Merge pull request #1 from XieBinghui/main

minor bug fix by Binghui

ColoredMNIST/pair.py

@@ -88,8 +88,8 @@ def step(self, closure=None):
                 grads.append(torch.cat(cur_grad))
 
             G = torch.stack(grads)
-            if self.get_grad_sim:
-                grad_sim = get_grad_sim(G,losses,preference=self.preference,is_G=True)
+            # if self.get_grad_sim:
+            #     grad_sim = get_grad_sim(G,losses,preference=self.preference,is_G=True)
             GG = G @ G.T
             moo_losses = np.stack([l.item() for l in losses])
             reset_optimizer = False
@@ -120,7 +120,6 @@ def step(self, closure=None):
 
 import numpy as np
 import cvxpy as cp
-import cvxopt
 
 class EPO(object):
     r"""

ColoredMNIST/pair_alg.py

@@ -99,7 +99,6 @@ def pair_train(mlp, topmlp, steps, envs, test_envs, lossf, \
             optimizer.zero_grad()
             optimizer.set_losses(losses=losses)
             pair_loss, moo_losses, mu_rl, alphas = optimizer.step()
-            pair_res = np.array([pair_loss, mu_rl, alphas])
         else:
             loss = erm_loss
 

ColoredMNIST/run_exp.py

@@ -173,10 +173,10 @@ def main(flags):
             np.save(os.path.join(flags.save_dir,'group%d.npy' % restart), group)
             np.save(os.path.join(flags.save_dir,'pseudolabel%d.npy' % restart), pseudolabel )
 
-    logs = np.array(logs)
+    # logs = np.array(logs)
 
-    if flags.save_dir is not None:
-        np.save(os.path.join(flags.save_dir,'logs.npy'), logs)
+    # if flags.save_dir is not None:
+    #     np.save(os.path.join(flags.save_dir,'logs.npy'), logs)
 
     return results, logs
 

PAIR/pair.py

@@ -88,8 +88,8 @@ def step(self, closure=None):
                 grads.append(torch.cat(cur_grad))
 
             G = torch.stack(grads)
-            if self.get_grad_sim:
-                grad_sim = get_grad_sim(G,losses,preference=self.preference,is_G=True)
+            # if self.get_grad_sim:
+            #     grad_sim = get_grad_sim(G,losses,preference=self.preference,is_G=True)
             GG = G @ G.T
             moo_losses = np.stack([l.item() for l in losses])
             reset_optimizer = False

WILDS/scripts/iwildcam.sh

@@ -4,5 +4,5 @@ exp_dir=""; # the experiment directory will be the same as data_dir by default
 
 for seed in {0..2};
 do
-CUDA_VISIBLE_DEVICES=${gpu} WANDB_MODE=offline python3 ./src/main.py --frozen --need_pretrain --use_old --data-dir ${data_dir} --dataset iwdilcam --algorithm pair -pc 0 -al -ac 1e-2 --seed ${seed} 
+CUDA_VISIBLE_DEVICES=${gpu} WANDB_MODE=offline python3 ./src/main.py --frozen --need_pretrain --use_old --data-dir ${data_dir} --dataset iwildcam --algorithm pair -pc 0 -al -ac 1e-2 --seed ${seed} 
 done;

WILDS/src/main.py

@@ -22,7 +22,7 @@
     Logger, return_predict_fn, return_criterion, fish_step
 
 # This is secret and shouldn't be checked into version control
-os.environ["WANDB_API_KEY"]=None
+# os.environ["WANDB_API_KEY"]=None
 # Name and notes optional
 # WANDB_NAME="My first run"
 # WANDB_NOTES="Smaller learning rate, more regularization."
@@ -192,6 +192,7 @@
 else:
     classifier = model.classifier
 trainable_params = classifier.parameters() if args.frozen else model.parameters()
+trainable_params = list(trainable_params)
 optimiserC = opt(trainable_params, **args.optimiser_args)
 predict_fn, criterion = return_predict_fn(args.dataset), return_criterion(args.dataset)
 

WILDS/src/models/__init__.py

@@ -1,9 +1,8 @@
 from .camelyon import Model as camelyon
-from .cdsprites import Model as cdsprites
 from .civil import Model as civil
 from .fmow import Model as fmow
 from .iwildcam import Model as iwildcam
 from .poverty import Model as poverty
 from .rxrx import Model as rxrx
 
-__all__ = [cdsprites,  iwildcam, camelyon, amazon, civil, fmow, poverty, rxrx]
+__all__ = [iwildcam, camelyon, civil, fmow, poverty, rxrx]

WILDS/src/pair.py

@@ -88,8 +88,8 @@ def step(self, closure=None):
                 grads.append(torch.cat(cur_grad))
 
             G = torch.stack(grads)
-            if self.get_grad_sim:
-                grad_sim = get_grad_sim(G,losses,preference=self.preference,is_G=True)
+            # if self.get_grad_sim:
+            #     grad_sim = get_grad_sim(G,losses,preference=self.preference,is_G=True)
             GG = G @ G.T
             moo_losses = np.stack([l.item() for l in losses])
             reset_optimizer = False
@@ -106,7 +106,10 @@ def step(self, closure=None):
                 alpha = self.preference / np.sum(self.preference)
 
             scales = torch.from_numpy(alpha).float().to(losses[-1].device)
-            pair_loss = scales.dot(losses)
+            pair_loss = 0.0
+            for i in range(len(scales)):
+                pair_loss += scales[i] * losses[i]
+            # pair_loss = scales.dot(losses)
             if reset_optimizer:
                 self.optimizer.param_groups[0]["lr"]/=5
                 # self.optimizer = torch.optim.Adam(self.params,lr=self.optimizer.param_groups[0]["lr"]/5)
@@ -120,7 +123,6 @@ def step(self, closure=None):
 
 import numpy as np
 import cvxpy as cp
-import cvxopt
 
 class EPO(object):
     r"""