├── experiments
├── __init__.py
└── adult
│ ├── __init__.py
│ ├── Dataloader.py
│ └── AdultExperiment.py
├── models
├── __init__.py
├── MultidimensionnalMonotonicNN.py
├── MonotonicNN.py
├── NeuralIntegral.py
├── ParallelNeuralIntegral.py
└── MaskedParallelNetwork.py
├── run_adult.py
├── run_adult.sh
├── README.md
├── LICENSE
└── PlayGround.ipynb
/experiments/__init__.py:
--------------------------------------------------------------------------------
1 |
--------------------------------------------------------------------------------
/experiments/adult/__init__.py:
--------------------------------------------------------------------------------
1 | from .Dataloader import AdultDataset
--------------------------------------------------------------------------------
/models/__init__.py:
--------------------------------------------------------------------------------
1 | from .MultidimensionnalMonotonicNN import SlowDMonotonicNN
2 |
--------------------------------------------------------------------------------
/run_adult.py:
--------------------------------------------------------------------------------
1 | from experiments.adult.AdultExperiment import *
2 |
3 | if __name__ == "__main__":
4 | print("starting experiment")
5 | run_adult_experiment()
--------------------------------------------------------------------------------
/run_adult.sh:
--------------------------------------------------------------------------------
1 | #!/usr/bin/env bash
2 | #
3 | # This file:
4 | #
5 | # -
6 | #
7 | # Slurm arguments.
8 | #
9 | #SBATCH --export=ALL
10 | #SBATCH --parsable
11 | #SBATCH --job-name "ADULT_MONOTONICITY"
12 | #SBATCH --output "ADULT_MONOTONICITY_%A.log"
13 | #SBATCH --requeue
14 | #SBATCH --cpus-per-task=4
15 | #SBATCH --mem-per-cpu=4000
16 | #SBATCH --ntasks=1
17 | #SBATCH --time="2-24:00:00"
18 | #SBATCH --gres=gpu:1
19 | #
20 |
21 | source activate UMNN
22 | python run_adult.py
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # generalized-UMNN
2 | Extension to multivariate unconstrained monotonic functions.
3 | Direct application of the underlying principle behind the Kolmogorov-Arnold theorem.
4 |
5 | This repository is a sketch of the possibility to model functions that are monotonic with respect to more than one input variable.
6 |
7 | For detailed information visit the [UMNN git repository](https://github.com/AWehenkel/UMNN).
8 |
9 | ## Cite
10 |
11 | If you make use of this code in your own work, please cite our paper:
12 |
13 | ```
14 | @inproceedings{wehenkel2019unconstrained,
15 | title={Unconstrained monotonic neural networks},
16 | author={Wehenkel, Antoine and Louppe, Gilles},
17 | booktitle={Advances in Neural Information Processing Systems},
18 | pages={1543--1553},
19 | year={2019}
20 | }
21 | ```
22 |
--------------------------------------------------------------------------------
/models/MultidimensionnalMonotonicNN.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import torch.nn as nn
3 | from .MonotonicNN import MonotonicNN
4 |
5 |
6 | class SlowDMonotonicNN(nn.Module):
7 | def __init__(self, mon_in, cond_in, hiddens, n_out=1, nb_steps=50, device="cpu"):
8 | super(SlowDMonotonicNN, self).__init__()
9 | self.inner_nets = []
10 | self.mon_in = mon_in
11 | for i in range(mon_in):
12 | self.inner_nets += [MonotonicNN(cond_in + 1, hiddens, nb_steps=nb_steps, dev=device)]
13 | self.weights = nn.Parameter(torch.randn(mon_in)).to(device)
14 | self.outer_net = MonotonicNN(1 + cond_in, hiddens, nb_steps=nb_steps, dev=device)
15 | self.device = device
16 |
17 | def to(self, device):
18 | for net in self.inner_nets:
19 | net.to(device)
20 | self.outer_net.to(device)
21 | self.weights.to(device)
22 | self.device = device
23 |
24 | def set_steps(self, nb_steps):
25 | for net in self.inner_nets:
26 | net.nb_steps = nb_steps
27 | self.outer_net.nb_steps = nb_steps
28 |
29 | def forward(self, mon_in, cond_in):
30 | inner_out = torch.zeros(mon_in.shape).to(self.device)
31 | for i in range(self.mon_in):
32 | inner_out[:, [i]] = self.inner_nets[i](mon_in[:, [i]], cond_in)
33 | inner_sum = (torch.exp(self.weights).unsqueeze(0).expand(mon_in.shape[0], -1) * inner_out).sum(1).unsqueeze(1)
34 | return self.outer_net(inner_sum, cond_in)
35 |
--------------------------------------------------------------------------------
/models/MonotonicNN.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import torch.nn as nn
3 | from .NeuralIntegral import NeuralIntegral
4 | from .ParallelNeuralIntegral import ParallelNeuralIntegral
5 |
6 |
7 | def _flatten(sequence):
8 | flat = [p.contiguous().view(-1) for p in sequence]
9 | return torch.cat(flat) if len(flat) > 0 else torch.tensor([])
10 |
11 |
12 | class IntegrandNN(nn.Module):
13 | def __init__(self, in_d, hidden_layers):
14 | super(IntegrandNN, self).__init__()
15 | self.net = []
16 | hs = [in_d] + hidden_layers + [1]
17 | for h0, h1 in zip(hs, hs[1:]):
18 | self.net.extend([
19 | nn.Linear(h0, h1),
20 | nn.ReLU(),
21 | ])
22 | self.net.pop() # pop the last ReLU for the output layer
23 | self.net.append(nn.ELU())
24 | self.net = nn.Sequential(*self.net)
25 |
26 | def to(self, device):
27 | self.net.to(device)
28 |
29 | def forward(self, x, h):
30 | return self.net(torch.cat((x, h), 1)) + 1.
31 |
32 | class MonotonicNN(nn.Module):
33 | def __init__(self, in_d, hidden_layers, nb_steps=50, dev="cpu"):
34 | super(MonotonicNN, self).__init__()
35 | self.integrand = IntegrandNN(in_d, hidden_layers)
36 | self.net = []
37 | hs = [in_d-1] + hidden_layers + [2]
38 | for h0, h1 in zip(hs[:-1], hs[1:]):
39 | self.net.extend([
40 | nn.Linear(h0, h1),
41 | nn.ReLU(),
42 | ])
43 | self.net.pop() # pop the last ReLU for the output layer
44 | # It will output the scaling and offset factors.
45 | self.net = nn.Sequential(*self.net)
46 | self.device = dev
47 | self.nb_steps = nb_steps
48 | self.to(dev)
49 |
50 | def to(self, device):
51 | self.net.to(device)
52 | self.integrand.to(device)
53 |
54 | '''
55 | The forward procedure takes as input x which is the variable for which the integration must be made, h is just other conditionning variables.
56 | '''
57 | def forward(self, x, h):
58 | x0 = torch.zeros(x.shape).to(self.device)
59 | out = self.net(h)
60 | offset = out[:, [0]]
61 | scaling = torch.exp(out[:, [1]])
62 | return scaling*ParallelNeuralIntegral.apply(x0, x, self.integrand, _flatten(self.integrand.parameters()), h, self.nb_steps) + offset
63 |
--------------------------------------------------------------------------------
/models/NeuralIntegral.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import numpy as np
3 | import math
4 |
5 |
6 | def _flatten(sequence):
7 | flat = [p.contiguous().view(-1) for p in sequence]
8 | return torch.cat(flat) if len(flat) > 0 else torch.tensor([])
9 |
10 |
11 | def compute_cc_weights(nb_steps):
12 | lam = np.arange(0, nb_steps + 1, 1).reshape(-1, 1)
13 | lam = np.cos((lam @ lam.T) * math.pi / nb_steps)
14 | lam[:, 0] = .5
15 | lam[:, -1] = .5 * lam[:, -1]
16 | lam = lam * 2 / nb_steps
17 | W = np.arange(0, nb_steps + 1, 1).reshape(-1, 1)
18 | W[np.arange(1, nb_steps + 1, 2)] = 0
19 | W = 2 / (1 - W ** 2)
20 | W[0] = 1
21 | W[np.arange(1, nb_steps + 1, 2)] = 0
22 | cc_weights = torch.tensor(lam.T @ W).float()
23 | steps = torch.tensor(np.cos(np.arange(0, nb_steps + 1, 1).reshape(-1, 1) * math.pi / nb_steps)).float()
24 |
25 | return cc_weights, steps
26 |
27 |
28 | def integrate(x0, nb_steps, step_sizes, integrand, h, compute_grad=False, x_tot=None):
29 | #Clenshaw-Curtis Quadrature Method
30 | cc_weights, steps = compute_cc_weights(nb_steps)
31 |
32 | device = x0.get_device() if x0.is_cuda else "cpu"
33 | cc_weights, steps = cc_weights.to(device), steps.to(device)
34 |
35 | if compute_grad:
36 | g_param = 0.
37 | g_h = 0.
38 | else:
39 | z = 0.
40 | xT = x0 + nb_steps*step_sizes
41 | for i in range(nb_steps + 1):
42 | x = (x0 + (xT - x0)*(steps[i] + 1)/2)
43 | if compute_grad:
44 | dg_param, dg_h = computeIntegrand(x, h, integrand, x_tot*(xT - x0)/2)
45 | g_param += cc_weights[i]*dg_param
46 | g_h += cc_weights[i]*dg_h
47 | else:
48 | dz = integrand(x, h)
49 | z = z + cc_weights[i]*dz
50 |
51 | if compute_grad:
52 | return g_param, g_h
53 |
54 | return z*(xT - x0)/2
55 |
56 |
57 | def computeIntegrand(x, h, integrand, x_tot):
58 | with torch.enable_grad():
59 | f = integrand.forward(x, h)
60 | g_param = _flatten(torch.autograd.grad(f, integrand.parameters(), x_tot, create_graph=True, retain_graph=True))
61 | g_h = _flatten(torch.autograd.grad(f, h, x_tot))
62 |
63 | return g_param, g_h
64 |
65 |
66 | class NeuralIntegral(torch.autograd.Function):
67 |
68 | @staticmethod
69 | def forward(ctx, x0, x, integrand, flat_params, h, nb_steps=20):
70 | with torch.no_grad():
71 | x_tot = integrate(x0, nb_steps, (x - x0)/nb_steps, integrand, h, False)
72 | # Save for backward
73 | ctx.integrand = integrand
74 | ctx.nb_steps = nb_steps
75 | ctx.save_for_backward(x0.clone(), x.clone(), h)
76 | return x_tot
77 |
78 | @staticmethod
79 | def backward(ctx, grad_output):
80 | x0, x, h = ctx.saved_tensors
81 | integrand = ctx.integrand
82 | nb_steps = ctx.nb_steps
83 | integrand_grad, h_grad = integrate(x0, nb_steps, x/nb_steps, integrand, h, True, grad_output)
84 | x_grad = integrand(x, h)
85 | x0_grad = integrand(x0, h)
86 | # Leibniz formula
87 | return -x0_grad*grad_output, x_grad*grad_output, None, integrand_grad, h_grad.view(h.shape), None
88 |
--------------------------------------------------------------------------------
/experiments/adult/Dataloader.py:
--------------------------------------------------------------------------------
1 | from torch.utils.data import Dataset
2 | import pandas as pd
3 |
4 |
5 | class AdultDataset(Dataset):
6 | def __init__(self, path, test=False, normalization=True):
7 | self.normalization = normalization
8 | self.df, self.y = self._prepare_data(path, test)
9 | self.test = test
10 |
11 | def _prepare_data(self, path, test):
12 | cols = [
13 | "age", "workclass", "fnlwgt", "education", "education_num",
14 | "marital_status", "occupation", "relationship", "race", "gender",
15 | "capital_gain", "capital_loss", "hours_per_week", "native_country",
16 | "income_bracket"
17 | ]
18 | x_columns = [
19 | "age", "workclass", "education", "education_num",
20 | "marital_status", "occupation", "relationship", "race", "gender",
21 | "capital_gain", "capital_loss", "hours_per_week", "native_country"
22 | ]
23 | df = pd.read_csv(path, names=cols, usecols=x_columns)
24 | df = df.iloc[1:, :]
25 | df["age"] = df["age"].astype(int)
26 |
27 | def one_hot_encode(df, col):
28 | s = df[col]
29 | encoded = pd.get_dummies(s)
30 | df.drop(columns=[col], inplace=True)
31 | df[encoded.columns] = encoded
32 |
33 | one_hot_encode(df, "workclass")
34 | one_hot_encode(df, "marital_status")
35 | one_hot_encode(df, "occupation")
36 | one_hot_encode(df, "relationship")
37 | one_hot_encode(df, "race")
38 | one_hot_encode(df, "gender")
39 | one_hot_encode(df, "native_country")
40 | one_hot_encode(df, "education")
41 |
42 | col = ['capital_gain', 'hours_per_week', 'education_num', ' Male', ' Separated', ' Tech-support',
43 | ' Without-pay', ' Widowed', ' Priv-house-serv', ' Puerto-Rico', ' Ecuador', ' Prof-specialty',
44 | ' France', ' Farming-fishing', ' Not-in-family', ' Other', ' Ireland', ' Black', ' Nicaragua',
45 | ' Philippines', ' Some-college', ' Hong', ' Prof-school', ' 12th', ' Germany', ' Adm-clerical',
46 | ' Assoc-acdm', ' Exec-managerial', ' Never-worked', ' Private', ' 10th', ' Doctorate',
47 | 'capital_loss', ' Transport-moving', ' Poland', ' Husband', ' Yugoslavia', ' HS-grad',
48 | ' Female', ' Haiti', ' Peru', ' Canada', ' White', ' India', ' South', ' Iran', ' Greece',
49 | ' Sales', ' Honduras', ' Hungary', ' China', ' Machine-op-inspct', ' Own-child', ' 1st-4th',
50 | ' Divorced', ' El-Salvador', ' Protective-serv', ' Preschool', ' Vietnam', ' Holand-Netherlands',
51 | ' Assoc-voc', ' 5th-6th', ' Italy', ' Japan', ' Wife', ' Craft-repair', ' Self-emp-inc',
52 | ' Outlying-US(Guam-USVI-etc)', ' 7th-8th', ' United-States', ' Unmarried', 'age',
53 | ' Married-AF-spouse', ' Taiwan', ' Trinadad&Tobago', ' Never-married', ' Jamaica',
54 | ' Other-service', ' Masters', ' Cambodia', ' Married-spouse-absent', ' Dominican-Republic',
55 | ' ?', ' Asian-Pac-Islander', ' Cuba', ' Portugal', ' England', ' State-gov', ' Armed-Forces',
56 | ' Married-civ-spouse', ' Amer-Indian-Eskimo', ' Guatemala', ' 11th', ' Columbia',
57 | ' Other-relative', ' Federal-gov', ' Local-gov', ' 9th', ' Self-emp-not-inc', ' Scotland',
58 | ' Laos', ' Bachelors', ' Thailand', ' Handlers-cleaners', ' Mexico']
59 | if test:
60 | df[' Holand-Netherlands'] = 0
61 |
62 | y = pd.get_dummies(pd.read_csv(path, names=cols, usecols=['income_bracket'])).iloc[1:, 1]
63 | #print(y.shape)
64 | #y = df[' Male']
65 | #print(y.shape)
66 | df = df[col]
67 | if self.normalization:
68 | self.mu = df.mean(0)
69 | self.std = df.std(0)
70 | df = (df - df.mean(0))/df.std(0)
71 | return df, y
72 |
73 | def normalize(self, mu, std):
74 | self.df = (self.df - mu)/std
75 |
76 | def __len__(self):
77 | return self.df.shape[0]
78 |
79 | def __getitem__(self, idx):
80 | return self.df.iloc[idx, :].to_numpy(), self.y.iloc[idx]
81 |
82 |
83 |
84 |
--------------------------------------------------------------------------------
/models/ParallelNeuralIntegral.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import numpy as np
3 | import math
4 |
5 |
6 | def _flatten(sequence):
7 | flat = [p.contiguous().view(-1) for p in sequence]
8 | return torch.cat(flat) if len(flat) > 0 else torch.tensor([])
9 |
10 |
11 | def compute_cc_weights(nb_steps):
12 | lam = np.arange(0, nb_steps + 1, 1).reshape(-1, 1)
13 | lam = np.cos((lam @ lam.T) * math.pi / nb_steps)
14 | lam[:, 0] = .5
15 | lam[:, -1] = .5 * lam[:, -1]
16 | lam = lam * 2 / nb_steps
17 | W = np.arange(0, nb_steps + 1, 1).reshape(-1, 1)
18 | W[np.arange(1, nb_steps + 1, 2)] = 0
19 | W = 2 / (1 - W ** 2)
20 | W[0] = 1
21 | W[np.arange(1, nb_steps + 1, 2)] = 0
22 | cc_weights = torch.tensor(lam.T @ W).float()
23 | steps = torch.tensor(np.cos(np.arange(0, nb_steps + 1, 1).reshape(-1, 1) * math.pi / nb_steps)).float()
24 |
25 | return cc_weights, steps
26 |
27 |
28 | def integrate(x0, nb_steps, step_sizes, integrand, h, compute_grad=False, x_tot=None):
29 | #Clenshaw-Curtis Quadrature Method
30 | cc_weights, steps = compute_cc_weights(nb_steps)
31 |
32 | device = x0.get_device() if x0.is_cuda else "cpu"
33 | cc_weights, steps = cc_weights.to(device), steps.to(device)
34 |
35 | xT = x0 + nb_steps*step_sizes
36 | if not compute_grad:
37 | x0_t = x0.unsqueeze(1).expand(-1, nb_steps + 1, -1)
38 | xT_t = xT.unsqueeze(1).expand(-1, nb_steps + 1, -1)
39 | h_steps = h.unsqueeze(1).expand(-1, nb_steps + 1, -1)
40 | steps_t = steps.unsqueeze(0).expand(x0_t.shape[0], -1, x0_t.shape[2])
41 | X_steps = x0_t + (xT_t-x0_t)*(steps_t + 1)/2
42 | X_steps = X_steps.contiguous().view(-1, x0_t.shape[2])
43 | h_steps = h_steps.contiguous().view(-1, h.shape[1])
44 | dzs = integrand(X_steps, h_steps)
45 | dzs = dzs.view(xT_t.shape[0], nb_steps+1, -1)
46 | dzs = dzs*cc_weights.unsqueeze(0).expand(dzs.shape)
47 | z_est = dzs.sum(1)
48 | return z_est*(xT - x0)/2
49 | else:
50 |
51 | x0_t = x0.unsqueeze(1).expand(-1, nb_steps + 1, -1)
52 | xT_t = xT.unsqueeze(1).expand(-1, nb_steps + 1, -1)
53 | x_tot = x_tot * (xT - x0) / 2
54 | x_tot_steps = x_tot.unsqueeze(1).expand(-1, nb_steps + 1, -1) * cc_weights.unsqueeze(0).expand(x_tot.shape[0], -1, x_tot.shape[1])
55 | h_steps = h.unsqueeze(1).expand(-1, nb_steps + 1, -1)
56 | steps_t = steps.unsqueeze(0).expand(x0_t.shape[0], -1, x0_t.shape[2])
57 | X_steps = x0_t + (xT_t - x0_t) * (steps_t + 1) / 2
58 | X_steps = X_steps.contiguous().view(-1, x0_t.shape[2])
59 | h_steps = h_steps.contiguous().view(-1, h.shape[1])
60 | x_tot_steps = x_tot_steps.contiguous().view(-1, x_tot.shape[1])
61 |
62 | g_param, g_h = computeIntegrand(X_steps, h_steps, integrand, x_tot_steps, nb_steps+1)
63 | return g_param, g_h
64 |
65 |
66 | def computeIntegrand(x, h, integrand, x_tot, nb_steps):
67 | h.requires_grad_(True)
68 | with torch.enable_grad():
69 | f = integrand.forward(x, h)
70 | g_param = _flatten(torch.autograd.grad(f, integrand.parameters(), x_tot, create_graph=True, retain_graph=True))
71 | g_h = _flatten(torch.autograd.grad(f, h, x_tot))
72 |
73 | return g_param, g_h.view(int(x.shape[0]/nb_steps), nb_steps, -1).sum(1)
74 |
75 |
76 | class ParallelNeuralIntegral(torch.autograd.Function):
77 |
78 | @staticmethod
79 | def forward(ctx, x0, x, integrand, flat_params, h, nb_steps=20):
80 | with torch.no_grad():
81 | x_tot = integrate(x0, nb_steps, (x - x0)/nb_steps, integrand, h, False)
82 | # Save for backward
83 | ctx.integrand = integrand
84 | ctx.nb_steps = nb_steps
85 | ctx.save_for_backward(x0.clone(), x.clone(), h)
86 | return x_tot
87 |
88 | @staticmethod
89 | def backward(ctx, grad_output):
90 | x0, x, h = ctx.saved_tensors
91 | integrand = ctx.integrand
92 | nb_steps = ctx.nb_steps
93 | integrand_grad, h_grad = integrate(x0, nb_steps, x/nb_steps, integrand, h, True, grad_output)
94 | x_grad = integrand(x, h)
95 | x0_grad = integrand(x0, h)
96 | # Leibniz formula
97 | return -x0_grad*grad_output, x_grad*grad_output, None, integrand_grad, h_grad.view(h.shape), None
98 |
--------------------------------------------------------------------------------
/experiments/adult/AdultExperiment.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import torch.nn as nn
3 | from torch.utils.data import DataLoader
4 | from torch.optim import Adam
5 | from experiments.adult.Dataloader import AdultDataset
6 | from models import SlowDMonotonicNN
7 | from tensorboardX import SummaryWriter
8 |
9 | class EmbeddingNet(nn.Module):
10 | def __init__(self, in_embedding, in_main, out_embedding, device):
11 | super(EmbeddingNet, self).__init__()
12 | self.embedding_net = nn.Sequential(nn.Linear(in_embedding, 200), nn.ReLU(),
13 | nn.Linear(200, 200), nn.ReLU(),
14 | nn.Linear(200, out_embedding), nn.ReLU()).to(device)
15 | self.umnn = SlowDMonotonicNN(in_main, out_embedding, [100, 100, 100], 1, 300, device)
16 |
17 | def set_steps(self, nb_steps):
18 | self.umnn.set_steps(nb_steps)
19 |
20 | def forward(self, x):
21 | h = self.embedding_net(x[:, 4:])
22 | return torch.sigmoid(self.umnn(x[:, :4], h))
23 |
24 |
25 | class SimpleMLP(nn.Module):
26 | def __init__(self, in_embedding, in_main, out_embedding, device):
27 | super(SimpleMLP, self).__init__()
28 | self.embedding_net = nn.Sequential(nn.Linear(in_embedding, 100), nn.ReLU(),
29 | nn.Linear(100, 100), nn.ReLU(),
30 | nn.Linear(100, out_embedding), nn.ReLU()).to(device)
31 | self.mlp = nn.Sequential(nn.Linear(in_main + out_embedding, 100), nn.ReLU(),
32 | nn.Linear(100, 100), nn.ReLU(),
33 | nn.Linear(100, 1), nn.Sigmoid()).to(device)
34 |
35 | def set_steps(self, nb_steps):
36 | return
37 |
38 | def forward(self, x):
39 | h = self.embedding_net(x[:, 4:])
40 | return self.mlp(torch.cat((x[:, :4], h), 1))
41 |
42 |
43 | def run_adult_experiment():
44 | writer = SummaryWriter()
45 | train_ds = AdultDataset("data/adult/adult.data", normalization=True)
46 | test_ds = AdultDataset("data/adult/adult.test", test=True, normalization=False)
47 |
48 | mu, std = train_ds.mu, train_ds.std
49 | test_ds.normalize(mu, std)
50 |
51 | train_dl = DataLoader(train_ds, 100, shuffle=True, num_workers=1)
52 | test_dl = DataLoader(test_ds, 100, shuffle=False, num_workers=1)
53 |
54 | x, y = train_ds[1]
55 | device = "cuda:0" if torch.cuda.is_available() else "cpu"
56 |
57 | net = EmbeddingNet(len(x) - 4, 4, 30, device)
58 | #net = SimpleMLP(len(x) - 4, 4, 30, device)
59 |
60 |
61 | if False:
62 | net.load_state_dict(torch.load("model.ckpt"))
63 | x = torch.randn(500, 4)
64 | h = torch.zeros(500, 100)
65 | with torch.no_grad():
66 | import matplotlib.pyplot as plt
67 | plt.subplot(221)
68 | y = net(torch.cat((x[:, [0]], torch.zeros(500, 3)), 1), h)
69 | plt.scatter(x[:, 0], y)
70 | plt.subplot(222)
71 | y = net(torch.cat((torch.zeros(500, 1), x[:, [1]], torch.zeros(500, 2)), 1), h)
72 | plt.scatter(x[:, 1], y)
73 | plt.subplot(223)
74 | y = net(torch.cat((torch.zeros(500, 2), x[:, [2]], torch.zeros(500, 1)), 1), h)
75 | plt.scatter(x[:, 2], y)
76 | plt.subplot(224)
77 | y = net(torch.cat((torch.zeros(500, 3), x[:, [3]]), 1), h)
78 | plt.scatter(x[:, 3], y)
79 | plt.show()
80 | exit()
81 | optim = Adam(net.parameters(), lr=.0001, weight_decay=1e-2)
82 | loss_f = nn.BCELoss()
83 |
84 | for epoch in range(1000):
85 | avg_loss = 0.
86 | i = 0
87 | avg_accuracy = 0.
88 | for x, y in train_dl:
89 | x,y = x.float().to(device), y.unsqueeze(1).float().to(device)
90 | y_est = net(x)
91 | loss = loss_f(y_est, y)
92 | optim.zero_grad()
93 | loss.backward()
94 | optim.step()
95 | avg_loss += loss.item()
96 | avg_accuracy += torch.abs((y_est.detach() > .5).float() == y).float().mean()
97 | net.set_steps(int(torch.randint(30, 60, [1])))
98 | #net.set_steps(100)
99 |
100 | i += 1
101 | if i % 100 == 0:
102 | print(i)
103 | writer.add_scalars("Adult/BCE", {"train": avg_loss / i}, epoch)
104 | writer.add_scalars("Adult/Accuracy", {"train": avg_accuracy / i}, epoch)
105 | print("train", epoch, avg_loss / i, avg_accuracy / i)
106 | avg_loss = 0.
107 | i = 0
108 | avg_accuracy = 0.
109 | net.set_steps(100)
110 | for x, y in test_dl:
111 | with torch.no_grad():
112 | x, y = x.float().to(device), y.unsqueeze(1).float().to(device)
113 | y_est = net(x)
114 | loss = loss_f(y_est, y)
115 | avg_loss += loss.item()
116 | avg_accuracy += torch.abs((y_est > .5).float() == y).float().mean()
117 | i += 1
118 | if i % 100 == 0:
119 | print(i)
120 | writer.add_scalars("Adult/BCE", {"test": avg_loss / i}, epoch)
121 | writer.add_scalars("Adult/Accuracy", {"test": avg_accuracy / i}, epoch)
122 | print("test", epoch, avg_loss / i, avg_accuracy / i)
123 | torch.save(net.state_dict(), "model.ckpt")
124 |
125 |
--------------------------------------------------------------------------------
/models/MaskedParallelNetwork.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import torch.nn as nn
3 | import torch.nn.functional as F
4 | import numpy as np
5 |
6 | class SlowParallellNetwork(nn.Module):
7 | def __init__(self, ind_in, cond_in, hidden_sizes, nout, device="cpu"):
8 | self.net = []
9 | for i in range(ind_in):
10 | net = []
11 | hs = [cond_in + 1] + hidden_sizes + [nout]
12 | for l_in, l_out in zip(hs[:-1], hs[1:]):
13 | net += [nn.Linear(l_in, l_out), nn.Relu()]
14 | net.pop()
15 | self.net += [nn.Sequential(net)]
16 | def forward(self, ind_in, cond_in):
17 | return
18 |
19 |
20 | class MaskedLinear(nn.Linear):
21 | """ same as Linear except has a configurable mask on the weights """
22 |
23 | def __init__(self, in_features, out_features, bias=True):
24 | super().__init__(in_features, out_features, bias)
25 | self.register_buffer('mask', torch.ones(out_features, in_features))
26 |
27 | def set_mask(self, mask):
28 | self.mask.data.copy_(torch.from_numpy(mask.astype(np.uint8).T))
29 |
30 | def forward(self, input):
31 | return F.linear(input, self.mask * self.weight, self.bias)
32 |
33 |
34 | class MADE(nn.Module):
35 | def __init__(self, ind_in, cond_in, hidden_sizes, nout, device="cpu"):
36 | """
37 | ind_in: integer; number of independant inputs
38 | cond_in: integer; number of conditionning inputs
39 | hidden sizes: a list of integers; number of units in hidden layers
40 | nout: integer; number of outputs per ind_in
41 | num_masks: can be used to train ensemble over orderings/connections
42 | natural_ordering: force natural ordering of dimensions, don't use random permutations
43 | """
44 |
45 | super().__init__()
46 | self.ind_in = ind_in
47 | self.cond_in = cond_in
48 | self.nout = nout
49 | self.device = device
50 | self.hidden_sizes = hidden_sizes
51 |
52 | # define a simple MLP neural net
53 | self.net = []
54 | hs = [ind_in + cond_in] + hidden_sizes + [nout * ind_in]
55 | for h0, h1 in zip(hs, hs[1:]):
56 | self.net.extend([
57 | MaskedLinear(h0, h1),
58 | nn.ReLU(),
59 | ])
60 | self.net.pop() # pop the last ReLU for the output layer
61 | self.net = nn.Sequential(*self.net).to(device)
62 |
63 | self.m = {}
64 | self.update_masks() # builds the initial self.m connectivity
65 | # note, we could also precompute the masks and cache them, but this
66 | # could get memory expensive for large number of masks.
67 |
68 | def update_masks(self):
69 | L = len(self.hidden_sizes)
70 |
71 | # sample the order of the inputs and the connectivity of all neurons
72 | self.m[-1] = np.arange(self.ind_in + self.cond_in)
73 | for l in range(L):
74 | self.m[l] = rng.randint(self.m[l - 1].min(), self.nin - 1, size=self.hidden_sizes[l])
75 |
76 |
77 | # construct the mask matrices
78 | masks = [self.m[l - 1][:, None] <= self.m[l][None, :] for l in range(L)]
79 | masks.append(self.m[L - 1][:, None] < self.m[-1][None, :])
80 |
81 | # handle the case where nout = nin * k, for integer k > 1
82 | if self.nout > self.nin:
83 | k = int(self.nout / self.nin)
84 | # replicate the mask across the other outputs
85 | masks[-1] = np.concatenate([masks[-1]] * k, axis=1)
86 |
87 | # set the masks in all MaskedLinear layers
88 | layers = [l for l in self.net.modules() if isinstance(l, MaskedLinear)]
89 | for l, m in zip(layers, masks):
90 | l.set_mask(m)
91 |
92 | # map between in_d and order
93 | self.i_map = self.m[-1].copy()
94 | for k in range(len(self.m[-1])):
95 | self.i_map[self.m[-1][k]] = k
96 |
97 | def forward(self, x, context=None):
98 | if self.nout == 2:
99 | transf = self.net(x)
100 | mu, sigma = transf[:, :self.nin], transf[:, self.nin:]
101 | z = (x - mu) * torch.exp(-sigma)
102 | return z
103 | return self.net(x)
104 |
105 | def computeLL(self, x):
106 | # Jac and x of MADE
107 | transf = self.net(x)
108 | mu, sigma = transf[:, :self.nin], transf[:, self.nin:]
109 | z = (x - mu) * torch.exp(-sigma)
110 |
111 | log_prob_gauss = -.5 * (torch.log(self.pi * 2) + z ** 2).sum(1)
112 | ll = - sigma.sum(1) + log_prob_gauss
113 |
114 | return ll, z
115 |
116 | def invert(self, z):
117 | if self.nin != self.nout / 2:
118 | return None
119 |
120 | # We suppose a Gaussian MADE
121 | u = torch.zeros(z.shape)
122 | for d in range(self.nin):
123 | transf = self.forward(u)
124 | mu, sigma = transf[:, self.i_map[d]], transf[:, self.nin + self.i_map[d]]
125 | u[:, self.i_map[d]] = z[:, self.i_map[d]] * torch.exp(sigma) + mu
126 | return u
127 |
128 |
129 | # ------------------------------------------------------------------------------
130 |
131 |
132 | class ConditionnalMADE(MADE):
133 |
134 | def __init__(self, nin, cond_in, hidden_sizes, nout, num_masks=1, natural_ordering=False, random=False,
135 | device="cpu"):
136 | """
137 | nin: integer; number of inputs
138 | hidden sizes: a list of integers; number of units in hidden layers
139 | nout: integer; number of outputs, which usually collectively parameterize some kind of 1D distribution
140 | note: if nout is e.g. 2x larger than nin (perhaps the mean and std), then the first nin
141 | will be all the means and the second nin will be stds. i.e. output dimensions depend on the
142 | same input dimensions in "chunks" and should be carefully decoded downstream appropriately.
143 | the output of running the tests for this file makes this a bit more clear with examples.
144 | num_masks: can be used to train ensemble over orderings/connections
145 | natural_ordering: force natural ordering of dimensions, don't use random permutations
146 | """
147 |
148 | super().__init__(nin + cond_in, hidden_sizes, nout, num_masks, natural_ordering, random, device)
149 | self.nin_non_cond = nin
150 | self.cond_in = cond_in
151 |
152 | def forward(self, x, context):
153 | out = super().forward(torch.cat((context, x), 1))
154 | out = out.contiguous().view(x.shape[0], int(out.shape[1] / self.nin), self.nin)[:, :,
155 | self.cond_in:].contiguous().view(x.shape[0], -1)
156 | return out
157 |
158 | def computeLL(self, x, context):
159 | # Jac and x of MADE
160 | transf = self.net(torch.cat((context, x), 1))
161 | transf = transf.contiguous().view(x.shape[0], int(transf.shape[1] / self.nin), self.nin)[:, :,
162 | self.cond_in:].contiguous().view(x.shape[0], -1)
163 | mu, sigma = transf[:, :self.nin], transf[:, self.nin:]
164 | z = (x - mu) * torch.exp(-sigma)
165 |
166 | log_prob_gauss = -.5 * (torch.log(self.pi * 2) + z ** 2).sum(1)
167 | ll = - sigma.sum(1) + log_prob_gauss
168 |
169 | return ll, z
170 |
171 | def invert(self, z, context):
172 | if self.nin != self.nout / 2:
173 | return None
174 |
175 | # We suppose a Gaussian MADE
176 | u = torch.zeros(z.shape)
177 | for d in range(self.nin):
178 | transf = self.net(torch.cat((context, x), 1))
179 | mu, sigma = transf[:, self.i_map[d]], transf[:, self.nin + self.i_map[d]]
180 | u[:, self.i_map[d]] = z[:, self.i_map[d]] * torch.exp(sigma) + mu
181 | return u
182 |
183 |
184 | if __name__ == '__main__':
185 | from torch.autograd import Variable
186 |
187 | # run a quick and dirty test for the autoregressive property
188 | D = 10
189 | rng = np.random.RandomState(14)
190 | x = (rng.rand(1, D) > 0.5).astype(np.float32)
191 |
192 | configs = [
193 | (D, [], D, False), # test various hidden sizes
194 | (D, [200], D, False),
195 | (D, [200, 220], D, False),
196 | (D, [200, 220, 230], D, False),
197 | (D, [200, 220], D, True), # natural ordering test
198 | (D, [200, 220], 2 * D, True), # test nout > nin
199 | (D, [200, 220], 3 * D, False), # test nout > nin
200 | ]
201 |
202 | for nin, hiddens, nout, natural_ordering in configs:
203 |
204 | print("checking nin %d, hiddens %s, nout %d, natural %s" %
205 | (nin, hiddens, nout, natural_ordering))
206 | model = MADE(nin, hiddens, nout, natural_ordering=natural_ordering)
207 | z = torch.randn(1, nin)
208 | model.invert(z)
209 | continue
210 | # run backpropagation for each dimension to compute what other
211 | # dimensions it depends on.
212 | res = []
213 | for k in range(nout):
214 | xtr = Variable(torch.from_numpy(x), requires_grad=True)
215 | xtrhat = model(xtr)
216 | loss = xtrhat[0, k]
217 | loss.backward()
218 |
219 | depends = (xtr.grad[0].numpy() != 0).astype(np.uint8)
220 | depends_ix = list(np.where(depends)[0])
221 | isok = k % nin not in depends_ix
222 |
223 | res.append((len(depends_ix), k, depends_ix, isok))
224 |
225 | # pretty print the dependencies
226 | res.sort()
227 | for nl, k, ix, isok in res:
228 | print("output %2d depends on inputs: %30s : %s" % (k, ix, "OK" if isok else "NOTOK"))
229 |
230 |
--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
1 | GNU GENERAL PUBLIC LICENSE
2 | Version 3, 29 June 2007
3 |
4 | Copyright (C) 2007 Free Software Foundation, Inc.
5 | Everyone is permitted to copy and distribute verbatim copies
6 | of this license document, but changing it is not allowed.
7 |
8 | Preamble
9 |
10 | The GNU General Public License is a free, copyleft license for
11 | software and other kinds of works.
12 |
13 | The licenses for most software and other practical works are designed
14 | to take away your freedom to share and change the works. By contrast,
15 | the GNU General Public License is intended to guarantee your freedom to
16 | share and change all versions of a program--to make sure it remains free
17 | software for all its users. We, the Free Software Foundation, use the
18 | GNU General Public License for most of our software; it applies also to
19 | any other work released this way by its authors. You can apply it to
20 | your programs, too.
21 |
22 | When we speak of free software, we are referring to freedom, not
23 | price. Our General Public Licenses are designed to make sure that you
24 | have the freedom to distribute copies of free software (and charge for
25 | them if you wish), that you receive source code or can get it if you
26 | want it, that you can change the software or use pieces of it in new
27 | free programs, and that you know you can do these things.
28 |
29 | To protect your rights, we need to prevent others from denying you
30 | these rights or asking you to surrender the rights. Therefore, you have
31 | certain responsibilities if you distribute copies of the software, or if
32 | you modify it: responsibilities to respect the freedom of others.
33 |
34 | For example, if you distribute copies of such a program, whether
35 | gratis or for a fee, you must pass on to the recipients the same
36 | freedoms that you received. You must make sure that they, too, receive
37 | or can get the source code. And you must show them these terms so they
38 | know their rights.
39 |
40 | Developers that use the GNU GPL protect your rights with two steps:
41 | (1) assert copyright on the software, and (2) offer you this License
42 | giving you legal permission to copy, distribute and/or modify it.
43 |
44 | For the developers' and authors' protection, the GPL clearly explains
45 | that there is no warranty for this free software. For both users' and
46 | authors' sake, the GPL requires that modified versions be marked as
47 | changed, so that their problems will not be attributed erroneously to
48 | authors of previous versions.
49 |
50 | Some devices are designed to deny users access to install or run
51 | modified versions of the software inside them, although the manufacturer
52 | can do so. This is fundamentally incompatible with the aim of
53 | protecting users' freedom to change the software. The systematic
54 | pattern of such abuse occurs in the area of products for individuals to
55 | use, which is precisely where it is most unacceptable. Therefore, we
56 | have designed this version of the GPL to prohibit the practice for those
57 | products. If such problems arise substantially in other domains, we
58 | stand ready to extend this provision to those domains in future versions
59 | of the GPL, as needed to protect the freedom of users.
60 |
61 | Finally, every program is threatened constantly by software patents.
62 | States should not allow patents to restrict development and use of
63 | software on general-purpose computers, but in those that do, we wish to
64 | avoid the special danger that patents applied to a free program could
65 | make it effectively proprietary. To prevent this, the GPL assures that
66 | patents cannot be used to render the program non-free.
67 |
68 | The precise terms and conditions for copying, distribution and
69 | modification follow.
70 |
71 | TERMS AND CONDITIONS
72 |
73 | 0. Definitions.
74 |
75 | "This License" refers to version 3 of the GNU General Public License.
76 |
77 | "Copyright" also means copyright-like laws that apply to other kinds of
78 | works, such as semiconductor masks.
79 |
80 | "The Program" refers to any copyrightable work licensed under this
81 | License. Each licensee is addressed as "you". "Licensees" and
82 | "recipients" may be individuals or organizations.
83 |
84 | To "modify" a work means to copy from or adapt all or part of the work
85 | in a fashion requiring copyright permission, other than the making of an
86 | exact copy. The resulting work is called a "modified version" of the
87 | earlier work or a work "based on" the earlier work.
88 |
89 | A "covered work" means either the unmodified Program or a work based
90 | on the Program.
91 |
92 | To "propagate" a work means to do anything with it that, without
93 | permission, would make you directly or secondarily liable for
94 | infringement under applicable copyright law, except executing it on a
95 | computer or modifying a private copy. Propagation includes copying,
96 | distribution (with or without modification), making available to the
97 | public, and in some countries other activities as well.
98 |
99 | To "convey" a work means any kind of propagation that enables other
100 | parties to make or receive copies. Mere interaction with a user through
101 | a computer network, with no transfer of a copy, is not conveying.
102 |
103 | An interactive user interface displays "Appropriate Legal Notices"
104 | to the extent that it includes a convenient and prominently visible
105 | feature that (1) displays an appropriate copyright notice, and (2)
106 | tells the user that there is no warranty for the work (except to the
107 | extent that warranties are provided), that licensees may convey the
108 | work under this License, and how to view a copy of this License. If
109 | the interface presents a list of user commands or options, such as a
110 | menu, a prominent item in the list meets this criterion.
111 |
112 | 1. Source Code.
113 |
114 | The "source code" for a work means the preferred form of the work
115 | for making modifications to it. "Object code" means any non-source
116 | form of a work.
117 |
118 | A "Standard Interface" means an interface that either is an official
119 | standard defined by a recognized standards body, or, in the case of
120 | interfaces specified for a particular programming language, one that
121 | is widely used among developers working in that language.
122 |
123 | The "System Libraries" of an executable work include anything, other
124 | than the work as a whole, that (a) is included in the normal form of
125 | packaging a Major Component, but which is not part of that Major
126 | Component, and (b) serves only to enable use of the work with that
127 | Major Component, or to implement a Standard Interface for which an
128 | implementation is available to the public in source code form. A
129 | "Major Component", in this context, means a major essential component
130 | (kernel, window system, and so on) of the specific operating system
131 | (if any) on which the executable work runs, or a compiler used to
132 | produce the work, or an object code interpreter used to run it.
133 |
134 | The "Corresponding Source" for a work in object code form means all
135 | the source code needed to generate, install, and (for an executable
136 | work) run the object code and to modify the work, including scripts to
137 | control those activities. However, it does not include the work's
138 | System Libraries, or general-purpose tools or generally available free
139 | programs which are used unmodified in performing those activities but
140 | which are not part of the work. For example, Corresponding Source
141 | includes interface definition files associated with source files for
142 | the work, and the source code for shared libraries and dynamically
143 | linked subprograms that the work is specifically designed to require,
144 | such as by intimate data communication or control flow between those
145 | subprograms and other parts of the work.
146 |
147 | The Corresponding Source need not include anything that users
148 | can regenerate automatically from other parts of the Corresponding
149 | Source.
150 |
151 | The Corresponding Source for a work in source code form is that
152 | same work.
153 |
154 | 2. Basic Permissions.
155 |
156 | All rights granted under this License are granted for the term of
157 | copyright on the Program, and are irrevocable provided the stated
158 | conditions are met. This License explicitly affirms your unlimited
159 | permission to run the unmodified Program. The output from running a
160 | covered work is covered by this License only if the output, given its
161 | content, constitutes a covered work. This License acknowledges your
162 | rights of fair use or other equivalent, as provided by copyright law.
163 |
164 | You may make, run and propagate covered works that you do not
165 | convey, without conditions so long as your license otherwise remains
166 | in force. You may convey covered works to others for the sole purpose
167 | of having them make modifications exclusively for you, or provide you
168 | with facilities for running those works, provided that you comply with
169 | the terms of this License in conveying all material for which you do
170 | not control copyright. Those thus making or running the covered works
171 | for you must do so exclusively on your behalf, under your direction
172 | and control, on terms that prohibit them from making any copies of
173 | your copyrighted material outside their relationship with you.
174 |
175 | Conveying under any other circumstances is permitted solely under
176 | the conditions stated below. Sublicensing is not allowed; section 10
177 | makes it unnecessary.
178 |
179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180 |
181 | No covered work shall be deemed part of an effective technological
182 | measure under any applicable law fulfilling obligations under article
183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184 | similar laws prohibiting or restricting circumvention of such
185 | measures.
186 |
187 | When you convey a covered work, you waive any legal power to forbid
188 | circumvention of technological measures to the extent such circumvention
189 | is effected by exercising rights under this License with respect to
190 | the covered work, and you disclaim any intention to limit operation or
191 | modification of the work as a means of enforcing, against the work's
192 | users, your or third parties' legal rights to forbid circumvention of
193 | technological measures.
194 |
195 | 4. Conveying Verbatim Copies.
196 |
197 | You may convey verbatim copies of the Program's source code as you
198 | receive it, in any medium, provided that you conspicuously and
199 | appropriately publish on each copy an appropriate copyright notice;
200 | keep intact all notices stating that this License and any
201 | non-permissive terms added in accord with section 7 apply to the code;
202 | keep intact all notices of the absence of any warranty; and give all
203 | recipients a copy of this License along with the Program.
204 |
205 | You may charge any price or no price for each copy that you convey,
206 | and you may offer support or warranty protection for a fee.
207 |
208 | 5. Conveying Modified Source Versions.
209 |
210 | You may convey a work based on the Program, or the modifications to
211 | produce it from the Program, in the form of source code under the
212 | terms of section 4, provided that you also meet all of these conditions:
213 |
214 | a) The work must carry prominent notices stating that you modified
215 | it, and giving a relevant date.
216 |
217 | b) The work must carry prominent notices stating that it is
218 | released under this License and any conditions added under section
219 | 7. This requirement modifies the requirement in section 4 to
220 | "keep intact all notices".
221 |
222 | c) You must license the entire work, as a whole, under this
223 | License to anyone who comes into possession of a copy. This
224 | License will therefore apply, along with any applicable section 7
225 | additional terms, to the whole of the work, and all its parts,
226 | regardless of how they are packaged. This License gives no
227 | permission to license the work in any other way, but it does not
228 | invalidate such permission if you have separately received it.
229 |
230 | d) If the work has interactive user interfaces, each must display
231 | Appropriate Legal Notices; however, if the Program has interactive
232 | interfaces that do not display Appropriate Legal Notices, your
233 | work need not make them do so.
234 |
235 | A compilation of a covered work with other separate and independent
236 | works, which are not by their nature extensions of the covered work,
237 | and which are not combined with it such as to form a larger program,
238 | in or on a volume of a storage or distribution medium, is called an
239 | "aggregate" if the compilation and its resulting copyright are not
240 | used to limit the access or legal rights of the compilation's users
241 | beyond what the individual works permit. Inclusion of a covered work
242 | in an aggregate does not cause this License to apply to the other
243 | parts of the aggregate.
244 |
245 | 6. Conveying Non-Source Forms.
246 |
247 | You may convey a covered work in object code form under the terms
248 | of sections 4 and 5, provided that you also convey the
249 | machine-readable Corresponding Source under the terms of this License,
250 | in one of these ways:
251 |
252 | a) Convey the object code in, or embodied in, a physical product
253 | (including a physical distribution medium), accompanied by the
254 | Corresponding Source fixed on a durable physical medium
255 | customarily used for software interchange.
256 |
257 | b) Convey the object code in, or embodied in, a physical product
258 | (including a physical distribution medium), accompanied by a
259 | written offer, valid for at least three years and valid for as
260 | long as you offer spare parts or customer support for that product
261 | model, to give anyone who possesses the object code either (1) a
262 | copy of the Corresponding Source for all the software in the
263 | product that is covered by this License, on a durable physical
264 | medium customarily used for software interchange, for a price no
265 | more than your reasonable cost of physically performing this
266 | conveying of source, or (2) access to copy the
267 | Corresponding Source from a network server at no charge.
268 |
269 | c) Convey individual copies of the object code with a copy of the
270 | written offer to provide the Corresponding Source. This
271 | alternative is allowed only occasionally and noncommercially, and
272 | only if you received the object code with such an offer, in accord
273 | with subsection 6b.
274 |
275 | d) Convey the object code by offering access from a designated
276 | place (gratis or for a charge), and offer equivalent access to the
277 | Corresponding Source in the same way through the same place at no
278 | further charge. You need not require recipients to copy the
279 | Corresponding Source along with the object code. If the place to
280 | copy the object code is a network server, the Corresponding Source
281 | may be on a different server (operated by you or a third party)
282 | that supports equivalent copying facilities, provided you maintain
283 | clear directions next to the object code saying where to find the
284 | Corresponding Source. Regardless of what server hosts the
285 | Corresponding Source, you remain obligated to ensure that it is
286 | available for as long as needed to satisfy these requirements.
287 |
288 | e) Convey the object code using peer-to-peer transmission, provided
289 | you inform other peers where the object code and Corresponding
290 | Source of the work are being offered to the general public at no
291 | charge under subsection 6d.
292 |
293 | A separable portion of the object code, whose source code is excluded
294 | from the Corresponding Source as a System Library, need not be
295 | included in conveying the object code work.
296 |
297 | A "User Product" is either (1) a "consumer product", which means any
298 | tangible personal property which is normally used for personal, family,
299 | or household purposes, or (2) anything designed or sold for incorporation
300 | into a dwelling. In determining whether a product is a consumer product,
301 | doubtful cases shall be resolved in favor of coverage. For a particular
302 | product received by a particular user, "normally used" refers to a
303 | typical or common use of that class of product, regardless of the status
304 | of the particular user or of the way in which the particular user
305 | actually uses, or expects or is expected to use, the product. A product
306 | is a consumer product regardless of whether the product has substantial
307 | commercial, industrial or non-consumer uses, unless such uses represent
308 | the only significant mode of use of the product.
309 |
310 | "Installation Information" for a User Product means any methods,
311 | procedures, authorization keys, or other information required to install
312 | and execute modified versions of a covered work in that User Product from
313 | a modified version of its Corresponding Source. The information must
314 | suffice to ensure that the continued functioning of the modified object
315 | code is in no case prevented or interfered with solely because
316 | modification has been made.
317 |
318 | If you convey an object code work under this section in, or with, or
319 | specifically for use in, a User Product, and the conveying occurs as
320 | part of a transaction in which the right of possession and use of the
321 | User Product is transferred to the recipient in perpetuity or for a
322 | fixed term (regardless of how the transaction is characterized), the
323 | Corresponding Source conveyed under this section must be accompanied
324 | by the Installation Information. But this requirement does not apply
325 | if neither you nor any third party retains the ability to install
326 | modified object code on the User Product (for example, the work has
327 | been installed in ROM).
328 |
329 | The requirement to provide Installation Information does not include a
330 | requirement to continue to provide support service, warranty, or updates
331 | for a work that has been modified or installed by the recipient, or for
332 | the User Product in which it has been modified or installed. Access to a
333 | network may be denied when the modification itself materially and
334 | adversely affects the operation of the network or violates the rules and
335 | protocols for communication across the network.
336 |
337 | Corresponding Source conveyed, and Installation Information provided,
338 | in accord with this section must be in a format that is publicly
339 | documented (and with an implementation available to the public in
340 | source code form), and must require no special password or key for
341 | unpacking, reading or copying.
342 |
343 | 7. Additional Terms.
344 |
345 | "Additional permissions" are terms that supplement the terms of this
346 | License by making exceptions from one or more of its conditions.
347 | Additional permissions that are applicable to the entire Program shall
348 | be treated as though they were included in this License, to the extent
349 | that they are valid under applicable law. If additional permissions
350 | apply only to part of the Program, that part may be used separately
351 | under those permissions, but the entire Program remains governed by
352 | this License without regard to the additional permissions.
353 |
354 | When you convey a copy of a covered work, you may at your option
355 | remove any additional permissions from that copy, or from any part of
356 | it. (Additional permissions may be written to require their own
357 | removal in certain cases when you modify the work.) You may place
358 | additional permissions on material, added by you to a covered work,
359 | for which you have or can give appropriate copyright permission.
360 |
361 | Notwithstanding any other provision of this License, for material you
362 | add to a covered work, you may (if authorized by the copyright holders of
363 | that material) supplement the terms of this License with terms:
364 |
365 | a) Disclaiming warranty or limiting liability differently from the
366 | terms of sections 15 and 16 of this License; or
367 |
368 | b) Requiring preservation of specified reasonable legal notices or
369 | author attributions in that material or in the Appropriate Legal
370 | Notices displayed by works containing it; or
371 |
372 | c) Prohibiting misrepresentation of the origin of that material, or
373 | requiring that modified versions of such material be marked in
374 | reasonable ways as different from the original version; or
375 |
376 | d) Limiting the use for publicity purposes of names of licensors or
377 | authors of the material; or
378 |
379 | e) Declining to grant rights under trademark law for use of some
380 | trade names, trademarks, or service marks; or
381 |
382 | f) Requiring indemnification of licensors and authors of that
383 | material by anyone who conveys the material (or modified versions of
384 | it) with contractual assumptions of liability to the recipient, for
385 | any liability that these contractual assumptions directly impose on
386 | those licensors and authors.
387 |
388 | All other non-permissive additional terms are considered "further
389 | restrictions" within the meaning of section 10. If the Program as you
390 | received it, or any part of it, contains a notice stating that it is
391 | governed by this License along with a term that is a further
392 | restriction, you may remove that term. If a license document contains
393 | a further restriction but permits relicensing or conveying under this
394 | License, you may add to a covered work material governed by the terms
395 | of that license document, provided that the further restriction does
396 | not survive such relicensing or conveying.
397 |
398 | If you add terms to a covered work in accord with this section, you
399 | must place, in the relevant source files, a statement of the
400 | additional terms that apply to those files, or a notice indicating
401 | where to find the applicable terms.
402 |
403 | Additional terms, permissive or non-permissive, may be stated in the
404 | form of a separately written license, or stated as exceptions;
405 | the above requirements apply either way.
406 |
407 | 8. Termination.
408 |
409 | You may not propagate or modify a covered work except as expressly
410 | provided under this License. Any attempt otherwise to propagate or
411 | modify it is void, and will automatically terminate your rights under
412 | this License (including any patent licenses granted under the third
413 | paragraph of section 11).
414 |
415 | However, if you cease all violation of this License, then your
416 | license from a particular copyright holder is reinstated (a)
417 | provisionally, unless and until the copyright holder explicitly and
418 | finally terminates your license, and (b) permanently, if the copyright
419 | holder fails to notify you of the violation by some reasonable means
420 | prior to 60 days after the cessation.
421 |
422 | Moreover, your license from a particular copyright holder is
423 | reinstated permanently if the copyright holder notifies you of the
424 | violation by some reasonable means, this is the first time you have
425 | received notice of violation of this License (for any work) from that
426 | copyright holder, and you cure the violation prior to 30 days after
427 | your receipt of the notice.
428 |
429 | Termination of your rights under this section does not terminate the
430 | licenses of parties who have received copies or rights from you under
431 | this License. If your rights have been terminated and not permanently
432 | reinstated, you do not qualify to receive new licenses for the same
433 | material under section 10.
434 |
435 | 9. Acceptance Not Required for Having Copies.
436 |
437 | You are not required to accept this License in order to receive or
438 | run a copy of the Program. Ancillary propagation of a covered work
439 | occurring solely as a consequence of using peer-to-peer transmission
440 | to receive a copy likewise does not require acceptance. However,
441 | nothing other than this License grants you permission to propagate or
442 | modify any covered work. These actions infringe copyright if you do
443 | not accept this License. Therefore, by modifying or propagating a
444 | covered work, you indicate your acceptance of this License to do so.
445 |
446 | 10. Automatic Licensing of Downstream Recipients.
447 |
448 | Each time you convey a covered work, the recipient automatically
449 | receives a license from the original licensors, to run, modify and
450 | propagate that work, subject to this License. You are not responsible
451 | for enforcing compliance by third parties with this License.
452 |
453 | An "entity transaction" is a transaction transferring control of an
454 | organization, or substantially all assets of one, or subdividing an
455 | organization, or merging organizations. If propagation of a covered
456 | work results from an entity transaction, each party to that
457 | transaction who receives a copy of the work also receives whatever
458 | licenses to the work the party's predecessor in interest had or could
459 | give under the previous paragraph, plus a right to possession of the
460 | Corresponding Source of the work from the predecessor in interest, if
461 | the predecessor has it or can get it with reasonable efforts.
462 |
463 | You may not impose any further restrictions on the exercise of the
464 | rights granted or affirmed under this License. For example, you may
465 | not impose a license fee, royalty, or other charge for exercise of
466 | rights granted under this License, and you may not initiate litigation
467 | (including a cross-claim or counterclaim in a lawsuit) alleging that
468 | any patent claim is infringed by making, using, selling, offering for
469 | sale, or importing the Program or any portion of it.
470 |
471 | 11. Patents.
472 |
473 | A "contributor" is a copyright holder who authorizes use under this
474 | License of the Program or a work on which the Program is based. The
475 | work thus licensed is called the contributor's "contributor version".
476 |
477 | A contributor's "essential patent claims" are all patent claims
478 | owned or controlled by the contributor, whether already acquired or
479 | hereafter acquired, that would be infringed by some manner, permitted
480 | by this License, of making, using, or selling its contributor version,
481 | but do not include claims that would be infringed only as a
482 | consequence of further modification of the contributor version. For
483 | purposes of this definition, "control" includes the right to grant
484 | patent sublicenses in a manner consistent with the requirements of
485 | this License.
486 |
487 | Each contributor grants you a non-exclusive, worldwide, royalty-free
488 | patent license under the contributor's essential patent claims, to
489 | make, use, sell, offer for sale, import and otherwise run, modify and
490 | propagate the contents of its contributor version.
491 |
492 | In the following three paragraphs, a "patent license" is any express
493 | agreement or commitment, however denominated, not to enforce a patent
494 | (such as an express permission to practice a patent or covenant not to
495 | sue for patent infringement). To "grant" such a patent license to a
496 | party means to make such an agreement or commitment not to enforce a
497 | patent against the party.
498 |
499 | If you convey a covered work, knowingly relying on a patent license,
500 | and the Corresponding Source of the work is not available for anyone
501 | to copy, free of charge and under the terms of this License, through a
502 | publicly available network server or other readily accessible means,
503 | then you must either (1) cause the Corresponding Source to be so
504 | available, or (2) arrange to deprive yourself of the benefit of the
505 | patent license for this particular work, or (3) arrange, in a manner
506 | consistent with the requirements of this License, to extend the patent
507 | license to downstream recipients. "Knowingly relying" means you have
508 | actual knowledge that, but for the patent license, your conveying the
509 | covered work in a country, or your recipient's use of the covered work
510 | in a country, would infringe one or more identifiable patents in that
511 | country that you have reason to believe are valid.
512 |
513 | If, pursuant to or in connection with a single transaction or
514 | arrangement, you convey, or propagate by procuring conveyance of, a
515 | covered work, and grant a patent license to some of the parties
516 | receiving the covered work authorizing them to use, propagate, modify
517 | or convey a specific copy of the covered work, then the patent license
518 | you grant is automatically extended to all recipients of the covered
519 | work and works based on it.
520 |
521 | A patent license is "discriminatory" if it does not include within
522 | the scope of its coverage, prohibits the exercise of, or is
523 | conditioned on the non-exercise of one or more of the rights that are
524 | specifically granted under this License. You may not convey a covered
525 | work if you are a party to an arrangement with a third party that is
526 | in the business of distributing software, under which you make payment
527 | to the third party based on the extent of your activity of conveying
528 | the work, and under which the third party grants, to any of the
529 | parties who would receive the covered work from you, a discriminatory
530 | patent license (a) in connection with copies of the covered work
531 | conveyed by you (or copies made from those copies), or (b) primarily
532 | for and in connection with specific products or compilations that
533 | contain the covered work, unless you entered into that arrangement,
534 | or that patent license was granted, prior to 28 March 2007.
535 |
536 | Nothing in this License shall be construed as excluding or limiting
537 | any implied license or other defenses to infringement that may
538 | otherwise be available to you under applicable patent law.
539 |
540 | 12. No Surrender of Others' Freedom.
541 |
542 | If conditions are imposed on you (whether by court order, agreement or
543 | otherwise) that contradict the conditions of this License, they do not
544 | excuse you from the conditions of this License. If you cannot convey a
545 | covered work so as to satisfy simultaneously your obligations under this
546 | License and any other pertinent obligations, then as a consequence you may
547 | not convey it at all. For example, if you agree to terms that obligate you
548 | to collect a royalty for further conveying from those to whom you convey
549 | the Program, the only way you could satisfy both those terms and this
550 | License would be to refrain entirely from conveying the Program.
551 |
552 | 13. Use with the GNU Affero General Public License.
553 |
554 | Notwithstanding any other provision of this License, you have
555 | permission to link or combine any covered work with a work licensed
556 | under version 3 of the GNU Affero General Public License into a single
557 | combined work, and to convey the resulting work. The terms of this
558 | License will continue to apply to the part which is the covered work,
559 | but the special requirements of the GNU Affero General Public License,
560 | section 13, concerning interaction through a network will apply to the
561 | combination as such.
562 |
563 | 14. Revised Versions of this License.
564 |
565 | The Free Software Foundation may publish revised and/or new versions of
566 | the GNU General Public License from time to time. Such new versions will
567 | be similar in spirit to the present version, but may differ in detail to
568 | address new problems or concerns.
569 |
570 | Each version is given a distinguishing version number. If the
571 | Program specifies that a certain numbered version of the GNU General
572 | Public License "or any later version" applies to it, you have the
573 | option of following the terms and conditions either of that numbered
574 | version or of any later version published by the Free Software
575 | Foundation. If the Program does not specify a version number of the
576 | GNU General Public License, you may choose any version ever published
577 | by the Free Software Foundation.
578 |
579 | If the Program specifies that a proxy can decide which future
580 | versions of the GNU General Public License can be used, that proxy's
581 | public statement of acceptance of a version permanently authorizes you
582 | to choose that version for the Program.
583 |
584 | Later license versions may give you additional or different
585 | permissions. However, no additional obligations are imposed on any
586 | author or copyright holder as a result of your choosing to follow a
587 | later version.
588 |
589 | 15. Disclaimer of Warranty.
590 |
591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599 |
600 | 16. Limitation of Liability.
601 |
602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610 | SUCH DAMAGES.
611 |
612 | 17. Interpretation of Sections 15 and 16.
613 |
614 | If the disclaimer of warranty and limitation of liability provided
615 | above cannot be given local legal effect according to their terms,
616 | reviewing courts shall apply local law that most closely approximates
617 | an absolute waiver of all civil liability in connection with the
618 | Program, unless a warranty or assumption of liability accompanies a
619 | copy of the Program in return for a fee.
620 |
621 | END OF TERMS AND CONDITIONS
622 |
623 | How to Apply These Terms to Your New Programs
624 |
625 | If you develop a new program, and you want it to be of the greatest
626 | possible use to the public, the best way to achieve this is to make it
627 | free software which everyone can redistribute and change under these terms.
628 |
629 | To do so, attach the following notices to the program. It is safest
630 | to attach them to the start of each source file to most effectively
631 | state the exclusion of warranty; and each file should have at least
632 | the "copyright" line and a pointer to where the full notice is found.
633 |
634 |
635 | Copyright (C)
636 |
637 | This program is free software: you can redistribute it and/or modify
638 | it under the terms of the GNU General Public License as published by
639 | the Free Software Foundation, either version 3 of the License, or
640 | (at your option) any later version.
641 |
642 | This program is distributed in the hope that it will be useful,
643 | but WITHOUT ANY WARRANTY; without even the implied warranty of
644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
645 | GNU General Public License for more details.
646 |
647 | You should have received a copy of the GNU General Public License
648 | along with this program. If not, see .
649 |
650 | Also add information on how to contact you by electronic and paper mail.
651 |
652 | If the program does terminal interaction, make it output a short
653 | notice like this when it starts in an interactive mode:
654 |
655 | Copyright (C)
656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657 | This is free software, and you are welcome to redistribute it
658 | under certain conditions; type `show c' for details.
659 |
660 | The hypothetical commands `show w' and `show c' should show the appropriate
661 | parts of the General Public License. Of course, your program's commands
662 | might be different; for a GUI interface, you would use an "about box".
663 |
664 | You should also get your employer (if you work as a programmer) or school,
665 | if any, to sign a "copyright disclaimer" for the program, if necessary.
666 | For more information on this, and how to apply and follow the GNU GPL, see
667 | .
668 |
669 | The GNU General Public License does not permit incorporating your program
670 | into proprietary programs. If your program is a subroutine library, you
671 | may consider it more useful to permit linking proprietary applications with
672 | the library. If this is what you want to do, use the GNU Lesser General
673 | Public License instead of this License. But first, please read
674 | .
675 |
--------------------------------------------------------------------------------
/PlayGround.ipynb:
--------------------------------------------------------------------------------
1 | {
2 | "cells": [
3 | {
4 | "cell_type": "code",
5 | "execution_count": 144,
6 | "metadata": {},
7 | "outputs": [],
8 | "source": [
9 | "import pandas as pd"
10 | ]
11 | },
12 | {
13 | "cell_type": "code",
14 | "execution_count": 145,
15 | "metadata": {},
16 | "outputs": [],
17 | "source": [
18 | "CSV_COLUMNS = [\n",
19 | " \"age\", \"workclass\", \"fnlwgt\", \"education\", \"education_num\",\n",
20 | " \"marital_status\", \"occupation\", \"relationship\", \"race\", \"gender\",\n",
21 | " \"capital_gain\", \"capital_loss\", \"hours_per_week\", \"native_country\",\n",
22 | " \"income_bracket\"\n",
23 | "]\n",
24 | "x_columns = [\n",
25 | " \"age\", \"workclass\", \"education\", \"education_num\",\n",
26 | " \"marital_status\", \"occupation\", \"relationship\", \"race\", \"gender\",\n",
27 | " \"capital_gain\", \"capital_loss\", \"hours_per_week\", \"native_country\"\n",
28 | "]"
29 | ]
30 | },
31 | {
32 | "cell_type": "code",
33 | "execution_count": 156,
34 | "metadata": {},
35 | "outputs": [
36 | {
37 | "data": {
38 | "text/html": [
39 | "\n",
40 | "\n",
53 | "
\n",
54 | " \n",
55 | " \n",
56 | " | \n",
57 | " age | \n",
58 | " workclass | \n",
59 | " education | \n",
60 | " education_num | \n",
61 | " marital_status | \n",
62 | " occupation | \n",
63 | " relationship | \n",
64 | " race | \n",
65 | " gender | \n",
66 | " capital_gain | \n",
67 | " capital_loss | \n",
68 | " hours_per_week | \n",
69 | " native_country | \n",
70 | "
\n",
71 | " \n",
72 | " \n",
73 | " \n",
74 | " | 1 | \n",
75 | " 50 | \n",
76 | " Self-emp-not-inc | \n",
77 | " Bachelors | \n",
78 | " 13 | \n",
79 | " Married-civ-spouse | \n",
80 | " Exec-managerial | \n",
81 | " Husband | \n",
82 | " White | \n",
83 | " Male | \n",
84 | " 0 | \n",
85 | " 0 | \n",
86 | " 13 | \n",
87 | " United-States | \n",
88 | "
\n",
89 | " \n",
90 | " | 2 | \n",
91 | " 38 | \n",
92 | " Private | \n",
93 | " HS-grad | \n",
94 | " 9 | \n",
95 | " Divorced | \n",
96 | " Handlers-cleaners | \n",
97 | " Not-in-family | \n",
98 | " White | \n",
99 | " Male | \n",
100 | " 0 | \n",
101 | " 0 | \n",
102 | " 40 | \n",
103 | " United-States | \n",
104 | "
\n",
105 | " \n",
106 | " | 3 | \n",
107 | " 53 | \n",
108 | " Private | \n",
109 | " 11th | \n",
110 | " 7 | \n",
111 | " Married-civ-spouse | \n",
112 | " Handlers-cleaners | \n",
113 | " Husband | \n",
114 | " Black | \n",
115 | " Male | \n",
116 | " 0 | \n",
117 | " 0 | \n",
118 | " 40 | \n",
119 | " United-States | \n",
120 | "
\n",
121 | " \n",
122 | " | 4 | \n",
123 | " 28 | \n",
124 | " Private | \n",
125 | " Bachelors | \n",
126 | " 13 | \n",
127 | " Married-civ-spouse | \n",
128 | " Prof-specialty | \n",
129 | " Wife | \n",
130 | " Black | \n",
131 | " Female | \n",
132 | " 0 | \n",
133 | " 0 | \n",
134 | " 40 | \n",
135 | " Cuba | \n",
136 | "
\n",
137 | " \n",
138 | " | 5 | \n",
139 | " 37 | \n",
140 | " Private | \n",
141 | " Masters | \n",
142 | " 14 | \n",
143 | " Married-civ-spouse | \n",
144 | " Exec-managerial | \n",
145 | " Wife | \n",
146 | " White | \n",
147 | " Female | \n",
148 | " 0 | \n",
149 | " 0 | \n",
150 | " 40 | \n",
151 | " United-States | \n",
152 | "
\n",
153 | " \n",
154 | "
\n",
155 | "
\n",
240 | "\n",
253 | "
\n",
254 | " \n",
255 | " \n",
256 | " | \n",
257 | " age | \n",
258 | " education_num | \n",
259 | " capital_gain | \n",
260 | " capital_loss | \n",
261 | " hours_per_week | \n",
262 | " ? | \n",
263 | " Federal-gov | \n",
264 | " Local-gov | \n",
265 | " Never-worked | \n",
266 | " Private | \n",
267 | " ... | \n",
268 | " 9th | \n",
269 | " Assoc-acdm | \n",
270 | " Assoc-voc | \n",
271 | " Bachelors | \n",
272 | " Doctorate | \n",
273 | " HS-grad | \n",
274 | " Masters | \n",
275 | " Preschool | \n",
276 | " Prof-school | \n",
277 | " Some-college | \n",
278 | "
\n",
279 | " \n",
280 | " \n",
281 | " \n",
282 | " | 1 | \n",
283 | " 50 | \n",
284 | " 13 | \n",
285 | " 0 | \n",
286 | " 0 | \n",
287 | " 13 | \n",
288 | " 0 | \n",
289 | " 0 | \n",
290 | " 0 | \n",
291 | " 0 | \n",
292 | " 0 | \n",
293 | " ... | \n",
294 | " 0 | \n",
295 | " 0 | \n",
296 | " 0 | \n",
297 | " 1 | \n",
298 | " 0 | \n",
299 | " 0 | \n",
300 | " 0 | \n",
301 | " 0 | \n",
302 | " 0 | \n",
303 | " 0 | \n",
304 | "
\n",
305 | " \n",
306 | " | 2 | \n",
307 | " 38 | \n",
308 | " 9 | \n",
309 | " 0 | \n",
310 | " 0 | \n",
311 | " 40 | \n",
312 | " 0 | \n",
313 | " 0 | \n",
314 | " 0 | \n",
315 | " 0 | \n",
316 | " 1 | \n",
317 | " ... | \n",
318 | " 0 | \n",
319 | " 0 | \n",
320 | " 0 | \n",
321 | " 0 | \n",
322 | " 0 | \n",
323 | " 1 | \n",
324 | " 0 | \n",
325 | " 0 | \n",
326 | " 0 | \n",
327 | " 0 | \n",
328 | "
\n",
329 | " \n",
330 | " | 3 | \n",
331 | " 53 | \n",
332 | " 7 | \n",
333 | " 0 | \n",
334 | " 0 | \n",
335 | " 40 | \n",
336 | " 0 | \n",
337 | " 0 | \n",
338 | " 0 | \n",
339 | " 0 | \n",
340 | " 1 | \n",
341 | " ... | \n",
342 | " 0 | \n",
343 | " 0 | \n",
344 | " 0 | \n",
345 | " 0 | \n",
346 | " 0 | \n",
347 | " 0 | \n",
348 | " 0 | \n",
349 | " 0 | \n",
350 | " 0 | \n",
351 | " 0 | \n",
352 | "
\n",
353 | " \n",
354 | " | 4 | \n",
355 | " 28 | \n",
356 | " 13 | \n",
357 | " 0 | \n",
358 | " 0 | \n",
359 | " 40 | \n",
360 | " 0 | \n",
361 | " 0 | \n",
362 | " 0 | \n",
363 | " 0 | \n",
364 | " 1 | \n",
365 | " ... | \n",
366 | " 0 | \n",
367 | " 0 | \n",
368 | " 0 | \n",
369 | " 1 | \n",
370 | " 0 | \n",
371 | " 0 | \n",
372 | " 0 | \n",
373 | " 0 | \n",
374 | " 0 | \n",
375 | " 0 | \n",
376 | "
\n",
377 | " \n",
378 | " | 5 | \n",
379 | " 37 | \n",
380 | " 14 | \n",
381 | " 0 | \n",
382 | " 0 | \n",
383 | " 40 | \n",
384 | " 0 | \n",
385 | " 0 | \n",
386 | " 0 | \n",
387 | " 0 | \n",
388 | " 1 | \n",
389 | " ... | \n",
390 | " 0 | \n",
391 | " 0 | \n",
392 | " 0 | \n",
393 | " 0 | \n",
394 | " 0 | \n",
395 | " 0 | \n",
396 | " 1 | \n",
397 | " 0 | \n",
398 | " 0 | \n",
399 | " 0 | \n",
400 | "
\n",
401 | " \n",
402 | " | ... | \n",
403 | " ... | \n",
404 | " ... | \n",
405 | " ... | \n",
406 | " ... | \n",
407 | " ... | \n",
408 | " ... | \n",
409 | " ... | \n",
410 | " ... | \n",
411 | " ... | \n",
412 | " ... | \n",
413 | " ... | \n",
414 | " ... | \n",
415 | " ... | \n",
416 | " ... | \n",
417 | " ... | \n",
418 | " ... | \n",
419 | " ... | \n",
420 | " ... | \n",
421 | " ... | \n",
422 | " ... | \n",
423 | " ... | \n",
424 | "
\n",
425 | " \n",
426 | " | 32556 | \n",
427 | " 27 | \n",
428 | " 12 | \n",
429 | " 0 | \n",
430 | " 0 | \n",
431 | " 38 | \n",
432 | " 0 | \n",
433 | " 0 | \n",
434 | " 0 | \n",
435 | " 0 | \n",
436 | " 1 | \n",
437 | " ... | \n",
438 | " 0 | \n",
439 | " 1 | \n",
440 | " 0 | \n",
441 | " 0 | \n",
442 | " 0 | \n",
443 | " 0 | \n",
444 | " 0 | \n",
445 | " 0 | \n",
446 | " 0 | \n",
447 | " 0 | \n",
448 | "
\n",
449 | " \n",
450 | " | 32557 | \n",
451 | " 40 | \n",
452 | " 9 | \n",
453 | " 0 | \n",
454 | " 0 | \n",
455 | " 40 | \n",
456 | " 0 | \n",
457 | " 0 | \n",
458 | " 0 | \n",
459 | " 0 | \n",
460 | " 1 | \n",
461 | " ... | \n",
462 | " 0 | \n",
463 | " 0 | \n",
464 | " 0 | \n",
465 | " 0 | \n",
466 | " 0 | \n",
467 | " 1 | \n",
468 | " 0 | \n",
469 | " 0 | \n",
470 | " 0 | \n",
471 | " 0 | \n",
472 | "
\n",
473 | " \n",
474 | " | 32558 | \n",
475 | " 58 | \n",
476 | " 9 | \n",
477 | " 0 | \n",
478 | " 0 | \n",
479 | " 40 | \n",
480 | " 0 | \n",
481 | " 0 | \n",
482 | " 0 | \n",
483 | " 0 | \n",
484 | " 1 | \n",
485 | " ... | \n",
486 | " 0 | \n",
487 | " 0 | \n",
488 | " 0 | \n",
489 | " 0 | \n",
490 | " 0 | \n",
491 | " 1 | \n",
492 | " 0 | \n",
493 | " 0 | \n",
494 | " 0 | \n",
495 | " 0 | \n",
496 | "
\n",
497 | " \n",
498 | " | 32559 | \n",
499 | " 22 | \n",
500 | " 9 | \n",
501 | " 0 | \n",
502 | " 0 | \n",
503 | " 20 | \n",
504 | " 0 | \n",
505 | " 0 | \n",
506 | " 0 | \n",
507 | " 0 | \n",
508 | " 1 | \n",
509 | " ... | \n",
510 | " 0 | \n",
511 | " 0 | \n",
512 | " 0 | \n",
513 | " 0 | \n",
514 | " 0 | \n",
515 | " 1 | \n",
516 | " 0 | \n",
517 | " 0 | \n",
518 | " 0 | \n",
519 | " 0 | \n",
520 | "
\n",
521 | " \n",
522 | " | 32560 | \n",
523 | " 52 | \n",
524 | " 9 | \n",
525 | " 15024 | \n",
526 | " 0 | \n",
527 | " 40 | \n",
528 | " 0 | \n",
529 | " 0 | \n",
530 | " 0 | \n",
531 | " 0 | \n",
532 | " 0 | \n",
533 | " ... | \n",
534 | " 0 | \n",
535 | " 0 | \n",
536 | " 0 | \n",
537 | " 0 | \n",
538 | " 0 | \n",
539 | " 1 | \n",
540 | " 0 | \n",
541 | " 0 | \n",
542 | " 0 | \n",
543 | " 0 | \n",
544 | "
\n",
545 | " \n",
546 | "
\n",
547 | "
32560 rows × 105 columns
\n",
548 | "
\n",
774 | "\n",
787 | "
\n",
788 | " \n",
789 | " \n",
790 | " | \n",
791 | " income_bracket | \n",
792 | "
\n",
793 | " \n",
794 | " \n",
795 | " \n",
796 | " | 0 | \n",
797 | " <=50K | \n",
798 | "
\n",
799 | " \n",
800 | " | 1 | \n",
801 | " <=50K | \n",
802 | "
\n",
803 | " \n",
804 | " | 2 | \n",
805 | " <=50K | \n",
806 | "
\n",
807 | " \n",
808 | " | 3 | \n",
809 | " <=50K | \n",
810 | "
\n",
811 | " \n",
812 | " | 4 | \n",
813 | " <=50K | \n",
814 | "
\n",
815 | " \n",
816 | " | ... | \n",
817 | " ... | \n",
818 | "
\n",
819 | " \n",
820 | " | 32556 | \n",
821 | " <=50K | \n",
822 | "
\n",
823 | " \n",
824 | " | 32557 | \n",
825 | " >50K | \n",
826 | "
\n",
827 | " \n",
828 | " | 32558 | \n",
829 | " <=50K | \n",
830 | "
\n",
831 | " \n",
832 | " | 32559 | \n",
833 | " <=50K | \n",
834 | "
\n",
835 | " \n",
836 | " | 32560 | \n",
837 | " >50K | \n",
838 | "
\n",
839 | " \n",
840 | "
\n",
841 | "
32561 rows × 1 columns
\n",
842 | "
\n",
921 | "\n",
934 | "
\n",
935 | " \n",
936 | " \n",
937 | " | \n",
938 | " age | \n",
939 | " education_num | \n",
940 | " capital_gain | \n",
941 | " capital_loss | \n",
942 | " hours_per_week | \n",
943 | " ? | \n",
944 | " Federal-gov | \n",
945 | " Local-gov | \n",
946 | " Never-worked | \n",
947 | " Private | \n",
948 | " ... | \n",
949 | " 9th | \n",
950 | " Assoc-acdm | \n",
951 | " Assoc-voc | \n",
952 | " Bachelors | \n",
953 | " Doctorate | \n",
954 | " HS-grad | \n",
955 | " Masters | \n",
956 | " Preschool | \n",
957 | " Prof-school | \n",
958 | " Some-college | \n",
959 | "
\n",
960 | " \n",
961 | " \n",
962 | " \n",
963 | " | 1 | \n",
964 | " 50 | \n",
965 | " 13 | \n",
966 | " 0 | \n",
967 | " 0 | \n",
968 | " 13 | \n",
969 | " 0 | \n",
970 | " 0 | \n",
971 | " 0 | \n",
972 | " 0 | \n",
973 | " 0 | \n",
974 | " ... | \n",
975 | " 0 | \n",
976 | " 0 | \n",
977 | " 0 | \n",
978 | " 1 | \n",
979 | " 0 | \n",
980 | " 0 | \n",
981 | " 0 | \n",
982 | " 0 | \n",
983 | " 0 | \n",
984 | " 0 | \n",
985 | "
\n",
986 | " \n",
987 | " | 2 | \n",
988 | " 38 | \n",
989 | " 9 | \n",
990 | " 0 | \n",
991 | " 0 | \n",
992 | " 40 | \n",
993 | " 0 | \n",
994 | " 0 | \n",
995 | " 0 | \n",
996 | " 0 | \n",
997 | " 1 | \n",
998 | " ... | \n",
999 | " 0 | \n",
1000 | " 0 | \n",
1001 | " 0 | \n",
1002 | " 0 | \n",
1003 | " 0 | \n",
1004 | " 1 | \n",
1005 | " 0 | \n",
1006 | " 0 | \n",
1007 | " 0 | \n",
1008 | " 0 | \n",
1009 | "
\n",
1010 | " \n",
1011 | " | 3 | \n",
1012 | " 53 | \n",
1013 | " 7 | \n",
1014 | " 0 | \n",
1015 | " 0 | \n",
1016 | " 40 | \n",
1017 | " 0 | \n",
1018 | " 0 | \n",
1019 | " 0 | \n",
1020 | " 0 | \n",
1021 | " 1 | \n",
1022 | " ... | \n",
1023 | " 0 | \n",
1024 | " 0 | \n",
1025 | " 0 | \n",
1026 | " 0 | \n",
1027 | " 0 | \n",
1028 | " 0 | \n",
1029 | " 0 | \n",
1030 | " 0 | \n",
1031 | " 0 | \n",
1032 | " 0 | \n",
1033 | "
\n",
1034 | " \n",
1035 | " | 4 | \n",
1036 | " 28 | \n",
1037 | " 13 | \n",
1038 | " 0 | \n",
1039 | " 0 | \n",
1040 | " 40 | \n",
1041 | " 0 | \n",
1042 | " 0 | \n",
1043 | " 0 | \n",
1044 | " 0 | \n",
1045 | " 1 | \n",
1046 | " ... | \n",
1047 | " 0 | \n",
1048 | " 0 | \n",
1049 | " 0 | \n",
1050 | " 1 | \n",
1051 | " 0 | \n",
1052 | " 0 | \n",
1053 | " 0 | \n",
1054 | " 0 | \n",
1055 | " 0 | \n",
1056 | " 0 | \n",
1057 | "
\n",
1058 | " \n",
1059 | " | 5 | \n",
1060 | " 37 | \n",
1061 | " 14 | \n",
1062 | " 0 | \n",
1063 | " 0 | \n",
1064 | " 40 | \n",
1065 | " 0 | \n",
1066 | " 0 | \n",
1067 | " 0 | \n",
1068 | " 0 | \n",
1069 | " 1 | \n",
1070 | " ... | \n",
1071 | " 0 | \n",
1072 | " 0 | \n",
1073 | " 0 | \n",
1074 | " 0 | \n",
1075 | " 0 | \n",
1076 | " 0 | \n",
1077 | " 1 | \n",
1078 | " 0 | \n",
1079 | " 0 | \n",
1080 | " 0 | \n",
1081 | "
\n",
1082 | " \n",
1083 | " | ... | \n",
1084 | " ... | \n",
1085 | " ... | \n",
1086 | " ... | \n",
1087 | " ... | \n",
1088 | " ... | \n",
1089 | " ... | \n",
1090 | " ... | \n",
1091 | " ... | \n",
1092 | " ... | \n",
1093 | " ... | \n",
1094 | " ... | \n",
1095 | " ... | \n",
1096 | " ... | \n",
1097 | " ... | \n",
1098 | " ... | \n",
1099 | " ... | \n",
1100 | " ... | \n",
1101 | " ... | \n",
1102 | " ... | \n",
1103 | " ... | \n",
1104 | " ... | \n",
1105 | "
\n",
1106 | " \n",
1107 | " | 32556 | \n",
1108 | " 27 | \n",
1109 | " 12 | \n",
1110 | " 0 | \n",
1111 | " 0 | \n",
1112 | " 38 | \n",
1113 | " 0 | \n",
1114 | " 0 | \n",
1115 | " 0 | \n",
1116 | " 0 | \n",
1117 | " 1 | \n",
1118 | " ... | \n",
1119 | " 0 | \n",
1120 | " 1 | \n",
1121 | " 0 | \n",
1122 | " 0 | \n",
1123 | " 0 | \n",
1124 | " 0 | \n",
1125 | " 0 | \n",
1126 | " 0 | \n",
1127 | " 0 | \n",
1128 | " 0 | \n",
1129 | "
\n",
1130 | " \n",
1131 | " | 32557 | \n",
1132 | " 40 | \n",
1133 | " 9 | \n",
1134 | " 0 | \n",
1135 | " 0 | \n",
1136 | " 40 | \n",
1137 | " 0 | \n",
1138 | " 0 | \n",
1139 | " 0 | \n",
1140 | " 0 | \n",
1141 | " 1 | \n",
1142 | " ... | \n",
1143 | " 0 | \n",
1144 | " 0 | \n",
1145 | " 0 | \n",
1146 | " 0 | \n",
1147 | " 0 | \n",
1148 | " 1 | \n",
1149 | " 0 | \n",
1150 | " 0 | \n",
1151 | " 0 | \n",
1152 | " 0 | \n",
1153 | "
\n",
1154 | " \n",
1155 | " | 32558 | \n",
1156 | " 58 | \n",
1157 | " 9 | \n",
1158 | " 0 | \n",
1159 | " 0 | \n",
1160 | " 40 | \n",
1161 | " 0 | \n",
1162 | " 0 | \n",
1163 | " 0 | \n",
1164 | " 0 | \n",
1165 | " 1 | \n",
1166 | " ... | \n",
1167 | " 0 | \n",
1168 | " 0 | \n",
1169 | " 0 | \n",
1170 | " 0 | \n",
1171 | " 0 | \n",
1172 | " 1 | \n",
1173 | " 0 | \n",
1174 | " 0 | \n",
1175 | " 0 | \n",
1176 | " 0 | \n",
1177 | "
\n",
1178 | " \n",
1179 | " | 32559 | \n",
1180 | " 22 | \n",
1181 | " 9 | \n",
1182 | " 0 | \n",
1183 | " 0 | \n",
1184 | " 20 | \n",
1185 | " 0 | \n",
1186 | " 0 | \n",
1187 | " 0 | \n",
1188 | " 0 | \n",
1189 | " 1 | \n",
1190 | " ... | \n",
1191 | " 0 | \n",
1192 | " 0 | \n",
1193 | " 0 | \n",
1194 | " 0 | \n",
1195 | " 0 | \n",
1196 | " 1 | \n",
1197 | " 0 | \n",
1198 | " 0 | \n",
1199 | " 0 | \n",
1200 | " 0 | \n",
1201 | "
\n",
1202 | " \n",
1203 | " | 32560 | \n",
1204 | " 52 | \n",
1205 | " 9 | \n",
1206 | " 15024 | \n",
1207 | " 0 | \n",
1208 | " 40 | \n",
1209 | " 0 | \n",
1210 | " 0 | \n",
1211 | " 0 | \n",
1212 | " 0 | \n",
1213 | " 0 | \n",
1214 | " ... | \n",
1215 | " 0 | \n",
1216 | " 0 | \n",
1217 | " 0 | \n",
1218 | " 0 | \n",
1219 | " 0 | \n",
1220 | " 1 | \n",
1221 | " 0 | \n",
1222 | " 0 | \n",
1223 | " 0 | \n",
1224 | " 0 | \n",
1225 | "
\n",
1226 | " \n",
1227 | "
\n",
1228 | "
32560 rows × 105 columns
\n",
1229 | "