fpn_clipvit-b_640x640_80k.py

_base_ = [
    "_base_/models/fpn_r50.py",
    "_base_/datasets/ade20k_clip_640.py",
    "_base_/default_runtime.py",
    "_base_/schedules/schedule_80k.py",
]

model = {
    "pretrained": "pretrained/ViT-B-16.pt",
    "backbone": {
        "type": "CLIPVisionTransformer",
        "patch_size": 16,
        "width": 768,
        "layers": 12,
        "out_indices": [3, 5, 7, 11],
        "input_resolution": 640,
        "style": "pytorch",
    },
    "neck": {"in_channels": [768, 768, 768, 768]},
    "decode_head": {"channels": 256, "num_classes": 150},
    "test_cfg": {"mode": "slide", "crop_size": (640, 640), "stride": (426, 426)},
}

lr_config = {
    "policy": "poly",
    "power": 0.9,
    "min_lr": 1e-6,
    "by_epoch": False,
    "warmup": "linear",
    "warmup_iters": 1500,
    "warmup_ratio": 1e-6,
}

optimizer = {
    "type": "AdamW",
    "lr": 0.0001,
    "weight_decay": 0.0001,
    "paramwise_cfg": {
        "custom_keys": {"backbone": {"lr_mult": 0.1}, "norm": {"decay_mult": 0.0}},
    },
}

data = {"samples_per_gpu": 4}