Merge branch 'master' of github.com:nimarb/pytorch_influence_functions

Author: nimarb

pytorch_influence_functions/calc_influence_function.py

@@ -466,7 +466,7 @@ def calc_img_wise(config, model, train_loader, test_loader):
 
         start_time = time.time()
         influence, harmful, helpful, _ = calc_influence_single(
-            model, train_loader, test_loader, test_id_num=i, gpu=0,
+            model, train_loader, test_loader, test_id_num=i, gpu=config['gpu'],
             recursion_depth=config['recursion_depth'], r=config['r_averaging'])
         end_time = time.time()
 

pytorch_influence_functions/influence_function.py

@@ -43,7 +43,8 @@ def s_test(z_test, t_test, model, z_loader, gpu=-1, damp=0.01, scale=25.0,
                 x, t = x.cuda(), t.cuda()
             y = model(x)
             loss = calc_loss(y, t)
-            hv = hvp(loss, list(model.parameters()), h_estimate)
+            params = [ p for p in model.parameters() if p.requires_grad ]
+            hv = hvp(loss, params, h_estimate)
             # Recursively caclulate h_estimate
             h_estimate = [
                 _v + (1 - damp) * _h_e - _hv / scale
@@ -93,7 +94,8 @@ def grad_z(z, t, model, gpu=-1):
     y = model(z)
     loss = calc_loss(y, t)
     # Compute sum of gradients from model parameters to loss
-    return list(grad(loss, list(model.parameters()), create_graph=True))
+    params = [ p for p in model.parameters() if p.requires_grad ]
+    return list(grad(loss, params, create_graph=True))
 
 
 def hvp(y, w, v):