Create configuration_nllb_clip.py

configuration_nllb_clip.py

@@ -0,0 +1,273 @@
+""" NLLB-CLIP model configuration"""
+
+import os
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Any, Mapping, Optional, Union
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import ProcessorMixin
+    from transformers.utils import TensorType
+
+from transformers import CLIPVisionConfig
+from transformers.configuration_utils import PretrainedConfig
+from transformers.onnx import OnnxConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class NLLBCLIPTextConfig(PretrainedConfig):
+    model_type = "clip_text_model"
+    attribute_map = {
+        "num_attention_heads": "encoder_attention_heads",
+        "hidden_size": "d_model",
+    }
+
+    def __init__(
+        self,
+        vocab_size=128112,
+        max_position_embeddings=1024,
+        encoder_layers=12,
+        encoder_ffn_dim=4096,
+        encoder_attention_heads=16,
+        encoder_layerdrop=0.05,
+        use_cache=True,
+        activation_function="relu",
+        d_model=1024,
+        dropout=0.1,
+        attention_dropout=0.1,
+        activation_dropout=0.0,
+        init_std=0.02,
+        scale_embedding=True,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        layer_norm_eps=1e-5,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.scale_embedding = scale_embedding
+        self.layer_norm_eps = layer_norm_eps
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs,
+        )
+
+    @classmethod
+    def from_pretrained(
+        cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
+    ) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs
+        )
+
+        # get the vision config dict if we are loading from CLIPConfig
+        if config_dict.get("model_type") == "clip":
+            config_dict = config_dict["text_config"]
+
+        if (
+            "model_type" in config_dict
+            and hasattr(cls, "model_type")
+            and config_dict["model_type"] != cls.model_type
+        ):
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class NLLBCLIPConfig(PretrainedConfig):
+    model_type = "clip"
+
+    def __init__(
+        self,
+        text_config=None,
+        vision_config=None,
+        projection_dim=512,
+        logit_scale_init_value=2.6592,
+        **kwargs,
+    ):
+        # If `_config_dict` exist, we use them for the backward compatibility.
+        # We pop out these 2 attributes before calling `super().__init__` to avoid them being saved (which causes a lot
+        # of confusion!).
+        text_config_dict = kwargs.pop("text_config_dict", None)
+        vision_config_dict = kwargs.pop("vision_config_dict", None)
+
+        super().__init__(**kwargs)
+
+        # Instead of simply assigning `[text|vision]_config_dict` to `[text|vision]_config`, we use the values in
+        # `[text|vision]_config_dict` to update the values in `[text|vision]_config`. The values should be same in most
+        # cases, but we don't want to break anything regarding `_config_dict` that existed before commit `8827e1b2`.
+        if text_config_dict is not None:
+            if text_config is None:
+                text_config = {}
+
+            # This is the complete result when using `text_config_dict`.
+            _text_config_dict = NLLBCLIPTextConfig(**text_config_dict).to_dict()
+
+            # Give a warning if the values exist in both `_text_config_dict` and `text_config` but being different.
+            for key, value in _text_config_dict.items():
+                if (
+                    key in text_config
+                    and value != text_config[key]
+                    and key not in ["transformers_version"]
+                ):
+                    # If specified in `text_config_dict`
+                    if key in text_config_dict:
+                        message = (
+                            f"`{key}` is found in both `text_config_dict` and `text_config` but with different values. "
+                            f'The value `text_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`text_config_dict` is provided which will be used to initialize `CLIPTextConfig`. The "
+                            f'value `text_config["{key}"]` will be overriden.'
+                        )
+                    logger.warning(message)
+
+            # Update all values in `text_config` with the ones in `_text_config_dict`.
+            text_config.update(_text_config_dict)
+
+        if vision_config_dict is not None:
+            if vision_config is None:
+                vision_config = {}
+
+            # This is the complete result when using `vision_config_dict`.
+            _vision_config_dict = CLIPVisionConfig(**vision_config_dict).to_dict()
+            # convert keys to string instead of integer
+            if "id2label" in _vision_config_dict:
+                _vision_config_dict["id2label"] = {
+                    str(key): value
+                    for key, value in _vision_config_dict["id2label"].items()
+                }
+
+            # Give a warning if the values exist in both `_vision_config_dict` and `vision_config` but being different.
+            for key, value in _vision_config_dict.items():
+                if (
+                    key in vision_config
+                    and value != vision_config[key]
+                    and key not in ["transformers_version"]
+                ):
+                    # If specified in `vision_config_dict`
+                    if key in vision_config_dict:
+                        message = (
+                            f"`{key}` is found in both `vision_config_dict` and `vision_config` but with different "
+                            f'values. The value `vision_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`vision_config_dict` is provided which will be used to initialize `CLIPVisionConfig`. "
+                            f'The value `vision_config["{key}"]` will be overriden.'
+                        )
+                    logger.warning(message)
+
+            # Update all values in `vision_config` with the ones in `_vision_config_dict`.
+            vision_config.update(_vision_config_dict)
+
+        if text_config is None:
+            text_config = {}
+            logger.info(
+                "`text_config` is `None`. Initializing the `NLLBCLIPTextConfig` with default values."
+            )
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info(
+                "`vision_config` is `None`. initializing the `CLIPVisionConfig` with default values."
+            )
+
+        self.text_config = NLLBCLIPTextConfig(**text_config)
+        self.vision_config = CLIPVisionConfig(**vision_config)
+
+        self.projection_dim = projection_dim
+        self.logit_scale_init_value = logit_scale_init_value
+        self.initializer_factor = 1.0
+
+    @classmethod
+    def from_text_vision_configs(
+        cls, text_config: NLLBCLIPTextConfig, vision_config: CLIPVisionConfig, **kwargs
+    ):
+        r"""
+        Instantiate a [`CLIPConfig`] (or a derived class) from clip text model configuration and clip vision model
+        configuration.
+        Returns:
+            [`CLIPConfig`]: An instance of a configuration object
+        """
+
+        return cls(
+            text_config=text_config.to_dict(),
+            vision_config=vision_config.to_dict(),
+            **kwargs,
+        )
+
+
+class CLIPOnnxConfig(OnnxConfig):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        return OrderedDict(
+            [
+                ("input_ids", {0: "batch", 1: "sequence"}),
+                ("attention_mask", {0: "batch", 1: "sequence"}),
+                (
+                    "pixel_values",
+                    {0: "batch", 1: "num_channels", 2: "height", 3: "width"},
+                ),
+            ]
+        )
+
+    @property
+    def outputs(self) -> Mapping[str, Mapping[int, str]]:
+        return OrderedDict(
+            [
+                ("logits_per_image", {0: "batch"}),
+                ("logits_per_text", {0: "batch"}),
+                ("text_embeds", {0: "batch"}),
+                ("image_embeds", {0: "batch"}),
+            ]
+        )
+
+    @property
+    def atol_for_validation(self) -> float:
+        return 1e-4
+
+    def generate_dummy_inputs(
+        self,
+        processor: "ProcessorMixin",
+        batch_size: int = -1,
+        seq_length: int = -1,
+        framework: Optional["TensorType"] = None,
+    ) -> Mapping[str, Any]:
+        text_input_dict = super().generate_dummy_inputs(
+            processor.tokenizer,
+            batch_size=batch_size,
+            seq_length=seq_length,
+            framework=framework,
+        )
+        image_input_dict = super().generate_dummy_inputs(
+            processor.image_processor, batch_size=batch_size, framework=framework
+        )
+        return {**text_input_dict, **image_input_dict}
+
+    @property
+    def default_onnx_opset(self) -> int:
+        return 14