ood_results = json.load(f)
ood_results = pd.DataFrame(ood_results)
ood_results = ood_results.set_index("model_name")
ood_results = ood_results.drop(
columns=["exp_name", "accuracy", "f1", "precision", "recall"]
)
ood_results.columns = ["seed", "Adversarial"]
ood_results_avg = ood_results.groupby(["model_name"]).mean()
ood_results_std = ood_results.groupby(["model_name"]).std()
st.write(
"""### Results Viewer 👇
## From Text to Source: Results in Detecting Large Language Model-Generated Content
### Wissam Antoun, Benoît Sagot, Djamé Seddah
##### ALMAnaCH, Inria
##### Paper: [https://arxiv.org/abs/2309.13322](https://arxiv.org/abs/2309.13322)
"""
)
# filters
show_diff = st.sidebar.checkbox("Show Diff", value=False)
sort_by_size = st.sidebar.checkbox("Sort by size", value=True)
split_chat_models = st.sidebar.checkbox("Split chat models", value=True)
split_quantized_models = st.sidebar.checkbox("Split quantized models", value=True)
split_watermarked_models = st.sidebar.checkbox("Split watermarked models", value=True)
add_mean = st.sidebar.checkbox("Add mean", value=False)
show_std = st.sidebar.checkbox("Show std", value=False)
filter_empty_col_row = st.sidebar.checkbox("Filter empty col/row", value=True)
model_size_train = st.sidebar.slider(
"Train Model Size in Billion", min_value=0, max_value=100, value=(0, 100), step=1
)
model_size_test = st.sidebar.slider(
"Test Model Size in Billion", min_value=0, max_value=100, value=(0, 100), step=1
)
is_chat_train = st.sidebar.selectbox("(Train) Is Chat?", [True, False, "Both"], index=2)
is_chat_test = st.sidebar.selectbox("(Test) Is Chat?", [True, False, "Both"], index=2)
is_quantized_train = st.sidebar.selectbox(
"(Train) Is Quantized?", [True, False, "Both"], index=1
)
is_quantized_test = st.sidebar.selectbox(
"(Test) Is Quantized?", [True, False, "Both"], index=1
)
is_watermarked_train = st.sidebar.selectbox(
"(Train) Is Watermark?", [True, False, "Both"], index=1
)
is_watermarked_test = st.sidebar.selectbox(
"(Test) Is Watermark?", [True, False, "Both"], index=1
)
model_family_train = st.sidebar.multiselect(
"Model Family Train",
MODEL_FAMILES,
default=MODEL_FAMILES,
)
model_family_test = st.sidebar.multiselect(
"Model Family Test",
list(MODEL_FAMILES) + ["Adversarial"],
default=MODEL_FAMILES,
)
show_values = st.sidebar.checkbox("Show Values", value=False)
add_adversarial = False
if "Adversarial" in model_family_test:
model_family_test.remove("Adversarial")
add_adversarial = True
sort_by_adversarial = False
if add_adversarial:
sort_by_adversarial = st.sidebar.checkbox("Sort by adversarial", value=False)
if st.sidebar.checkbox("Use default color scale", value=False):
color_scale = "Viridis_r"
else:
color_scale = viridis_rgb
is_debug = st.sidebar.checkbox("Debug", value=False)
if show_std:
selected_df = df_std.copy()
else:
selected_df = df.copy()
filtered_df = filter_df(
selected_df,
model_family_train,
model_family_test,
model_size_train,
model_size_test,
is_chat_train,
is_chat_test,
is_quantized_train,
is_quantized_test,
is_watermarked_train,
is_watermarked_test,
sort_by_size,
split_chat_models,
split_quantized_models,
split_watermarked_models,
filter_empty_col_row,
is_debug,
)
if show_diff:
# get those 3 columns {'model_size', 'model_family', 'is_chat'}
diag = filtered_df.values.diagonal()
filtered_df = filtered_df.sub(diag, axis=1)
# subtract each row by the diagonal
if add_adversarial:
if show_diff:
index = filtered_df.index
ood_results_avg = ood_results_avg.loc[index]
filtered_df = filtered_df.join(ood_results_avg.sub(diag, axis=0))
else:
filtered_df = filtered_df.join(ood_results_avg)
if add_mean:
col_mean = filtered_df.mean(axis=1)
row_mean = filtered_df.mean(axis=0)
diag = filtered_df.values.diagonal()
filtered_df["mean"] = col_mean
filtered_df.loc["mean"] = row_mean
filtered_df = filtered_df * 100
filtered_df = filtered_df.round(0)
# sort by the column called Adversarial
if sort_by_adversarial:
filtered_df = filtered_df.sort_values(by=["Adversarial"], ascending=False)
# check if the df has columns and rows
if filtered_df.shape[0] == 0:
st.write("No results found")
st.stop()
if filtered_df.shape[1] == 0:
st.write("No results found")
st.stop()
fig = px.imshow(
filtered_df.values,
x=list(filtered_df.columns),
y=list(filtered_df.index),
color_continuous_scale=color_scale,
contrast_rescaling=None,
text_auto=show_values,
aspect="auto",
)
# width = st.sidebar.text_input("Width", "1920")
# height = st.sidebar.text_input("Height", "1080")
# scale = st.sidebar.text_input("Scale", "1.0")
# margin = st.sidebar.text_input("Margin[l,r,b,t]", "200,100,100,100")
fig.update_traces(textfont_size=9)
fig.update_layout(
xaxis={"side": "top"},
yaxis={"side": "left"},
# margin=dict(
# l=int(margin.split(",")[0]),
# r=int(margin.split(",")[1]),
# b=int(margin.split(",")[2]),
# t=int(margin.split(",")[3]),
# ),
font=dict(size=10),
)
fig.update_xaxes(tickangle=45)
fig.update_xaxes(tickmode="linear")
fig.update_yaxes(tickmode="linear")
# change the font in the heatmap
st.plotly_chart(fig, use_container_width=True)
# if st.sidebar.button("save", key="save"):
# fig.write_image(
# "fig1.pdf",
# width=int(width),
# height=int(height),
# validate=True,
# scale=float(scale),
# )
# plot the col mean vs model size
if add_mean and not show_diff:
# check if any of the chat models are in the filtered df columns and index
if len([x for x in CHAT_MODELS if x in filtered_df.columns]) > 0 or len(
[x for x in CHAT_MODELS if x in filtered_df.index]
):
st.warning(
"Chat models are in the filtered df columns or index."
"This will cause the mean graph to be skewed."
)
fig3 = px.scatter(
y=row_mean,
x=[MODELS[x]["model_size"] for x in filtered_df.columns if x not in ["mean"]],
# hover_data=[x for x in filtered_df.index if x not in ["mean"]],
color=[
MODELS[x]["model_family"] for x in filtered_df.columns if x not in ["mean"]
],
color_discrete_sequence=px.colors.qualitative.Plotly,
title="",
# x axis title
labels={
"x": "Target Model Size",
"y": "Average ROC AUC",
"color": "Model Family",
},
log_x=True,
trendline="ols",
)
fig4 = px.scatter(
y=diag,
x=[MODELS[x]["model_size"] for x in filtered_df.columns if x not in ["mean"]],
# hover_data=[x for x in filtered_df.index if x not in ["mean"]],
color=[
MODELS[x]["model_family"] for x in filtered_df.columns if x not in ["mean"]
],
color_discrete_sequence=px.colors.qualitative.Plotly,
title="",
# x axis title
labels={
"x": "Target Model Size",
"y": "Self ROC AUC",
"color": "Model Family",
},
log_x=True,
trendline="ols",
)
# put the two plots side by side
fig_subplot = make_subplots(
rows=1,
cols=2,
shared_yaxes=False,
subplot_titles=("Self Detection ROC AUC", "Average Target ROC AUC"),
)
for i, figure in enumerate([fig4, fig3]):
for trace in range(len(figure["data"])):
trace_data = figure["data"][trace]
if i == 1:
trace_data["showlegend"] = False
fig_subplot.append_trace(trace_data, row=1, col=i + 1)
fig_subplot.update_xaxes(type="log")
# y axis range
fig_subplot.update_yaxes(range=[0.90, 1])
fig_subplot.update_layout(
height=500,
width=1200,
)
# put the legend on the bottom
fig_subplot.update_layout(
legend=dict(orientation="h", yanchor="bottom", y=-0.2, x=0.09)
)
st.plotly_chart(fig_subplot, use_container_width=True)
fig2 = px.scatter(
y=col_mean,
x=[MODELS_SIZE_MAPPING[x] for x in filtered_df.index if x not in ["mean"]],
# hover_data=[x for x in filtered_df.index if x not in ["mean"]],
color=[
MODELS_FAMILY_MAPPING[x] for x in filtered_df.index if x not in ["mean"]
],
color_discrete_sequence=px.colors.qualitative.Plotly,
title="Mean vs Train Model Size",
log_x=True,
trendline="ols",
)
fig2.update_layout(
height=600,
width=900,
)
st.plotly_chart(fig2, use_container_width=False)