Spaces:
Sleeping
Sleeping
File size: 17,133 Bytes
e75a247 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 |
import wandb
import numpy as np
import torch
from sklearn.metrics import roc_curve, roc_auc_score
import json
import dgl
import matplotlib.pyplot as plt
from sklearn.decomposition import PCA
from torch_scatter import scatter_max
from matplotlib.cm import ScalarMappable
from matplotlib.colors import Normalize
def log_wandb_init(args, data_config={}):
"""log information about the run in the config section of wandb
Currently wandb is only initialized in training mode
Args:
args (_type_): parsed args from training
"""
if args.regression_mode:
wandb.config.regression_mode = True
else:
wandb.config.classification_mode = True
wandb.config.num_epochs = args.num_epochs
wandb.config.args = vars(args)
wandb.config.graph_config = data_config.graph_config
wandb.config.custom_model_kwargs = data_config.custom_model_kwargs
def log_confussion_matrix_wandb(y_true, y_score, epoch):
"""function to log confussion matrix in the wandb.ai website
Args:
y_true (_type_): labels (B,)
y_score (_type_): probabilities (B,num_classes)
epoch (_type_): epoch of training so that maybe we can build slider in wandb
"""
if y_score.ndim == 1:
y_pred = y_score > 0.5
else:
y_pred = y_score.argmax(1)
cm = wandb.plot.confusion_matrix(y_score, y_true=y_true)
wandb.log({"confussion matrix": cm})
# we could also log multiple cm during training but no sliding for now.
def log_roc_curves(y_true, y_score, epoch):
# 5 classes G(0),Q(1),S(2),C(3),B(4)
# b tagging (b/g, b/ud, b/c)
_bg = create_binary_rocs(4, 0, y_true, y_score)
_bud = create_binary_rocs(4, 1, y_true, y_score)
_bc = create_binary_rocs(4, 3, y_true, y_score)
if len(_bg) > 0 and len(_bud) > 0 and len(_bc) > 0:
# this if checks if all elements are not of the same class
calculate_and_log_tpr_1_10_percent(_bg[0], _bg[1], "b", "g")
calculate_and_log_tpr_1_10_percent(_bud[0], _bud[1], "b", "ud")
calculate_and_log_tpr_1_10_percent(_bc[0], _bc[1], "b", "c")
columns = ["b vs g", "b vs ud", "b vs c"]
xs = [_bg[1], _bud[1], _bc[1]]
ys = [_bg[0], _bud[0], _bc[0]]
auc_ = [_bg[2], _bud[2], _bc[2]]
title_log = "roc b"
title_plot = "b tagging"
wandb_log_multiline_rocs(xs, ys, title_log, title_plot, columns)
wandb_log_auc(auc_, ["b_g", "b_ud", "b_c"])
else:
print("all batch from the same class in b", len(_bg), len(_bud), len(_bc))
# c tagging (c/g, c/ud, c/b)
_cg = create_binary_rocs(3, 0, y_true, y_score)
_cud = create_binary_rocs(3, 1, y_true, y_score)
_cb = create_binary_rocs(3, 4, y_true, y_score)
if len(_cg) > 0 and len(_cud) > 0 and len(_cb) > 0:
calculate_and_log_tpr_1_10_percent(_cg[0], _cg[1], "c", "g")
calculate_and_log_tpr_1_10_percent(_cud[0], _cud[1], "c", "ud")
calculate_and_log_tpr_1_10_percent(_cb[0], _cb[1], "c", "b")
columns = ["c vs g", "c vs ud", "c vs b"]
xs = [_cg[1], _cud[1], _cb[1]]
ys = [_cg[0], _cud[0], _cb[0]]
auc_ = [_cg[2], _cud[2], _cb[2]]
title_log = "roc c"
title_plot = "c tagging"
wandb_log_multiline_rocs(xs, ys, title_log, title_plot, columns)
wandb_log_auc(auc_, ["c_g", "c_ud", "c_b"])
else:
print("all batch from the same class in c", len(_cg), len(_cud), len(_cb))
# s tagging (s/g, s/ud, s/c, s/b)
_sg = create_binary_rocs(2, 0, y_true, y_score)
_sud = create_binary_rocs(2, 1, y_true, y_score)
_sc = create_binary_rocs(2, 3, y_true, y_score)
_sb = create_binary_rocs(2, 4, y_true, y_score)
if len(_sg) > 0 and len(_sud) > 0 and len(_sc) > 0 and len(_sb) > 0:
calculate_and_log_tpr_1_10_percent(_sg[0], _sg[1], "s", "g")
calculate_and_log_tpr_1_10_percent(_sud[0], _sud[1], "s", "ud")
calculate_and_log_tpr_1_10_percent(_sc[0], _sc[1], "s", "c")
calculate_and_log_tpr_1_10_percent(_sb[0], _sb[1], "s", "b")
columns = ["s vs g", "s vs ud", "s vs c", "s vs b"]
xs = [_sg[1], _sud[1], _sc[1], _sb[1]]
ys = [_sg[0], _sud[0], _sc[0], _sb[0]]
auc_ = [_sg[2], _sud[2], _sb[2]]
title_log = "roc s"
title_plot = "s tagging"
wandb_log_multiline_rocs(xs, ys, title_log, title_plot, columns)
wandb_log_auc(auc_, ["s_g", "s_ud", "s_c", "s_b"])
else:
print(
"all batch from the same class in s",
len(_sg),
len(_sud),
len(_sc),
len(_sb),
)
# g tagging (g/ud, g/s, g/c, g/b)
_gud = create_binary_rocs(0, 1, y_true, y_score)
_gs = create_binary_rocs(0, 2, y_true, y_score)
_gc = create_binary_rocs(0, 3, y_true, y_score)
_gb = create_binary_rocs(0, 4, y_true, y_score)
if len(_gud) > 0 and len(_gs) > 0 and len(_gc) > 0 and len(_gb) > 0:
calculate_and_log_tpr_1_10_percent(_gud[0], _gud[1], "g", "ud")
calculate_and_log_tpr_1_10_percent(_gs[0], _gs[1], "g", "s")
calculate_and_log_tpr_1_10_percent(_gc[0], _gc[1], "g", "c")
calculate_and_log_tpr_1_10_percent(_gb[0], _gb[1], "g", "b")
columns = ["g vs ud", "g vs s", "g vs c", "g vs b"]
xs = [_gud[1], _gs[1], _gc[1], _gb[1]]
ys = [_gud[0], _gs[0], _gc[0], _gb[0]]
auc_ = [_gud[2], _gs[2], _gc[2], _gb[2]]
title_log = "roc g"
title_plot = "g tagging"
wandb_log_multiline_rocs(xs, ys, title_log, title_plot, columns)
wandb_log_auc(auc_, ["g_ud", "g_s", "g_c", "g_b"])
else:
print(
"all batch from the same class in g",
len(_gud),
len(_gs),
len(_gc),
len(_gb),
)
# def tagging_at_xpercent_misstag():
def log_histograms(y_true_wandb, scores_wandb, counts_particles, epoch):
print("logging hist func")
y_pred = np.argmax(scores_wandb, axis=1)
errors_class_examples = y_true_wandb != y_pred
correct_class_examples = y_true_wandb == y_pred
errors_number_count = counts_particles[errors_class_examples]
correct_number_count = counts_particles[correct_class_examples]
#print("count", errors_number_count.shape, correct_number_count.shape)
data_correct = [
[i, correct_number_count[i]] for i in range(0, len(correct_number_count))
]
data_errors = [
[i, errors_number_count[i]] for i in range(0, len(errors_number_count))
]
table_correct = wandb.Table(
data=data_correct, columns=["IDs", "correct_number_count"]
)
table_errors = wandb.Table(data=data_errors, columns=["IDs", "errors_number_count"])
wandb.log({"hist_errors_count": wandb.Histogram(errors_number_count)})
# wandb.log({'hist_errors_count': wandb.plot.histogram(table_errors, "errors_number_count",
# title="Histogram errors number const")})
def wandb_log_auc(auc_, names):
for i in range(0, len(auc_)):
name = "auc/" + names[i]
# logging 1-auc because we are looking at roc with flipped axis
wandb.log({name: 1 - auc_[i]})
return auc_
def wandb_log_multiline_rocs(xs, ys, title_log, title_plot, columns):
ys_log = [np.log10(j + 1e-8) for j in ys]
wandb.log(
{
title_log: wandb.plot.line_series(
xs=xs,
ys=ys_log,
keys=columns,
title=title_plot,
xname="jet tagging efficiency",
)
}
)
def find_nearest(a, a0):
"Element in nd array `a` closest to the scalar value `a0`"
idx = np.abs(a - a0).argmin()
return idx
def create_binary_rocs(positive, negative, y_true, y_score):
mask_positive = y_true == positive
mask_negative = y_true == negative
# print(y_true.shape, np.sum(mask_positive), np.sum(mask_negative), positive, negative)
number_positive = len(y_true[mask_positive])
number_negative = len(y_true[mask_negative])
if number_positive > 0 and number_negative > 0:
# print('s',positive,negative,number_positive,number_negative)
y_true_positive = torch.reshape(torch.ones([number_positive]), (-1,))
y_true_negative = torch.reshape(torch.zeros([number_negative]), (-1,))
y_true_ = torch.cat((y_true_positive, y_true_negative), dim=0)
y_score_positive = torch.tensor(y_score[mask_positive])
y_score_negative = torch.tensor(y_score[mask_negative])
indices = torch.tensor([negative, positive])
y_score_positive_ = torch.index_select(y_score_positive, 1, indices)
y_score_negative_ = torch.index_select(y_score_negative, 1, indices)
y_scores_pos_prob = torch.exp(y_score_positive_) / torch.sum(
torch.exp(y_score_positive_), keepdim=True, dim=1
)
y_scores_neg_prob = torch.exp(y_score_negative_) / torch.sum(
torch.exp(y_score_negative_), keepdim=True, dim=1
)
y_prob_positiveclass = torch.reshape(y_scores_pos_prob[:, 1], (-1,))
y_prob_positiveclass_neg = torch.reshape(y_scores_neg_prob[:, 1], (-1,))
y_prob_positive = torch.cat(
(y_prob_positiveclass, y_prob_positiveclass_neg), dim=0
)
fpr, tpr, thrs = roc_curve(
y_true_.numpy(), y_prob_positive.numpy(), pos_label=1
)
auc_score = roc_auc_score(y_true_.numpy(), y_prob_positive.numpy())
return [fpr, tpr, auc_score]
else:
return []
def calculate_and_log_tpr_1_10_percent(fpr, tpr, name_pos, name_neg):
idx_10_percent = find_nearest(fpr, 0.1)
idx_1_percent = find_nearest(fpr, 0.01)
tpr_10_percent = tpr[idx_10_percent]
tpr_1_percent = tpr[idx_1_percent]
name_10 = "te/" + name_pos + "_vs_" + name_neg + "_10%"
name_1 = "te/" + name_pos + "_vs_" + name_neg + "_1%"
wandb.log({name_10: tpr_10_percent, name_1: tpr_1_percent})
def plot_clust(g, q, xj, title_prefix="", y=None, radius=None, betas=None, loss_e_frac=None):
graph_list = dgl.unbatch(g)
node_counter = 0
particle_counter = 0
fig, ax = plt.subplots(12, 10, figsize=(33, 40))
for i in range(0, min(12, len(graph_list))):
graph_eval = graph_list[i]
# print([g.num_nodes() for g in graph_list])
non = graph_eval.number_of_nodes()
assert non == graph_eval.ndata["h"].shape[0]
n_part = graph_eval.ndata["particle_number"].max().long().item()
particle_number = graph_eval.ndata["particle_number"]
# if particle_number.max() > 1:
# print("skipping one, only plotting events with 2 particles")
# continue
q_graph = q[node_counter : node_counter + non].flatten()
if betas != None:
beta_graph = betas[node_counter : node_counter + non].flatten()
hit_type = torch.argmax(graph_eval.ndata["hit_type"], dim=1).view(-1)
part_num = graph_eval.ndata["particle_number"].view(-1).to(torch.long)
q_alpha, index_alpha = scatter_max(
q_graph.cpu().view(-1), part_num.cpu() - 1
)
# print(part_num.unique())
xj_graph = xj[node_counter : node_counter + non, :].detach().cpu()
if len(index_alpha) == 1:
index_alpha = index_alpha.item()
clr = graph_eval.ndata["particle_number"]
ax[i, 2].set_title("x and y of hits")
xhits, yhits = (
graph_eval.ndata["h"][:, 0].detach().cpu(),
graph_eval.ndata["h"][:, 1].detach().cpu(),
)
hittype = torch.argmax(graph_eval.ndata["h"][:, [3, 4, 5, 6]], dim=1).view(
-1
)
clr_energy = torch.log10(graph_eval.ndata["h"][:, 7].detach().cpu())
ax[i, 2].scatter(xhits, yhits, c=clr.tolist(), alpha=0.2)
ax[i, 3].scatter(xhits, yhits, c=clr_energy.tolist(), alpha=0.2)
ax[i, 3].set_title("x and y of hits colored by log10 energy")
ax[i, 4].scatter(xhits, yhits, c=hittype.tolist(), alpha=0.2)
ax[i, 4].set_title("x and y of hits colored by hit type (ecal/hcal)")
if betas != None:
ax[i, 5].scatter(xhits, yhits, c=beta_graph.detach().cpu(), alpha=0.2)
ax[i, 5].set_title("hits coloored by beta")
fig.colorbar(
ScalarMappable(norm=Normalize(vmin=0, vmax=1)), ax=ax[i, 5]
).set_label("beta")
ax[i, 6].hist(beta_graph.detach().cpu(), bins=100, range=(0, 1))
ax[i, 6].set_title("beta distr.")
fig.colorbar(
ScalarMappable(norm=Normalize(vmin=0.5, vmax=1)), ax=ax[i, 7]
).set_label("beta > 0.5")
no_objects = len(np.unique(part_num.cpu()))
ax[i, 7].scatter(
xj_graph[:, 0][beta_graph.detach().cpu() > 0.5],
xj_graph[:, 1][beta_graph.detach().cpu() > 0.5],
c=beta_graph[beta_graph.detach().cpu() > 0.5].detach().cpu(),
alpha=0.2
)
# plot no_objects highest betas
index_highest = np.argsort(beta_graph.detach().cpu())[-no_objects:]
ax[i, 7].scatter(
xj_graph[:, 0][index_highest],
xj_graph[:, 1][index_highest],
marker="*",
c="red"
)
ax[i, 7].set_title("hits with beta > 0.5")
ax[i, 8].set_title("hits of particles that have a low loss_e_frac")
if loss_e_frac is not None:
if not isinstance(loss_e_frac, torch.Tensor):
loss_e_frac = torch.cat(loss_e_frac)
loss_e_frac_batch = loss_e_frac[particle_counter : particle_counter + n_part]
particle_counter += n_part
low_filter = torch.nonzero(loss_e_frac_batch < 0.05).flatten()
if not len(low_filter):
continue
ax[i, 8].set_title(loss_e_frac_batch[low_filter[0]])
particle_number_low = part_num[low_filter[0]]
# filter to particle numbers contained in particle_number_low
low_filter = torch.nonzero(part_num == particle_number_low).flatten().detach().cpu()
ax[i, 8].scatter(
xj_graph[:, 0],
xj_graph[:, 1],
c="gray",
alpha=0.2
)
ax[i, 9].scatter(
xj_graph[:, 0],
xj_graph[:, 2],
c="gray",
alpha=0.2
)
ax[i, 8].set_xlabel("X")
ax[i, 8].set_ylabel("Y")
ax[i, 9].set_xlabel("X")
ax[i, 9].set_ylabel("Z")
ax[i, 8].scatter(
xj_graph[:, 0][low_filter],
xj_graph[:, 1][low_filter],
c="blue",
alpha=0.2
)
ax[i, 9].scatter(
xj_graph[:, 0][low_filter],
xj_graph[:, 2][low_filter],
c="blue",
alpha=0.2
)
ia1 = torch.zeros(xj_graph.shape[0]).long()
ia2 = torch.zeros_like(ia1)
ia1[index_alpha] = 1.
ia2[low_filter] = 1.
ax[i, 8].scatter(
xj_graph[ia1, 0],
xj_graph[ia1, 1],
marker="*",
c="r",
alpha=1.0,
)
ax[i, 8].scatter(
xj_graph[ia1*ia2, 0],
xj_graph[ia1*ia2, 1],
marker="*",
c="g",
alpha=1.0,
)
ax[i, 9].scatter(
xj_graph[ia1, 0],
xj_graph[ia1, 2],
marker="*",
c="r",
alpha=1.0,
)
ax[i, 9].scatter(
xj_graph[ia1 * ia2, 0],
xj_graph[ia1 * ia2, 2],
marker="*",
c="g",
alpha=1.0,
)
ax[i, 0].set_title(
title_prefix
+ " "
+ str(np.unique(part_num.cpu()))
+ " "
+ str(len(np.unique(part_num.cpu())))
)
ax[i, 1].set_title("PCA of node features")
ax[i, 0].scatter(xj_graph[:, 0], xj_graph[:, 1], c=clr.tolist(), alpha=0.2)
if non > 1:
PCA_2d_node_feats = PCA(n_components=2).fit_transform(
graph_eval.ndata["h"].detach().cpu().numpy()
)
ax[i, 1].scatter(
PCA_2d_node_feats[:, 0],
PCA_2d_node_feats[:, 1],
c=clr.tolist(),
alpha=0.2,
)
ax[i, 0].scatter(
xj_graph[index_alpha, 0],
xj_graph[index_alpha, 1],
marker="*",
c="r",
alpha=1.0,
)
pos = graph_eval.ndata["pos_hits_norm"]
node_counter += non
return fig, ax
|