suggestions for training on entire sen12

[QiuWYZju]

I have downloaded the entire SEN12 dataset, which contains approximately 200,000 samples. I trained the model using the default hyperparameters on 8 A100 GPUs, but the results are not satisfactory. Although the FID is low, both PSNR and SSIM are somewhat low as well. Do you have any suggestions?

[DeepSARRS]

Which stage? Could you provide the configuration and results here?

[QiuWYZju]

------------------stage1:-------------------------
cfg:

`{
"name": "S1_freq0_lp0_gan0_lr5e-5_OnSen12ALL",
"phase": "train", // train or val
"gpu_ids": [
0,1,2,3,4,5,6,7
],
"stage": 1, // 1 for modality transform 2 for local refinement
"ddim_steps": null, // set null for stage 1, 1 for step 2
"loss_w":
{
"fft_w": 0,
"lpips_w": 0,
"lambda_gan": 0,
"lcondition_w": 0
},

"path": { //set the path
    "log": "logs",
    
    "tb_logger": "tb_logger",
    "results": "results",
    "checkpoint": "checkpoint",
    "resume_state": null
    // "resume_state": "experiments/S1_freq0_lp0_gan0_lr5e-5_240908_215924/checkpoint/I380000_E390" //pretrain model or training state
},
"datasets": {
    "train": {
        "name": "SEN12_ALL",
        "mode": "translate", // whether need LR img
        "dataroot": "/mnt/si000082zwb2/default/data/SEN12",
        "datatype": "img", //lmdb or img, path of img files
        "l_resolution": 256, // low resolution need to super_resolution
        "r_resolution": 256, // high resolution
        "batch_size": 16,  //16
        "num_workers": 4,
        "use_shuffle": true,
        "data_len": -1 // -1 represents all data used in train
    },
    "val": {
        "name": "SEN12_ALL",
        "mode": "translate",
        "dataroot": "/mnt/si000082zwb2/default/data/SEN12",
        "datatype": "img", //lmdb or img, path of img files
        "l_resolution": 256,
        "r_resolution": 256,
        "data_len": 30
    }
},
"model": {
    "which_model_G": "sr3", // use the ddpm or sr3 network structure
    "finetune_norm": false,
    "unet": {
        "in_channel": 3,   
        // "condition_channel": 2,
        "out_channel": 3,
        "condition_ch": 3,
        "inner_channel": 64,
        "norm_groups": 16,   //GroupNorm groups
        "channel_multiplier": [
            1,
            2,
            4,
            8,
            // 8,
            // 16,
            16
        ],
        "attn_res": [
            // 16
        ],
        "res_blocks": 1,
        "dropout": 0
    },
    "beta_schedule": { // use munual beta_schedule for acceleration
        "train": {
            "schedule": "linear",   //linear
            "n_timestep": 1000,
            "linear_start": 1e-6,
            "linear_end": 1e-2,
            "ddim": 0,  // set 1 for ddim, 0 for ddpm
            "lq_noiselevel": 0   // 0 by default

        },
        "val": {
            "schedule": "linear",  //linear
            "n_timestep": 50,
            "linear_start": 1e-6,
            "linear_end": 1e-2,
            "ddim": 1,  //set 1 for ddim, 0 for ddpm
            "lq_noiselevel": 0   // 0 by default
        }
    },
    "diffusion": {
        "image_size": 256,
        "channels": 3, //sample channel
        "conditional": true // unconditional generation or unconditional generation(super_resolution)
    }
},
"train": {
    "n_iter": 600000,
    "val_freq": 5000,   //20000
    "save_checkpoint_freq": 10000,
    "print_freq": 50,
    "optimizer": {
        "type": "adam",
        "lr": 5e-5    
    },
    "ema_scheduler": { // not used now
        "step_start_ema": 5000,
        "update_ema_every": 1,
        "ema_decay": 0.9999
    },

    "scheduler": {
        "type": "MultiStepLR",
        "milestones": [],   // 540000
        "gamma": 0.5
    }
        
},
"wandb": {
    "project": "SEN12_S1"
}

}
`
result: random 100 SEN12 images

stage1:FID分数: 187.74 SSIM平均值 (灰度图): 0.2174 ± 0.1449 PSNR平均值 (灰度图): 12.0821 ± 3.0162 dB

-----------------stage2:-------------------------
cfg:
`{
"name": "S2_1step_freq10_lp5_gan0.5_lr5e-5_OnSen12ALL",
"phase": "train", // train or val
"gpu_ids": [
0,1,2,3
],
"stage": 2, // 1 for modality transform 2 for local refinement
"ddim_steps": 1, // ddim steps for sample
"loss_w":
{
"fft_w": 10,
"lpips_w": 5,
"lambda_gan": 0.5,
"lcondition_w": 0
},

"path": { //set the path
    "log": "logs",
    
    "tb_logger": "tb_logger",
    "results": "results",
    "checkpoint": "checkpoint",
    // "resume_state": null
    "resume_state": "/mnt/si000082zwb2/default/qwy/E3Diff/experiments/S1_freq0_lp0_gan0_lr5e-5_OnSen12ALL_251106_090510/checkpoint/I200000_E14" //pretrain model or training state
},
"datasets": {
    "train": {
        "name": "SEN12_ALL",
        "mode": "translate", // whether need LR img
        "dataroot": "/mnt/si000082zwb2/default/data/SEN12",
        "datatype": "img", //lmdb or img, path of img files
        "l_resolution": 256, // low resolution need to super_resolution
        "r_resolution": 256, // high resolution
        "batch_size": 16,  //16
        "num_workers": 4,
        "use_shuffle": true,
        "data_len": -1 // -1 represents all data used in train
    },
    "val": {
        "name": "SEN12_ALL",
        "mode": "translate",
        "dataroot": "/mnt/si000082zwb2/default/data/SEN12",
        "datatype": "img", //lmdb or img, path of img files
        "l_resolution": 256,
        "r_resolution": 256,
        "data_len": 30
    }
},
"model": {
    "which_model_G": "sr3", // use the ddpm or sr3 network structure
    "finetune_norm": false,
    "unet": {
        "in_channel": 3,   
        // "condition_channel": 2,
        "out_channel": 3,
        "condition_ch": 3,
        "inner_channel": 64,
        "norm_groups": 16,   //GroupNorm参数
        "channel_multiplier": [
            1,
            2,
            4,
            8,
            // 8,
            // 16,
            16
        ],
        "attn_res": [
            // 16
        ],
        "res_blocks": 1,
        "dropout": 0
    },
    "beta_schedule": { // use munual beta_schedule for acceleration

        "train": {
            "schedule": "linear",   //linear
            "n_timestep": 1000,
            "linear_start": 1e-6,
            "linear_end": 1e-2,
            "ddim": 0,  //1 for ddim, 0 for ddpm
            "lq_noiselevel": 0   //0 by default

        },
        "val": {
            "schedule": "linear",  //linear
            "n_timestep": 1,  // ddim steps for sample
            "linear_start": 1e-6,
            "linear_end": 1e-2,
            "ddim": 1,  //1 for ddim, 0 for ddpm
            "lq_noiselevel": 0   //0 by default
        }
    },
    "diffusion": {
        "image_size": 256,
        "channels": 3, //sample channel
        "conditional": true // unconditional generation or unconditional generation(super_resolution)
    }
},
"train": {
    "n_iter": 800000,
    "val_freq": 500,   //20000
    "save_checkpoint_freq": 10000,
    "print_freq": 50,
    "optimizer": {
        "type": "adam",
        "lr": 5e-5    
    },
    "ema_scheduler": { // not used now
        "step_start_ema": 5000,
        "update_ema_every": 1,
        "ema_decay": 0.9999
    },

    "scheduler": {
        "type": "MultiStepLR",
        "milestones": [ ],   // 540000
        "gamma": 0.5
    }
        
},
"wandb": {
    "project": "SEN12_S2"
}

}`
result: random 100 SEN12 images

stage2:FID分数: 162.85 SSIM平均值 (灰度图): 0.1964 ± 0.0936 PSNR平均值 (灰度图): 14.9670 ± 4.0532 dB

[deepffff]

The first stage doesn’t seem to have converged yet.