-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
311 lines (259 loc) · 11.2 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
import torch
from torch.distributions.multivariate_normal import MultivariateNormal
from matplotlib import pyplot as plt
import regressor
import torch.nn.functional as F
from arl import ARL
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
from cqr import helper
from nonconformist.nc import RegressorNc
from nonconformist.nc import QuantileRegErrFunc
import numpy as np
import random
def seed_everything(seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Use your desired seed here
#import argparse and use it
import argparse
parser = argparse.ArgumentParser(description='Process some integers.')
parser.add_argument('--seed', type=int, default=6)
args = parser.parse_args()
seed_everything(args.seed)
def sample_from_2d_gaussian(mean, cov,num_samples=1):
mv_normal = MultivariateNormal(mean, cov)
sample = mv_normal.sample((num_samples,))
return sample
def freeze_model(model):
for param in model.parameters():
param.requires_grad = False
def unfreeze_model(model):
for param in model.parameters():
param.requires_grad = True
def create_icp(n_train,x_train,y_train):
# divide the data into proper training set and calibration set
idx = np.random.permutation(n_train)
n_half = int(np.floor(n_train/2))
idx_train, idx_cal = idx[:n_half], idx[n_half:2*n_half]
in_shape = x_train.shape[1]
alpha = 0.1
# zero mean and unit variance scaling
scalerX = StandardScaler()
scalerX = scalerX.fit(x_train[idx_train])
# scale
x_train = scalerX.transform(x_train)
# scale the labels by dividing each by the mean absolute response
mean_y_train = np.mean(np.abs(y_train[idx_train]))
y_train = np.squeeze(y_train)/mean_y_train
#########################################################
# Quantile random forests parameters
# (See QuantileForestRegressorAdapter class in helper.py)
#########################################################
# when tuning the two QRF quantile levels one may
# ask for a prediction band with smaller average coverage
# to avoid too conservative estimation of the prediction band
# This would be equal to coverage_factor*(quantiles[1] - quantiles[0])
# coverage_factor = 0.85
# ratio of held-out data, used in cross-validation
cv_test_ratio = 0.05
# seed for splitting the data in cross-validation.
# Also used as the seed in quantile random forests function
cv_random_state = 1
# determines the lowest and highest quantile level parameters.
# This is used when tuning the quanitle levels by cross-validation.
# The smallest value is equal to quantiles[0] - range_vals.
# Similarly, the largest value is equal to quantiles[1] + range_vals.
# cv_range_vals = 30
# sweep over a grid of length num_vals when tuning QRF's quantile parameters
# cv_num_vals = 10
# pytorch's optimizer object
nn_learn_func = torch.optim.Adam
# number of epochs
epochs = 50
# learning rate
lr = 0.0005
# mini-batch size
batch_size = 64
# hidden dimension of the network
hidden_size = 16
# dropout regularization rate
dropout = 0.1
# weight decay regularization
wd = 1e-6
# Ask for a reduced coverage when tuning the network parameters by
# cross-validataion to avoid too concervative initial estimation of the
# prediction interval. This estimation will be conformalized by CQR.
quantiles_net = [0.05, 0.95]
# define quantile neural network model
quantile_estimator = helper.AllQNet_RegressorAdapter(model=None,
fit_params=None,
in_shape=in_shape,
hidden_size=hidden_size,
quantiles=quantiles_net,
learn_func=nn_learn_func,
epochs=epochs,
batch_size=batch_size,
dropout=dropout,
lr=lr,
wd=wd,
test_ratio=cv_test_ratio,
random_state=cv_random_state,
use_rearrangement=False)
# define a CQR object, computes the absolute residual error of points
# located outside the estimated quantile neural network band
nc = RegressorNc(quantile_estimator, QuantileRegErrFunc())
# run CQR procedure
icp = helper.train_icp(nc, x_train, y_train, idx_train, idx_cal, alpha)
# print(icp)
return icp
# Example usage
centroids = torch.tensor([(1,1),(1,1),(1,1)]).float()
cov = torch.tensor([[[2.0, 1], [1, 2.0]],[[2.0, 1], [1, 2.0]],[[100,10],[10,30]]])
n_samples = [10000,10000,200]
samples = []
target = []
for i,centroid in enumerate(centroids):
num_samples = n_samples[i]
samples_cur = sample_from_2d_gaussian(centroid, cov[i], num_samples)
samples = samples + [samples_cur]
target += [centroid.unsqueeze(0)]*num_samples
total_targets = torch.cat(target, dim=0)
total_samples = torch.cat(samples, dim=0)
from sklearn.model_selection import train_test_split
# Concatenate the targets and samples into one tensor
data = torch.cat((total_samples, total_targets), dim=1)
# Convert to numpy for splitting
data_np = data.numpy()
# Perform an 80-10 split first to separate the training set
train_val_data, test_data = train_test_split(data_np, test_size=0.1, random_state=42)
# Then split the training set further into training and validation sets (80-10-10 split overall)
train_data, val_data = train_test_split(train_val_data, test_size=1/9, random_state=42)
# Convert back to torch.Tensor
train_data = torch.tensor(train_data, dtype=torch.float32)
val_data = torch.tensor(val_data, dtype=torch.float32)
test_data = torch.tensor(test_data, dtype=torch.float32)
# Separate the features (samples) from the targets
X_train, y_train = train_data[:, :2], train_data[:, 2:]
X_val, y_val = val_data[:, :2], val_data[:, 2:]
X_test, y_test = test_data[:, :2], test_data[:, 2:]
icp = create_icp(X_train.shape[0],X_train.cpu().numpy(),y_train.cpu().numpy())
# Now you can convert your data to PyTorch DataLoader, which makes it easier to handle mini-batches
from torch.utils.data import TensorDataset, DataLoader
batch_size = 256 # Adjust based on your needs
train_dataset = TensorDataset(X_train, y_train)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True,drop_last=True)
val_dataset = TensorDataset(X_val, y_val)
val_loader = DataLoader(val_dataset, batch_size=batch_size)
test_dataset = TensorDataset(X_test, y_test)
test_loader = DataLoader(test_dataset, batch_size=batch_size)
# Initialize model and optimizer
model = ARL()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Transfer model to correct device.
model = model.to(device)
# Adagrad is the defeault optimizer.
optimizer_learner = torch.optim.Adagrad(
model.learner.parameters(), lr=1e-1
)
optimizer_adv = torch.optim.Adagrad(
model.adversary.parameters(), lr=1e-1
)
batch_losses = []
epoch_losses = []
val_epoch_losses = []
total_steps = 0
from tqdm import tqdm
for epoch in tqdm(range(30)):
for step, (train_point, train_target) in enumerate(train_loader):
# Transfer data to GPU if possible.
train_point = train_point.to(device)
train_target = train_target.to(device)
total_steps += 1
out = model.learner_step(train_point)
adv_weights = model.adversary_step(out).squeeze()
conformal_margins = torch.from_numpy(icp.predict(train_point.cpu().numpy(),significance=0.05)).to(device="cuda")
lower_bound = conformal_margins[:,:2]
upper_bound = conformal_margins[:,2:]
certainty = torch.mean((upper_bound - lower_bound)**2,dim=-1)
loss_learner = (((out-train_target)**2).sum(dim=-1)*adv_weights/certainty).sum()
optimizer_learner.zero_grad()
optimizer_adv.zero_grad()
# freeze_model(model.adversary)
# unfreeze_model(model.learner)
loss_learner.backward(retain_graph=True)
optimizer_learner.step()
# if epoch>=1:
loss_adv = -(((out-train_target)**2).sum(dim=-1)*adv_weights).sum()
# freeze_model(model.learner)
# unfreeze_model(model.adversary)
loss_adv.backward()
optimizer_adv.step()
# if epoch>=2:
# print(loss_learner)
# Learner update step.
# batch_losses.append(loss.item())
# Average loss for this epoch
# avg_loss = sum(batch_losses[-len(train_loader):]) / len(train_loader)
# epoch_losses.append(avg_loss)
# # Calculate validation loss
# with torch.no_grad():
# val_losses = [F.mse_loss(regressor(X_val), y_val) for X_val, y_val in val_loader]
# avg_val_loss = sum(val_losses) / len(val_losses)
# val_epoch_losses.append(avg_val_loss)
# Calculate test loss
out = []
errors = []
dict_loss = {}
with torch.no_grad():
for batch in test_loader:
X_test, y_test = batch
X_test, y_test = X_test.to(device="cuda"), y_test.to(device="cuda")
# dict_loss[y_test.cpu().numpy()] = []
yhat = model.learner(X_test)
error = torch.sqrt(((yhat-y_test)**2).sum(dim=-1))
for i,centroid in enumerate(y_test):
centroid = centroid.cpu().numpy()
if (centroid[0],centroid[1]) not in dict_loss:
dict_loss[(centroid[0],centroid[1])] = []
dict_loss[(centroid[0],centroid[1])]+= [error[i].item()]
out += [yhat]
errors += [error]
out = torch.cat(out, dim=0).cpu().numpy()
errors = torch.cat(errors, dim=0).cpu().numpy() # collect errors
for key in dict_loss:
dict_loss[key] = np.mean(dict_loss[key])
print(f"{errors.mean()}",end='')
x = out[:,0]
y = out[:,1]
plt.scatter(x, y, c=errors, cmap='viridis') # use errors for color
plt.colorbar(label='Error')
plt.savefig("test_arl_cqr.png")
# test_losses = [F.mse_loss(regressor(X_test.to(device="cuda")), y_test.to(device="cuda")).cpu() for X_test, y_test in test_loader]
# avg_test_loss = sum(test_losses) / len(test_losses)
# print("Average test loss: ", avg_test_loss)
# # Plot the batch losses
# plt.figure(figsize=(18, 6))
# plt.subplot(1, 3, 1)
# plt.plot(batch_losses)
# plt.title("Batch Loss over All Batches")
# plt.xlabel("Batch")
# plt.ylabel("Loss")
# # Plot the average epoch losses
# plt.subplot(1, 3, 2)
# plt.plot(epoch_losses)
# plt.title("Average Training Loss over Epochs")
# plt.xlabel("Epoch")
# plt.ylabel("Average Loss")
# # Plot the average validation losses
# plt.subplot(1, 3, 3)
# plt.plot(val_epoch_losses)
# plt.title("Average Validation Loss over Epochs")
# plt.xlabel("Epoch")
# plt.ylabel("Average Loss")
# plt.tight_layout()
# plt.savefig("losses.png")