File size: 12,764 Bytes
4d16332 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 |
from torch import nn
import torch
from root_gnn_base import utils
import numpy as np
class MaskedLoss():
def __init__(self, mask = []):
self.mask = mask
def make_mask(self, targets):
mask = torch.ones_like(targets[:,0])
for m in self.mask:
if m['op'] == 'eq':
mask[targets[:,m['idx']] == m['val']] = 0
elif m['op'] == 'gt':
mask[targets[:,m['idx']] > m['val']] = 0
elif m['op'] == 'lt':
mask[targets[:,m['idx']] < m['val']] = 0
elif m['op'] == 'ge':
mask[targets[:,m['idx']] >= m['val']] = 0
elif m['op'] == 'le':
mask[targets[:,m['idx']] <= m['val']] = 0
elif m['op'] == 'ne':
mask[targets[:,m['idx']] != m['val']] = 0
else:
raise ValueError(f'Unknown mask op {m["op"]}')
return mask == 1
class MaskedL1Loss(MaskedLoss):
def __init__(self, mask = [], index = 0):
super().__init__(mask)
self.index = index
self.loss = nn.L1Loss()
def __call__(self, logits, targets):
mask = self.make_mask(targets)
return self.loss(logits[mask], targets[mask][:,self.index])
class BCEWithLogitsLoss():
def __init__(self, weight=None, reduction='mean'):
self.loss = nn.BCEWithLogitsLoss(weight=weight, reduction=reduction)
def __call__(self, logits, targets):
return self.loss(logits[:,0], targets.float())
class MultiScore():
def __init__(self, scores):
self. score_fcns = []
self.start_idx = []
self.end_idx = []
for score in scores:
self.score_fcns.append(utils.buildFromConfig(score))
self.start_idx.append(score['start_idx'])
self.end_idx.append(score['end_idx'])
def __call__(self, last_layer):
scores = []
for i in range(len(self.score_fcns)):
scores.append(self.score_fcns[i](last_layer[:, self.start_idx[i]:self.end_idx[i]]))
return torch.cat(scores, dim=1)
class MultiLoss():
def __init__(self, losses):
self.loss_fcns = []
self.label_start_idx = []
self.label_end_idx = []
self.output_start_idx = []
self.output_end_idx = []
self.weights = []
self.label_types = []
for loss in losses:
self.loss_fcns.append(utils.buildFromConfig(loss))
self.label_start_idx.append(loss['label_start_idx'])
self.label_end_idx.append(loss['label_end_idx'])
self.output_start_idx.append(loss['output_start_idx'])
self.output_end_idx.append(loss['output_end_idx'])
self.weights.append(loss.get('weight', 1.0))
self.label_types.append(loss.get('label_type', 'float'))
def __call__(self, logits, targets):
loss = 0
# print(logits.shape, targets.shape)
for i in range(len(self.loss_fcns)):
if self.label_types[i] == 'int':
# print('loss', i, self.label_start_idx[i], self.label_end_idx[i], self.output_start_idx[i], self.output_end_idx[i])
# print(logits[:, self.output_start_idx[i]:self.output_end_idx[i]].shape, targets[:, self.label_start_idx[i]].shape)
loss += self.weights[i] * self.loss_fcns[i](logits[:, self.output_start_idx[i]:self.output_end_idx[i]], targets[:, self.label_start_idx[i]].to(int))
elif self.label_end_idx[i] - self.label_start_idx[i] == 1:
loss += self.weights[i] * self.loss_fcns[i](logits[:, self.output_start_idx[i]:self.output_end_idx[i]], targets[:, self.label_start_idx[i]])
else:
# print('loos', i, self.label_start_idx[i], self.label_end_idx[i], self.output_start_idx[i], self.output_end_idx[i])
# print(logits[:, self.output_start_idx[i]:self.output_end_idx[i]].shape, targets[:, self.label_start_idx[i]:self.label_end_idx[i]].shape)
loss += self.weights[i] * self.loss_fcns[i](logits[:, self.output_start_idx[i]:self.output_end_idx[i]], targets[:, self.label_start_idx[i]:self.label_end_idx[i]])
return loss
class AdvLoss():
def __init__(self, loss, adv_loss, adv_weight=1.0):
self.loss_fcn = utils.buildFromConfig(loss)
self.adv_loss_fcn = utils.buildFromConfig(adv_loss)
self.adv_weight = adv_weight
def __call__(self, logits, targets):
mask = targets[:,0] == 0
loss = self.loss_fcn(logits[:,0], targets[:,0])
adv_loss = self.adv_loss_fcn(logits[mask][:,1], targets[mask])
return loss - self.adv_weight * adv_loss
class MassWindowAdvLoss(AdvLoss):
def __call__(self, logits, targets):
mask = (targets[:,0] == 0) & (targets[:,1] > 5) & (targets[:,1] < 25)
print(mask, mask.shape, mask.sum())
loss = self.loss_fcn(logits[:,0], targets[:,0])
print(loss)
adv_loss = self.adv_loss_fcn(logits[mask][:,1], targets[mask][:,1])
print(adv_loss)
return loss - self.adv_weight * adv_loss
class KDELoss(MaskedLoss):
def __init__(self, mask = [], index = 0):
self.index = index
super().__init__(mask)
def __call__(self, logits, targets):
mask = self.make_mask(targets)
logits = logits[mask]
targets = targets[mask][:,self.index]
N = logits.shape[0]
masses = targets / torch.sqrt(torch.mean(targets**2))
scores = logits[:,0] / torch.sqrt(torch.mean(logits**2))
factor_2d = (1.0*N) ** (-2/6)
covs = (factor_2d * torch.var(masses), factor_2d * torch.var(scores))
m_diffs = torch.unsqueeze(masses, 1) - torch.unsqueeze(masses, 0)
s_diffs = torch.unsqueeze(scores, 1) - torch.unsqueeze(scores, 0)
ymm = torch.exp(- (m_diffs**2) / (4 * covs[0]))
yss = torch.exp(- (s_diffs**2) / (4 * covs[1]))
integral_rho_2d_rho_2d = torch.einsum('ij,ij->', ymm, yss)
integral_rho_1d_rho_1d = torch.einsum('ij,kl->', ymm, yss)
integral_rho_2d_rho_1d = torch.einsum('ij,ik->', ymm, yss)
raw_integral = integral_rho_2d_rho_2d - 2 * integral_rho_2d_rho_1d / N + integral_rho_1d_rho_1d / N**2
return raw_integral / (4 * torch.pi * N**2)
class MultiLabelLoss():
def __init__(self, label_names, label_types, label_weights = None):
self.loss_fcn = []
if (label_weights):
self.weights = torch.tensor(label_weights)
else:
self.weights = torch.ones(len(label_types))
for type in label_types:
if (type == "r"):
self.loss_fcn.append(torch.nn.MSELoss(reduce=False))
elif (type == "c"):
self.loss_fcn.append(torch.nn.BCEWithLogitsLoss())
print(f"self.weights = {self.weights}")
def __call__(self, logits, targets):
targets = targets.float()
loss = torch.zeros(len(logits[:, 0]), device = logits.get_device())
for i in range(len(self.loss_fcn)):
loss += self.weights[i] * self.loss_fcn[i](logits[:, i], targets[:, i])
return torch.mean(loss)
class MultiLabelFinish():
def __init__(self, label_names, label_types):
self.finish_fcn = []
for type in label_types:
if (type == "r"):
self.finish_fcn.append(None)
elif (type == "c"):
self.finish_fcn.append(torch.special.expit)
def __call__(self, logits):
for i in range(len(self.finish_fcn)):
if (self.finish_fcn[i]):
logits[:, i] = self.finish_fcn[i](logits[:, i].to(torch.long))
return logits
class ContrastiveClusterLoss():
def __init__(self, k=10, temperature=1, alpha=1):
self.k = k
self.temperature = temperature
self.alpha = alpha
def __call__(self, logits, targets):
targets = targets.float()
logits_combined = logits.float()
hid_size = int(len(logits[0]) / 2)
logits = normalize_embeddings(logits_combined[:, :hid_size])
logits_augmented = normalize_embeddings(logits_combined[:, hid_size:])
contrastive = contrastive_loss(logits, logits_augmented, self.temperature)
clustering, _ = clustering_loss(logits, self.k)
variance_loss = variance_regularization(logits) + variance_regularization(logits_augmented)
return torch.mean(contrastive + clustering + self.alpha * variance_loss)
class ContrastiveClusterFinish():
def __init__(self, k = 10, temperature = 1, max_cluster_iterations = 10):
self.k = k
self.temperature = temperature
self.max_cluster_iterations = max_cluster_iterations
print(f"ContrastiveClusterFinish: k = {k}, temperature = {temperature}")
def __call__(self, logits):
logits_combined = logits.float()
hid_size = int(len(logits[0]) / 2)
logits = logits_combined[:, :hid_size]
logits_augmented = logits_combined[:, hid_size:]
contrastive = contrastive_loss(logits, logits_augmented, self.temperature)
clustering, _ = clustering_loss(logits, self.k, self.max_cluster_iterations)
variance = variance_regularization(logits) + variance_regularization(logits_augmented)
return contrastive, clustering, variance
def s(z_i, z_j):
z_i = torch.tensor(z_i) if not isinstance(z_i, torch.Tensor) else z_i
z_j = torch.tensor(z_j) if not isinstance(z_j, torch.Tensor) else z_j
return torch.cdist(z_i, z_j, p=2)
# dot_product = torch.dot(z_i, z_j)
# norm_i = torch.linalg.norm(z_i)
# norm_j = torch.linalg.norm(z_j)
# return dot_product / (norm_i * norm_j)
def contrastive_loss(logits, logits_augmented, temperature=1, margin=1.0):
logits = torch.tensor(logits) if not isinstance(logits, torch.Tensor) else logits
logits_augmented = torch.tensor(logits_augmented) if not isinstance(logits_augmented, torch.Tensor) else logits_augmented
z = torch.cat((logits, logits_augmented), dim=0)
similarity_matrix = torch.mm(z, z.t()) / temperature
norms = torch.linalg.norm(z, dim=1)
norm_matrix = torch.ger(norms, norms)
similarity_matrix = similarity_matrix / norm_matrix
mask = torch.eye(similarity_matrix.size(0), dtype=torch.bool)
loss = 0
for k in range(len(logits)):
numerator = torch.exp(similarity_matrix[k, k + len(logits)])
denominator = torch.sum(torch.exp(similarity_matrix[k, ~mask[k]]))
loss += -torch.log(numerator / denominator)
return loss
def clustering_loss(logits, k=10, max_iterations=10):
# Step 1: Initialize cluster means
indices = torch.randperm(logits.size(0))[:k]
cluster_means = logits[indices]
prev_assignments = None
assignment_history = []
iteration = 0
while iteration < max_iterations:
iteration += 1
# Step 2: Assign each data point to the nearest cluster mean
distances = torch.cdist(logits, cluster_means, p=2) # Compute distances between logits and cluster means
cluster_assignments = torch.argmin(distances, dim=1) # Assign each point to the nearest cluster mean
# Check for convergence: if assignments do not change, break the loop
if prev_assignments is not None and torch.equal(cluster_assignments, prev_assignments):
break
# Check for cycles: if assignments have been seen before, break the loop
if any(torch.equal(cluster_assignments, prev) for prev in assignment_history):
break
assignment_history.append(cluster_assignments.clone())
prev_assignments = cluster_assignments.clone()
# Step 3: Update cluster means based on assignments
new_cluster_means = torch.zeros_like(cluster_means)
for i in range(k):
assigned_points = logits[cluster_assignments == i]
if assigned_points.size(0) > 0:
new_cluster_means[i] = assigned_points.mean(dim=0)
else:
# If no points are assigned to the cluster, reinitialize the mean randomly
new_cluster_means[i] = logits[torch.randint(0, logits.size(0), (1,)).item()]
cluster_means = new_cluster_means
# Step 4: Compute the clustering loss
distances = torch.cdist(logits, cluster_means, p=2)
min_distances = torch.min(distances, dim=1)[0]
loss = torch.sum(min_distances ** 2)
return loss, cluster_means
def normalize_embeddings(embeddings):
return embeddings / embeddings.norm(dim=1, keepdim=True)
def variance_regularization(embeddings):
mean_embedding = embeddings.mean(dim=0)
variance = ((embeddings - mean_embedding) ** 2).mean()
return variance
|