Spaces:
Running
Running
Changed submission ID and added more informations from submission
Browse files
app.py
CHANGED
@@ -19,29 +19,37 @@ from server import LeaderboardServer
|
|
19 |
leaderboard_server = LeaderboardServer()
|
20 |
|
21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
def on_submit_pressed():
|
23 |
return gr.update(value='Processing submission...', interactive=False)
|
24 |
|
25 |
|
26 |
-
def validate_submission_inputs(
|
27 |
-
if
|
28 |
raise ValueError('Please fill in all fields')
|
29 |
-
if not os.path.exists(submission_file):
|
30 |
raise ValueError('File does not exist')
|
31 |
|
32 |
-
|
33 |
-
def process_submission(team_name, submission_id, description, link_to_model, submission_file):
|
34 |
try:
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
"link_to_model": link_to_model,
|
41 |
-
}
|
42 |
gr.Info('Submission valid, running local tournament...')
|
43 |
|
44 |
-
leaderboard_server.prepare_model_for_submission(submission_file, metadata)
|
45 |
except ValueError as err:
|
46 |
gr.Warning(str(err))
|
47 |
return (
|
@@ -226,15 +234,22 @@ with gr.Blocks(theme=gr.themes.Soft(text_size=text_md), css=custom_css) as main:
|
|
226 |
with gr.Tab('Submission'):
|
227 |
with gr.Column():
|
228 |
gr.Markdown(SUBMISSION_TAB_TITLE_MARKDOWN)
|
|
|
|
|
|
|
229 |
with gr.Row():
|
230 |
-
|
231 |
-
|
|
|
|
|
232 |
|
233 |
with gr.Row():
|
234 |
-
|
235 |
-
|
|
|
236 |
|
237 |
-
|
|
|
238 |
pre_submission_btn = gr.Button(value='Pre-submit model', interactive=True)
|
239 |
|
240 |
submit_prompt = gr.Markdown(
|
@@ -263,13 +278,7 @@ with gr.Blocks(theme=gr.themes.Soft(text_size=text_md), css=custom_css) as main:
|
|
263 |
outputs=[pre_submission_btn],
|
264 |
).then(
|
265 |
fn=process_submission,
|
266 |
-
inputs=
|
267 |
-
submission_team_name_tb,
|
268 |
-
submission_id_tb,
|
269 |
-
description_tb,
|
270 |
-
link_to_model_tb,
|
271 |
-
submission_file_path,
|
272 |
-
],
|
273 |
outputs=[
|
274 |
pre_submission_btn,
|
275 |
submit_prompt,
|
|
|
19 |
leaderboard_server = LeaderboardServer()
|
20 |
|
21 |
|
22 |
+
SUBMISSION_INPUTS = dict.fromkeys((
|
23 |
+
"team_name",
|
24 |
+
"model_name",
|
25 |
+
"model_type",
|
26 |
+
"parameters",
|
27 |
+
"precision",
|
28 |
+
"description",
|
29 |
+
"link_to_model",
|
30 |
+
"submission_file",
|
31 |
+
)).keys()
|
32 |
+
|
33 |
def on_submit_pressed():
|
34 |
return gr.update(value='Processing submission...', interactive=False)
|
35 |
|
36 |
|
37 |
+
def validate_submission_inputs(**kwargs):
|
38 |
+
if SUBMISSION_INPUTS - set(kwargs) != set():
|
39 |
raise ValueError('Please fill in all fields')
|
40 |
+
if not os.path.exists(kwargs["submission_file"]):
|
41 |
raise ValueError('File does not exist')
|
42 |
|
43 |
+
def process_submission(*inputs):
|
|
|
44 |
try:
|
45 |
+
inputs = dict(zip(SUBMISSION_INPUTS, inputs))
|
46 |
+
validate_submission_inputs(**inputs)
|
47 |
+
metadata = SUBMISSION_INPUTS - {"submission_file"}
|
48 |
+
metadata = {key: inputs[key] for key in metadata}
|
49 |
+
|
|
|
|
|
50 |
gr.Info('Submission valid, running local tournament...')
|
51 |
|
52 |
+
leaderboard_server.prepare_model_for_submission(inputs["submission_file"], metadata)
|
53 |
except ValueError as err:
|
54 |
gr.Warning(str(err))
|
55 |
return (
|
|
|
234 |
with gr.Tab('Submission'):
|
235 |
with gr.Column():
|
236 |
gr.Markdown(SUBMISSION_TAB_TITLE_MARKDOWN)
|
237 |
+
|
238 |
+
submission_inputs = dict.fromkeys(SUBMISSION_INPUTS)
|
239 |
+
|
240 |
with gr.Row():
|
241 |
+
submission_inputs["team_name"] = gr.Textbox(label='Team name')
|
242 |
+
submission_inputs["model_name"] = gr.Textbox(label='Model name')
|
243 |
+
submission_inputs["model_type"] = gr.Textbox(label='Model type')
|
244 |
+
submission_inputs["parameters"] = gr.Textbox(label='Parameters')
|
245 |
|
246 |
with gr.Row():
|
247 |
+
submission_inputs["precision"] = gr.Textbox(label='Precision')
|
248 |
+
submission_inputs["description"] = gr.Textbox(label='Description', type='text')
|
249 |
+
submission_inputs["link_to_model"] = gr.Textbox(label='Link to model', type='text')
|
250 |
|
251 |
+
submission_inputs["submission_file"] = gr.File(label='Upload your results', type='filepath')
|
252 |
+
|
253 |
pre_submission_btn = gr.Button(value='Pre-submit model', interactive=True)
|
254 |
|
255 |
submit_prompt = gr.Markdown(
|
|
|
278 |
outputs=[pre_submission_btn],
|
279 |
).then(
|
280 |
fn=process_submission,
|
281 |
+
inputs=list(submission_inputs.values()),
|
|
|
|
|
|
|
|
|
|
|
|
|
282 |
outputs=[
|
283 |
pre_submission_btn,
|
284 |
submit_prompt,
|
server.py
CHANGED
@@ -2,6 +2,7 @@ import copy
|
|
2 |
import glob
|
3 |
import json
|
4 |
import os
|
|
|
5 |
|
6 |
import gradio as gr
|
7 |
import pandas as pd
|
@@ -18,7 +19,6 @@ REPO = f"{ORG}/LLM_benchmark_data"
|
|
18 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
19 |
TASKS_METADATA_PATH = "./tasks_metadata.json"
|
20 |
|
21 |
-
|
22 |
class LeaderboardServer:
|
23 |
def __init__(self):
|
24 |
self.server_address = REPO
|
@@ -29,7 +29,7 @@ class LeaderboardServer:
|
|
29 |
token=HF_TOKEN,
|
30 |
local_dir="./",
|
31 |
)
|
32 |
-
self.
|
33 |
self.tasks_metadata = json.load(open(TASKS_METADATA_PATH))
|
34 |
self.tasks_categories = {self.tasks_metadata[task]["category"] for task in self.tasks_metadata}
|
35 |
self.submission_ids = set()
|
@@ -57,15 +57,15 @@ class LeaderboardServer:
|
|
57 |
|
58 |
def fetch_existing_models(self):
|
59 |
# Models data
|
60 |
-
for
|
61 |
-
data = json.load(open(
|
62 |
metadata = data.get('metadata')
|
63 |
if metadata is None:
|
64 |
continue
|
65 |
-
submission_id = metadata["
|
66 |
self.submission_ids.add(submission_id)
|
67 |
|
68 |
-
self.
|
69 |
|
70 |
def get_leaderboard(self, tournament_results=None):
|
71 |
results = tournament_results if tournament_results else self.tournament_results
|
@@ -74,24 +74,26 @@ class LeaderboardServer:
|
|
74 |
return pd.DataFrame(columns=['No submissions yet'])
|
75 |
else:
|
76 |
processed_results = []
|
77 |
-
for
|
78 |
-
path = self.
|
79 |
if path is None:
|
80 |
-
if self.pre_submit and
|
81 |
data = json.load(open(self.pre_submit[2]))
|
82 |
else:
|
83 |
-
raise gr.Error(f"Internal error: Submission [{
|
84 |
elif path:
|
85 |
data = json.load(open(path))
|
86 |
else:
|
87 |
-
raise gr.Error(f"Submission [{
|
88 |
-
|
|
|
|
|
89 |
|
90 |
local_results = {}
|
91 |
for task in self.tasks_metadata.keys():
|
92 |
local_results[task] = 0
|
93 |
-
for model in results[
|
94 |
-
if results[
|
95 |
local_results[task] += 1
|
96 |
for metric in VISIBLE_METRICS:
|
97 |
metric_value = data['results'][task].get(metric)
|
@@ -100,7 +102,7 @@ class LeaderboardServer:
|
|
100 |
|
101 |
local_results["submission_id"] = submission_id
|
102 |
|
103 |
-
if self.pre_submit and
|
104 |
processed_results.insert(0, local_results)
|
105 |
else:
|
106 |
processed_results.append(local_results)
|
@@ -120,42 +122,71 @@ class LeaderboardServer:
|
|
120 |
)
|
121 |
return dataframe
|
122 |
|
123 |
-
def start_tournament(self,
|
124 |
new_tournament = copy.deepcopy(self.tournament_results)
|
125 |
-
new_tournament[
|
126 |
-
new_tournament[
|
127 |
task: False for task in self.tasks_metadata.keys()
|
128 |
}
|
129 |
|
130 |
-
for
|
131 |
-
res = check_significance(new_model_file, self.
|
132 |
-
res_inverse = check_significance(self.
|
133 |
-
new_tournament[
|
134 |
task: data["significant"] for task, data in res.items()
|
135 |
}
|
136 |
-
new_tournament[
|
137 |
task: data["significant"] for task, data in res_inverse.items()
|
138 |
}
|
139 |
return new_tournament
|
140 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
141 |
def prepare_model_for_submission(self, file, metadata) -> None:
|
142 |
with open(file, "r") as f:
|
143 |
data = json.load(f)
|
|
|
144 |
data["metadata"] = metadata
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
145 |
with open(file, "w") as f:
|
146 |
json.dump(data, f, separators=(',', ':')) # compact JSON
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
self.pre_submit = tournament_results, model_id, file
|
151 |
|
152 |
def save_pre_submit(self):
|
153 |
if self.pre_submit:
|
154 |
-
tournament_results,
|
155 |
-
filename = os.path.basename(file)
|
156 |
api.upload_file(
|
157 |
path_or_fileobj=file,
|
158 |
-
path_in_repo=f"data/{
|
159 |
repo_id=self.server_address,
|
160 |
repo_type=self.repo_type,
|
161 |
token=HF_TOKEN,
|
@@ -175,7 +206,7 @@ class LeaderboardServer:
|
|
175 |
)
|
176 |
|
177 |
def get_model_detail(self, submission_id):
|
178 |
-
path = self.
|
179 |
if path is None:
|
180 |
raise gr.Error(f"Submission [{submission_id}] not found")
|
181 |
data = json.load(open(path))
|
|
|
2 |
import glob
|
3 |
import json
|
4 |
import os
|
5 |
+
import hashlib
|
6 |
|
7 |
import gradio as gr
|
8 |
import pandas as pd
|
|
|
19 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
20 |
TASKS_METADATA_PATH = "./tasks_metadata.json"
|
21 |
|
|
|
22 |
class LeaderboardServer:
|
23 |
def __init__(self):
|
24 |
self.server_address = REPO
|
|
|
29 |
token=HF_TOKEN,
|
30 |
local_dir="./",
|
31 |
)
|
32 |
+
self.submission_id_to_file = {} # Map submission ids to file paths
|
33 |
self.tasks_metadata = json.load(open(TASKS_METADATA_PATH))
|
34 |
self.tasks_categories = {self.tasks_metadata[task]["category"] for task in self.tasks_metadata}
|
35 |
self.submission_ids = set()
|
|
|
57 |
|
58 |
def fetch_existing_models(self):
|
59 |
# Models data
|
60 |
+
for submission_file in glob.glob(os.path.join(self.local_leaderboard, "data") + "/*.json"):
|
61 |
+
data = json.load(open(submission_file))
|
62 |
metadata = data.get('metadata')
|
63 |
if metadata is None:
|
64 |
continue
|
65 |
+
submission_id = metadata["submission_id"]
|
66 |
self.submission_ids.add(submission_id)
|
67 |
|
68 |
+
self.submission_id_to_file[submission_id] = submission_file
|
69 |
|
70 |
def get_leaderboard(self, tournament_results=None):
|
71 |
results = tournament_results if tournament_results else self.tournament_results
|
|
|
74 |
return pd.DataFrame(columns=['No submissions yet'])
|
75 |
else:
|
76 |
processed_results = []
|
77 |
+
for submission_id in results.keys():
|
78 |
+
path = self.submission_id_to_file.get(submission_id)
|
79 |
if path is None:
|
80 |
+
if self.pre_submit and submission_id == self.pre_submit[1]:
|
81 |
data = json.load(open(self.pre_submit[2]))
|
82 |
else:
|
83 |
+
raise gr.Error(f"Internal error: Submission [{submission_id}] not found")
|
84 |
elif path:
|
85 |
data = json.load(open(path))
|
86 |
else:
|
87 |
+
raise gr.Error(f"Submission [{submission_id}] not found")
|
88 |
+
|
89 |
+
if submission_id != data["metadata"]["submission_id"]:
|
90 |
+
raise gr.Error(f"Proper submission [{submission_id}] not found")
|
91 |
|
92 |
local_results = {}
|
93 |
for task in self.tasks_metadata.keys():
|
94 |
local_results[task] = 0
|
95 |
+
for model in results[submission_id].keys():
|
96 |
+
if results[submission_id][model][task]:
|
97 |
local_results[task] += 1
|
98 |
for metric in VISIBLE_METRICS:
|
99 |
metric_value = data['results'][task].get(metric)
|
|
|
102 |
|
103 |
local_results["submission_id"] = submission_id
|
104 |
|
105 |
+
if self.pre_submit and submission_id == self.pre_submit[1]:
|
106 |
processed_results.insert(0, local_results)
|
107 |
else:
|
108 |
processed_results.append(local_results)
|
|
|
122 |
)
|
123 |
return dataframe
|
124 |
|
125 |
+
def start_tournament(self, new_submission_id, new_model_file):
|
126 |
new_tournament = copy.deepcopy(self.tournament_results)
|
127 |
+
new_tournament[new_submission_id] = {}
|
128 |
+
new_tournament[new_submission_id][new_submission_id] = {
|
129 |
task: False for task in self.tasks_metadata.keys()
|
130 |
}
|
131 |
|
132 |
+
for submission_id in self.submission_ids:
|
133 |
+
res = check_significance(new_model_file, self.submission_id_to_file[submission_id])
|
134 |
+
res_inverse = check_significance(self.submission_id_to_file[submission_id], new_model_file)
|
135 |
+
new_tournament[new_submission_id][submission_id] = {
|
136 |
task: data["significant"] for task, data in res.items()
|
137 |
}
|
138 |
+
new_tournament[submission_id][new_submission_id] = {
|
139 |
task: data["significant"] for task, data in res_inverse.items()
|
140 |
}
|
141 |
return new_tournament
|
142 |
|
143 |
+
@staticmethod
|
144 |
+
def create_submission_id(metadata):
|
145 |
+
# Délka ID můsí být omezena, protože se používá v názvu souboru
|
146 |
+
submission_id = "_".join([metadata[key][:7] for key in (
|
147 |
+
"team_name",
|
148 |
+
"model_name",
|
149 |
+
"model_predictions_sha256",
|
150 |
+
"model_results_sha256",
|
151 |
+
)])
|
152 |
+
return submission_id
|
153 |
+
|
154 |
+
@staticmethod
|
155 |
+
def get_sha256_hexdigest(obj):
|
156 |
+
data = json.dumps(
|
157 |
+
obj,
|
158 |
+
separators=(',', ':'),
|
159 |
+
sort_keys=True,
|
160 |
+
ensure_ascii=True,
|
161 |
+
).encode()
|
162 |
+
result = hashlib.sha256(data).hexdigest()
|
163 |
+
return result
|
164 |
+
|
165 |
def prepare_model_for_submission(self, file, metadata) -> None:
|
166 |
with open(file, "r") as f:
|
167 |
data = json.load(f)
|
168 |
+
|
169 |
data["metadata"] = metadata
|
170 |
+
|
171 |
+
metadata["model_predictions_sha256"] = self.get_sha256_hexdigest(data["predictions"])
|
172 |
+
metadata["model_results_sha256"] = self.get_sha256_hexdigest(data["results"])
|
173 |
+
|
174 |
+
# Délka ID můsí být omezena, protože se používá v názvu souboru
|
175 |
+
submission_id = self.create_submission_id(metadata)
|
176 |
+
metadata["submission_id"] = submission_id
|
177 |
+
|
178 |
with open(file, "w") as f:
|
179 |
json.dump(data, f, separators=(',', ':')) # compact JSON
|
180 |
+
|
181 |
+
tournament_results = self.start_tournament(submission_id, file)
|
182 |
+
self.pre_submit = tournament_results, submission_id, file
|
|
|
183 |
|
184 |
def save_pre_submit(self):
|
185 |
if self.pre_submit:
|
186 |
+
tournament_results, submission_id, file = self.pre_submit
|
|
|
187 |
api.upload_file(
|
188 |
path_or_fileobj=file,
|
189 |
+
path_in_repo=f"data/{submission_id}.json",
|
190 |
repo_id=self.server_address,
|
191 |
repo_type=self.repo_type,
|
192 |
token=HF_TOKEN,
|
|
|
206 |
)
|
207 |
|
208 |
def get_model_detail(self, submission_id):
|
209 |
+
path = self.submission_id_to_file.get(submission_id)
|
210 |
if path is None:
|
211 |
raise gr.Error(f"Submission [{submission_id}] not found")
|
212 |
data = json.load(open(path))
|