Hbsun/feature iss205 (#210)

* fix #205: remove act2fn

* fix pytest

docs/index.rst

@@ -16,6 +16,7 @@ Welcome to MMOCR's documentation!
    textdet_models.md
    textrecog_models.md
    kie_models.md
+   ner_models.md
 
 .. toctree::
    :maxdepth: 2

mmocr/models/ner/encoders/bert_encoder.py

@@ -25,7 +25,7 @@ class BertEncoder(nn.Module):
         attention_probs_dropout_prob (float): The dropout probability
             of attention.
         intermediate_size (int): The size of intermediate layer.
-        hidden_act (str):  Hidden layer activation.
+        hidden_act_cfg (dict):  Hidden layer activation.
     """
 
     def __init__(self,
@@ -42,7 +42,7 @@ class BertEncoder(nn.Module):
                  num_attention_heads=12,
                  attention_probs_dropout_prob=0.1,
                  intermediate_size=3072,
-                 hidden_act='gelu_new',
+                 hidden_act_cfg=dict(type='GeluNew'),
                  pretrained=None):
         super().__init__()
         self.bert = BertModel(
@@ -59,7 +59,7 @@ class BertEncoder(nn.Module):
             num_attention_heads=num_attention_heads,
             attention_probs_dropout_prob=attention_probs_dropout_prob,
             intermediate_size=intermediate_size,
-            hidden_act=hidden_act)
+            hidden_act_cfg=hidden_act_cfg)
         self.init_weights(pretrained=pretrained)
 
     def forward(self, results):

mmocr/models/ner/utils/__init__.py

@@ -1,4 +1,4 @@
-from .activations import ACT2FN
+from .activations import GeluNew
 from .bert import BertModel
 
-__all__ = ['BertModel', 'ACT2FN']
+__all__ = ['BertModel', 'GeluNew']

mmocr/models/ner/utils/activations.py

@@ -6,34 +6,26 @@
 import math
 
 import torch
-from mmcv.cnn import Swish
+import torch.nn as nn
+from mmcv.cnn import ACTIVATION_LAYERS
 
 
-def gelu(x):
+@ACTIVATION_LAYERS.register_module()
-    """Original Implementation of the gelu activation function in Google Bert
+class GeluNew(nn.Module):
-    repo when initially created. For information: OpenAI GPT's gelu is slightly
-    different (and gives slightly different results):
-
-    0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) *
-    (x + 0.044715 * torch.pow(x, 3))))
-    Also see https://arxiv.org/abs/1606.08415
-    """
-    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
-
-
-def gelu_new(x):
     """Implementation of the gelu activation function currently in Google Bert
     repo (identical to OpenAI GPT).
 
     Also see https://arxiv.org/abs/1606.08415
     """
-    return 0.5 * x * (1 + torch.tanh(
-        math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
 
+    def forward(self, x):
+        """Forward function.
 
-ACT2FN = {
+        Args:
-    'gelu': gelu,
+            x (torch.Tensor): The input tensor.
-    'relu': torch.nn.functional.relu,
+
-    'swish': Swish,
+        Returns:
-    'gelu_new': gelu_new
+            torch.Tensor: Activated tensor.
-}
+        """
+        return 0.5 * x * (1 + torch.tanh(
+            math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))

mmocr/models/ner/utils/bert.py

@@ -7,8 +7,7 @@ import math
 
 import torch
 import torch.nn as nn
-
+from mmcv.cnn import build_activation_layer
-from mmocr.models.ner.utils.activations import ACT2FN
 
 
 class BertModel(nn.Module):
@@ -31,7 +30,7 @@ class BertModel(nn.Module):
             for the attention probabilities normalized from
             the attention scores.
         intermediate_size (int): The size of intermediate layer.
-        hidden_act (str):  hidden layer activation
+        hidden_act_cfg (str):  hidden layer activation
     """
 
     def __init__(self,
@@ -48,7 +47,7 @@ class BertModel(nn.Module):
                  num_attention_heads=12,
                  attention_probs_dropout_prob=0.1,
                  intermediate_size=3072,
-                 hidden_act='gelu_new'):
+                 hidden_act_cfg=dict(type='GeluNew')):
         super().__init__()
         self.embeddings = BertEmbeddings(
             vocab_size=vocab_size,
@@ -67,7 +66,7 @@ class BertModel(nn.Module):
             layer_norm_eps=layer_norm_eps,
             hidden_dropout_prob=hidden_dropout_prob,
             intermediate_size=intermediate_size,
-            hidden_act=hidden_act)
+            hidden_act_cfg=hidden_act_cfg)
         self.pooler = BertPooler(hidden_size=hidden_size)
         self.num_hidden_layers = num_hidden_layers
         self.initializer_range = initializer_range
@@ -205,7 +204,7 @@ class BertEncoder(nn.Module):
                  layer_norm_eps=1e-12,
                  hidden_dropout_prob=0.1,
                  intermediate_size=3072,
-                 hidden_act='gelu_new'):
+                 hidden_act_cfg=dict(type='GeluNew')):
         super().__init__()
         self.output_attentions = output_attentions
         self.output_hidden_states = output_hidden_states
@@ -218,7 +217,8 @@ class BertEncoder(nn.Module):
                 layer_norm_eps=layer_norm_eps,
                 hidden_dropout_prob=hidden_dropout_prob,
                 intermediate_size=intermediate_size,
-                hidden_act=hidden_act) for _ in range(num_hidden_layers)
+                hidden_act_cfg=hidden_act_cfg)
+            for _ in range(num_hidden_layers)
         ])
 
     def forward(self, hidden_states, attention_mask=None, head_mask=None):
@@ -278,7 +278,7 @@ class BertLayer(nn.Module):
                  layer_norm_eps=1e-12,
                  hidden_dropout_prob=0.1,
                  intermediate_size=3072,
-                 hidden_act='gelu_new'):
+                 hidden_act_cfg=dict(type='GeluNew')):
         super().__init__()
         self.attention = BertAttention(
             hidden_size=hidden_size,
@@ -290,7 +290,7 @@ class BertLayer(nn.Module):
         self.intermediate = BertIntermediate(
             hidden_size=hidden_size,
             intermediate_size=intermediate_size,
-            hidden_act=hidden_act)
+            hidden_act_cfg=hidden_act_cfg)
         self.output = BertOutput(
             intermediate_size=intermediate_size,
             hidden_size=hidden_size,
@@ -448,13 +448,11 @@ class BertIntermediate(nn.Module):
     def __init__(self,
                  hidden_size=768,
                  intermediate_size=3072,
-                 hidden_act='gelu_new'):
+                 hidden_act_cfg=dict(type='GeluNew')):
         super().__init__()
+
         self.dense = nn.Linear(hidden_size, intermediate_size)
-        if isinstance(hidden_act, str):
+        self.intermediate_act_fn = build_activation_layer(hidden_act_cfg)
-            self.intermediate_act_fn = ACT2FN[hidden_act]
-        else:
-            self.intermediate_act_fn = hidden_act
 
     def forward(self, hidden_states):
         hidden_states = self.dense(hidden_states)

tests/test_models/test_ner_model.py

@@ -6,7 +6,6 @@ import pytest
 import torch
 
 from mmocr.models import build_detector
-from mmocr.models.ner.utils.activations import gelu, gelu_new
 
 
 def _create_dummy_vocab_file(vocab_file):
@@ -35,7 +34,7 @@ def _get_detector_cfg(fname):
 
 @pytest.mark.parametrize(
     'cfg_file', ['configs/ner/bert_softmax/bert_softmax_cluener_18e.py'])
-def test_encoder_decoder_pipeline(cfg_file):
+def test_bert_softmax(cfg_file):
     # prepare data
     texts = ['中'] * 47
     img = [31] * 47
@@ -77,7 +76,3 @@ def test_encoder_decoder_pipeline(cfg_file):
         batch_results = []
         result = detector.forward(None, img_metas, return_loss=False)
         batch_results.append(result)
-
-    # Test activations
-    gelu(torch.tensor(0.5))
-    gelu_new(torch.tensor(0.5))