Spaces:
Sleeping
Sleeping
Commit
•
6679087
1
Parent(s):
44d1791
Optimize data loading and filter in display
Browse files- app.py +80 -139
- requirements.txt +0 -2
app.py
CHANGED
@@ -12,26 +12,13 @@ EXCLUDED_KEYS = {
|
|
12 |
"chat_template",
|
13 |
"group_subtasks",
|
14 |
}
|
15 |
-
EXCLUDED_RESULTS_KEYS = {
|
16 |
-
|
17 |
-
}
|
18 |
-
EXCLUDED_RESULTS_LEADERBOARDS_KEYS = {
|
19 |
-
|
20 |
-
}
|
21 |
|
22 |
-
DEFAULT_HTML_TABLE = """
|
23 |
-
<table>
|
24 |
-
<thead>
|
25 |
-
<tr>
|
26 |
-
<th>Parameters</th>
|
27 |
-
<th>Model-1</th>
|
28 |
-
<th>Model-2</th>
|
29 |
-
</tr>
|
30 |
-
</thead>
|
31 |
-
<tbody>
|
32 |
-
</tbody>
|
33 |
-
</table>
|
34 |
-
"""
|
35 |
|
36 |
TASKS = {
|
37 |
"leaderboard_arc_challenge": ("ARC", "leaderboard_arc_challenge"),
|
@@ -71,98 +58,64 @@ def load_data(result_path) -> pd.DataFrame:
|
|
71 |
return data
|
72 |
|
73 |
|
74 |
-
def
|
75 |
result_path = get_result_path_from_model(model_id, latest_result_path_per_model)
|
76 |
data = load_data(result_path)
|
77 |
-
|
78 |
-
result = [
|
79 |
-
# to_vertical(df),
|
80 |
-
to_vertical(filter_results(df)),
|
81 |
-
to_vertical(filter_configs(df)),
|
82 |
-
]
|
83 |
-
return result
|
84 |
-
|
85 |
-
|
86 |
-
def to_vertical(df):
|
87 |
-
df = df.T.rename_axis("Parameters")
|
88 |
-
df.index = df.index.str.join(".")
|
89 |
-
return df
|
90 |
-
|
91 |
-
|
92 |
-
def to_dataframe(data):
|
93 |
df = pd.json_normalize([{key: value for key, value in data.items() if key not in EXCLUDED_KEYS}])
|
94 |
# df.columns = df.columns.str.split(".") # .split return a list instead of a tuple
|
95 |
-
df.
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
#
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
return
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
|
|
124 |
)
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
135 |
)
|
136 |
-
|
137 |
-
|
138 |
-
def concat_result_2(result_2, results):
|
139 |
-
results = pd.read_html(io.StringIO(results))[0]
|
140 |
-
df = (
|
141 |
-
pd.concat([results.iloc[:, [0, 1]].set_index("Parameters"), result_2], axis=1)
|
142 |
-
.reset_index()
|
143 |
-
)
|
144 |
-
return df
|
145 |
-
|
146 |
-
|
147 |
-
def render_result_1(model_id, task, *results):
|
148 |
-
result = load_result(model_id)
|
149 |
-
concat_results = [concat_result_1(*result_args) for result_args in zip(result, results)]
|
150 |
-
if task:
|
151 |
-
concat_results = [df[df["Parameters"].str.startswith(task[len("leaderboard_"):])] for df in concat_results]
|
152 |
-
return [display_dataframe(df) for df in concat_results]
|
153 |
-
|
154 |
-
|
155 |
-
def render_result_2(model_id, task, *results):
|
156 |
-
result = load_result(model_id)
|
157 |
-
concat_results = [concat_result_2(*result_args) for result_args in zip(result, results)]
|
158 |
-
if task:
|
159 |
-
concat_results = [df[df["Parameters"].str.startswith(task[len("leaderboard_"):])] for df in concat_results]
|
160 |
-
return [display_dataframe(df) for df in concat_results]
|
161 |
-
|
162 |
-
|
163 |
-
def render_results(model_id_1, model_id_2, task, *results):
|
164 |
-
results = render_result_1(model_id_1, task, *results)
|
165 |
-
return render_result_2(model_id_2, task, *results)
|
166 |
|
167 |
|
168 |
# if __name__ == "__main__":
|
@@ -176,9 +129,11 @@ with gr.Blocks(fill_height=True) as demo:
|
|
176 |
with gr.Column():
|
177 |
model_id_1 = gr.Dropdown(choices=list(latest_result_path_per_model.keys()), label="Results")
|
178 |
load_btn_1 = gr.Button("Load")
|
|
|
179 |
with gr.Column():
|
180 |
model_id_2 = gr.Dropdown(choices=list(latest_result_path_per_model.keys()), label="Results")
|
181 |
load_btn_2 = gr.Button("Load")
|
|
|
182 |
with gr.Row():
|
183 |
task = gr.Radio(
|
184 |
["All"] + list(TASKS.values()),
|
@@ -187,50 +142,36 @@ with gr.Blocks(fill_height=True) as demo:
|
|
187 |
value="All",
|
188 |
)
|
189 |
|
190 |
-
results = []
|
191 |
with gr.Row():
|
192 |
# with gr.Tab("All"):
|
193 |
-
#
|
194 |
-
# # label="Results",
|
195 |
-
# # headers=["Parameters", "Model-1", "Model-2"],
|
196 |
-
# # interactive=False,
|
197 |
-
# # column_widths=["30%", "30%", "30%"],
|
198 |
-
# # wrap=True,
|
199 |
-
# # ))
|
200 |
-
# results.append(gr.HTML(value=DEFAULT_HTML_TABLE))
|
201 |
with gr.Tab("Results"):
|
202 |
-
|
203 |
-
# label="Results",
|
204 |
-
# headers=["Parameters", "Model-1", "Model-2"],
|
205 |
-
# interactive=False,
|
206 |
-
# column_widths=["30%", "30%", "30%"],
|
207 |
-
# wrap=True,
|
208 |
-
# ))
|
209 |
-
results.append(gr.HTML(value=DEFAULT_HTML_TABLE))
|
210 |
with gr.Tab("Configs"):
|
211 |
-
|
212 |
-
# label="Results",
|
213 |
-
# headers=["Parameters", "Model-1", "Model-2"],
|
214 |
-
# interactive=False,
|
215 |
-
# column_widths=["30%", "30%", "30%"],
|
216 |
-
# wrap=True,
|
217 |
-
# ))
|
218 |
-
results.append(gr.HTML(value=DEFAULT_HTML_TABLE))
|
219 |
|
220 |
load_btn_1.click(
|
221 |
-
fn=
|
222 |
-
inputs=
|
223 |
-
outputs=
|
|
|
|
|
|
|
|
|
224 |
)
|
225 |
load_btn_2.click(
|
226 |
-
fn=
|
227 |
-
inputs=
|
228 |
-
outputs=
|
|
|
|
|
|
|
|
|
229 |
)
|
230 |
task.change(
|
231 |
-
fn=
|
232 |
-
inputs=[
|
233 |
-
outputs=[
|
234 |
)
|
235 |
|
236 |
demo.launch()
|
|
|
12 |
"chat_template",
|
13 |
"group_subtasks",
|
14 |
}
|
15 |
+
# EXCLUDED_RESULTS_KEYS = {
|
16 |
+
# "leaderboard",
|
17 |
+
# }
|
18 |
+
# EXCLUDED_RESULTS_LEADERBOARDS_KEYS = {
|
19 |
+
# "alias",
|
20 |
+
# }
|
21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
|
23 |
TASKS = {
|
24 |
"leaderboard_arc_challenge": ("ARC", "leaderboard_arc_challenge"),
|
|
|
58 |
return data
|
59 |
|
60 |
|
61 |
+
def load_result_dataframe(model_id):
|
62 |
result_path = get_result_path_from_model(model_id, latest_result_path_per_model)
|
63 |
data = load_data(result_path)
|
64 |
+
model_name = data.get("model_name", "Model")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
65 |
df = pd.json_normalize([{key: value for key, value in data.items() if key not in EXCLUDED_KEYS}])
|
66 |
# df.columns = df.columns.str.split(".") # .split return a list instead of a tuple
|
67 |
+
return df.set_index(pd.Index([model_name])).reset_index()
|
68 |
+
|
69 |
+
|
70 |
+
def display_results(df_1, df_2, task):
|
71 |
+
df = pd.concat([df.set_index("index") for df in [df_1, df_2] if "index" in df.columns])
|
72 |
+
df = df.T.rename_axis(columns=None) # index="Parameters", # .reset_index()
|
73 |
+
# return display_dataframe(df)
|
74 |
+
# d = df.set_index(df.index.str.split(".")) # .split return a list instead of a tuple
|
75 |
+
# results = d.loc[d.index.str[0] == "results"]
|
76 |
+
# results.index = results.index.str.join(".")
|
77 |
+
# configs = d.loc[d.index.str[0] == "configs"]
|
78 |
+
# configs.index = configs.index.str.join(".")
|
79 |
+
# return display_dataframe(results), display_dataframe(configs)
|
80 |
+
return display_results_tab(df, task), display_configs_tab(df, task)
|
81 |
+
|
82 |
+
|
83 |
+
def display_results_tab(df, task):
|
84 |
+
df = df.style.format(na_rep="")
|
85 |
+
df.hide(
|
86 |
+
[
|
87 |
+
row
|
88 |
+
for row in df.index
|
89 |
+
if (
|
90 |
+
not row.startswith("results.")
|
91 |
+
or row.startswith("results.leaderboard.")
|
92 |
+
or row.endswith(".alias")
|
93 |
+
or (not row.startswith(f"results.{task}") if task != "All" else False)
|
94 |
+
)
|
95 |
+
],
|
96 |
+
axis="index",
|
97 |
)
|
98 |
+
df.format_index(lambda idx: idx[len("results.leaderboard_"):].removesuffix(",none"), axis="index")
|
99 |
+
return df.to_html()
|
100 |
+
|
101 |
+
|
102 |
+
def display_configs_tab(df, task):
|
103 |
+
df = df.style.format(na_rep="")
|
104 |
+
df.hide(
|
105 |
+
[
|
106 |
+
row
|
107 |
+
for row in df.index
|
108 |
+
if (
|
109 |
+
not row.startswith("configs.")
|
110 |
+
or row.startswith("configs.leaderboard.")
|
111 |
+
or row.endswith(".alias")
|
112 |
+
or (not row.startswith(f"configs.{task}") if task != "All" else False)
|
113 |
+
)
|
114 |
+
],
|
115 |
+
axis="index",
|
116 |
)
|
117 |
+
df.format_index(lambda idx: idx[len("configs.leaderboard_"):], axis="index")
|
118 |
+
return df.to_html()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
|
120 |
|
121 |
# if __name__ == "__main__":
|
|
|
129 |
with gr.Column():
|
130 |
model_id_1 = gr.Dropdown(choices=list(latest_result_path_per_model.keys()), label="Results")
|
131 |
load_btn_1 = gr.Button("Load")
|
132 |
+
dataframe_1 = gr.Dataframe(visible=False)
|
133 |
with gr.Column():
|
134 |
model_id_2 = gr.Dropdown(choices=list(latest_result_path_per_model.keys()), label="Results")
|
135 |
load_btn_2 = gr.Button("Load")
|
136 |
+
dataframe_2 = gr.Dataframe(visible=False)
|
137 |
with gr.Row():
|
138 |
task = gr.Radio(
|
139 |
["All"] + list(TASKS.values()),
|
|
|
142 |
value="All",
|
143 |
)
|
144 |
|
|
|
145 |
with gr.Row():
|
146 |
# with gr.Tab("All"):
|
147 |
+
# pass
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
148 |
with gr.Tab("Results"):
|
149 |
+
results = gr.HTML()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
150 |
with gr.Tab("Configs"):
|
151 |
+
configs = gr.HTML()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
|
153 |
load_btn_1.click(
|
154 |
+
fn=load_result_dataframe,
|
155 |
+
inputs=model_id_1,
|
156 |
+
outputs=dataframe_1,
|
157 |
+
).then(
|
158 |
+
fn=display_results,
|
159 |
+
inputs=[dataframe_1, dataframe_2, task],
|
160 |
+
outputs=[results, configs],
|
161 |
)
|
162 |
load_btn_2.click(
|
163 |
+
fn=load_result_dataframe,
|
164 |
+
inputs=model_id_2,
|
165 |
+
outputs=dataframe_2,
|
166 |
+
).then(
|
167 |
+
fn=display_results,
|
168 |
+
inputs=[dataframe_1, dataframe_2, task],
|
169 |
+
outputs=[results, configs],
|
170 |
)
|
171 |
task.change(
|
172 |
+
fn=display_results,
|
173 |
+
inputs=[dataframe_1, dataframe_2, task],
|
174 |
+
outputs=[results, configs],
|
175 |
)
|
176 |
|
177 |
demo.launch()
|
requirements.txt
DELETED
@@ -1,2 +0,0 @@
|
|
1 |
-
lxml
|
2 |
-
|
|
|
|
|
|