fix amp bug

tools/program.py

@@ -162,18 +162,18 @@ def to_float32(preds):
         for k in preds:
             if isinstance(preds[k], dict) or isinstance(preds[k], list):
                 preds[k] = to_float32(preds[k])
-            else:
-                preds[k] = paddle.to_tensor(preds[k], dtype='float32')
+            elif isinstance(preds[k], paddle.Tensor):
+                preds[k] = preds[k].astype(paddle.float32)
     elif isinstance(preds, list):
         for k in range(len(preds)):
             if isinstance(preds[k], dict):
                 preds[k] = to_float32(preds[k])
             elif isinstance(preds[k], list):
                 preds[k] = to_float32(preds[k])
-            else:
-                preds[k] = paddle.to_tensor(preds[k], dtype='float32')
-    else:
-        preds = paddle.to_tensor(preds, dtype='float32')
+            elif isinstance(preds[k], paddle.Tensor):
+                preds[k] = preds[k].astype(paddle.float32)
+    elif isinstance(preds, paddle.Tensor):
+            preds = preds.astype(paddle.float32)
     return preds
 
 
@@ -190,7 +190,8 @@ def train(config,
           pre_best_model_dict,
           logger,
           log_writer=None,
-          scaler=None):
+          scaler=None,
+          amp_level='O2'):
     cal_metric_during_train = config['Global'].get('cal_metric_during_train',
                                                    False)
     calc_epoch_interval = config['Global'].get('calc_epoch_interval', 1)
@@ -276,7 +277,7 @@ def train(config,
                 model_average = True
             # use amp
             if scaler:
-                with paddle.amp.auto_cast(level='O2'):
+                with paddle.amp.auto_cast(level=amp_level):
                     if model_type == 'table' or extra_input:
                         preds = model(images, data=batch[1:])
                     elif model_type in ["kie", 'vqa']:
@@ -514,6 +515,7 @@ def eval(model,
                                 sum_images, i), fm_lr)
                     else:
                         preds = model(images)
+                preds = to_float32(preds)
             else:
                 if model_type == 'table' or extra_input:
                     preds = model(images, data=batch[1:])

tools/train.py

@@ -147,6 +147,7 @@ def main(config, device, logger, vdl_writer):
             len(valid_dataloader)))
 
     use_amp = config["Global"].get("use_amp", False)
+    amp_level = config["Global"].get("amp_level", 'O2')
     if use_amp:
         AMP_RELATED_FLAGS_SETTING = {
             'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
@@ -159,8 +160,9 @@ def main(config, device, logger, vdl_writer):
         scaler = paddle.amp.GradScaler(
             init_loss_scaling=scale_loss,
             use_dynamic_loss_scaling=use_dynamic_loss_scaling)
-        model, optimizer = paddle.amp.decorate(
-            models=model, optimizers=optimizer, level='O2', master_weight=True)
+        if amp_level == "O2":
+            model, optimizer = paddle.amp.decorate(
+                models=model, optimizers=optimizer, level=amp_level, master_weight=True)
     else:
         scaler = None
 
@@ -169,7 +171,7 @@ def main(config, device, logger, vdl_writer):
     # start train
     program.train(config, train_dataloader, valid_dataloader, device, model,
                   loss_class, optimizer, lr_scheduler, post_process_class,
-                  eval_class, pre_best_model_dict, logger, vdl_writer, scaler)
+                  eval_class, pre_best_model_dict, logger, vdl_writer, scaler,amp_level)
 
 
 def test_reader(config, device, logger):