code/detection_trt/configs/fvit_tensorrt_fp16.py

# TensorRT FP16 部署配置
# 用于 DeCLIP (csa模式) 和 CLIP (vanilla模式) 检测器

# 后端配置
backend_config = dict(
    type='tensorrt',
    common_config=dict(
        fp16_mode=True,  # 启用 FP16 加速
        max_workspace_size=1 << 30,  # 1GB workspace
    ),
    model_inputs=[
        dict(
            input_shapes=dict(
                # 图像输入: [batch, channels, height, width]
                input=dict(
                    min_shape=[1, 3, 560, 560],
                    opt_shape=[1, 3, 560, 560],
                    max_shape=[1, 3, 800, 1333],
                )
            )
        )
    ]
)

# ONNX 配置
onnx_config = dict(
    type='onnx',
    export_params=True,
    keep_initializers_as_inputs=False,
    opset_version=17,
    save_file='model.onnx',
    input_names=['input'],
    output_names=['roi_features', 'dense_features'],
    input_shape=None,
    optimize=True,
    dynamic_axes={
        'input': {
            0: 'batch',
            2: 'height', 
            3: 'width'
        },
        'roi_features': {
            0: 'num_rois',
        },
        'dense_features': {
            0: 'batch',
            2: 'feat_height',
            3: 'feat_width'
        }
    }
)

# 代码库配置 (用于自定义模块)
codebase_config = dict(
    type='mmdet',  # 使用 mmdet 代码库
    task='ObjectDetection',
    model_type='end2end',
    post_processing=dict(
        score_threshold=0.05,
        confidence_threshold=0.005,
        iou_threshold=0.5,
        max_output_boxes_per_class=200,
        pre_top_k=5000,
        keep_top_k=100,
        background_label_id=-1,
    )
)

# EVA-CLIP ViT 特定配置
evaclip_config = dict(
    # 模型变体
    model_name='EVA02-CLIP-B-16',  # 或 'EVA02-CLIP-L-14-336'
    
    # 输入尺寸
    image_size=560,  # 检测任务使用的图像尺寸
    patch_size=16,
    
    # 特征模式
    feature_mode='csa',  # 'vanilla' for CLIP, 'csa' for DeCLIP
    
    # RoPE 位置编码
    use_rope=True,
    
    # xformers (需要在 ONNX 导出时禁用)
    disable_xformers=True,
)

# 模型路径配置
model_paths = dict(
    # EVA-CLIP 预训练权重
    evaclip_b16='/mnt/SSD8T/home/wjj/code/my_CLIPSelf/checkpoints/EVA02_CLIP_B_psz16_s8B.pt',
    evaclip_l14_336='/mnt/SSD8T/home/wjj/code/my_CLIPSelf/checkpoints/EVA02_CLIP_L_psz14_336_s8B.pt',
    
    # DeCLIP 微调权重 (需要根据实际路径修改)
    declip_csa_b16=None,  # 待填写
    declip_csa_l14_336=None,  # 待填写
)

# 数据配置
data_config = dict(
    # COCO Panoptic 数据集路径
    coco_root='/mnt/SSD8T/home/wjj/dataset/standard_coco',
    train_ann='annotations/instances_train2017.json',
    val_ann='annotations/panoptic_val2017.json',
    train_img='train2017',
    val_img='val2017',
    panoptic_segm='annotations/panoptic_val2017',
    
    # 文本嵌入缓存
    embed_path='metadata/coco_panoptic_clip_hand_craft_EVACLIP_ViTB16.npy',
)

# Benchmark 配置
benchmark_config = dict(
    # 测试轮数
    warmup_rounds=10,
    test_rounds=100,
    
    # 批处理大小
    batch_sizes=[1, 4, 8],
    
    # 输入尺寸
    input_sizes=[(560, 560), (640, 640), (800, 1333)],
    
    # 是否同步 CUDA
    sync_cuda=True,
)