- [MINOR] [SOURCE] [UPDATE] 1. update app.py

app.py

@@ -23,6 +23,18 @@ mean_clip, std_clip = (
     torch.FloatTensor([68.50, 66.63, 70.32])
 )
 
+sample_interval = 30
+
+def get_sampler_params(video_path):
+    vr = VideoReader(video_path)
+    total_frames = len(vr)
+    clip_len = (total_frames + sample_interval // 2) // sample_interval
+    if clip_len == 0:
+        clip_len = 1
+    t_frag = clip_len
+    
+    return total_frames, clip_len, t_frag
+
 def fuse_results(results: list):
     x = (results[0] + results[1] + results[2])
     return {
@@ -42,7 +54,19 @@ def inference_one_video(input_video):
     
     dopt = opt["data"]["val-ytugc"]["args"]
     temporal_samplers = {}
+    
+    # auto decision of parameters of sampler
+    total_frames, clip_len, t_frag = get_sampler_params(input_video)
+                    
     for stype, sopt in dopt["sample_types"].items():
+        sopt["clip_len"] = clip_len
+        sopt["t_frag"] = t_frag
+        if stype == 'technical' or stype == 'aesthetic':
+            if total_frames > 1:
+                sopt["clip_len"] = clip_len * 2
+            if stype == 'technical':
+                sopt["aligned"] = sopt["clip_len"]
+        print(sopt["clip_len"], sopt["t_frag"])
         temporal_samplers[stype] = UnifiedFrameSampler(
             sopt["clip_len"] // sopt["t_frag"],
             sopt["t_frag"],