fixing bugs in transform, expert demo processing, main train function, and behavior cloning class. need to get bc class parameters to return nonempty list

.gitignore

@@ -140,6 +140,7 @@ expert_data/
 
 # Results
 experiments/results/
+output/
 
 # Dependencies
 InteractionSimulator/

src/bc/bc.py

@@ -2,7 +2,7 @@ import torch
 import torch.nn as nn
 from torch.utils.data import DataLoader, RandomSampler
 import pickle
-from torch.utils.tensorboard import SummaryWriter
+#from torch.utils.tensorboard import SummaryWriter
 
 from src.policies import DeepSetsPolicy
 from src.util.transform import SciKitMinMaxScaler
@@ -13,7 +13,7 @@ class BehaviorCloningPolicy():
     Class for (continuous) behavior cloning policy
     """
     
-    def __init__(self, config: dict transforms: dict={}):
+    def __init__(self, config: dict, transforms: dict={}):
         """
         Initialize BehaviorCloningPolicy
         Args:
@@ -29,7 +29,7 @@ class BehaviorCloningPolicy():
         return self._transforms
     
     @transforms.setter
-    def transforms(self, transforms)
+    def transforms(self, transforms):
         self._transforms=transforms
 
     def __call__(self, ob):
@@ -93,11 +93,11 @@ def generate_transforms(dataset):
         dataset (Dataset): dataset of demo observations and actions
     """
     transforms = {
-        'action': SciKitMinMaxScaler()
-        'state': SciKitMinMaxScaler()
-        'relative_state': SciKitMinMaxScaler(reduce_dim=2)
-        'path_x': SciKitMinMaxScaler(reduce_dim=2)
-        'path_y': SciKitMinMaxScaler(reduce_dim=2)
+        'action': SciKitMinMaxScaler(),
+        'state': SciKitMinMaxScaler(),
+        'relative_state': SciKitMinMaxScaler(reduce_dim=2),
+        'path_x': SciKitMinMaxScaler(reduce_dim=2),
+        'path_y': SciKitMinMaxScaler(reduce_dim=2),
     }
     for key in transforms.keys():
         transforms[key].fit(dataset[:][key])
@@ -126,6 +126,8 @@ def train(train_dataset, cv_dataset, policy, filestr, **kwargs):
 
     # generate loss function, optimizer
     loss_fn = nn.HuberLoss(reduction='sum')
+    import pdb
+    pdb.set_trace()
     optimizer = torch.optim.Adam(policy.parameters(), lr=learning_rate, weight_decay=weight_decay)
 
     for i in train_epochs:

src/expert_data.py

@@ -43,7 +43,8 @@ def generate_expert_data(path: str='expert_data', loc: int = 0, track:int = 0, *
         env_state = env.projected_state
         nni = ~torch.isnan(env_state[:,0])
         norms = torch.norm(env_state[nni,:2]-states[i,nni,:2], dim=1)
-        max_devs.append(norms.max())
+        if len(norms)>0:
+            max_devs.append(norms.max())
 
         # propagate environment
         ob, r, done, info = env.step(env.target_state(svt.simstate[i+1]))

src/main.py

@@ -37,7 +37,7 @@ def main(method='bc', train=False, test=False, loc=0, config_path=None, **kwargs
     # make prefix of output files
     outdir = opj('output',method,'loc%02i'%(loc))
     if not os.path.isdir(outdir):
-        os.mkdir(outdir)
+        os.makedirs(outdir)
     filestr = opj(outdir, basestr(**kwargs))
     
     # load config
@@ -60,7 +60,7 @@ def main(method='bc', train=False, test=False, loc=0, config_path=None, **kwargs
 
         # make policy, train and test datasets, and send to 
         policy = policy_class(config)
-        train_dataset = InteractionDatasetSingleAgent(loc=loc, tracks=[0,1,2])
+        train_dataset = InteractionDatasetSingleAgent(loc=loc, tracks=[0])#,1,2])
         cv_dataset = InteractionDatasetSingleAgent(loc=loc, tracks=[3])
         train_fn(train_dataset, cv_dataset, policy, filestr, **kwargs)
     
@@ -71,11 +71,11 @@ def main(method='bc', train=False, test=False, loc=0, config_path=None, **kwargs
         policy.eval()
 
         # simulate policy
-        test_track = 4
+        track = 4
         simulate_policy(policy, loc=loc, track=track, filestr=filestr)
 
         # run test metrics
-        test_dataset = InteractionDatasetSingleAgent(loc=loc, tracks=[4])
+        test_dataset = InteractionDatasetSingleAgent(loc=loc, tracks=[track])
         metrics(filestr, test_dataset, policy)
 
 
@@ -132,16 +132,14 @@ def parse_args():
         default='config/networks.json5', type=str)
     parser.add_argument('--seed', default=0, type=int,
         help='seed')    
-    parser.add_argument()              
-    parser.add_argument()
     args = parser.parse_args()
     kwargs = {
-        'train'=args.train, 
-        'test'=args.test, 
-        'method'=args.method,
-        'loc'=args.loc,
-        'config'=args.config,
-        'seed'=args.seed
+        'train':args.train, 
+        'test':args.test, 
+        'method':args.method,
+        'loc':args.loc,
+        'config_path':args.config,
+        'seed':args.seed
     }
     return kwargs
 

src/metrics.py

@@ -35,7 +35,7 @@ def average_velocity(x):
     """
     pass
 
-def divergence(p, q, type='kl')
+def divergence(p, q, type='kl'):
     """
     Calculate a divergence between p and q
     Args:

src/util/transform.py

@@ -52,27 +52,27 @@ class SciKitTransform(Transform):
                 e.g. with reduce_dim=2, (A, B, C, D, E) will be reshaped to (A*B, C*D*E)
         """
         self.tf = tf
-        self.reduce_dim
+        self.reduce_dim = reduce_dim
         super(SciKitTransform, self).__init__()
 
     def fit(self, X):
         nd = X.ndim
         if self.reduce_dim:
-            self.nfeatures = X.shape[reduce_dim:].prod()
+            self.nfeatures = int(torch.tensor(X.shape[self.reduce_dim:]).prod())
         else:
             assert nd==2, 'Invalid ndim'
             self.nfeatures = X.shape[1]
 
-        self.tf.fit(X.reshape((-1,selfnfeatures)))
+        self.tf.fit(X.reshape((-1,self.nfeatures)))
 
     def transform(self, X):
         shape = X.shape
-        t = torch.tensor(self.tf.transform(X.reshape((-1,selfnfeatures))), dtype=torch.float)
+        t = torch.tensor(self.tf.transform(X.reshape((-1,self.nfeatures))), dtype=torch.float)
         return t.reshape(shape)
 
     def inverse_transform(self, X):
         shape = X.shape
-        it =  torch.tensor(self.tf.inverse_transform(X.reshape((-1,selfnfeatures))), dtype=torch.float)
+        it =  torch.tensor(self.tf.inverse_transform(X.reshape((-1,self.nfeatures))), dtype=torch.float)
         return it.reshape(shape)
 
 class SciKitStandardScaler(SciKitTransform):