Spaces:
Running
Running
using gdwon as a downloader
Browse files- app.py +194 -150
- requirements.txt +2 -1
app.py
CHANGED
@@ -1,97 +1,112 @@
|
|
1 |
-
import numpy as np
|
2 |
-
import torch
|
3 |
-
from tqdm import tqdm
|
4 |
-
import clip
|
5 |
-
from glob import glob
|
6 |
-
import gradio as gr
|
7 |
import os
|
8 |
-
import torchvision
|
9 |
import pickle
|
10 |
from collections import Counter
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
from SimSearch import FaissCosineNeighbors
|
13 |
|
14 |
# DOWNLOAD THE DATASET and Files
|
15 |
-
|
16 |
-
|
17 |
-
|
|
|
|
|
|
|
|
|
|
|
18 |
|
19 |
# EXTRACT
|
20 |
-
torchvision.datasets.utils.extract_archive(
|
21 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
|
23 |
# Initialize CLIP model
|
24 |
clip.available_models()
|
25 |
|
26 |
# # Searcher
|
27 |
-
|
28 |
class GamePhysicsSearcher:
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
self.all_embeddings = glob(f'{EMBEDDING_PATH}{self.GAME_NAME}/*.npy')
|
35 |
-
|
36 |
-
self.filenames = [os.path.basename(x) for x in self.all_embeddings]
|
37 |
-
self.file_to_class_id = {x:i for i, x in enumerate(self.filenames)}
|
38 |
-
self.class_id_to_file = {i:x for i, x in enumerate(self.filenames)}
|
39 |
-
self.build_index()
|
40 |
-
|
41 |
-
def read_features(self, file_path):
|
42 |
-
with open(file_path, 'rb') as f:
|
43 |
-
video_features = pickle.load(f)
|
44 |
-
return video_features
|
45 |
-
|
46 |
-
def read_all_features(self):
|
47 |
-
features = {}
|
48 |
-
filenames_extended = []
|
49 |
-
|
50 |
-
X_train = []
|
51 |
-
y_train = []
|
52 |
-
|
53 |
-
for i, vfile in enumerate(tqdm(self.all_embeddings)):
|
54 |
-
vfeatures = self.read_features(vfile)
|
55 |
-
features[vfile.split('/')[-1]] = vfeatures
|
56 |
-
X_train.extend(vfeatures)
|
57 |
-
y_train.extend([i]*vfeatures.shape[0])
|
58 |
-
filenames_extended.extend(vfeatures.shape[0]*[vfile.split('/')[-1]])
|
59 |
-
|
60 |
-
X_train = np.asarray(X_train)
|
61 |
-
y_train = np.asarray(y_train)
|
62 |
-
|
63 |
-
return X_train, y_train
|
64 |
-
|
65 |
-
def build_index(self):
|
66 |
-
X_train, y_train = self.read_all_features()
|
67 |
-
self.simsearcher.fit(X_train, y_train)
|
68 |
-
|
69 |
-
def text_to_vector(self, query):
|
70 |
-
text_tokens = clip.tokenize(query)
|
71 |
-
with torch.no_grad():
|
72 |
-
text_features = self.CLIP_MODEL.encode_text(text_tokens).float()
|
73 |
-
text_features /= text_features.norm(dim=-1, keepdim=True)
|
74 |
-
return text_features
|
75 |
-
|
76 |
-
# Source: https://stackoverflow.com/a/480227
|
77 |
-
def f7(self, seq):
|
78 |
-
seen = set()
|
79 |
-
seen_add = seen.add # This is for performance improvement, don't remove
|
80 |
-
return [x for x in seq if not (x in seen or seen_add(x))]
|
81 |
-
|
82 |
-
def search_top_k(self, q, k=5, pool_size=1000, search_mod='Majority'):
|
83 |
-
q = self.text_to_vector(q)
|
84 |
-
nearest_data_points = self.simsearcher.get_nearest_labels(q, pool_size)
|
85 |
-
|
86 |
-
if search_mod == 'Majority':
|
87 |
-
topKs = [x[0] for x in Counter(nearest_data_points[0]).most_common(k)]
|
88 |
-
elif search_mod == 'Top-K':
|
89 |
-
topKs = list(self.f7(nearest_data_points[0]))[:k]
|
90 |
-
|
91 |
-
video_filename = [f'./Videos/{self.GAME_NAME}/' + self.class_id_to_file[x].replace('npy', 'mp4') for x in topKs]
|
92 |
-
|
93 |
-
return video_filename
|
94 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
|
96 |
|
97 |
################ SEARCH CORE ################
|
@@ -100,77 +115,106 @@ vit_model, vit_preprocess = clip.load("ViT-B/32")
|
|
100 |
vit_model.eval()
|
101 |
|
102 |
saved_searchers = {}
|
|
|
|
|
103 |
def gradio_search(query, game_name, selected_model, aggregator, pool_size, k=6):
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
else:
|
108 |
-
if selected_model == 'ViT-B/32':
|
109 |
-
model = vit_model
|
110 |
-
searcher = GamePhysicsSearcher(CLIP_MODEL=model, GAME_NAME=game_name)
|
111 |
else:
|
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 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import os
|
|
|
2 |
import pickle
|
3 |
from collections import Counter
|
4 |
+
from glob import glob
|
5 |
+
|
6 |
+
import clip
|
7 |
+
import gdown
|
8 |
+
import gradio as gr
|
9 |
+
import numpy as np
|
10 |
+
import torch
|
11 |
+
import torchvision
|
12 |
+
from tqdm import tqdm
|
13 |
|
14 |
from SimSearch import FaissCosineNeighbors
|
15 |
|
16 |
# DOWNLOAD THE DATASET and Files
|
17 |
+
gdown.download(
|
18 |
+
id="1kB1vNdVaNS1OGZ3K8BspBUKkPACCsnrG", output="GTAV-Videos.zip", quiet=False
|
19 |
+
)
|
20 |
+
gdown.download(
|
21 |
+
id="1IF1ljcoFd31C-PA2SO8F5fEblDYf0Bw6",
|
22 |
+
output="GTAV-embedding-vit32.zip",
|
23 |
+
quiet=False,
|
24 |
+
)
|
25 |
|
26 |
# EXTRACT
|
27 |
+
torchvision.datasets.utils.extract_archive(
|
28 |
+
from_path="GTAV-embedding-vit32.zip",
|
29 |
+
to_path="Embeddings/VIT32/",
|
30 |
+
remove_finished=False,
|
31 |
+
)
|
32 |
+
torchvision.datasets.utils.extract_archive(
|
33 |
+
from_path="GTAV-Videos.zip", to_path="Videos/", remove_finished=False
|
34 |
+
)
|
35 |
|
36 |
# Initialize CLIP model
|
37 |
clip.available_models()
|
38 |
|
39 |
# # Searcher
|
|
|
40 |
class GamePhysicsSearcher:
|
41 |
+
def __init__(self, CLIP_MODEL, GAME_NAME, EMBEDDING_PATH="./Embeddings/VIT32/"):
|
42 |
+
self.CLIP_MODEL = CLIP_MODEL
|
43 |
+
self.GAME_NAME = GAME_NAME
|
44 |
+
self.simsearcher = FaissCosineNeighbors()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
|
46 |
+
self.all_embeddings = glob(f"{EMBEDDING_PATH}{self.GAME_NAME}/*.npy")
|
47 |
+
|
48 |
+
self.filenames = [os.path.basename(x) for x in self.all_embeddings]
|
49 |
+
self.file_to_class_id = {x: i for i, x in enumerate(self.filenames)}
|
50 |
+
self.class_id_to_file = {i: x for i, x in enumerate(self.filenames)}
|
51 |
+
self.build_index()
|
52 |
+
|
53 |
+
def read_features(self, file_path):
|
54 |
+
with open(file_path, "rb") as f:
|
55 |
+
video_features = pickle.load(f)
|
56 |
+
return video_features
|
57 |
+
|
58 |
+
def read_all_features(self):
|
59 |
+
features = {}
|
60 |
+
filenames_extended = []
|
61 |
+
|
62 |
+
X_train = []
|
63 |
+
y_train = []
|
64 |
+
|
65 |
+
for i, vfile in enumerate(tqdm(self.all_embeddings)):
|
66 |
+
vfeatures = self.read_features(vfile)
|
67 |
+
features[vfile.split("/")[-1]] = vfeatures
|
68 |
+
X_train.extend(vfeatures)
|
69 |
+
y_train.extend([i] * vfeatures.shape[0])
|
70 |
+
filenames_extended.extend(vfeatures.shape[0] * [vfile.split("/")[-1]])
|
71 |
+
|
72 |
+
X_train = np.asarray(X_train)
|
73 |
+
y_train = np.asarray(y_train)
|
74 |
+
|
75 |
+
return X_train, y_train
|
76 |
+
|
77 |
+
def build_index(self):
|
78 |
+
X_train, y_train = self.read_all_features()
|
79 |
+
self.simsearcher.fit(X_train, y_train)
|
80 |
+
|
81 |
+
def text_to_vector(self, query):
|
82 |
+
text_tokens = clip.tokenize(query)
|
83 |
+
with torch.no_grad():
|
84 |
+
text_features = self.CLIP_MODEL.encode_text(text_tokens).float()
|
85 |
+
text_features /= text_features.norm(dim=-1, keepdim=True)
|
86 |
+
return text_features
|
87 |
+
|
88 |
+
# Source: https://stackoverflow.com/a/480227
|
89 |
+
def f7(self, seq):
|
90 |
+
seen = set()
|
91 |
+
seen_add = seen.add # This is for performance improvement, don't remove
|
92 |
+
return [x for x in seq if not (x in seen or seen_add(x))]
|
93 |
+
|
94 |
+
def search_top_k(self, q, k=5, pool_size=1000, search_mod="Majority"):
|
95 |
+
q = self.text_to_vector(q)
|
96 |
+
nearest_data_points = self.simsearcher.get_nearest_labels(q, pool_size)
|
97 |
+
|
98 |
+
if search_mod == "Majority":
|
99 |
+
topKs = [x[0] for x in Counter(nearest_data_points[0]).most_common(k)]
|
100 |
+
elif search_mod == "Top-K":
|
101 |
+
topKs = list(self.f7(nearest_data_points[0]))[:k]
|
102 |
+
|
103 |
+
video_filename = [
|
104 |
+
f"./Videos/{self.GAME_NAME}/"
|
105 |
+
+ self.class_id_to_file[x].replace("npy", "mp4")
|
106 |
+
for x in topKs
|
107 |
+
]
|
108 |
+
|
109 |
+
return video_filename
|
110 |
|
111 |
|
112 |
################ SEARCH CORE ################
|
|
|
115 |
vit_model.eval()
|
116 |
|
117 |
saved_searchers = {}
|
118 |
+
|
119 |
+
|
120 |
def gradio_search(query, game_name, selected_model, aggregator, pool_size, k=6):
|
121 |
+
# print(query, game_name, selected_model, aggregator, pool_size)
|
122 |
+
if f"{game_name}_{selected_model}" in saved_searchers.keys():
|
123 |
+
searcher = saved_searchers[f"{game_name}_{selected_model}"]
|
|
|
|
|
|
|
|
|
124 |
else:
|
125 |
+
if selected_model == "ViT-B/32":
|
126 |
+
model = vit_model
|
127 |
+
searcher = GamePhysicsSearcher(CLIP_MODEL=model, GAME_NAME=game_name)
|
128 |
+
else:
|
129 |
+
raise
|
130 |
+
|
131 |
+
saved_searchers[f"{game_name}_{selected_model}"] = searcher
|
132 |
+
|
133 |
+
results = []
|
134 |
+
relevant_videos = searcher.search_top_k(
|
135 |
+
query, k=k, pool_size=pool_size, search_mod=aggregator
|
136 |
+
)
|
137 |
+
|
138 |
+
params = ", ".join(
|
139 |
+
map(str, [query, game_name, selected_model, aggregator, pool_size])
|
140 |
+
)
|
141 |
+
results.append(params)
|
142 |
+
|
143 |
+
for v in relevant_videos:
|
144 |
+
results.append(v)
|
145 |
+
sid = v.split("/")[-1].split(".")[0]
|
146 |
+
results.append(f"https://www.reddit.com/r/GamePhysics/comments/{sid}/")
|
147 |
+
return results
|
148 |
+
|
149 |
+
|
150 |
+
def main():
|
151 |
+
list_of_games = ["Grand Theft Auto V"]
|
152 |
+
|
153 |
+
title = "CLIP + GamePhysics - Searching dataset of Gameplay bugs"
|
154 |
+
description = "Enter your query and select the game you want to search. The results will be displayed in the console."
|
155 |
+
article = """
|
156 |
+
This demo shows how to use the CLIP model to search for gameplay bugs in a video game.
|
157 |
+
"""
|
158 |
+
|
159 |
+
# GRADIO APP
|
160 |
+
iface = gr.Interface(
|
161 |
+
fn=gradio_search,
|
162 |
+
inputs=[
|
163 |
+
gr.inputs.Textbox(
|
164 |
+
lines=1,
|
165 |
+
placeholder="Search Query",
|
166 |
+
default="A person flying in the air",
|
167 |
+
label=None,
|
168 |
+
),
|
169 |
+
gr.inputs.Radio(list_of_games, label="Game To Search"),
|
170 |
+
gr.inputs.Radio(["ViT-B/32"], label="MODEL"),
|
171 |
+
gr.inputs.Radio(["Majority", "Top-K"], label="Aggregator"),
|
172 |
+
gr.inputs.Slider(300, 2000, label="Pool Size", default=1000),
|
173 |
+
],
|
174 |
+
outputs=[
|
175 |
+
gr.outputs.Textbox(type="auto", label="Search Params"),
|
176 |
+
gr.outputs.Video(type="mp4", label="Result 1"),
|
177 |
+
gr.outputs.Textbox(type="auto", label="Submission URL - Result 1"),
|
178 |
+
gr.outputs.Video(type="mp4", label="Result 2"),
|
179 |
+
gr.outputs.Textbox(type="auto", label="Submission URL - Result 2"),
|
180 |
+
gr.outputs.Video(type="mp4", label="Result 3"),
|
181 |
+
gr.outputs.Textbox(type="auto", label="Submission URL - Result 3"),
|
182 |
+
gr.outputs.Video(type="mp4", label="Result 4"),
|
183 |
+
gr.outputs.Textbox(type="auto", label="Submission URL - Result 4"),
|
184 |
+
gr.outputs.Video(type="mp4", label="Result 5"),
|
185 |
+
gr.outputs.Textbox(type="auto", label="Submission URL - Result 5"),
|
186 |
+
],
|
187 |
+
examples=[
|
188 |
+
["A red car", list_of_games[0], "ViT-B/32", "Top-K", 1000],
|
189 |
+
["A person wearing pink", list_of_games[0], "ViT-B/32", "Top-K", 1000],
|
190 |
+
["A car flying in the air", list_of_games[0], "ViT-B/32", "Majority", 1000],
|
191 |
+
[
|
192 |
+
"A person flying in the air",
|
193 |
+
list_of_games[0],
|
194 |
+
"ViT-B/32",
|
195 |
+
"Majority",
|
196 |
+
1000,
|
197 |
+
],
|
198 |
+
[
|
199 |
+
"A car in vertical position",
|
200 |
+
list_of_games[0],
|
201 |
+
"ViT-B/32",
|
202 |
+
"Majority",
|
203 |
+
1000,
|
204 |
+
],
|
205 |
+
["A bike inside a car", list_of_games[0], "ViT-B/32", "Majority", 1000],
|
206 |
+
["A bike on a wall", list_of_games[0], "ViT-B/32", "Majority", 1000],
|
207 |
+
["A car stuck in a rock", list_of_games[0], "ViT-B/32", "Majority", 1000],
|
208 |
+
["A car stuck in a tree", list_of_games[0], "ViT-B/32", "Majority", 1000],
|
209 |
+
],
|
210 |
+
title=title,
|
211 |
+
description=description,
|
212 |
+
article=article,
|
213 |
+
enable_queue=True,
|
214 |
+
)
|
215 |
+
|
216 |
+
iface.launch()
|
217 |
+
|
218 |
+
|
219 |
+
if __name__ == "__main__":
|
220 |
+
main()
|
requirements.txt
CHANGED
@@ -6,4 +6,5 @@ scikit-image
|
|
6 |
gdown
|
7 |
torchvision
|
8 |
git+https://github.com/openai/CLIP.git
|
9 |
-
faiss-cpu
|
|
|
|
6 |
gdown
|
7 |
torchvision
|
8 |
git+https://github.com/openai/CLIP.git
|
9 |
+
faiss-cpu
|
10 |
+
gdown
|