Modified training example to follow hyperparameter

selection format of other examples.
arnab39 · Mar 12, 2024 · 4c6bfcf · 4c6bfcf
1 parent c42c473
commit 4c6bfcf
Show file tree

Hide file tree

Showing 4 changed files with 39 additions and 64 deletions.
diff --git a/equiadapt/nbody/canonicalization/continuous_group.py b/equiadapt/nbody/canonicalization/continuous_group.py
@@ -40,7 +40,6 @@ def canonicalize(self, nodes, loc, edges, vel, edge_attr, charges):
         self.device = nodes.device
 
         group_element_dict = self.get_groupelement(nodes, loc, edges, vel, edge_attr, charges)
-        rotation_matrix = group_element_dict["rotation_matrix"]
         translation_vectors = group_element_dict["translation_vectors"]
         rotation_matrix_inverse = group_element_dict["rotation_matrix_inverse"]
 

diff --git a/examples/nbody/model.py b/examples/nbody/model.py
@@ -11,8 +11,8 @@ class NBodyPipeline(pl.LightningModule):
     def __init__(self, hyperparams: DictConfig):
         super().__init__()
         self.hyperparams = hyperparams
-        self.prediction_network = get_prediction_network(hyperparams)
-        canonicalization_network = get_canonicalization_network(hyperparams)
+        self.prediction_network = get_prediction_network(hyperparams.pred_hyperparams)
+        canonicalization_network = get_canonicalization_network(hyperparams.canon_hyperparams)
 
         self.canonicalizer = ContinuousGroupNBody(canonicalization_network, hyperparams)
 

diff --git a/examples/nbody/model_utils.py b/examples/nbody/model_utils.py
@@ -3,54 +3,27 @@
 from omegaconf import OmegaConf
 import torch
 
+
 def get_canonicalization_network(hyperparams):
-    architecture = hyperparams.canon_model_type
-    model_hyperparams = {
-            "num_layers": hyperparams.canon_num_layers,
-            "hidden_dim": hyperparams.canon_hidden_dim,
-            "layer_pooling": hyperparams.canon_layer_pooling,
-            "final_pooling": hyperparams.canon_final_pooling,
-            "out_dim": 4,
-            "batch_size": hyperparams.batch_size,
-            "nonlinearity": hyperparams.canon_nonlinearity,
-            "canon_feature": hyperparams.canon_feature,
-            "canon_translation": hyperparams.canon_translation,
-            "angular_feature": hyperparams.canon_angular_feature,
-            "dropout": hyperparams.canon_dropout,
-        }
-    model_hyperparams = OmegaConf.create(dict(model_hyperparams))
-
+    architecture = hyperparams.architecture
     model_dict = {
-    #"EGNN": lambda: EGNN_vel(hyperparams),
-    "vndeepsets": lambda: VNDeepSets(model_hyperparams),
+    "vndeepsets": lambda: VNDeepSets(hyperparams),
     }
 
-    if architecture not in model_dict:
-        raise ValueError(f'{architecture} is not implemented as prediction network for now.')
-
     return model_dict[architecture]()
 
 
 def get_prediction_network(hyperparams):
-    architecture = hyperparams.pred_model_type
-    model_hyperparams = {
-            "num_layers": hyperparams.num_layers,
-            "hidden_dim": hyperparams.hidden_dim,
-            "input_dim": hyperparams.input_dim,
-            "in_node_nf": hyperparams.in_node_nf,
-            "in_edge_nf": hyperparams.in_edge_nf,
-        }
-    model_hyperparams = OmegaConf.create(dict(model_hyperparams))
+    architecture = hyperparams.architecture
     model_dict = {
-            "GNN": lambda: GNN(model_hyperparams),
-            "EGNN": lambda: EGNN_vel(model_hyperparams),
-            "vndeepsets": lambda: VNDeepSets(model_hyperparams),
-            "Transformer": lambda: Transformer(model_hyperparams),
+            "GNN": lambda: GNN(hyperparams),
+            "EGNN": lambda: EGNN_vel(hyperparams),
+            "vndeepsets": lambda: VNDeepSets(hyperparams),
+            "Transformer": lambda: Transformer(hyperparams),
         }
-
     if architecture not in model_dict:
         raise ValueError(f'{architecture} is not implemented as prediction network for now.')
-    
+
     return model_dict[architecture]()
 
 def get_edges(batch_size, n_nodes):

diff --git a/examples/nbody/train.py b/examples/nbody/train.py
@@ -12,48 +12,51 @@
 
 HYPERPARAMS = {"model": "NBodyPipeline", 
                "canon_model_type": "vndeepsets", 
-               "pred_model_type": "GNN", 
+               "pred_model_type": "Transformer", 
                "batch_size": 100, 
                "dryrun": False, 
                "use_wandb": False, 
                "checkpoint": False, 
                "num_epochs": 1000, 
                "num_workers":0, 
                "auto_tune":False, 
-               "seed": 0}
-
+               "seed": 0,
+                "learning_rate": 1e-3, #1e-3
+                "weight_decay": 1e-12,
+                "patience": 1000,
+}
 
-NBODY_HYPERPARAMS = {
-    "learning_rate": 1e-3, #1e-3
-    "weight_decay": 1e-12,
-    "patience": 1000,
-    "hidden_dim": 32, #32
-    "input_dim": 6,
-    "in_node_nf": 1,
-    "in_edge_nf": 2,
-    "num_layers": 4, #4
+CANON_HYPERPARAMS = {
+    "architecture": "vndeepsets",
+    "num_layers": 4,
+    "hidden_dim": 16,
+    "layer_pooling": "mean",
+    "final_pooling": "mean",
     "out_dim": 4,
-    "canon_num_layers": 4,
-    "canon_hidden_dim": 16,
-    "canon_layer_pooling": "mean",
-    "canon_final_pooling": "mean",
-    "canon_nonlinearity": "relu",
+    "batch_size": 100,
+    "nonlinearity": "relu",
     "canon_feature": "p",
     "canon_translation": False,
-    "canon_angular_feature": 0,
-    "canon_dropout": 0.5,
-    "freeze_canon": False,
-    "layer_pooling": "sum",
-    "final_pooling": "mean",
-    "nonlinearity": "relu",
     "angular_feature": "pv",
-    "dropout": 0, #0
+    "dropout": 0.5,
+}
+
+PRED_HYPERPARAMS = {
+    "architecture": "Transformer",
+    "num_layers": 4,
+    "hidden_dim": 32,
+    "input_dim": 6,
+    "in_node_nf": 1,
+    "in_edge_nf": 2,
     "nheads": 8,
     "ff_hidden": 32
 }
 
+HYPERPARAMS["canon_hyperparams"] = CANON_HYPERPARAMS
+HYPERPARAMS["pred_hyperparams"] = PRED_HYPERPARAMS
+
 def train_nbody():
-    hyperparams = HYPERPARAMS | NBODY_HYPERPARAMS
+    hyperparams = HYPERPARAMS
 
     if not hyperparams["use_wandb"]:
         print('Wandb disable for logging.')