Fang Yunhao
commited on
Commit
·
d2dbd7f
1
Parent(s):
246b4eb
Update evaluation script.
Browse files- evaluation.py +16 -12
evaluation.py
CHANGED
@@ -34,12 +34,12 @@ if __name__ == "__main__":
|
|
34 |
help="Path to judge model checkpoint.",
|
35 |
)
|
36 |
parser.add_argument(
|
37 |
-
"--
|
38 |
type=str,
|
39 |
help="Path to the generated video directory.",
|
40 |
)
|
41 |
parser.add_argument(
|
42 |
-
"--
|
43 |
type=str,
|
44 |
help="Path to save evaluation results.",
|
45 |
)
|
@@ -47,25 +47,27 @@ if __name__ == "__main__":
|
|
47 |
args = parser.parse_args()
|
48 |
|
49 |
validation_set = load("./worldmodelbench.json")
|
50 |
-
save_name = args.judge.replace("/", "_")
|
51 |
if args.cot:
|
52 |
-
save_name += "_cot"
|
53 |
results = None
|
54 |
-
if os.path.exists(save_name):
|
55 |
-
results = load(save_name)
|
56 |
-
|
57 |
-
|
|
|
|
|
|
|
58 |
else:
|
59 |
model = llava.load(args.judge)
|
60 |
-
video_root = args.video_path
|
61 |
|
62 |
preds = dict()
|
63 |
accs = defaultdict(list)
|
64 |
for vid, v_i in tqdm(enumerate(validation_set), total=len(validation_set)):
|
|
|
65 |
video_name = v_i["first_frame"].split("/")[-1].split(".")[0]
|
66 |
-
video = os.path.join(
|
67 |
video = llava.Video(video)
|
68 |
-
##
|
69 |
for k in ["instruction", "physical_laws", "commonsense"]:
|
70 |
preds_i = []
|
71 |
prompt_template = PROMPT_TEMPLATES[k]
|
@@ -83,6 +85,7 @@ if __name__ == "__main__":
|
|
83 |
).replace("Let's analyze step-by-step and conclude with", "Answer with")
|
84 |
pred = model.generate_content([video, text_prompt])
|
85 |
preds_i.append(pred)
|
|
|
86 |
accs_i.append("no" in pred.lower())
|
87 |
accs[k].append(np.mean(accs_i))
|
88 |
else:
|
@@ -107,7 +110,8 @@ if __name__ == "__main__":
|
|
107 |
print(f"{k} accuracy: {np.mean(v) * 100}%.")
|
108 |
else:
|
109 |
print(f"{k} accuracy: {v}%.")
|
|
|
110 |
if results is None:
|
111 |
results = {"preds": preds, "accs": accs}
|
112 |
-
dump(results, f"./{save_name}.json", indent=4)
|
113 |
|
|
|
34 |
help="Path to judge model checkpoint.",
|
35 |
)
|
36 |
parser.add_argument(
|
37 |
+
"--video_dir",
|
38 |
type=str,
|
39 |
help="Path to the generated video directory.",
|
40 |
)
|
41 |
parser.add_argument(
|
42 |
+
"--save_name",
|
43 |
type=str,
|
44 |
help="Path to save evaluation results.",
|
45 |
)
|
|
|
47 |
args = parser.parse_args()
|
48 |
|
49 |
validation_set = load("./worldmodelbench.json")
|
|
|
50 |
if args.cot:
|
51 |
+
args.save_name += "_cot"
|
52 |
results = None
|
53 |
+
if os.path.exists(args.save_name):
|
54 |
+
results = load(args.save_name)
|
55 |
+
try:
|
56 |
+
preds = results["preds"]
|
57 |
+
accs = results["accs"]
|
58 |
+
except:
|
59 |
+
raise "Expected keys are not found in the results."
|
60 |
else:
|
61 |
model = llava.load(args.judge)
|
|
|
62 |
|
63 |
preds = dict()
|
64 |
accs = defaultdict(list)
|
65 |
for vid, v_i in tqdm(enumerate(validation_set), total=len(validation_set)):
|
66 |
+
## Load video
|
67 |
video_name = v_i["first_frame"].split("/")[-1].split(".")[0]
|
68 |
+
video = os.path.join(args.video_dir, video_name + ".mp4")
|
69 |
video = llava.Video(video)
|
70 |
+
## Traverse criterions
|
71 |
for k in ["instruction", "physical_laws", "commonsense"]:
|
72 |
preds_i = []
|
73 |
prompt_template = PROMPT_TEMPLATES[k]
|
|
|
85 |
).replace("Let's analyze step-by-step and conclude with", "Answer with")
|
86 |
pred = model.generate_content([video, text_prompt])
|
87 |
preds_i.append(pred)
|
88 |
+
## Always ask for violations, so a "No" is preferred!
|
89 |
accs_i.append("no" in pred.lower())
|
90 |
accs[k].append(np.mean(accs_i))
|
91 |
else:
|
|
|
110 |
print(f"{k} accuracy: {np.mean(v) * 100}%.")
|
111 |
else:
|
112 |
print(f"{k} accuracy: {v}%.")
|
113 |
+
## Save results
|
114 |
if results is None:
|
115 |
results = {"preds": preds, "accs": accs}
|
116 |
+
dump(results, f"./{args.save_name}.json", indent=4)
|
117 |
|