Update acwm_dataset.py

acwm_dataset.py

@@ -62,6 +62,8 @@ class ACWMDataset(Dataset):
         force_clip=8.0,
         contact_gain=5.0,
         eps=1e-6,
+        repeat=1,
+        max_samples=None,
     ):
         self.source_root = source_root
         self.height = height
@@ -69,6 +71,7 @@ class ACWMDataset(Dataset):
         self.force_clip = force_clip
         self.contact_gain = contact_gain
         self.eps = eps
+        self.repeat = max(1, int(repeat))
 
         # Required by launch_training_task: it checks `dataset.load_from_cache`
         # to decide model({}, inputs=data) vs model(data). We never use the
@@ -77,7 +80,11 @@ class ACWMDataset(Dataset):
 
         with open(pipeline_json) as f:
             blob = json.load(f)
-        self.samples = blob["samples"] if isinstance(blob, dict) else blob
+        samples = blob["samples"] if isinstance(blob, dict) else blob
+        # optional subset for overfit experiments
+        if max_samples is not None:
+            samples = samples[:max_samples]
+        self.samples = samples
 
         # ---- contact/force pixel-space normalization stats (VAE-matched) ----
         with open(physical_stats_json) as f:
@@ -100,7 +107,7 @@ class ACWMDataset(Dataset):
             np.asarray(astats["std"], dtype=np.float32), eps)            # (7,)
 
     def __len__(self):
-        return len(self.samples)
+        return len(self.samples) * self.repeat
 
     # ------------------------------------------------------------------ #
     # loaders
@@ -147,7 +154,7 @@ class ACWMDataset(Dataset):
     # ------------------------------------------------------------------ #
 
     def __getitem__(self, idx):
-        s = self.samples[idx]
+        s = self.samples[idx % len(self.samples)]
 
         obs_png = s["observation_frame"]
         future_pngs = s["frames"]                       # 16 frames