Spaces:
Running
Running
| from pixrender.properties import (NB_EPOCHS, TRAIN, VALIDATION, SCHEDULER, REDUCELRONPLATEAU, | |
| MODEL, ARCHITECTURE, ID, NAME, SCHEDULER_CONFIGURATION, OPTIMIZER, PARAMS, LR, | |
| LOSS, LOSS_MSE, DATALOADER, BATCH_SIZE, SCENE, NB_POINTS, SEED, PSEUDO_COLOR_DIMENSION, | |
| SCALE_LIST, RATIO_TRAIN, MULTISCALE_SUPERVISION) | |
| from pixrender.synthesis.world_simulation import STAIRCASE | |
| from typing import List | |
| def model_configurations(config, model_preset="UNet", k_size=1, depth=2, h_dim=8) -> dict: | |
| if model_preset == "UNet": | |
| config[MODEL] = { | |
| ARCHITECTURE: dict( | |
| in_channels=config[PSEUDO_COLOR_DIMENSION], | |
| width=64, | |
| enc_blk_nums=[1, 1, 1, 28], | |
| middle_blk_num=1, | |
| dec_blk_nums=[1, 1, 1, 1], | |
| ), | |
| NAME: model_preset | |
| } | |
| elif model_preset == "StackedConvolutions": | |
| config[MODEL] = { | |
| ARCHITECTURE: dict( | |
| in_channels=config[PSEUDO_COLOR_DIMENSION], | |
| n_scales=len(config[SCALE_LIST]), | |
| k_size=k_size, | |
| depth=depth, | |
| h_dim=h_dim | |
| ), | |
| NAME: model_preset | |
| } | |
| elif model_preset == "Bypass": | |
| config[MODEL] = { | |
| ARCHITECTURE: dict( | |
| in_channels=config[PSEUDO_COLOR_DIMENSION], | |
| n_scales=len(config[SCALE_LIST]), | |
| k_size=k_size | |
| ), | |
| NAME: model_preset | |
| } | |
| elif model_preset == "TrueBypass": | |
| assert config[PSEUDO_COLOR_DIMENSION] == 3, f"TrueBypass requires in_channels == out_channels" | |
| config[MODEL] = { | |
| ARCHITECTURE: dict( | |
| in_channels=config[PSEUDO_COLOR_DIMENSION], | |
| n_scales=len(config[SCALE_LIST]), | |
| ), | |
| NAME: model_preset | |
| } | |
| else: | |
| raise ValueError(f"Unknown model preset {model_preset}") | |
| def presets_experiments( | |
| exp: int, | |
| b: int = 16, | |
| n: int = 50, | |
| model_preset: str = "UNet", | |
| scene: str = STAIRCASE, | |
| nb_points: int = 20000, | |
| seed: int = 42, | |
| pseudo_color_dimension: int = 3, | |
| scale_list: List[int] = [0, 1, 2, 3], | |
| lr: float = 1e-3, | |
| k_size: int = 3, | |
| ratio_train: float = 0.8, | |
| depth: int = 2, | |
| h_dim: int = 8, | |
| ms_supervision: bool = True | |
| ) -> dict: | |
| config = { | |
| ID: exp, | |
| NAME: f"{exp:04d}", | |
| NB_EPOCHS: n | |
| } | |
| config[OPTIMIZER] = { | |
| NAME: "Adam", | |
| PARAMS: { | |
| LR: lr | |
| } | |
| } | |
| config[DATALOADER] = { | |
| BATCH_SIZE: { | |
| TRAIN: b, | |
| VALIDATION: b | |
| }, | |
| } | |
| config[MULTISCALE_SUPERVISION] = ms_supervision | |
| config[RATIO_TRAIN] = ratio_train | |
| config[PSEUDO_COLOR_DIMENSION] = pseudo_color_dimension | |
| config[SCALE_LIST] = scale_list | |
| model_configurations(config, model_preset=model_preset, k_size=k_size, depth=depth, h_dim=h_dim) | |
| config[SCHEDULER] = REDUCELRONPLATEAU | |
| config[SCHEDULER_CONFIGURATION] = { | |
| "factor": 0.8, | |
| "patience": 5 | |
| } | |
| config[LOSS] = LOSS_MSE | |
| config[SCENE] = scene | |
| config[NB_POINTS] = nb_points | |
| config[SEED] = seed | |
| return config | |
| def get_experiment_from_id(exp: int): | |
| if exp == 0: | |
| conf = presets_experiments(exp, b=4, n=100, model_preset="TrueBypass", | |
| scene=STAIRCASE, pseudo_color_dimension=3, lr=0.3, k_size=1) | |
| if exp == 1: | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene=STAIRCASE, pseudo_color_dimension=3, lr=0.3, k_size=1) | |
| if exp == 2: | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene=STAIRCASE, pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98) # Chekc | |
| if exp == 3: | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene="volleyball", pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98) # Backface culling disabled! | |
| if exp == 4: | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene="volleyball", pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98) # Optimize with backface culling | |
| if exp == 5: | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene="volleyball", pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98) # Optimize with backface culling + 12 views! | |
| if exp == 6: | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="StackedConvolutions", | |
| scene="volleyball", pseudo_color_dimension=3, lr=0.001, k_size=5, ratio_train=0.98) | |
| if exp == 7: | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="Bypass", | |
| scene="volleyball", pseudo_color_dimension=3, lr=0.01, k_size=5, ratio_train=0.98) | |
| if exp == 10: | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene="volleyball", pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98) | |
| elif exp == 11: | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="StackedConvolutions", | |
| scene="volleyball", pseudo_color_dimension=3, lr=0.1, k_size=5, ratio_train=0.98) | |
| elif exp == 12: # 29.5db | |
| conf = presets_experiments(exp, b=16, n=100, model_preset="StackedConvolutions", | |
| scene="volleyball", pseudo_color_dimension=8, lr=0.01, k_size=5, ratio_train=0.98, | |
| depth=2, h_dim=8) | |
| elif exp == 13: # 29.2db | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="StackedConvolutions", | |
| scene="volleyball", pseudo_color_dimension=8, lr=0.01, k_size=5, ratio_train=0.98, | |
| depth=4, h_dim=8) | |
| elif exp == 14: # 31.5db | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="StackedConvolutions", | |
| scene="volleyball", pseudo_color_dimension=8, lr=0.001, k_size=5, ratio_train=0.98, | |
| depth=4, h_dim=8) | |
| elif exp == 15: # 31.6db | |
| conf = presets_experiments(exp, b=8, n=400, model_preset="StackedConvolutions", | |
| scene="volleyball", pseudo_color_dimension=8, lr=0.005, k_size=5, ratio_train=0.98, | |
| depth=4, h_dim=8) | |
| # CHAIR IS TOO SMALL - alpha used in depth test probably hinders the learning process | |
| elif exp == 20: | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene="chair", pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98) | |
| elif exp == 21: | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="StackedConvolutions", | |
| scene="chair", pseudo_color_dimension=8, lr=0.005, k_size=5, ratio_train=0.98, | |
| depth=4, h_dim=8) | |
| elif exp == 30: # 20db | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene="material_balls", pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98) | |
| elif exp == 31: | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="StackedConvolutions", | |
| scene="material_balls", pseudo_color_dimension=8, lr=0.005, k_size=5, ratio_train=0.98, | |
| depth=4, h_dim=8) | |
| elif exp == 32: | |
| conf = presets_experiments(exp, b=4, n=2000, model_preset="StackedConvolutions", | |
| scene="material_balls", pseudo_color_dimension=8, lr=0.005, k_size=7, ratio_train=0.98, | |
| depth=4, h_dim=8) | |
| elif exp == 40: # 20.5db | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene="ficus", pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98) | |
| elif exp == 41: # 22.7db | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene="ficus", pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98, | |
| nb_points=100000) | |
| elif exp == 42: # ?? | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="Bypass", | |
| scene="ficus", pseudo_color_dimension=8, lr=0.01, k_size=5, ratio_train=0.98, | |
| nb_points=100000) | |
| elif exp == 43: | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="StackedConvolutions", | |
| scene="ficus", pseudo_color_dimension=8, lr=0.01, k_size=3, ratio_train=0.98, | |
| depth=2, h_dim=8, | |
| nb_points=20000) | |
| elif exp == 44: | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="StackedConvolutions", | |
| scene="ficus", pseudo_color_dimension=8, lr=0.01, k_size=3, ratio_train=0.98, | |
| depth=2, h_dim=8, | |
| nb_points=100000) | |
| elif exp == 45: | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="StackedConvolutions", | |
| scene="ficus", pseudo_color_dimension=8, lr=0.01, k_size=3, ratio_train=0.98, | |
| depth=4, h_dim=8, | |
| nb_points=100000) | |
| elif exp == 46: | |
| conf = presets_experiments(exp, b=8, n=300, model_preset="StackedConvolutions", | |
| scene="ficus", pseudo_color_dimension=8, lr=0.01, k_size=3, ratio_train=0.98, | |
| depth=8, h_dim=8, | |
| nb_points=100000) | |
| elif exp == 500: # 16.7db | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene="old_chair", pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98, | |
| nb_points=100000, | |
| ms_supervision=False) | |
| elif exp == 50: # 16.7db | |
| conf = presets_experiments(exp, b=32, n=100, model_preset="TrueBypass", | |
| scene="old_chair", pseudo_color_dimension=3, lr=0.3, k_size=1, ratio_train=0.98, | |
| nb_points=100000) | |
| elif exp == 51: # 21.2db ?? | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="Bypass", | |
| scene="old_chair", pseudo_color_dimension=8, lr=0.01, k_size=5, ratio_train=0.98, | |
| nb_points=100000) | |
| elif exp == 52: # 28.6db -> not bad | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="StackedConvolutions", | |
| scene="old_chair", pseudo_color_dimension=8, lr=0.01, k_size=3, ratio_train=0.98, | |
| depth=2, h_dim=8, | |
| nb_points=100000) | |
| elif exp == 53: # Redo 52 - longer - 27.7db | |
| conf = presets_experiments(exp, b=8, n=600, model_preset="StackedConvolutions", | |
| scene="old_chair", pseudo_color_dimension=8, lr=0.01, k_size=3, ratio_train=0.98, | |
| depth=2, h_dim=8, | |
| nb_points=100000) | |
| elif exp == 54: # Redo 52 - longer + slower LR - 29.1db | |
| conf = presets_experiments(exp, b=8, n=600, model_preset="StackedConvolutions", | |
| scene="old_chair", pseudo_color_dimension=8, lr=0.001, k_size=3, ratio_train=0.98, | |
| depth=2, h_dim=8, | |
| nb_points=100000) | |
| elif exp == 55: # 29.9db | |
| conf = presets_experiments(exp, b=8, n=600, model_preset="StackedConvolutions", | |
| scene="old_chair", pseudo_color_dimension=8, lr=0.001, k_size=3, ratio_train=0.98, | |
| depth=4, h_dim=8, | |
| nb_points=100000) | |
| elif exp == 56: | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="Bypass", | |
| scene="old_chair", pseudo_color_dimension=8, lr=0.01, | |
| k_size=1, | |
| ratio_train=0.98, | |
| nb_points=100000) | |
| elif exp == 57: | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="Bypass", | |
| scene="old_chair", pseudo_color_dimension=8, lr=0.3, | |
| k_size=1, | |
| ratio_train=0.98, | |
| nb_points=100000) | |
| elif exp == 58: # 21.5db - 400k points! | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="TrueBypass", | |
| scene="old_chair", pseudo_color_dimension=3, lr=0.3, | |
| k_size=1, | |
| ratio_train=0.98, | |
| nb_points=400000) | |
| elif exp == 59: # 23dB - 400k points -> no MS supervision | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="TrueBypass", | |
| scene="old_chair", pseudo_color_dimension=3, lr=0.3, | |
| k_size=1, | |
| ratio_train=0.98, | |
| nb_points=400000, | |
| ms_supervision=False) | |
| elif exp == 60: # 22.8 - 400k points + LR 0.01 -> no MS supervision | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="TrueBypass", | |
| scene="old_chair", pseudo_color_dimension=3, lr=0.01, | |
| k_size=1, | |
| ratio_train=0.98, | |
| nb_points=400000, | |
| ms_supervision=False) | |
| elif exp == 61: # 25db - 800k points + LR 0.01 -> no MS supervision | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="TrueBypass", | |
| scene="old_chair", pseudo_color_dimension=3, lr=0.01, | |
| k_size=1, | |
| ratio_train=0.98, | |
| nb_points=800000, | |
| ms_supervision=False) | |
| elif exp == 62: # 25db - 800k points + LR 0.01 -> no MS supervision | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="TrueBypass", | |
| scene="old_chair", pseudo_color_dimension=3, lr=0.01, | |
| k_size=1, | |
| ratio_train=0.98, | |
| nb_points=100000, | |
| ms_supervision=False) | |
| elif exp == 63: # 19.6?? | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="StackedConvolutions", | |
| scene="old_chair", pseudo_color_dimension=8, lr=0.001, k_size=3, ratio_train=0.98, | |
| depth=2, h_dim=8, | |
| nb_points=100000, | |
| ms_supervision=False) | |
| elif exp == 70: # 26.5dB | |
| conf = presets_experiments(exp, b=8, n=100, model_preset="TrueBypass", | |
| scene="material_balls", pseudo_color_dimension=3, lr=0.01, | |
| k_size=1, | |
| ratio_train=0.98, | |
| nb_points=800000, | |
| ms_supervision=False) | |
| print(conf) | |
| return conf | |