Rectify code of the LayoutLM series models and adjust it to amp_level mode.

Author: Bourn3z

configs/kie/layoutlmv3/README.md

@@ -144,7 +144,7 @@ Apart from the dataset setting, please also check the following important args:
 system:
   mode:
   distribute: False                                                     # `True` for distributed training, `False` for standalone training
-  amp_level: 'O0'
+  amp_level: 'O3'
   seed: 42
   val_while_train: True                                                 # Validate while training
   drop_overflow_update: False
@@ -157,7 +157,6 @@ model:
     name: TokenClassificationHead
     num_classes: 7
     use_visual_backbone: True
-    use_float16: True
   pretrained:
 ...
 train:

configs/kie/layoutlmv3/ser_layoutlmv3_xfund_zh.yaml

@@ -1,7 +1,7 @@
 system:
   mode: 0 # 0 for graph mode, 1 for pynative mode in MindSpore
   distribute: False
-  amp_level: "O0"
+  amp_level: "O3"
   seed: 42
   log_interval: 10
   val_start_epoch: 50
@@ -17,7 +17,6 @@ model:
     name: TokenClassificationHead
     num_classes: 7
     use_visual_backbone: True
-    use_float16: True
   pretrained:
 
 postprocess:

configs/kie/layoutxlm/re_layoutxlm_xfund_zh.yaml

@@ -1,7 +1,7 @@
 system:
   mode: 0 # 0 for graph mode, 1 for pynative mode in MindSpore
   distribute: False
-  amp_level: 'O0'
+  amp_level: 'O3'
   seed: 42
   log_interval: 10
   val_while_train: True
@@ -16,11 +16,9 @@ model:
     pretrained: True
     num_classes: &num_classes 7
     use_visual_backbone: True
-    use_float16: True
   head:
     name: RelationExtractionHead
     use_visual_backbone: True
-    use_float16: True
   pretrained:
 
 postprocess:

configs/kie/layoutxlm/ser_layoutxlm_xfund_zh.yaml

@@ -1,7 +1,7 @@
 system:
   mode: 0 # 0 for graph mode, 1 for pynative mode in MindSpore
   distribute: False
-  amp_level: 'O0'
+  amp_level: 'O3'
   seed: 42
   log_interval: 10
   val_while_train: True
@@ -15,12 +15,10 @@ model:
     pretrained: True
     num_classes: &num_classes 7
     use_visual_backbone: True
-    use_float16: True
   head :
     name: TokenClassificationHead
     num_classes: 7
     use_visual_backbone: True
-    use_float16: True
   pretrained:
 
 postprocess:

configs/kie/vi_layoutxlm/README.md

@@ -159,7 +159,7 @@ Apart from the dataset setting, please also check the following important args:
 system:
   mode:
   distribute: False                                                     # `True` for distributed training, `False` for standalone training
-  amp_level: 'O0'
+  amp_level: 'O3'
   seed: 42
   val_while_train: True                                                 # Validate while training
   drop_overflow_update: False
@@ -171,12 +171,10 @@ model:
     pretrained: True
     num_classes: &num_classes 7
     use_visual_backbone: False
-    use_float16: True
   head :
     name: TokenClassificationHead
     num_classes: 7
     use_visual_backbone: False
-    use_float16: True
   pretrained:
 ...
 train:

configs/kie/vi_layoutxlm/README_CN.md

@@ -156,7 +156,7 @@ eval:
 system:
   mode:
   distribute: False                                                     # 分布式训练为True，单卡训练为False
-  amp_level: 'O0'
+  amp_level: 'O3'
   seed: 42
   val_while_train: True                                                 # 边训练边验证
   drop_overflow_update: False
@@ -168,12 +168,10 @@ model:
     pretrained: True
     num_classes: &num_classes 7
     use_visual_backbone: False
-    use_float16: True
   head :
     name: TokenClassificationHead
     num_classes: 7
     use_visual_backbone: False
-    use_float16: True
   pretrained:
 ...
 train:

configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yaml

@@ -1,7 +1,7 @@
 system:
   mode: 0 # 0 for graph mode, 1 for pynative mode in MindSpore
   distribute: False
-  amp_level: "O0"
+  amp_level: "O3"
   seed: 42
   log_interval: 10
   val_while_train: True
@@ -16,11 +16,9 @@ model:
     pretrained: True
     num_classes: &num_classes 7
     use_visual_backbone: False
-    use_float16: True
   head:
     name: RelationExtractionHead
     use_visual_backbone: False
-    use_float16: True
   pretrained:
 
 postprocess:
@@ -90,11 +88,11 @@ train:
         "bbox",
         "attention_mask",
         "token_type_ids",
-        "image",
         "question",
         "question_label",
         "answer",
         "answer_label",
+        "image",
         "relation_label",
       ]
     net_input_column_index: [0, 1, 2, 3, 4, 5, 6, 7, 8] # input indices for network forward func in output_columns

configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yaml

@@ -1,7 +1,7 @@
 system:
   mode: 0 # 0 for graph mode, 1 for pynative mode in MindSpore
   distribute: False
-  amp_level: 'O0'
+  amp_level: 'O3'
   seed: 42
   log_interval: 10
   val_while_train: True
@@ -15,12 +15,10 @@ model:
     pretrained: True
     num_classes: &num_classes 7
     use_visual_backbone: False
-    use_float16: True
   head :
     name: TokenClassificationHead
     num_classes: 7
     use_visual_backbone: False
-    use_float16: True
   pretrained:
 
 postprocess:

mindocr/models/backbones/layoutlmv3/configuration.py

@@ -3,9 +3,8 @@
 
 @dataclass
 class LayoutLMv3PretrainedConfig:
-    def __init__(self, use_float16=False):
+    def __init__(self):
         pretrained_config = {
-            "use_float16": use_float16,
             "fast_qkv": False,
             "vocab_size": 250002,
             "hidden_size": 768,

mindocr/models/backbones/layoutlmv3/layoutlmv3.py

@@ -175,7 +175,7 @@ def construct(
 
         if attention_mask is not None:
             # Apply the attention mask is (precomputed for all layers in RobertaModel forward() function)
-            attention_scores = attention_scores + attention_mask.astype(self.dense_dtype)
+            attention_scores = attention_scores + attention_mask.astype(attention_scores.dtype)
 
         # Normalize the attention scores to probabilities.
         # Use the trick of the CogView paper to stablize training
@@ -227,11 +227,8 @@ def __init__(self, config):
         self.has_relative_attention_bias = config.has_relative_attention_bias
         self.has_spatial_attention_bias = config.has_spatial_attention_bias
         self.patch_size = config.patch_size
-        self.use_float16 = config.use_float16
-        self.dense_dtype = mstype.float32
-        if self.use_float16 is True:
-            self.dense_dtype = mstype.float16
-        self.min = finfo(self.dense_dtype)
+        self.float32_min = finfo(mstype.float32)
+        self.float16_min = finfo(mstype.float16)
         self.out_channels = 1
         self.use_visual_backbone = True
 
@@ -342,7 +339,13 @@ def get_extended_attention_mask(self, attention_mask: Tensor, input_shape, dtype
         # Since we are adding it to the raw scores before the softmax, this is
         # effectively the same as removing these entirely.  # fp16 compatibility
         extended_attention_mask = extended_attention_mask.astype(dtype)
-        extended_attention_mask = (1.0 - extended_attention_mask) * self.min
+
+        if dtype == mstype.float32:
+            minimum = self.float32_min
+        elif dtype == mstype.float16:
+            minimum = self.float16_min
+
+        extended_attention_mask = (1.0 - extended_attention_mask) * minimum
         return extended_attention_mask
 
     def get_head_mask(self, head_mask, num_hidden_layers: int, is_attention_chunked: bool = False):
@@ -518,7 +521,7 @@ def construct(
 
 
 @register_backbone
-def layoutlmv3(use_float16: bool = True, **kwargs):
-    pretrained_config = LayoutLMv3PretrainedConfig(use_float16)
+def layoutlmv3(**kwargs):
+    pretrained_config = LayoutLMv3PretrainedConfig()
     model = LayoutLMv3Model(pretrained_config)
     return model

mindocr/models/backbones/layoutxlm/configuration.py

@@ -3,13 +3,11 @@
 
 @dataclass
 class LayoutXLMPretrainedConfig:
-    def __init__(self, use_visual_backbone=True, use_float16=False):
+    def __init__(self, use_visual_backbone=True):
         pretrained_config = {
             "use_visual_backbone": use_visual_backbone,
-            "use_float16": use_float16,
             "attention_probs_dropout_prob": 0.1,
             "use_visual_backbone": use_visual_backbone,
-            "use_float16": use_float16,
             "bos_token_id": 0,
             "coordinate_size": 128,
             "eos_token_id": 2,

mindocr/models/backbones/layoutxlm/layoutxlm.py

@@ -99,18 +99,12 @@ def __init__(self, config):
         self.has_visual_segment_embedding = config.has_visual_segment_embedding
         self.embeddings = LayoutXLMEmbeddings(config)
         self.use_visual_backbone = config.use_visual_backbone
-        self.use_float16 = config.use_float16
-        self.dense_dtype = mstype.float32
-        if self.use_float16 is True:
-            self.dense_dtype = mstype.float16
 
         if self.use_visual_backbone is True:
             set_context(jit_syntax_level=0)
             self.visual = VisualBackbone(config)
             self.visual.freeze()
-            self.visual_proj = nn.Dense(config.image_feature_pool_shape[-1], config.hidden_size).to_float(
-                self.dense_dtype
-            )
+            self.visual_proj = nn.Dense(config.image_feature_pool_shape[-1], config.hidden_size)
         if self.has_visual_segment_embedding:
             self.visual_segment_embedding = Parameter(nn.Embedding(1, config.hidden_size).embedding_table[0])
         self.visual_LayerNorm = nn.LayerNorm((config.hidden_size,), epsilon=config.layer_norm_eps)
@@ -302,8 +296,8 @@ def construct(
 
 
 @register_backbone
-def layoutxlm(pretrained: bool = True, use_visual_backbone: bool = True, use_float16: bool = False, **kwargs):
-    pretrained_config = LayoutXLMPretrainedConfig(use_visual_backbone, use_float16)
+def layoutxlm(pretrained: bool = True, use_visual_backbone: bool = True, **kwargs):
+    pretrained_config = LayoutXLMPretrainedConfig(use_visual_backbone)
     model = LayoutXLMModel(pretrained_config)
     if pretrained:
         if use_visual_backbone is True:

mindocr/models/backbones/layoutxlm/visual_backbone.py

@@ -130,7 +130,7 @@ def construct(self, x):
                 top_block_in_feature = bottom_up_features[self.top_block.in_feature]
             else:
                 top_block_in_feature = results[self._out_features.index(self.top_block.in_feature)]
-            results.extend(self.top_block(top_block_in_feature.astype(ms.float16)))
+            results.extend(self.top_block(top_block_in_feature))
 
         assert len(self._out_features) == len(results)