Merge pull request #223 from ar-nowaczynski/unclamped_fape_per_sample

Set clamped vs unclamped FAPE for each sample in batch independently

openfold/data/data_modules.py

@@ -440,11 +440,6 @@ def _prep_batch_properties_probs(self):
         stage_cfg = self.config[self.stage]
 
         max_iters = self.config.common.max_recycling_iters
-        if(stage_cfg.supervised):
-            clamp_prob = self.config.supervised.clamp_prob
-            keyed_probs.append(
-                ("use_clamped_fape", [1 - clamp_prob, clamp_prob])
-            )
 
         if(stage_cfg.uniform_recycling):
             recycling_probs = [

openfold/data/feature_pipeline.py

@@ -94,6 +94,21 @@ def np_example_to_features(
             cfg[mode],
         )
 
+    if mode == "train":
+        p = torch.rand(1).item()
+        use_clamped_fape_value = float(p < cfg.supervised.clamp_prob)
+        features["use_clamped_fape"] = torch.full(
+            size=[cfg.common.max_recycling_iters + 1],
+            fill_value=use_clamped_fape_value,
+            dtype=torch.float32,
+        )
+    else:
+        features["use_clamped_fape"] = torch.full(
+            size=[cfg.common.max_recycling_iters + 1],
+            fill_value=0.0,
+            dtype=torch.float32,
+        )
+
     return {k: v for k, v in features.items()}