[PretrainedConfig] Bring back deprecation warning and fix redundant warnings (#4264)

sijunhe · wj-Mcat · web-flow · commit 2c59e39c105a · 2022-12-29T09:41:07.000+08:00
* commits

* wip

* remove all hasattr

* codegen

* turn on test

* add standard config map

* standard config map

* Update paddlenlp/transformers/bert/configuration.py

Co-authored-by: 骑马小猫 &lt;1435130236@qq.com&gt;

Co-authored-by: 骑马小猫 &lt;1435130236@qq.com&gt;
diff --git a/paddlenlp/transformers/bart/modeling.py b/paddlenlp/transformers/bart/modeling.py
@@ -23,8 +23,7 @@
 from paddle import Tensor
 from paddle.nn import Embedding, Layer, MultiHeadAttention
 
-from paddlenlp.utils.env import CONFIG_NAME
-
+from ...utils.env import CONFIG_NAME
 from ...utils.log import logger
 from .. import PretrainedModel, register_base_model
 from ..model_outputs import (
@@ -92,7 +91,7 @@ def init_weights(self, layer):
                 layer.weight.set_value(
                     paddle.tensor.normal(
                         mean=0.0,
-                        std=self.init_std if hasattr(self, "init_std") else self.bart.config["init_std"],
+                        std=self.config.init_std,
                         shape=layer.weight.shape,
                     )
                 )
@@ -1127,7 +1126,4 @@ def __getattr__(self, name):
         try:
             return super().__getattr__(name)
         except AttributeError:
-            try:
-                return getattr(getattr(self, self.base_model_prefix), name)
-            except AttributeError:
-                return getattr(getattr(self, self.base_model_prefix).config, name)
+            return getattr(getattr(self, self.base_model_prefix), name)
diff --git a/paddlenlp/transformers/bert/configuration.py b/paddlenlp/transformers/bert/configuration.py
@@ -15,7 +15,8 @@
 from __future__ import annotations
 
 from typing import Dict
-from paddlenlp.transformers.configuration_utils import PretrainedConfig, attribute_map
+
+from paddlenlp.transformers.configuration_utils import PretrainedConfig
 
 __all__ = ["BERT_PRETRAINED_INIT_CONFIGURATION", "BertConfig", "BERT_PRETRAINED_RESOURCE_FILES_MAP"]
 
@@ -364,7 +365,7 @@ class BertConfig(PretrainedConfig):
     >>> configuration = model.config
     ```"""
     model_type = "bert"
-    attribute_map: Dict[str, str] = {"num_classes": "num_labels", "dropout": "classifier_dropout"}
+    standard_config_map: Dict[str, str] = {"dropout": "classifier_dropout"}
     pretrained_init_configuration = BERT_PRETRAINED_INIT_CONFIGURATION
 
     def __init__(
diff --git a/paddlenlp/transformers/bert/modeling.py b/paddlenlp/transformers/bert/modeling.py
@@ -30,6 +30,7 @@
 
 from paddlenlp.transformers.model_utils import PretrainedModel, register_base_model
 
+from ...utils.env import CONFIG_NAME
 from ..model_outputs import (
     BaseModelOutputWithPoolingAndCrossAttentions,
     MaskedLMOutput,
@@ -138,7 +139,7 @@ class BertPretrainedModel(PretrainedModel):
     See :class:`~paddlenlp.transformers.model_utils.PretrainedModel` for more details.
     """
 
-    model_config_file = "config.json"
+    model_config_file = CONFIG_NAME
     config_class = BertConfig
     resource_files_names = {"model_state": "model_state.pdparams"}
     base_model_prefix = "bert"
@@ -155,9 +156,7 @@ def init_weights(self, layer):
                 layer.weight.set_value(
                     paddle.tensor.normal(
                         mean=0.0,
-                        std=self.initializer_range
-                        if hasattr(self, "initializer_range")
-                        else self.config.initializer_range,
+                        std=self.config.initializer_range,
                         shape=layer.weight.shape,
                     )
                 )
diff --git a/paddlenlp/transformers/codegen/modeling.py b/paddlenlp/transformers/codegen/modeling.py
@@ -21,8 +21,7 @@
 from paddle import Tensor
 from paddle.nn import Layer
 
-from paddlenlp.utils.env import CONFIG_NAME
-
+from ...utils.env import CONFIG_NAME
 from ...utils.log import logger
 from .. import PretrainedModel, register_base_model
 from ..model_outputs import (
@@ -295,16 +294,14 @@ def init_weights(self, layer):
                 layer.weight.set_value(
                     paddle.tensor.normal(
                         mean=0.0,
-                        std=self.initializer_range
-                        if hasattr(self, "initializer_range")
-                        else self.transformer.config.initializer_range,
+                        std=self.config.initializer_range,
                         shape=layer.weight.shape,
                     )
                 )
         elif isinstance(layer, nn.LayerNorm):
             layer.bias.set_value(paddle.zeros_like(layer.bias))
             layer.weight.set_value(paddle.full_like(layer.weight, 1.0))
-            layer._epsilon = getattr(self, "layer_norm_epsilon", 1e-05)
+            layer._epsilon = self.config.layer_norm_epsilon
         if isinstance(layer, nn.Linear) and layer.bias is not None:
             layer.bias.set_value(paddle.zeros_like(layer.bias))
 
@@ -696,7 +693,4 @@ def __getattr__(self, name):
         try:
             return super().__getattr__(name)
         except AttributeError:
-            try:
-                return getattr(getattr(self, self.base_model_prefix), name)
-            except AttributeError:
-                return getattr(getattr(self, self.base_model_prefix).config, name)
+            return getattr(getattr(self, self.base_model_prefix), name)
diff --git a/paddlenlp/transformers/configuration_utils.py b/paddlenlp/transformers/configuration_utils.py
@@ -441,21 +441,25 @@ class PretrainedConfig:
     # global attribute mapping
     attribute_map: Dict[str, str] = {"num_classes": "num_labels"}
 
-    # map hf attribute to paddle attribute
-    # { "standard_field": "paddle_field", ... }
+    # model-specific attribute map from hf attribute to paddle attribute
+    # { "paddle_field": "standard_field", ... }
     standard_config_map: Dict[str, str] = {}
 
     _auto_class: Optional[str] = None
 
     def __setattr__(self, key, value):
         if key in super().__getattribute__("attribute_map"):
             key = super().__getattribute__("attribute_map")[key]
+        elif key in super().__getattribute__("standard_config_map"):
+            key = super().__getattribute__("standard_config_map")[key]
         super().__setattr__(key, value)
         assert hasattr(self, key)
 
     def __getattribute__(self, key):
         if key != "attribute_map" and key in super().__getattribute__("attribute_map"):
             key = super().__getattribute__("attribute_map")[key]
+        elif key != "standard_config_map" and key in super().__getattribute__("standard_config_map"):
+            key = super().__getattribute__("standard_config_map")[key]
         return super().__getattribute__(key)
 
     def __getitem__(self, key):
diff --git a/paddlenlp/transformers/ernie/configuration.py b/paddlenlp/transformers/ernie/configuration.py
@@ -1065,7 +1065,7 @@ class ErnieConfig(PretrainedConfig):
     >>> configuration = model.config
     ```"""
     model_type = "ernie"
-    attribute_map: Dict[str, str] = {"num_classes": "num_labels", "dropout": "classifier_dropout"}
+    standard_config_map: Dict[str, str] = {"dropout": "classifier_dropout"}
     pretrained_init_configuration = ERNIE_PRETRAINED_INIT_CONFIGURATION
 
     def __init__(
diff --git a/paddlenlp/transformers/ernie/modeling.py b/paddlenlp/transformers/ernie/modeling.py
@@ -20,6 +20,7 @@
 import paddle.nn.functional as F
 from paddle import Tensor
 
+from ...utils.env import CONFIG_NAME
 from .. import PretrainedModel, register_base_model
 from ..model_outputs import (
     BaseModelOutputWithPoolingAndCrossAttentions,
@@ -148,7 +149,7 @@ class ErniePretrainedModel(PretrainedModel):
 
     """
 
-    model_config_file = "model_config.json"
+    model_config_file = CONFIG_NAME
     config_class = ErnieConfig
     resource_files_names = {"model_state": "model_state.pdparams"}
     base_model_prefix = "ernie"
@@ -165,9 +166,7 @@ def init_weights(self, layer):
                 layer.weight.set_value(
                     paddle.tensor.normal(
                         mean=0.0,
-                        std=self.initializer_range
-                        if hasattr(self, "initializer_range")
-                        else self.ernie.config["initializer_range"],
+                        std=self.config.initializer_range,
                         shape=layer.weight.shape,
                     )
                 )
diff --git a/paddlenlp/transformers/ernie_layout/configuration.py b/paddlenlp/transformers/ernie_layout/configuration.py
@@ -150,7 +150,7 @@ class ErnieLayoutConfig(PretrainedConfig):
     >>> configuration = model.config
     ```"""
     model_type = "ernie_layout"
-    attribute_map: Dict[str, str] = {"num_classes": "num_labels", "dropout": "classifier_dropout"}
+    standard_config_map: Dict[str, str] = {"dropout": "classifier_dropout"}
     pretrained_init_configuration = ERNIE_LAYOUT_PRETRAINED_INIT_CONFIGURATION
 
     def __init__(
diff --git a/paddlenlp/transformers/ernie_layout/modeling.py b/paddlenlp/transformers/ernie_layout/modeling.py
@@ -23,6 +23,7 @@
 
 from paddlenlp.utils.log import logger
 
+from ...utils.env import CONFIG_NAME
 from .. import PretrainedModel, register_base_model
 from .configuration import (
     ERNIE_LAYOUT_PRETRAINED_INIT_CONFIGURATION,
@@ -171,7 +172,7 @@ def forward(self, input_ids, bbox=None, token_type_ids=None, position_ids=None):
 
 
 class ErnieLayoutPretrainedModel(PretrainedModel):
-    model_config_file = "config.json"
+    model_config_file = CONFIG_NAME
     pretrained_init_configuration = ERNIE_LAYOUT_PRETRAINED_INIT_CONFIGURATION
     pretrained_resource_files_map = ERNIE_LAYOUT_PRETRAINED_RESOURCE_FILES_MAP
     base_model_prefix = "ernie_layout"
@@ -184,9 +185,7 @@ def init_weights(self, layer):
                 layer.weight.set_value(
                     paddle.tensor.normal(
                         mean=0.0,
-                        std=self.pretrained_init_configuration["initializer_range"]
-                        if "initializer_range" in self.pretrained_init_configuration
-                        else 0.02,
+                        std=self.config.initializer_range,
                         shape=layer.weight.shape,
                     )
                 )
diff --git a/paddlenlp/transformers/ernie_m/configuration.py b/paddlenlp/transformers/ernie_m/configuration.py
@@ -144,7 +144,7 @@ class ErnieMConfig(PretrainedConfig):
         >>> configuration = model.config
         ```"""
     model_type = "ernie_m"
-    attribute_map: Dict[str, str] = {"num_classes": "num_labels", "dropout": "classifier_dropout"}
+    standard_config_map: Dict[str, str] = {"dropout": "classifier_dropout"}
     pretrained_init_configuration = ERNIE_M_PRETRAINED_INIT_CONFIGURATION
 
     def __init__(
diff --git a/paddlenlp/transformers/ernie_m/modeling.py b/paddlenlp/transformers/ernie_m/modeling.py
@@ -133,9 +133,7 @@ def init_weights(self, layer):
                 layer.weight.set_value(
                     paddle.tensor.normal(
                         mean=0.0,
-                        std=self.initializer_range
-                        if hasattr(self, "initializer_range")
-                        else self.ernie_m.config["initializer_range"],
+                        std=self.config.initializer_range,
                         shape=layer.weight.shape,
                     )
                 )
diff --git a/paddlenlp/transformers/mbart/modeling.py b/paddlenlp/transformers/mbart/modeling.py
@@ -22,8 +22,7 @@
 from paddle import Tensor
 from paddle.nn import Embedding, Layer, MultiHeadAttention
 
-from paddlenlp.utils.env import CONFIG_NAME
-
+from ...utils.env import CONFIG_NAME
 from ...utils.log import logger
 from .. import PretrainedModel, register_base_model
 from ..model_outputs import (
@@ -94,7 +93,7 @@ def init_weights(self, layer):
                 layer.weight.set_value(
                     paddle.tensor.normal(
                         mean=0.0,
-                        std=self.init_std if hasattr(self, "init_std") else self.mbart.config.init_std,
+                        std=self.config.init_std,
                         shape=layer.weight.shape,
                     )
                 )
@@ -1132,7 +1131,4 @@ def __getattr__(self, name):
         try:
             return super().__getattr__(name)
         except AttributeError:
-            try:
-                return getattr(getattr(self, self.base_model_prefix), name)
-            except AttributeError:
-                return getattr(getattr(self, self.base_model_prefix).config, name)
+            return getattr(getattr(self, self.base_model_prefix), name)
diff --git a/paddlenlp/transformers/model_utils.py b/paddlenlp/transformers/model_utils.py
@@ -294,11 +294,10 @@ def __getattr__(self, name):
         except AttributeError:
             result = getattr(self.config, name)
 
-            # FIXME(wj-Mcat): there are a lot of consistent errors, so temporary disable.
-            # logger.warning(
-            #     f"Do not access config from `model.{name}` which will be deprecated after v2.6.0, "
-            #     f"Instead, do `model.config.{name}`"
-            # )
+            logger.warning(
+                f"Accessing `{name}` through `model.{name}` will be deprecated after v2.6.0. "
+                f"Instead, do `model.config.{name}`"
+            )
             return result
 
     @property
diff --git a/paddlenlp/transformers/t5/configuration.py b/paddlenlp/transformers/t5/configuration.py
@@ -217,7 +217,7 @@ class T5Config(PretrainedConfig):
 
     """
     model_type = "t5"
-    attribute_map: Dict[str, str] = {
+    standard_config_map: Dict[str, str] = {
         "hidden_size": "d_model",
         "num_attention_heads": "num_heads",
         "num_hidden_layers": "num_layers",
diff --git a/paddlenlp/transformers/t5/modeling.py b/paddlenlp/transformers/t5/modeling.py
@@ -593,13 +593,12 @@ def init_weights(self):
 
     def _init_weights(self, layer):
         """Initialize the weights"""
-        factor = (
-            self.initializer_factor if hasattr(self, "initializer_factor") else self.t5.config["initializer_factor"]
-        )  # Used for testing weights initialization
-        d_model = self.d_model if hasattr(self, "d_model") else self.t5.config["d_model"]
-        d_ff = self.d_ff if hasattr(self, "d_ff") else self.t5.config["d_ff"]
-        n_heads = self.num_heads if hasattr(self, "num_heads") else self.t5.config["num_heads"]
-        key_value_proj_dim = self.d_kv if hasattr(self, "d_kv") else self.t5.config["d_kv"]
+        # Used for testing weights initialization
+        factor = self.config.initializer_factor
+        d_model = self.config.d_model
+        d_ff = self.config.d_ff
+        n_heads = self.config.num_heads
+        key_value_proj_dim = self.config.d_kv
 
         if isinstance(layer, T5LayerNorm):
             layer.weight.set_value(paddle.ones_like(layer.weight) * factor)
@@ -681,8 +680,8 @@ def _init_weights(self, layer):
                 )
 
     def _shift_right(self, input_ids):
-        bos_token_id = self.bos_token_id if hasattr(self, "bos_token_id") else self.t5.config["bos_token_id"]
-        pad_token_id = self.pad_token_id if hasattr(self, "pad_token_id") else self.t5.config["pad_token_id"]
+        bos_token_id = self.config.bos_token_id
+        pad_token_id = self.config.pad_token_id
 
         assert (
             bos_token_id is not None
diff --git a/paddlenlp/transformers/unified_transformer/modeling.py b/paddlenlp/transformers/unified_transformer/modeling.py
@@ -20,8 +20,7 @@
 import paddle.nn.functional as F
 from paddle import Tensor
 
-from paddlenlp.utils.env import CONFIG_NAME
-
+from ...utils.env import CONFIG_NAME
 from ...utils.log import logger
 from .. import PretrainedModel, register_base_model
 from ..model_outputs import CausalLMOutputWithCrossAttentions
@@ -66,9 +65,7 @@ def init_weights(self, layer):
                     # big models.
                     paddle.tensor.normal(
                         mean=0.0,
-                        std=self.initializer_range
-                        if hasattr(self, "initializer_range")
-                        else self.unified_transformer.config.initializer_range,
+                        std=self.config.initializer_range,
                         shape=layer.weight.shape,
                     )
                 )
diff --git a/paddlenlp/transformers/unimo/modeling.py b/paddlenlp/transformers/unimo/modeling.py
@@ -20,8 +20,7 @@
 import paddle.nn.functional as F
 from paddle import Tensor
 
-from paddlenlp.utils.env import CONFIG_NAME
-
+from ...utils.env import CONFIG_NAME
 from ...utils.log import logger
 from .. import PretrainedModel, register_base_model
 from ..model_outputs import CausalLMOutputWithCrossAttentions
@@ -64,9 +63,7 @@ def init_weights(self, layer):
                 layer.weight.set_value(
                     paddle.tensor.normal(
                         mean=0.0,
-                        std=self.initializer_range
-                        if hasattr(self, "initializer_range")
-                        else self.unimo.config.initializer_range,
+                        std=self.config.initializer_range,
                         shape=layer.weight.shape,
                     )
                 )
@@ -542,10 +539,7 @@ def __getattr__(self, name):
         try:
             return super().__getattr__(name)
         except AttributeError:
-            try:
-                return getattr(getattr(self, self.base_model_prefix), name)
-            except AttributeError:
-                return getattr(getattr(self, self.base_model_prefix).config, name)
+            return getattr(getattr(self, self.base_model_prefix), name)
 
 
 UNIMOForMaskedLM = UNIMOLMHeadModel
diff --git a/tests/transformers/test_configuration_utils.py b/tests/transformers/test_configuration_utils.py
@@ -150,7 +150,7 @@ class FakeBertConfig(BertConfig):
         config = FakeBertConfig.from_pretrained("__internal_testing__/bert")
         hidden_size = config.hidden_size
 
-        FakeBertConfig.standard_config_map = {"hidden_size": "fake_field"}
+        FakeBertConfig.standard_config_map = {"fake_field": "hidden_size"}
 
         loaded_config = FakeBertConfig.from_pretrained("__internal_testing__/bert")
         fake_field = loaded_config.fake_field
@@ -181,7 +181,7 @@ class FakeBertConfig(BertConfig):
             # rename `config.json` -> `model_config.json`
             shutil.move(os.path.join(tempdir, CONFIG_NAME), os.path.join(tempdir, LEGACY_CONFIG_NAME))
 
-            FakeBertConfig.standard_config_map = {"hidden_size": "fake_field"}
+            FakeBertConfig.standard_config_map = {"fake_field": "hidden_size"}
 
             loaded_config = FakeBertConfig.from_pretrained(tempdir)
             self.assertEqual(loaded_config.fake_field, config.hidden_size)
diff --git a/tests/transformers/test_modeling_common.py b/tests/transformers/test_modeling_common.py
@@ -67,7 +67,7 @@ class ModelTesterMixin:
     test_resize_position_embeddings = False
     test_mismatched_shapes = True
     test_missing_keys = True
-    test_model_compatibility_keys = False
+    test_model_compatibility_keys = True
     use_test_inputs_embeds = False
     use_test_model_name_list = True
     is_encoder_decoder = False
