ProDiff/conf/config.py

"""
Configuration settings for the trajectory interpolation project.

This file defines a function `load_config()` which returns a dictionary 
containing various parameters grouped by their purpose (e.g., data, model, 
diffusion, training, sampling).
"""
from types import SimpleNamespace
import torch

def load_config():
    config_args = {
        'data': {
            'traj_length': 256, # 修正：增加长度以容纳两个1小时(120点)的遮蔽和上下文
            'dataset': 'TKY_temporal',
            'traj_path1': './data/',
            'num_workers': 16,  # 增加数据加载线程，减少GPU等待
        },
        'train': {
            'batch_size': 512,  # 显著降低batch_size以适应增长的traj_length，避免显存溢出
            'n_epochs': 50,
            'n_iters': 5000000,
            'snapshot_freq': 5000,
            'validation_freq': 5,
            'dis_gpu': False,
        },
        'trans': {
            'input_dim': 3,
            'embed_dim': 512,
            'num_layers': 4,
            'num_heads': 8,
            'forward_dim': 256,
            'dropout': 0.1,
            'N_CLUSTER': 20,
        },
                       'test': {
                   'batch_size': 256,  # 同样降低测试batch size以适应增长的traj_length
                   'last_only': True,
               },
        'diffusion': {
            'beta_schedule': 'linear',
            'beta_start': 0.0001,
            'beta_end': 0.05,
            'num_diffusion_timesteps': 500,
        },
        'model': {
            'type': "simple",
            'attr_dim': 8,
            'guidance_scale': 2,
            'in_channels': 3,
            'out_ch': 3,
            'ch': 128,
            'ch_mult': [1, 2, 2, 2],
            'num_res_blocks': 2,
            'attn_resolutions': [16],
            'dropout': 0.1,
            'var_type': 'fixedlarge',
            'resamp_with_conv': True,
        },
        'data_source': 'TKY',
        'data_dir': './data/TKY/manually_split/',
        'normalization_params_file': './data/TKY/normalization_params.json',
    }
    
    # Create nested config structure to maintain compatibility
    config = SimpleNamespace()
    config.training = SimpleNamespace(**config_args['train'])
    config.test = SimpleNamespace(**config_args['test'])
    config.diffusion = SimpleNamespace(**config_args['diffusion'])
    config.model = SimpleNamespace(**config_args['model'])
    config.sampling = SimpleNamespace(**config_args['test'])  # Use test config for sampling
    # Add DDIM sampling configuration for faster testing
    config.sampling.type = 'ddim'  # Use DDIM instead of DDPM for 10x faster testing
    config.sampling.ddim_steps = 50  # 50 steps instead of 500, 10x speedup
    config.sampling.ddim_eta = 0.0  # Deterministic sampling
    config.data = SimpleNamespace(**config_args['data'])
    config.trans = SimpleNamespace(**config_args['trans'])
    
    config.device = 'cuda' if torch.cuda.is_available() else 'cpu'
    config.masking_strategy = 'multi_segment' # 新增：'multi_segment'
    config.mask_segments = [60, 60] # 修正：基于“一分钟一点”，每段遮蔽60个点
    config.mask_ratio = 0.2
    config.mask_points_per_hour = 60 # 修正：基于“一分钟一点”的先验知识
    config.z_score_normalization = False
    config.dis_gpu = False  # Distributed GPU training
    
    # Add missing top-level config fields  
    config.learning_rate = 1.5e-4  # 降低学习率，提高训练稳定性
    config.batch_size = config_args['train']['batch_size']
    config.n_epochs = config_args['train']['n_epochs']
    config.validation_freq = config_args['train']['validation_freq']
    config.warmup_epochs = 5  # 减少warmup epochs，加速训练
    config.contrastive_margin = 1.0
    config.kmeans_memory_size = 15  # 增加K-means缓存，提高聚类效率
    config.contrastive_loss_weight = 0.1
    config.ce_loss_weight = 0.1
    config.diffusion_loss_weight = 1.0
    config.device_id = 0
    config.use_amp = True  # 启用混合精度训练
    config.normalization_params_file = config_args['normalization_params_file']
    config.data_source = config_args['data_source']
    config.data_dir = config_args['data_dir']
    config.traj_length = config_args['data']['traj_length']
    
    return config