configuration_efficientnet.py

from transformers.configuration_utils import PretrainedConfig

from optimum.utils.normalized_config import NormalizedVisionConfig
from optimum.utils.input_generators import DummyVisionInputGenerator
from optimum.exporters.onnx.model_configs import ViTOnnxConfig

from typing import OrderedDict, Dict

MODEL_NAMES = [
        'efficientnet_b0',
        'efficientnet_b1',
        'efficientnet_b2',
        'efficientnet_b3',
        'efficientnet_b4',
        'efficientnet_b5',
        'efficientnet_b6',
        'efficientnet_b7',
        'efficientnet_b8',
        'efficientnet_l2'
    ]

class EfficientNetConfig(PretrainedConfig):
    model_type = 'efficientnet'

    def __init__(
            self,
            model_name: str = 'efficientnet_b0',
            pretrained: bool = False,
            **kwargs
            ):
        
        if model_name not in MODEL_NAMES:
            raise ValueError(f'`model_name` must be one of these: {MODEL_NAMES}, but got {model_name}')
        
        self.model_name = model_name
        self.pretrained = pretrained

        # add attributes
        # "batch_norm_eps": 0.001,
        # "batch_norm_momentum": 0.99,
        # "depth_coefficient": 3.1,
        # "depth_divisor": 8,
        # "depthwise_padding": [],
        # "drop_connect_rate": 0.2,
        # "dropout_rate": 0.5,
        # "expand_ratios": [
        # 1,
        # 6,
        # 6
        # ],
        # "hidden_act": "gelu",
        # "hidden_dim": 2560,
        # "id2label": { IMAGE NET DATASET
        # "0": "LABEL_0",
        # "1": "LABEL_1",
        # "2": "LABEL_2",
        # ...
        # },
        # "image_size": 600,
        # "in_channels": [
        # 32,
        # 16,
        # 24
        # ],
        # "initializer_range": 0.02,
        # "kernel_sizes": [
        # 3,
        # 3,
        # 5
        # ],
        # "label2id": {
        # "LABEL_0": 0,
        # "LABEL_1": 1,
        # "LABEL_2": 2,
        # ...
        # },
        # "model_type": "efficientnet",
        # "num_block_repeats": [
        # 1,
        # 1,
        # 2
        # ],
        # "num_channels": 3,
        # "num_hidden_layers": 16,
        # "out_channels": [
        # 16,
        # 24,
        # 40
        # ],
        # "pooling_type": "mean",
        # "squeeze_expansion_ratio": 0.25,
        # "strides": [
        # 1,
        # 1,
        # 2
        # ],
        # "width_coefficient": 2.0
        # }
       
        super().__init__(**kwargs)

class EfficientNetOnnxConfig(ViTOnnxConfig):
    @property
    def outputs(self) -> Dict[str, Dict[int, str]]:
        common_outputs = super().outputs 

        if self.task == "image-classification":
            common_outputs["logits"] = {0: "batch_size", 1: "num_classes"}
        
        return common_outputs
    
__all__ = [
    'MODEL_NAMES',
    'EfficientNetConfig',
    'EfficientNetOnnxConfig'
]