wissamantoun
commited on
Commit
β’
7cf7655
1
Parent(s):
a1925cb
added watermarking and quantization exp
Browse files
app.py
CHANGED
@@ -1,4 +1,5 @@
|
|
1 |
import json
|
|
|
2 |
|
3 |
import numpy as np
|
4 |
import pandas as pd
|
@@ -11,8 +12,6 @@ from plotly.subplots import make_subplots
|
|
11 |
from exp_utils import MODELS
|
12 |
from visualize_utils import viridis_rgb
|
13 |
|
14 |
-
#
|
15 |
-
|
16 |
st.set_page_config(
|
17 |
page_title="Results Viewer",
|
18 |
page_icon="π",
|
@@ -23,14 +22,35 @@ st.set_page_config(
|
|
23 |
MODELS_SIZE_MAPPING = {k: v["model_size"] for k, v in MODELS.items()}
|
24 |
MODELS_FAMILY_MAPPING = {k: v["model_family"] for k, v in MODELS.items()}
|
25 |
MODEL_FAMILES = set([model["model_family"] for model in MODELS.values()])
|
26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
MODEL_NAMES_SORTED_BY_NAME_AND_SIZE = sorted(
|
29 |
-
MODEL_NAMES,
|
|
|
|
|
|
|
|
|
30 |
)
|
31 |
|
32 |
MODEL_NAMES_SORTED_BY_SIZE = sorted(
|
33 |
-
MODEL_NAMES,
|
|
|
|
|
|
|
|
|
34 |
)
|
35 |
|
36 |
|
@@ -43,7 +63,11 @@ MODELS_SIZE_MAPPING = {
|
|
43 |
MODELS_SIZE_MAPPING_LIST = list(MODELS_SIZE_MAPPING.keys())
|
44 |
|
45 |
|
46 |
-
CHAT_MODELS = [
|
|
|
|
|
|
|
|
|
47 |
|
48 |
|
49 |
def clean_dataframe(df: pd.DataFrame) -> pd.DataFrame:
|
@@ -66,7 +90,11 @@ def clean_dataframe(df: pd.DataFrame) -> pd.DataFrame:
|
|
66 |
df.columns = df.columns.str.replace("_roc_auc", "")
|
67 |
df.columns = df.columns.str.replace("eval_", "")
|
68 |
|
69 |
-
df["model_family"] = df["model_name"].
|
|
|
|
|
|
|
|
|
70 |
# create a dict with the model_name and the model_family
|
71 |
model_family_dict = {
|
72 |
k: v
|
@@ -84,8 +112,16 @@ def clean_dataframe(df: pd.DataFrame) -> pd.DataFrame:
|
|
84 |
df_std = df_std.drop(columns=["exp_seed"])
|
85 |
df_avg["model_family"] = df_avg.index.map(model_family_dict)
|
86 |
df_std["model_family"] = df_std.index.map(model_family_dict)
|
87 |
-
df_avg["model_size"] = df_avg.index.map(
|
88 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
|
90 |
# sort rows by model family then model size
|
91 |
df_avg = df_avg.sort_values(
|
@@ -101,10 +137,15 @@ def clean_dataframe(df: pd.DataFrame) -> pd.DataFrame:
|
|
101 |
availables_rows = [x for x in df_std.columns if x in df_std.index]
|
102 |
df_std = df_std.reindex(availables_rows)
|
103 |
|
|
|
|
|
|
|
|
|
|
|
104 |
return df_avg, df_std
|
105 |
|
106 |
|
107 |
-
def get_data(path):
|
108 |
df, df_std = clean_dataframe(pd.read_csv(path, index_col=0))
|
109 |
return df, df_std
|
110 |
|
@@ -117,8 +158,15 @@ def filter_df(
|
|
117 |
model_size_test: tuple,
|
118 |
is_chat_train: bool,
|
119 |
is_chat_test: bool,
|
|
|
|
|
|
|
|
|
120 |
sort_by_size: bool,
|
121 |
split_chat_models: bool,
|
|
|
|
|
|
|
122 |
is_debug: bool,
|
123 |
) -> pd.DataFrame:
|
124 |
# remove all columns and rows that have "pythia-70m" in the name
|
@@ -143,6 +191,16 @@ def filter_df(
|
|
143 |
if is_debug:
|
144 |
st.write("Filter is chat train")
|
145 |
st.write(df)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
146 |
|
147 |
# filter columns
|
148 |
if is_debug:
|
@@ -150,8 +208,13 @@ def filter_df(
|
|
150 |
st.write(df)
|
151 |
columns_to_keep = []
|
152 |
for column in df.columns:
|
153 |
-
if
|
154 |
-
|
|
|
|
|
|
|
|
|
|
|
155 |
if (
|
156 |
model_size >= model_size_test[0] * 1e9
|
157 |
and model_size <= model_size_test[1] * 1e9
|
@@ -167,7 +230,12 @@ def filter_df(
|
|
167 |
columns_to_keep = []
|
168 |
for column in df.columns:
|
169 |
for model_family in model_family_test:
|
170 |
-
if
|
|
|
|
|
|
|
|
|
|
|
171 |
columns_to_keep.append(column)
|
172 |
df = df[list(sorted(list(set(columns_to_keep))))]
|
173 |
if is_debug:
|
@@ -178,13 +246,44 @@ def filter_df(
|
|
178 |
# filter columns
|
179 |
columns_to_keep = []
|
180 |
for column in df.columns:
|
181 |
-
if
|
|
|
|
|
|
|
|
|
|
|
182 |
columns_to_keep.append(column)
|
183 |
df = df[list(sorted(list(set(columns_to_keep))))]
|
184 |
if is_debug:
|
185 |
st.write("Filter is chat test")
|
186 |
st.write(df)
|
187 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
df = df.select_dtypes(include="number")
|
189 |
if is_debug:
|
190 |
st.write("Select dtypes to be only numbers")
|
@@ -227,10 +326,121 @@ def filter_df(
|
|
227 |
if is_debug:
|
228 |
st.write("Split chat models")
|
229 |
st.write(df)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
230 |
return df
|
231 |
|
232 |
|
233 |
df, df_std = get_data("./deberta_results.csv")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
234 |
|
235 |
with open("./ood_results.json", "r") as f:
|
236 |
ood_results = json.load(f)
|
@@ -258,11 +468,14 @@ st.write(
|
|
258 |
)
|
259 |
|
260 |
# filters
|
261 |
-
|
262 |
-
sort_by_size = st.sidebar.checkbox("Sort by size", value=
|
263 |
-
split_chat_models = st.sidebar.checkbox("Split chat models", value=
|
|
|
|
|
264 |
add_mean = st.sidebar.checkbox("Add mean", value=False)
|
265 |
show_std = st.sidebar.checkbox("Show std", value=False)
|
|
|
266 |
model_size_train = st.sidebar.slider(
|
267 |
"Train Model Size in Billion", min_value=0, max_value=100, value=(0, 100), step=1
|
268 |
)
|
@@ -271,6 +484,18 @@ model_size_test = st.sidebar.slider(
|
|
271 |
)
|
272 |
is_chat_train = st.sidebar.selectbox("(Train) Is Chat?", [True, False, "Both"], index=2)
|
273 |
is_chat_test = st.sidebar.selectbox("(Test) Is Chat?", [True, False, "Both"], index=2)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
274 |
model_family_train = st.sidebar.multiselect(
|
275 |
"Model Family Train",
|
276 |
MODEL_FAMILES,
|
@@ -282,6 +507,8 @@ model_family_test = st.sidebar.multiselect(
|
|
282 |
default=MODEL_FAMILES,
|
283 |
)
|
284 |
|
|
|
|
|
285 |
add_adversarial = False
|
286 |
if "Adversarial" in model_family_test:
|
287 |
model_family_test.remove("Adversarial")
|
@@ -304,14 +531,6 @@ if show_std:
|
|
304 |
else:
|
305 |
selected_df = df.copy()
|
306 |
|
307 |
-
if show_diff:
|
308 |
-
# get those 3 columns {'model_size', 'model_family', 'is_chat'}
|
309 |
-
columns_to_keep = ["model_size", "model_family", "is_chat"]
|
310 |
-
to_be_added = selected_df[columns_to_keep]
|
311 |
-
selected_df = selected_df.drop(columns=columns_to_keep)
|
312 |
-
selected_df = selected_df.sub(selected_df.values.diagonal(), axis=1)
|
313 |
-
selected_df = selected_df.join(to_be_added)
|
314 |
-
|
315 |
|
316 |
filtered_df = filter_df(
|
317 |
selected_df,
|
@@ -321,18 +540,32 @@ filtered_df = filter_df(
|
|
321 |
model_size_test,
|
322 |
is_chat_train,
|
323 |
is_chat_test,
|
|
|
|
|
|
|
|
|
324 |
sort_by_size,
|
325 |
split_chat_models,
|
|
|
|
|
|
|
326 |
is_debug,
|
327 |
)
|
328 |
|
329 |
|
330 |
-
|
|
|
|
|
|
|
331 |
|
332 |
-
#
|
333 |
-
# filtered_df = filtered_df.sub(filtered_df.values.diagonal(), axis=1)
|
334 |
if add_adversarial:
|
335 |
-
|
|
|
|
|
|
|
|
|
|
|
336 |
|
337 |
if add_mean:
|
338 |
col_mean = filtered_df.mean(axis=1)
|
@@ -341,7 +574,6 @@ if add_mean:
|
|
341 |
filtered_df["mean"] = col_mean
|
342 |
filtered_df.loc["mean"] = row_mean
|
343 |
|
344 |
-
|
345 |
filtered_df = filtered_df * 100
|
346 |
filtered_df = filtered_df.round(0)
|
347 |
|
@@ -364,7 +596,7 @@ fig = px.imshow(
|
|
364 |
y=list(filtered_df.index),
|
365 |
color_continuous_scale=color_scale,
|
366 |
contrast_rescaling=None,
|
367 |
-
text_auto=
|
368 |
aspect="auto",
|
369 |
)
|
370 |
|
|
|
1 |
import json
|
2 |
+
from typing import Tuple
|
3 |
|
4 |
import numpy as np
|
5 |
import pandas as pd
|
|
|
12 |
from exp_utils import MODELS
|
13 |
from visualize_utils import viridis_rgb
|
14 |
|
|
|
|
|
15 |
st.set_page_config(
|
16 |
page_title="Results Viewer",
|
17 |
page_icon="π",
|
|
|
22 |
MODELS_SIZE_MAPPING = {k: v["model_size"] for k, v in MODELS.items()}
|
23 |
MODELS_FAMILY_MAPPING = {k: v["model_family"] for k, v in MODELS.items()}
|
24 |
MODEL_FAMILES = set([model["model_family"] for model in MODELS.values()])
|
25 |
+
Q_W_MODELS = [
|
26 |
+
"llama-7b",
|
27 |
+
"llama-2-7b",
|
28 |
+
"llama-13b",
|
29 |
+
"llama-2-13b",
|
30 |
+
"llama-30b",
|
31 |
+
"llama-65b",
|
32 |
+
"llama-2-70b",
|
33 |
+
]
|
34 |
+
Q_W_MODELS = [f"{model}_quantized" for model in Q_W_MODELS] + [
|
35 |
+
f"{model}_watermarked" for model in Q_W_MODELS
|
36 |
+
]
|
37 |
+
|
38 |
+
MODEL_NAMES = list(MODELS.keys()) + Q_W_MODELS
|
39 |
|
40 |
MODEL_NAMES_SORTED_BY_NAME_AND_SIZE = sorted(
|
41 |
+
MODEL_NAMES,
|
42 |
+
key=lambda x: (
|
43 |
+
MODELS[x.replace("_quantized", "").replace("_watermarked", "")]["model_family"],
|
44 |
+
MODELS[x.replace("_quantized", "").replace("_watermarked", "")]["model_size"],
|
45 |
+
),
|
46 |
)
|
47 |
|
48 |
MODEL_NAMES_SORTED_BY_SIZE = sorted(
|
49 |
+
MODEL_NAMES,
|
50 |
+
key=lambda x: (
|
51 |
+
MODELS[x.replace("_quantized", "").replace("_watermarked", "")]["model_size"],
|
52 |
+
MODELS[x.replace("_quantized", "").replace("_watermarked", "")]["model_family"],
|
53 |
+
),
|
54 |
)
|
55 |
|
56 |
|
|
|
63 |
MODELS_SIZE_MAPPING_LIST = list(MODELS_SIZE_MAPPING.keys())
|
64 |
|
65 |
|
66 |
+
CHAT_MODELS = [
|
67 |
+
x
|
68 |
+
for x in MODEL_NAMES_SORTED_BY_NAME_AND_SIZE
|
69 |
+
if MODELS[x.replace("_quantized", "").replace("_watermarked", "")]["is_chat"]
|
70 |
+
]
|
71 |
|
72 |
|
73 |
def clean_dataframe(df: pd.DataFrame) -> pd.DataFrame:
|
|
|
90 |
df.columns = df.columns.str.replace("_roc_auc", "")
|
91 |
df.columns = df.columns.str.replace("eval_", "")
|
92 |
|
93 |
+
df["model_family"] = df["model_name"].apply(
|
94 |
+
lambda x: MODELS_FAMILY_MAPPING[
|
95 |
+
x.replace("_quantized", "").replace("_watermarked", "")
|
96 |
+
]
|
97 |
+
)
|
98 |
# create a dict with the model_name and the model_family
|
99 |
model_family_dict = {
|
100 |
k: v
|
|
|
112 |
df_std = df_std.drop(columns=["exp_seed"])
|
113 |
df_avg["model_family"] = df_avg.index.map(model_family_dict)
|
114 |
df_std["model_family"] = df_std.index.map(model_family_dict)
|
115 |
+
df_avg["model_size"] = df_avg.index.map(
|
116 |
+
lambda x: MODELS_SIZE_MAPPING[
|
117 |
+
x.replace("_quantized", "").replace("_watermarked", "")
|
118 |
+
]
|
119 |
+
)
|
120 |
+
df_std["model_size"] = df_std.index.map(
|
121 |
+
lambda x: MODELS_SIZE_MAPPING[
|
122 |
+
x.replace("_quantized", "").replace("_watermarked", "")
|
123 |
+
]
|
124 |
+
)
|
125 |
|
126 |
# sort rows by model family then model size
|
127 |
df_avg = df_avg.sort_values(
|
|
|
137 |
availables_rows = [x for x in df_std.columns if x in df_std.index]
|
138 |
df_std = df_std.reindex(availables_rows)
|
139 |
|
140 |
+
df_avg["is_quantized"] = df_avg.index.str.contains("quantized")
|
141 |
+
df_avg["is_watermarked"] = df_avg.index.str.contains("watermarked")
|
142 |
+
df_std["is_quantized"] = df_std.index.str.contains("quantized")
|
143 |
+
df_std["is_watermarked"] = df_std.index.str.contains("watermarked")
|
144 |
+
|
145 |
return df_avg, df_std
|
146 |
|
147 |
|
148 |
+
def get_data(path) -> Tuple[pd.DataFrame, pd.DataFrame]:
|
149 |
df, df_std = clean_dataframe(pd.read_csv(path, index_col=0))
|
150 |
return df, df_std
|
151 |
|
|
|
158 |
model_size_test: tuple,
|
159 |
is_chat_train: bool,
|
160 |
is_chat_test: bool,
|
161 |
+
is_quantized_train: bool,
|
162 |
+
is_quantized_test: bool,
|
163 |
+
is_watermarked_train: bool,
|
164 |
+
is_watermarked_test: bool,
|
165 |
sort_by_size: bool,
|
166 |
split_chat_models: bool,
|
167 |
+
split_quantized_models: bool,
|
168 |
+
split_watermarked_models: bool,
|
169 |
+
filter_empty_col_row: bool,
|
170 |
is_debug: bool,
|
171 |
) -> pd.DataFrame:
|
172 |
# remove all columns and rows that have "pythia-70m" in the name
|
|
|
191 |
if is_debug:
|
192 |
st.write("Filter is chat train")
|
193 |
st.write(df)
|
194 |
+
if is_quantized_train != "Both":
|
195 |
+
df = df.loc[df["is_quantized"] == is_quantized_train]
|
196 |
+
if is_debug:
|
197 |
+
st.write("Filter is quantized train")
|
198 |
+
st.write(df)
|
199 |
+
if is_watermarked_train != "Both":
|
200 |
+
df = df.loc[df["is_watermarked"] == is_watermarked_train]
|
201 |
+
if is_debug:
|
202 |
+
st.write("Filter is watermark train")
|
203 |
+
st.write(df)
|
204 |
|
205 |
# filter columns
|
206 |
if is_debug:
|
|
|
208 |
st.write(df)
|
209 |
columns_to_keep = []
|
210 |
for column in df.columns:
|
211 |
+
if (
|
212 |
+
column.replace("_quantized", "").replace("_watermarked", "")
|
213 |
+
in MODELS.keys()
|
214 |
+
):
|
215 |
+
model_size = MODELS[
|
216 |
+
column.replace("_quantized", "").replace("_watermarked", "")
|
217 |
+
]["model_size"]
|
218 |
if (
|
219 |
model_size >= model_size_test[0] * 1e9
|
220 |
and model_size <= model_size_test[1] * 1e9
|
|
|
230 |
columns_to_keep = []
|
231 |
for column in df.columns:
|
232 |
for model_family in model_family_test:
|
233 |
+
if (
|
234 |
+
model_family
|
235 |
+
== MODELS[column.replace("_quantized", "").replace("_watermarked", "")][
|
236 |
+
"model_family"
|
237 |
+
]
|
238 |
+
):
|
239 |
columns_to_keep.append(column)
|
240 |
df = df[list(sorted(list(set(columns_to_keep))))]
|
241 |
if is_debug:
|
|
|
246 |
# filter columns
|
247 |
columns_to_keep = []
|
248 |
for column in df.columns:
|
249 |
+
if (
|
250 |
+
MODELS[column.replace("_quantized", "").replace("_watermarked", "")][
|
251 |
+
"is_chat"
|
252 |
+
]
|
253 |
+
== is_chat_test
|
254 |
+
):
|
255 |
columns_to_keep.append(column)
|
256 |
df = df[list(sorted(list(set(columns_to_keep))))]
|
257 |
if is_debug:
|
258 |
st.write("Filter is chat test")
|
259 |
st.write(df)
|
260 |
|
261 |
+
if is_quantized_test != "Both":
|
262 |
+
# filter columns
|
263 |
+
columns_to_keep = []
|
264 |
+
for column in df.columns:
|
265 |
+
if "quantized" in column and is_quantized_test:
|
266 |
+
columns_to_keep.append(column)
|
267 |
+
elif "quantized" not in column and not is_quantized_test:
|
268 |
+
columns_to_keep.append(column)
|
269 |
+
df = df[list(sorted(list(set(columns_to_keep))))]
|
270 |
+
if is_debug:
|
271 |
+
st.write("Filter is quantized test")
|
272 |
+
st.write(df)
|
273 |
+
|
274 |
+
if is_watermarked_test != "Both":
|
275 |
+
# filter columns
|
276 |
+
columns_to_keep = []
|
277 |
+
for column in df.columns:
|
278 |
+
if "watermark" in column and is_watermarked_test:
|
279 |
+
columns_to_keep.append(column)
|
280 |
+
elif "watermark" not in column and not is_watermarked_test:
|
281 |
+
columns_to_keep.append(column)
|
282 |
+
df = df[list(sorted(list(set(columns_to_keep))))]
|
283 |
+
if is_debug:
|
284 |
+
st.write("Filter is watermark test")
|
285 |
+
st.write(df)
|
286 |
+
|
287 |
df = df.select_dtypes(include="number")
|
288 |
if is_debug:
|
289 |
st.write("Select dtypes to be only numbers")
|
|
|
326 |
if is_debug:
|
327 |
st.write("Split chat models")
|
328 |
st.write(df)
|
329 |
+
|
330 |
+
if split_quantized_models:
|
331 |
+
# put chat models at the end of the columns
|
332 |
+
quantized_models = [
|
333 |
+
x for x in Q_W_MODELS if x in df.columns and "quantized" in x
|
334 |
+
]
|
335 |
+
# sort chat models by size
|
336 |
+
quantized_models = sorted(
|
337 |
+
quantized_models,
|
338 |
+
key=lambda x: MODELS[
|
339 |
+
x.replace("_quantized", "").replace("_watermarked", "")
|
340 |
+
]["model_size"],
|
341 |
+
)
|
342 |
+
df = df[[x for x in df.columns if x not in quantized_models] + quantized_models]
|
343 |
+
|
344 |
+
# put chat models at the end of the rows
|
345 |
+
quantized_models = [x for x in Q_W_MODELS if x in df.index and "quantized" in x]
|
346 |
+
# sort chat models by size
|
347 |
+
quantized_models = sorted(
|
348 |
+
quantized_models,
|
349 |
+
key=lambda x: MODELS[
|
350 |
+
x.replace("_quantized", "").replace("_watermarked", "")
|
351 |
+
]["model_size"],
|
352 |
+
)
|
353 |
+
df = df.reindex(
|
354 |
+
[x for x in df.index if x not in quantized_models] + quantized_models
|
355 |
+
)
|
356 |
+
|
357 |
+
if split_watermarked_models:
|
358 |
+
# put chat models at the end of the columns
|
359 |
+
watermarked_models = [
|
360 |
+
x for x in Q_W_MODELS if x in df.columns and "watermarked" in x
|
361 |
+
]
|
362 |
+
# sort chat models by size
|
363 |
+
watermarked_models = sorted(
|
364 |
+
watermarked_models,
|
365 |
+
key=lambda x: MODELS[
|
366 |
+
x.replace("_quantized", "").replace("_watermarked", "")
|
367 |
+
]["model_size"],
|
368 |
+
)
|
369 |
+
df = df[
|
370 |
+
[x for x in df.columns if x not in watermarked_models] + watermarked_models
|
371 |
+
]
|
372 |
+
|
373 |
+
# put chat models at the end of the rows
|
374 |
+
watermarked_models = [
|
375 |
+
x for x in Q_W_MODELS if x in df.index and "watermarked" in x
|
376 |
+
]
|
377 |
+
# sort chat models by size
|
378 |
+
watermarked_models = sorted(
|
379 |
+
watermarked_models,
|
380 |
+
key=lambda x: MODELS[
|
381 |
+
x.replace("_quantized", "").replace("_watermarked", "")
|
382 |
+
]["model_size"],
|
383 |
+
)
|
384 |
+
df = df.reindex(
|
385 |
+
[x for x in df.index if x not in watermarked_models] + watermarked_models
|
386 |
+
)
|
387 |
+
|
388 |
+
if is_debug:
|
389 |
+
st.write("Split chat models")
|
390 |
+
st.write(df)
|
391 |
+
|
392 |
+
if filter_empty_col_row:
|
393 |
+
# remove all for which the row and column are Nan
|
394 |
+
df = df.dropna(axis=0, how="all")
|
395 |
+
df = df.dropna(axis=1, how="all")
|
396 |
return df
|
397 |
|
398 |
|
399 |
df, df_std = get_data("./deberta_results.csv")
|
400 |
+
df_q_w, df_std_q_w = get_data("./results_qantized_watermarked.csv")
|
401 |
+
|
402 |
+
df = df.merge(
|
403 |
+
df_q_w[
|
404 |
+
df_q_w.columns[
|
405 |
+
df_q_w.columns.str.contains("quantized|watermarked", case=False, regex=True)
|
406 |
+
]
|
407 |
+
],
|
408 |
+
how="outer",
|
409 |
+
left_index=True,
|
410 |
+
right_index=True,
|
411 |
+
)
|
412 |
+
df_std = df_std.merge(
|
413 |
+
df_std_q_w[
|
414 |
+
df_std_q_w.columns[
|
415 |
+
df_std_q_w.columns.str.contains(
|
416 |
+
"quantized|watermarked", case=False, regex=True
|
417 |
+
)
|
418 |
+
]
|
419 |
+
],
|
420 |
+
how="outer",
|
421 |
+
left_index=True,
|
422 |
+
right_index=True,
|
423 |
+
)
|
424 |
+
|
425 |
+
|
426 |
+
df.columns = df.columns.str.replace("_y", "", regex=True)
|
427 |
+
df_std.columns = df_std.columns.str.replace("_y", "", regex=True)
|
428 |
+
|
429 |
+
df = df.drop(columns=["is_quantized_x", "is_watermarked_x"])
|
430 |
+
|
431 |
+
|
432 |
+
df.update(df_q_w)
|
433 |
+
df_std.update(df_std_q_w)
|
434 |
+
|
435 |
+
|
436 |
+
df["is_chat"].fillna(False, inplace=True)
|
437 |
+
df_std["is_chat"].fillna(False, inplace=True)
|
438 |
+
|
439 |
+
df["is_watermarked"].fillna(False, inplace=True)
|
440 |
+
df_std["is_watermarked"].fillna(False, inplace=True)
|
441 |
+
|
442 |
+
df["is_quantized"].fillna(False, inplace=True)
|
443 |
+
df_std["is_quantized"].fillna(False, inplace=True)
|
444 |
|
445 |
with open("./ood_results.json", "r") as f:
|
446 |
ood_results = json.load(f)
|
|
|
468 |
)
|
469 |
|
470 |
# filters
|
471 |
+
how_diff = st.sidebar.checkbox("Show Diff", value=False)
|
472 |
+
sort_by_size = st.sidebar.checkbox("Sort by size", value=True)
|
473 |
+
split_chat_models = st.sidebar.checkbox("Split chat models", value=True)
|
474 |
+
split_quantized_models = st.sidebar.checkbox("Split quantized models", value=True)
|
475 |
+
split_watermarked_models = st.sidebar.checkbox("Split watermarked models", value=True)
|
476 |
add_mean = st.sidebar.checkbox("Add mean", value=False)
|
477 |
show_std = st.sidebar.checkbox("Show std", value=False)
|
478 |
+
filter_empty_col_row = st.sidebar.checkbox("Filter empty col/row", value=True)
|
479 |
model_size_train = st.sidebar.slider(
|
480 |
"Train Model Size in Billion", min_value=0, max_value=100, value=(0, 100), step=1
|
481 |
)
|
|
|
484 |
)
|
485 |
is_chat_train = st.sidebar.selectbox("(Train) Is Chat?", [True, False, "Both"], index=2)
|
486 |
is_chat_test = st.sidebar.selectbox("(Test) Is Chat?", [True, False, "Both"], index=2)
|
487 |
+
is_quantized_train = st.sidebar.selectbox(
|
488 |
+
"(Train) Is Quantized?", [True, False, "Both"], index=1
|
489 |
+
)
|
490 |
+
is_quantized_test = st.sidebar.selectbox(
|
491 |
+
"(Test) Is Quantized?", [True, False, "Both"], index=1
|
492 |
+
)
|
493 |
+
is_watermarked_train = st.sidebar.selectbox(
|
494 |
+
"(Train) Is Watermark?", [True, False, "Both"], index=1
|
495 |
+
)
|
496 |
+
is_watermarked_test = st.sidebar.selectbox(
|
497 |
+
"(Test) Is Watermark?", [True, False, "Both"], index=1
|
498 |
+
)
|
499 |
model_family_train = st.sidebar.multiselect(
|
500 |
"Model Family Train",
|
501 |
MODEL_FAMILES,
|
|
|
507 |
default=MODEL_FAMILES,
|
508 |
)
|
509 |
|
510 |
+
show_values = st.sidebar.checkbox("Show Values", value=False)
|
511 |
+
|
512 |
add_adversarial = False
|
513 |
if "Adversarial" in model_family_test:
|
514 |
model_family_test.remove("Adversarial")
|
|
|
531 |
else:
|
532 |
selected_df = df.copy()
|
533 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
534 |
|
535 |
filtered_df = filter_df(
|
536 |
selected_df,
|
|
|
540 |
model_size_test,
|
541 |
is_chat_train,
|
542 |
is_chat_test,
|
543 |
+
is_quantized_train,
|
544 |
+
is_quantized_test,
|
545 |
+
is_watermarked_train,
|
546 |
+
is_watermarked_test,
|
547 |
sort_by_size,
|
548 |
split_chat_models,
|
549 |
+
split_quantized_models,
|
550 |
+
split_watermarked_models,
|
551 |
+
filter_empty_col_row,
|
552 |
is_debug,
|
553 |
)
|
554 |
|
555 |
|
556 |
+
if show_diff:
|
557 |
+
# get those 3 columns {'model_size', 'model_family', 'is_chat'}
|
558 |
+
diag = filtered_df.values.diagonal()
|
559 |
+
filtered_df = filtered_df.sub(diag, axis=1)
|
560 |
|
561 |
+
# subtract each row by the diagonal
|
|
|
562 |
if add_adversarial:
|
563 |
+
if show_diff:
|
564 |
+
index = filtered_df.index
|
565 |
+
ood_results_avg = ood_results_avg.loc[index]
|
566 |
+
filtered_df = filtered_df.join(ood_results_avg.sub(diag, axis=0))
|
567 |
+
else:
|
568 |
+
filtered_df = filtered_df.join(ood_results_avg)
|
569 |
|
570 |
if add_mean:
|
571 |
col_mean = filtered_df.mean(axis=1)
|
|
|
574 |
filtered_df["mean"] = col_mean
|
575 |
filtered_df.loc["mean"] = row_mean
|
576 |
|
|
|
577 |
filtered_df = filtered_df * 100
|
578 |
filtered_df = filtered_df.round(0)
|
579 |
|
|
|
596 |
y=list(filtered_df.index),
|
597 |
color_continuous_scale=color_scale,
|
598 |
contrast_rescaling=None,
|
599 |
+
text_auto=show_values,
|
600 |
aspect="auto",
|
601 |
)
|
602 |
|