diff --git a/data/mtl_ds.py b/data/mtl_ds.py
index 80df3c6..670dc58 100644
--- a/data/mtl_ds.py
+++ b/data/mtl_ds.py
@@ -22,13 +22,12 @@
 
 from utils import mkdir_if_missing
 from torch.utils.data import DataLoader
-from torch._six import string_classes
 from collections.abc import Mapping, Sequence
 from torchvision import transforms
 from easydict import EasyDict as edict
 from skimage.morphology import thin
 
-
+string_classes = (str,)
 int_classes = int
 _use_shared_memory = False
 r"""Whether to use shared memory in default_collate"""
diff --git a/evaluation/jaccard.py b/evaluation/jaccard.py
index a9b0b60..a36a3e3 100644
--- a/evaluation/jaccard.py
+++ b/evaluation/jaccard.py
@@ -17,9 +17,9 @@ def jaccard(gt, pred, void_pixels=None):
         void_pixels = np.zeros_like(gt)
     assert(void_pixels.shape == gt.shape)
 
-    gt = gt.astype(np.bool)
-    pred = pred.astype(np.bool)
-    void_pixels = void_pixels.astype(np.bool)
+    gt = gt.astype(bool)
+    pred = pred.astype(bool)
+    void_pixels = void_pixels.astype(bool)
     if np.isclose(np.sum(gt & np.logical_not(void_pixels)), 0) and np.isclose(np.sum(pred & np.logical_not(void_pixels)), 0):
         return 1
 
@@ -33,9 +33,9 @@ def precision_recall(gt, pred, void_pixels=None):
     if void_pixels is None:
         void_pixels = np.zeros_like(gt)
 
-    gt = gt.astype(np.bool)
-    pred = pred.astype(np.bool)
-    void_pixels = void_pixels.astype(np.bool)
+    gt = gt.astype(bool)
+    pred = pred.astype(bool)
+    void_pixels = void_pixels.astype(bool)
 
     tp = ((pred & gt) & ~void_pixels).sum()
     fn = ((~pred & gt) & ~void_pixels).sum()
diff --git a/main.py b/main.py
index 55e18a3..bc632bc 100644
--- a/main.py
+++ b/main.py
@@ -445,7 +445,7 @@ def train_one_epoch(config, model, criterion, data_loader, optimizer, epoch, mix
         if mixup_fn is not None:
             samples, targets = mixup_fn(samples, targets)
         teacher_outputs = None
-        with torch.cuda.amp.autocast(enabled=config.AMP_ENABLE):
+        with torch.amp.autocast('cuda', enabled=config.AMP_ENABLE):
             if config.TRAIN.CONTROLLERS_PRETRAIN or config.TRAIN.MTL_MULTI_OBJECTIVE_TRAIN:
                 outputs, NBs, NTs, nTsPerTask, nBsPerTask = model(
                     samples, return_activation_stats=True, task=task)
@@ -620,7 +620,7 @@ def flatten(l):
                 task_batch, dim=0) for task, task_batch in labels_batch.items()}
 
             # Measure performance
-            with torch.cuda.amp.autocast(enabled=config.AMP_ENABLE):
+            with torch.amp.autocast('cuda', enabled=config.AMP_ENABLE):
                 loss, loss_dict = criterion(
                     output_batch_tesnor, label_batch_tesnor)
                 loss_meter.update(loss.item())
@@ -638,7 +638,7 @@ def flatten(l):
             task_batch, dim=0) for task, task_batch in labels_batch.items()}
 
         # Measure performance
-        with torch.cuda.amp.autocast(enabled=config.AMP_ENABLE):
+        with torch.amp.autocast('cuda', enabled=config.AMP_ENABLE):
             loss, loss_dict = criterion(
                 output_batch_tesnor, label_batch_tesnor)
             loss_meter.update(loss.item())
diff --git a/utils.py b/utils.py
index 458b405..9cd07de 100644
--- a/utils.py
+++ b/utils.py
@@ -8,7 +8,7 @@
 import os
 import torch
 import torch.distributed as dist
-from torch._six import inf
+from math import inf
 import errno
 
 from PIL import Image
@@ -235,7 +235,7 @@ class NativeScalerWithGradNormCount:
     state_dict_key = "amp_scaler"
 
     def __init__(self):
-        self._scaler = torch.cuda.amp.GradScaler()
+        self._scaler = torch.amp.GradScaler('cuda')
 
     def __call__(self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True):
         self._scaler.scale(loss).backward(create_graph=create_graph)