DDPM: per-timestep affine eps-model trained on the noise-prediction objective (hmddpm)

Usage

morie_geron_ddpm(
  X,
  T = 10,
  beta_schedule = "linear",
  epochs = 200,
  lr = 0.05,
  seed = 0
)

Arguments

X: Training data (m, d).
T: Diffusion steps.
beta_schedule: "linear"/"cosine"/vector.
epochs, lr, seed: Training config.

Value

List with loss_history, final_loss, loss_by_t, A, b, alpha_bar, sample.

Examples

V <- c(1, 2, 3, 4, 5, 6, 7, 8)
morie_geron_ddpm(V)
#> $loss_history
#>   [1] 1.146176e+00 1.265423e+00 1.572123e+00 2.342635e+00 4.263296e+00
#>   [6] 9.041201e+00 2.092866e+01 5.053361e+01 1.243570e+02 3.086984e+02
#>  [11] 7.696554e+02 1.923917e+03 4.818238e+03 1.208564e+04 3.035783e+04
#>  [16] 7.635936e+04 1.923207e+05 4.850076e+05 1.224668e+06 3.096169e+06
#>  [21] 7.837123e+06 1.986112e+07 5.039125e+07 1.279971e+08 3.254825e+08
#>  [26] 8.285655e+08 2.111478e+09 5.386363e+09 1.375444e+10 3.515743e+10
#>  [31] 8.995120e+10 2.303570e+11 5.904582e+11 1.514814e+12 3.889562e+12
#>  [36] 9.995456e+12 2.570717e+13 6.616759e+13 1.704375e+14 4.393433e+14
#>  [41] 1.133318e+15 2.925489e+15 7.556750e+15 1.953225e+16 5.051743e+16
#>  [46] 1.307354e+17 3.385320e+17 8.771078e+17 2.273761e+18 5.897501e+18
#>  [51] 1.530435e+19 3.973552e+19 1.032174e+20 2.682444e+20 6.974384e+20
#>  [56] 1.814147e+21 4.720900e+21 1.229015e+22 3.200845e+22 8.339538e+22
#>  [61] 2.173623e+23 5.667440e+23 1.478242e+24 3.857052e+24 1.006730e+25
#>  [66] 2.628539e+25 6.865236e+25 1.793629e+26 4.687515e+26 1.225412e+27
#>  [71] 3.204407e+27 8.381782e+27 2.193033e+28 5.739466e+28 1.502492e+29
#>  [76] 3.934277e+29 1.030450e+30 2.699579e+30 7.074074e+30 1.854152e+31
#>  [81] 4.860950e+31 1.274661e+32 3.343215e+32 8.770563e+32 2.301349e+33
#>  [86] 6.039865e+33 1.585478e+34 4.162738e+34 1.093157e+35 2.871235e+35
#>  [91] 7.542859e+35 1.981904e+36 5.208433e+36 1.369014e+37 3.599011e+37
#>  [96] 9.463064e+37 2.488583e+38 6.545501e+38 1.721879e+39 4.530330e+39
#> [101] 1.192129e+40 3.137484e+40 8.258544e+40 2.174141e+41 5.724441e+41
#> [106] 1.507434e+42 3.970105e+42 1.045739e+43 2.754868e+43 7.258277e+43
#> [111] 1.912583e+44 5.040341e+44 1.328469e+45 3.501818e+45 9.231784e+45
#> [116] 2.434032e+46 6.418221e+46 1.692582e+47 4.464065e+47 1.177487e+48
#> [121] 3.106172e+48 8.194787e+48 2.162180e+49 5.705415e+49 1.505645e+50
#> [126] 3.973723e+50 1.048844e+51 2.768611e+51 7.308861e+51 1.929628e+52
#> [131] 5.094865e+52 1.345322e+53 3.552660e+53 9.382399e+53 2.478029e+54
#> [136] 6.545314e+54 1.728962e+55 4.567414e+55 1.206660e+56 3.188072e+56
#> [141] 8.423634e+56 2.225863e+57 5.881988e+57 1.554448e+58 4.108222e+58
#> [146] 1.085817e+59 2.870014e+59 7.586392e+59 2.005441e+60 5.301605e+60
#> [151] 1.401610e+61 3.705687e+61 9.797868e+61 2.590689e+62 6.850452e+62
#> [156] 1.811520e+63 4.790559e+63 1.266917e+64 3.350648e+64 8.861912e+64
#> [161] 2.343924e+65 6.199789e+65 1.639937e+66 4.338040e+66 1.147562e+67
#> [166] 3.035807e+67 8.031331e+67 2.124789e+68 5.621581e+68 1.487357e+69
#> [171] 3.935374e+69 1.041287e+70 2.755292e+70 7.290845e+70 1.929304e+71
#> [176] 5.105469e+71 1.351085e+72 3.575537e+72 9.462616e+72 2.504334e+73
#> [181] 6.628026e+73 1.754231e+74 4.643008e+74 1.228916e+75 3.252782e+75
#> [186] 8.609883e+75 2.279024e+76 6.032669e+76 1.596905e+77 4.227245e+77
#> [191] 1.119036e+78 2.962369e+78 7.842277e+78 2.076123e+79 5.496316e+79
#> [196] 1.455117e+80 3.852398e+80 1.019933e+81 2.700344e+81 7.149462e+81
#> 
#> $final_loss
#> [1] 1.892928e+82
#> 
#> $loss_by_t
#>  [1] 1.846091e+83 3.812394e+81 8.462487e+79 1.020204e+76 4.954265e+79
#>  [6] 7.369048e+80 2.546293e+77 8.354322e+67 2.007857e+67 1.480795e+76
#> 
#> $A
#>  [1]  8.252301e+40  1.189975e+40  1.776355e+39  1.965751e+37 -1.360093e+39
#>  [6] -5.239063e+39 -9.794097e+37 -1.799280e+33 -8.838968e+32 -2.367931e+37
#> 
#> $b
#>                [,1]
#>  [1,]  1.468204e+40
#>  [2,]  2.121407e+39
#>  [3,]  3.166604e+38
#>  [4,]  3.515605e+36
#>  [5,] -2.416864e+38
#>  [6,] -9.349161e+38
#>  [7,] -1.767883e+37
#>  [8,] -3.328886e+32
#>  [9,] -1.640623e+32
#> [10,] -4.357508e+36
#> 
#> $alpha_bar
#>  [1] 0.9999000 0.9975891 0.9930778 0.9863911 0.9775684 0.9666630 0.9537420
#>  [8] 0.9388848 0.9221831 0.9037394
#> 
#> $betas
#>  [1] 0.000100000 0.002311111 0.004522222 0.006733333 0.008944444 0.011155556
#>  [7] 0.013366667 0.015577778 0.017788889 0.020000000
#> 
#> $sample
#>      [,1]
#> [1,]  NaN
#> 
#> $monotone
#> [1] FALSE
#> 
#> $T
#> [1] 10
#> 
#> $estimate
#> [1] 1.892928e+82
#> 
#> $n
#> [1] 8
#> 
#> $method
#> [1] "DDPM with a per-timestep affine eps-model trained on the noise-prediction objective"
#>