Use mps device in eval.py

eval.py

@@ -33,7 +33,7 @@ if __name__ == '__main__':
     parser.add_argument('--CUDA_VISIBLE_DEVICES', default='-1', type=str, help="GPU设置")
     parser.add_argument('--dataset_path', default='dataset/HW-hand-write/HW_Chinese/*/*.[j|p]*', type=str,
                         help="img path")
-    parser.add_argument('--random_state', default=None, type=int, help="用于训练集划分的随机数")
+    parser.add_argument('--random_state', default=10086, type=int, help="用于训练集划分的随机数")
 
     args = parser.parse_args()
     os.environ["CUDA_VISIBLE_DEVICES"] = args.CUDA_VISIBLE_DEVICES
@@ -51,9 +51,10 @@ if __name__ == '__main__':
     vocab = processor.tokenizer.get_vocab()
 
     vocab_inp = {vocab[key]: key for key in vocab}
+    mps_device = torch.device("mps")
     model = VisionEncoderDecoderModel.from_pretrained(args.cust_data_init_weights_path)
     model.eval()
-    model.cuda()
+    model.to(mps_device)
 
     vocab = processor.tokenizer.get_vocab()
     vocab_inp = {vocab[key]: key for key in vocab}
@@ -67,7 +68,7 @@ if __name__ == '__main__':
         pixel_values = processor([img], return_tensors="pt").pixel_values
 
         with torch.no_grad():
-            generated_ids = model.generate(pixel_values[:, :, :].cuda())
+            generated_ids = model.generate(pixel_values[:, :, :].to(mps_device))
 
         generated_text = decode_text(generated_ids[0].cpu().numpy(), vocab, vocab_inp)
         pred_str.append(generated_text)