csHR2acdcHR_ciss.py

# Licensed under the CC BY-NC 4.0 license (https://creativecommons.org/licenses/by-nc/4.0/)
# Default CISS configuration for Cityscapes->ACDC
_base_ = [
    '../_base_/default_runtime.py',
    # DAFormer network architecture
    '../_base_/models/daformer_sepaspp_mitb5.py',
    # Cityscapes->ACDC High-Resolution Data Loading with FDA-based stylization
    '../_base_/datasets/uda_cityscapesHR_to_acdcHR_1024x1024_fda.py',
    # Training with labels on source dataset, pseudo-labels on target dataset and feature invariance losses on both datasets
    '../_base_/uda/dacs_a999_fdthings_ciss_src_ceorig_inv_trg_ceorigorig_invorigorigstylizedstylized.py',
    # AdamW Optimizer
    '../_base_/schedules/adamw.py',
    # Linear Learning Rate Warmup with Subsequent Linear Decay
    '../_base_/schedules/poly10warm.py'
]
# Random Seed
seed = 0
# CISS Configuration
model = dict(
    type='HRDAEncoderDecoder',
    decode_head=dict(
        type='HRDAHead',
        # Use the DAFormer decoder for each scale.
        single_scale_head='DAFormerHead',
        # Learn a scale attention for each class channel of the prediction.
        attention_classwise=True,
        # Set the detail loss weight $\lambda_d=0.1$.
        hr_loss_weight=0.1),
    # Use the full resolution for the detail crop and half the resolution for
    # the context crop.
    scales=[1, 0.5],
    # Use a relative crop size of 0.5 (=512/1024) for the detail crop.
    hr_crop_size=[512, 512],
    # Use LR features for the Feature Distance as in the original DAFormer.
    feature_scale=0.5,
    # Make the crop coordinates divisible by 8 (output stride = 4,
    # downscale factor = 2) to ensure alignment during fusion.
    crop_coord_divisible=8,
    # Use overlapping slide inference for detail crops for pseudo-labels.
    hr_slide_inference=True,
    # Use overlapping slide inference for fused crops during test time.
    test_cfg=dict(
        mode='slide',
        batched_slide=True,
        stride=[512, 512],
        crop_size=[1024, 1024]))
uda = dict(
    # FDA-based stylization.
    stylization='fda',
    stylize=dict(
        inv_loss=dict(
            # Feature invariance loss weights set to 50 for source Cityscapes domain and 20 for target ACDC domain.
            weight=50.0,
            weight_target=20.0
        )
    )
)
data = dict(
    train=dict(
        # Rare Class Sampling
        rare_class_sampling=dict(
            min_pixels=3000, class_temp=0.01, min_crop_ratio=2.0)),
    # Use several separate threads/workers for data loading to account for on-the-fly stylization.
    workers_per_gpu=16,
    # Keep batch size the same as in HRDA.
    samples_per_gpu=2)
# Optimizer Hyperparameters
optimizer_config = None
optimizer = dict(
    lr=6e-05,
    paramwise_cfg=dict(
        custom_keys=dict(
            head=dict(lr_mult=10.0),
            pos_block=dict(decay_mult=0.0),
            norm=dict(decay_mult=0.0))))
n_gpus = 1
gpu_model = 'NVIDIAA100_PCIE_80GB'
runner = dict(type='IterBasedRunner', max_iters=40000)
# Logging Configuration
checkpoint_config = dict(by_epoch=False, interval=40000, max_keep_ckpts=1)
evaluation = dict(interval=4000, metric='mIoU', distributed_eval=False, pre_eval=False)
# Meta Information for Result Analysis
name = 'csHR2acdcHR_ciss_s0'
exp = 'basic'
name_dataset = 'cityscapesHR2acdcHR_1024x1024'
name_architecture = 'hrda1-512-0.1_daformer_sepaspp_sl_mitb5'
name_encoder = 'mitb5'
name_decoder = 'hrda1-512-0.1_daformer_sepaspp_sl'
name_uda = 'dacs_a999_fdthings_ciss_src_ceorig_inv_trg_ceorigorig_invorigorigstylizedstylized_rcs0.01-2.0'
name_opt = 'adamw_6e-05_pmTrue_poly10warm_1x2_40k'
first_run = True
resume_from = None

# For the other configurations used in the paper, please refer to experiments.py