EMRDM/configs/example_training/sentinel.yaml at main · Ly403/EMRDM · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
model:
  base_learning_rate: 1.0e-4
  target: sgm.models.diffusion.ResidualDiffusionEngine
  params:
    input_key: "target"
    mean_key: "S2"
    compile_model: False
    use_ema: True
    use_flash_attn2: False
    # ckpt_path: "" # your checkpoint path

    sigma_st_config:
      target: sgm.modules.diffusionmodules.sigma2st.EDMSigma2St
      params:
        alpha: 3.0

    denoiser_config:
      target: sgm.modules.diffusionmodules.denoiser.ResidualDenoiser
      params:
        scaling_config:
          target: sgm.modules.diffusionmodules.denoiser_scaling.ResidualEDMScaling
          params:
            sigma_input: 1.0
            sigma_mu: 1.0

    network_wrapper: "sgm.modules.diffusionmodules.wrappers.CloudRemovalWrapper"

    network_config:
      target: sgm.modules.diffusionmodules.k_diffusion.image_transformer.ImageTransformerDenoiserModelInterface
      params:
        in_channels:  28
        out_channels: 13
        patch_size: [1,1]
        widths: [128,256,384,768]
        depths: [2,2,2,2]
        d_ffs: [256,512,768,1536]
        self_attns: [
            {"type": "neighborhood", "d_head": 64, "kernel_size": 7},
            {"type": "neighborhood", "d_head": 64, "kernel_size": 7},
            {"type": "global", "d_head": 64},
            {"type": "global", "d_head": 64},
        ]
        dropout_rate: [0.0,0.0,0.0,0.1]
        mapping_depth: 2
        mapping_width: 768
        mapping_d_ff: 1536
        mapping_dropout_rate: 0.1

    conditioner_config:
      target: sgm.modules.GeneralConditioner
      params:
        emb_models:
          - is_trainable: True
            input_key:  "S1S2"
            ucg_rate: 0.0
            target: sgm.modules.encoders.modules.IndentityEmbedder
          # - is_trainable: True
          #   input_key:  "S1"
          #   ucg_rate: 0.0
          #   target: sgm.modules.encoders.modules.ImageTransformerEncoderInterface
          #   params:
          #     in_channels: 2
          #     patch_size: [4,4]
          #     widths: [256,512,768]
          #     depths: [2,2,16]
          #     d_ffs: [512,1024,1536]
          #     self_attns: [
          #         {"type": "neighborhood", "d_head": 64, "kernel_size": 7},
          #         {"type": "neighborhood", "d_head": 64, "kernel_size": 7},
          #         {"type": "global", "d_head": 64},
          #     ]
          #     dropout_rate: [0.0,0.0,0.0]

    first_stage_config:
      target: sgm.models.autoencoder.IdentityFirstStage

    loss_fn_config:
      target: sgm.modules.diffusionmodules.loss.ResidualDiffusionLoss
      params:
        loss_weighting_config:
          target: sgm.modules.diffusionmodules.loss_weighting.ResidualEDMWeighting # ResidualSoftMinSnrWeighting
          params:
            sigma_input: 1.0
            sigma_mu: 1.0
        sigma_sampler_config:
          target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling
          params:
            p_mean: -1.2
            p_std: 1.2

    sampler_config:
      target: sgm.modules.diffusionmodules.sampling.ResidualEulerEDMSampler
      params:
        num_steps: 5
        s_churn: 5.0
        s_tmin: 0.0
        s_tmax: 100000000.0
        s_noise: 1.023

        discretization_config:
          target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization
          params:
            sigma_min: 0.001
            sigma_max: 100.0

    to_rgb_config:
      target: sgm.util.S1andS2_to_rgb

data:
  target: sgm.data.base.DataModuleFromConfig
  params:
    batch_size: 1
    num_workers: 8
    wrap: True
    train:
      target: sgm.data.sentinel.sentinel.SEN12MSCRInterface
      params:
        root: "/remote-home/share/dmb_nas2/liuyi/SEN12MSCR"
        split: "train"
        region: "all"
        rescale: True
        cloud_masks:

    validation:
      target: sgm.data.sentinel.sentinel.SEN12MSCRInterface
      params:
        root: "/remote-home/share/dmb_nas2/liuyi/SEN12MSCR"
        split: "val"
        region: "all"
        rescale: True
        cloud_masks:

    test:
      target: sgm.data.sentinel.sentinel.SEN12MSCRInterface
      params:
        root: "/remote-home/share/dmb_nas2/liuyi/SEN12MSCR"
        split: "test"
        region: "all"
        rescale: True
        cloud_masks:

    predict:
      target: sgm.data.sentinel.sentinel.SEN12MSCRInterface
      params:
        root: "/remote-home/share/dmb_nas2/liuyi/SEN12MSCR"
        split: "test"
        region: "all"
        rescale: True
        cloud_masks:

lightning:
  modelcheckpoint:
    params:
      every_n_train_steps: 5000
      monitor: "RMSE"

  callbacks:
    metrics_over_trainsteps_checkpoint:
      params:
        every_n_train_steps: 25000

    image_logger:
      target: main.ImageLogger
      params:
        disabled: False
        batch_frequency: 2000
        max_images: 64
        increase_log_steps: False
        log_first_step: True
        log_images_kwargs:
          use_ema_scope: False
          N: 64
          n_rows: 16
          return_intermediate: True
          return_denoised: True
          return_add_mu: True
          return_add_noise: True
          return_cond: True

  trainer:
    devices: 2,4
    num_sanity_val_steps: 2
    benchmark: True
    accumulate_grad_batches: 1
    max_epochs: 100