-
Notifications
You must be signed in to change notification settings - Fork 0
/
Train_TrackWalker3D.py
443 lines (401 loc) · 26.7 KB
/
Train_TrackWalker3D.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
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
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
import torch
import torchvision
# from torch.utils.tensorboard import SummaryWriter
from torchvision import datasets, transforms
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
from MiscFunctions import get_loss_weights_v2, unflatten_pos, calc_logger_stats
from MiscFunctions import make_log_stat_dict, reravel_array, make_prediction_vector
from MiscFunctions import blockPrint, enablePrint, get_writers
# from DataLoader import get_net_inputs_mc, DataLoader_MC
# from ReformattedDataLoader import ReformattedDataLoader_MC
# from ComplexReformattedDataLoader import ComplexReformattedDataLoader_MC
from DataLoader3D import DataLoader3D
from ModelFunctions import LSTMTagger, GRUTagger, run_validation_pass
import random
import ROOT
import os
def prepare_sequence(seq, to_ix):
idxs = [to_ix[w] for w in seq]
return torch.tensor(idxs, dtype=torch.long)
def prepare_sequence_steps(seq,long=False):
full_np = np.stack(seq,axis=0)
if not long:
return torch.tensor(full_np, dtype=torch.float)
else:
return torch.tensor(full_np, dtype=torch.long)
PARAMS = {}
PARAMS['USE_CONV_IM'] = True
# PARAMS['LARMATCH_CKPT'] = '/home/jmills/workdir/TrackWalker/larmatch_ckpt/checkpoint.1974000th.tar'
PARAMS['MASK_WC'] = False
PARAMS['RAND_FLIP_INPUT'] = False
PARAMS['MIN_TRACK_LENGTH'] = 3.0
PARAMS['HIDDEN_DIM'] =1024
PARAMS['PADDING'] = 10
PARAMS['VOXCUBESIDE'] = 3
PARAMS['APPEND_WIREIM'] = True
PARAMS['NDIMENSIONS'] = 3 #Not configured to have 3 yet.
PARAMS['NFEATS'] = 17
PARAMS['EMBEDDING_DIM'] =(PARAMS['PADDING']*2+1)*(PARAMS['PADDING']*2+1)*PARAMS['NFEATS']*3 # N_Features*3planes
PARAMS['CENTERPOINT_ISEND'] = True
PARAMS['NUM_CLASSES'] = PARAMS['VOXCUBESIDE']**3
PARAMS['TRACKEND_CLASS'] = (PARAMS['NUM_CLASSES']-1)/2
# PARAMS['INFILE'] ="/home/jmills/workdir/TrackWalker/inputfiles/merged_dlreco_75e9707a-a05b-4cb7-a246-bedc2982ff7e.root"
# PARAMS['INFILE'] ="/home/jmills/workdir/TrackWalker/inputfiles/mcc9_v29e_dl_run3b_bnb_nu_overlay_nocrtmerge_TrackWalker_traindata_198files.root"
# PARAMS['INFILE'] = "/home/jmills/workdir/TrackWalker/inputfiles/ReformattedInput/Reformat_LArMatch_Pad_010.root"
# Small Files
# PARAMS['INFILE_TRAIN'] = "/home/jmills/workdir/TrackWalker/TEST3DReformat/0/Reformat_LArMatch_ComplexTrackIdx_TESTSPARSEALL.root"
# PARAMS['INFILE_VAL'] = "/home/jmills/workdir/TrackWalker/TEST3DReformat/0/Reformat_LArMatch_ComplexTrackIdx_TESTSPARSEVAL.root"
PARAMS['INFILE_TRAIN'] = "/home/jmills/workdir/TrackWalker/inputfiles/Formatted3D/train3D_OldLArMatch_1008Files.root"
PARAMS['INFILE_VAL'] = "/home/jmills/workdir/TrackWalker/inputfiles/Formatted3D/val3D_OldLArMatch_330Files.root"
PARAMS['ALWAYS_EDGE'] = True # True points are always placed at the edge of the Padded Box
PARAMS['CLASSIFIER_NOT_DISTANCESHIFTER'] =True # True -> Predict Output Pixel to step to, False -> Predict X,Y shift to next point
PARAMS['DO_TENSORLOG'] = True
PARAMS['TENSORDIR'] = None # Default runs/DATE_TIME Deprecated
PARAMS['TWOWRITERS'] = True
PARAMS['SAVE_MODEL'] = True #should the network save the model?
PARAMS['CHECKPOINT_EVERY_N_TRACKS'] = 20000 # if not saving then this doesn't matter
PARAMS['EPOCHS'] = 5
PARAMS['STOP_AFTER_NTRACKS'] = 999999999999999999999
PARAMS['VALIDATION_EPOCH_LOGINTERVAL'] = 1
PARAMS['VALIDATION_TRACKIDX_LOGINTERVAL'] = 100
PARAMS['TRAIN_EPOCH_LOGINTERVAL'] = 1
PARAMS['TRAIN_TRACKIDX_LOGINTERVAL'] = 100
PARAMS['DEVICE'] = 'cuda:3'
PARAMS['LOAD_SIZE'] = 100 #Number of Entries to Load training tracks from
PARAMS['TRAIN_EPOCH_SIZE'] = -1 #500 # Number of Training Tracks to use (load )
PARAMS['VAL_EPOCH_SIZE'] = -1 #int(0.8*PARAMS['TRAIN_EPOCH_SIZE'])
PARAMS['VAL_SAMPLE_SIZE'] = 50
PARAMS['SHUFFLE_DATASET'] = False
PARAMS['VAL_IS_TRAIN'] = False # This will set the validation set equal to the training set
PARAMS['AREA_TARGET'] = True # Change network to be predicting
PARAMS['TARGET_BUFFER'] = 2
PARAMS['DO_CROPSHIFT'] = False
PARAMS['CROPSHIFT_MAXAMT'] = 2
PARAMS['LEARNING_RATES'] = [(000000, 0.0001),(120000, 0.00001)] #(Step, New LR), place steps in order.
PARAMS['WEIGHT_DECAY'] = 1e-03
# PARAMS['LEARNING_RATES'] = [(0, 0.001), (10000, 0.0001), (100000, 1e-05), (1000000, 1e-06)] #(Step, New LR), place steps in order.
PARAMS['NEXTSTEP_LOSS_WEIGHT'] = 5.0
PARAMS['ENDSTEP_LOSS_WEIGHT'] = 1.0
PARAMS['LOAD_PREVIOUS_CHECKPOINT'] = '/home/jmills/workdir/TrackWalker/runs/new_runs/Oct04_13-37-26_mayer/TrackerCheckPoint_2_120000.pt'
PARAMS['STARTING_ENTRY'] = 120000
def main():
print("Let's Get Started.")
torch.manual_seed(1)
nbins = PARAMS['PADDING']*2+1
pred_h = ROOT.TH2D("Prediction Steps Heatmap","Prediction Steps Heatmap",nbins,-0.5,nbins+0.5,nbins,-0.5,nbins+0.5)
targ_h = ROOT.TH2D("Target Steps Heatmap","Target Steps Heatmap",nbins,-0.5,nbins+0.5,nbins,-0.5,nbins+0.5)
ComplexReformattedDataLoader_Train = DataLoader3D(PARAMS,all_train=True)
ComplexReformattedDataLoader_Val = DataLoader3D(PARAMS,all_valid=True)
PARAMS['TRAIN_EPOCH_SIZE'] = ComplexReformattedDataLoader_Train.nentries_train #500 # Number of Training Tracks to use (load )
PARAMS['VAL_EPOCH_SIZE'] = ComplexReformattedDataLoader_Val.nentries_val #int(0.8*PARAMS['TRAIN_EPOCH_SIZE'])
writer_train = None
writer_val = None
writer_dir = None
if PARAMS['DO_TENSORLOG']:
writer_train, writer_val, writer_dir = get_writers(PARAMS)
print("/////////////////////////////////////////////////")
print("/////////////////////////////////////////////////")
print("/////////////////////////////////////////////////")
print("/////////////////////////////////////////////////")
print("Initializing Model")
output_dim = None
loss_function_next_step = None
loss_function_endpoint = None
if PARAMS['AREA_TARGET']:
output_dim = PARAMS['NUM_CLASSES']
loss_function_next_step = nn.MSELoss(reduction='none')
loss_function_endpoint = nn.NLLLoss(reduction='none')
elif PARAMS['CLASSIFIER_NOT_DISTANCESHIFTER']:
output_dim = PARAMS['NUM_CLASSES'] # nPixels in crop + 1 for 'end of track'
loss_function_next_step = nn.NLLLoss(reduction='none')
loss_function_endpoint = nn.NLLLoss(reduction='none')
else:
output_dim = PARAMS['NDIMENSIONS'] # Shift X, Shift Y
loss_function_next_step = nn.MSELoss(reduction='sum')
loss_function_endpoint = nn.NLLLoss(reduction='none')
model = LSTMTagger(PARAMS['EMBEDDING_DIM'], PARAMS['HIDDEN_DIM'], output_dim).to(torch.device(PARAMS['DEVICE']))
# model = GRUTagger(PARAMS['EMBEDDING_DIM'], PARAMS['HIDDEN_DIM'], output_dim).to(torch.device(PARAMS['DEVICE']))
step_counter = 0
startEpoch = 0
startEpochEntryOffset = 0
if PARAMS['LOAD_PREVIOUS_CHECKPOINT'] != '':
step_counter = PARAMS['STARTING_ENTRY']
trainEntry = step_counter%ComplexReformattedDataLoader_Train.nentries_train
valEntry = step_counter%ComplexReformattedDataLoader_Val.nentries_val
startEpoch = int(step_counter/ComplexReformattedDataLoader_Train.nentries_train)
startEpochEntryOffset = step_counter%ComplexReformattedDataLoader_Train.nentries_train
ComplexReformattedDataLoader_Train.set_current_entry(trainEntry)
ComplexReformattedDataLoader_Val.set_current_entry(valEntry)
if (PARAMS['DEVICE'] != 'cpu'):
model.load_state_dict(torch.load(PARAMS['LOAD_PREVIOUS_CHECKPOINT'], map_location={'cpu':PARAMS['DEVICE'],'cuda:0':PARAMS['DEVICE'],'cuda:1':PARAMS['DEVICE'],'cuda:2':PARAMS['DEVICE'],'cuda:3':PARAMS['DEVICE']}))
else:
model.load_state_dict(torch.load(PARAMS['LOAD_PREVIOUS_CHECKPOINT'], map_location={'cpu':'cpu','cuda:0':'cpu','cuda:1':'cpu','cuda:2':'cpu','cuda:3':'cpu'}))
CURRENT_LR_IDX = 0
NEXT_LR_TUPLE = PARAMS['LEARNING_RATES'][CURRENT_LR_IDX]
PARAMS['CURRENT_LR'] = NEXT_LR_TUPLE[1]
optimizer = optim.Adam(model.parameters(), lr=NEXT_LR_TUPLE[1],weight_decay=PARAMS['WEIGHT_DECAY'])
if len(PARAMS['LEARNING_RATES']) > CURRENT_LR_IDX+1:
CURRENT_LR_IDX += 1
NEXT_LR_TUPLE = PARAMS['LEARNING_RATES'][CURRENT_LR_IDX]
is_long = PARAMS['CLASSIFIER_NOT_DISTANCESHIFTER'] and not PARAMS['AREA_TARGET']
# Make our trusty loggers
if not PARAMS['TWOWRITERS']:
# To Log Stats every N epoch
log_stats_dict_epoch_train = make_log_stat_dict('epoch_train_')
log_stats_dict_epoch_val = make_log_stat_dict('epoch_val_')
# To Log Stats Every N Tracks Looked At
log_stats_dict_step_train = make_log_stat_dict('step_train_')
log_stats_dict_step_val = make_log_stat_dict('step_val_')
else:
log_stats_dict_epoch_train = make_log_stat_dict('epoch_')
log_stats_dict_epoch_val = make_log_stat_dict('epoch_')
# To Log Stats Every N Tracks Looked At
log_stats_dict_step_train = make_log_stat_dict('step_')
log_stats_dict_step_val = make_log_stat_dict('step_')
firstEpoch = True
for epoch in range(startEpoch, PARAMS['EPOCHS']): # again, normally you would NOT do 300 epochs, it is toy data
print("\n-----------------------------------\nEpoch:",epoch,"\n")
train_idx = -1
number_train_loaded_so_far = 0
if firstEpoch:
number_train_loaded_so_far = startEpochEntryOffset
firstEpoch = False
n_to_load = PARAMS['LOAD_SIZE']
while number_train_loaded_so_far < PARAMS['TRAIN_EPOCH_SIZE']:
if (PARAMS['TRAIN_EPOCH_SIZE'] - number_train_loaded_so_far) < n_to_load:
n_to_load = PARAMS['TRAIN_EPOCH_SIZE'] - number_train_loaded_so_far
print()
training_data = ComplexReformattedDataLoader_Train.get_train_data(n_to_load)
number_train_loaded_so_far += len(training_data)
print(number_train_loaded_so_far, "Tracks loaded total this epoch.")
for step_images, targ_next_step_idx, targ_area_next_step in training_data:
model.train()
# Change LR
if step_counter == NEXT_LR_TUPLE[0]:
print("\n\nUpdating Learning Rate:")
print("Old LR:",PARAMS['CURRENT_LR'])
print("New LR:",NEXT_LR_TUPLE[1],"\n\n")
for pgroup in optimizer.param_groups:
pgroup['lr'] = NEXT_LR_TUPLE[1]
PARAMS['CURRENT_LR'] = NEXT_LR_TUPLE[1]
if len(PARAMS['LEARNING_RATES']) > CURRENT_LR_IDX+1:
CURRENT_LR_IDX += 1
NEXT_LR_TUPLE = PARAMS['LEARNING_RATES'][CURRENT_LR_IDX]
step_counter += 1
train_idx += 1
# Remember that Pytorch accumulates gradients.
# We need to clear them out before each instance
model.zero_grad()
step_images_in = prepare_sequence_steps(step_images).to(torch.device(PARAMS['DEVICE']))
n_steps = step_images_in.shape[0]
np_targ_endpt = np.zeros((n_steps))
np_targ_endpt[n_steps-1] = 1
endpoint_targ_t = torch.tensor(np_targ_endpt).to(torch.device(PARAMS['DEVICE']),dtype=torch.long)
targets_next_step_area = None
targets_onept = prepare_sequence_steps(targ_next_step_idx,long=is_long)
if PARAMS['AREA_TARGET']:
targets_next_step_area = prepare_sequence_steps(targ_area_next_step,long=is_long).to(torch.device(PARAMS['DEVICE']))
else:
targets_next_step_area = targets_onept.to(torch.device(PARAMS['DEVICE']))
# Step 3. Run our forward pass.
next_steps_pred_scores, endpoint_scores, hidden_n, cell_n = model(step_images_in) # _ is hidden state, no need to hold onto
np_pred = None
np_targ = None
np_pred_endpt = None
if PARAMS['AREA_TARGET']:
npts = next_steps_pred_scores.cpu().detach().numpy().shape[0]
np_pred = next_steps_pred_scores.cpu().detach().numpy().reshape(npts,PARAMS['VOXCUBESIDE'],PARAMS['VOXCUBESIDE'],PARAMS['VOXCUBESIDE'])
np_targ = targets_onept.cpu().detach().numpy()
np_pred_endpt = np.argmax(endpoint_scores.cpu().detach().numpy(),axis=1)
elif PARAMS['CLASSIFIER_NOT_DISTANCESHIFTER']:
np_pred = np.argmax(next_steps_pred_scores.cpu().detach().numpy(),axis=1)
np_targ = targets_onept.cpu().detach().numpy()
else:
np_pred = np.rint(next_steps_pred_scores.cpu().detach().numpy()) # Rounded to integers
np_targ = targets_onept.cpu().detach().numpy()
# loss_weights = torch.tensor(get_loss_weights_v2(targets_next_step_area.cpu().detach().numpy(),np_pred,PARAMS),dtype=torch.float).to(torch.device(PARAMS['DEVICE']))
# print("Pred Targ")
# print(np_pred)
# print()
# print(targets_next_step_area)
# print()
loss_next_steps = loss_function_next_step(next_steps_pred_scores, targets_next_step_area)
vals_per_step = loss_next_steps.shape[1]
loss_next_steps_per_step =torch.mean(torch.div(torch.sum(loss_next_steps, dim=1),vals_per_step))*PARAMS['NEXTSTEP_LOSS_WEIGHT']
loss_endpoint = torch.mean(loss_function_endpoint(endpoint_scores, endpoint_targ_t))*PARAMS['ENDSTEP_LOSS_WEIGHT']
loss_total = loss_next_steps_per_step + loss_endpoint
loss_total.backward()
optimizer.step()
np_idx_v = make_prediction_vector(PARAMS, np_pred)
for ixx in range(np_idx_v.shape[0]):
pred_x, pred_y, pred_z = unflatten_pos(np_idx_v[ixx], PARAMS['VOXCUBESIDE'])
pred_h.Fill(pred_x, pred_y, pred_z)
for ixx in range(np_targ.shape[0]):
targ_x, targ_y, targ_z = unflatten_pos(np_targ[ixx], PARAMS['VOXCUBESIDE'])
targ_h.Fill(targ_x,targ_y,targ_z)
if PARAMS['SAVE_MODEL'] and step_counter%PARAMS['CHECKPOINT_EVERY_N_TRACKS'] == 0:
# print("CANT SAVE NEED TO SPECIFY SUBFOLDER")
torch.save(model.state_dict(), writer_dir+"TrackerCheckPoint_"+str(epoch)+"_"+str(step_counter)+".pt")
if PARAMS['DO_TENSORLOG']:
calc_logger_stats(log_stats_dict_epoch_train, PARAMS, np_pred, np_targ, loss_total, loss_endpoint, loss_next_steps_per_step, PARAMS['TRAIN_EPOCH_SIZE'], np_pred_endpt, np_targ_endpt, is_train=True,is_epoch=True)
if PARAMS['TRAIN_TRACKIDX_LOGINTERVAL']!=-1:
calc_logger_stats(log_stats_dict_step_train, PARAMS, np_pred, np_targ, loss_total, loss_endpoint, loss_next_steps_per_step, PARAMS['TRAIN_TRACKIDX_LOGINTERVAL'], np_pred_endpt, np_targ_endpt, is_train=True,is_epoch=False)
if step_counter%PARAMS['TRAIN_TRACKIDX_LOGINTERVAL']== 0:
print("Logging Train Step",step_counter)
if not PARAMS['TWOWRITERS']:
writer_train.add_scalar('Step/train_loss_total', log_stats_dict_step_train['step_train_loss_average'], step_counter)
writer_train.add_scalar('Step/train_loss_endpointnet', log_stats_dict_step_train['step_train_loss_endptnet'], step_counter)
writer_train.add_scalar('Step/train_loss_stepnet', log_stats_dict_step_train['step_train_loss_stepnet'], step_counter)
writer_train.add_scalar('Step/train_acc_endpoint', log_stats_dict_step_train['step_train_acc_endpoint'], step_counter)
writer_train.add_scalar('Step/train_acc_exact', log_stats_dict_step_train['step_train_acc_exact'], step_counter)
writer_train.add_scalar('Step/train_acc_2dist', log_stats_dict_step_train['step_train_acc_2dist'], step_counter)
writer_train.add_scalar('Step/train_acc_5dist', log_stats_dict_step_train['step_train_acc_5dist'], step_counter)
writer_train.add_scalar('Step/train_acc_10dist', log_stats_dict_step_train['step_train_acc_10dist'], step_counter)
writer_train.add_scalar('Step/train_num_correct_exact', log_stats_dict_step_train['step_train_num_correct_exact'], step_counter)
writer_train.add_scalar("Step/train_average_off_distance", log_stats_dict_step_train['step_train_average_distance_off'],step_counter)
writer_train.add_scalar("Step/train_frac_misIDas_endpoint", log_stats_dict_step_train['step_train_frac_misIDas_endpoint'],step_counter)
writer_train.add_scalar("Step/train_lr", PARAMS['CURRENT_LR'],step_counter)
log_stats_dict_step_train = make_log_stat_dict('step_train')
else:
writer_train.add_scalar('Step/loss_total', log_stats_dict_step_train['step_loss_average'], step_counter)
writer_train.add_scalar('Step/loss_endpointnet', log_stats_dict_step_train['step_loss_endptnet'], step_counter)
writer_train.add_scalar('Step/loss_stepnet', log_stats_dict_step_train['step_loss_stepnet'], step_counter)
writer_train.add_scalar('Step/acc_endpoint', log_stats_dict_step_train['step_acc_endpoint'], step_counter)
writer_train.add_scalar('Step/acc_exact', log_stats_dict_step_train['step_acc_exact'], step_counter)
writer_train.add_scalar('Step/acc_2dist', log_stats_dict_step_train['step_acc_2dist'], step_counter)
writer_train.add_scalar('Step/acc_5dist', log_stats_dict_step_train['step_acc_5dist'], step_counter)
writer_train.add_scalar('Step/acc_10dist', log_stats_dict_step_train['step_acc_10dist'], step_counter)
writer_train.add_scalar('Step/num_correct_exact', log_stats_dict_step_train['step_num_correct_exact'], step_counter)
writer_train.add_scalar("Step/average_off_distance", log_stats_dict_step_train['step_average_distance_off'],step_counter)
writer_train.add_scalar("Step/frac_misIDas_endpoint", log_stats_dict_step_train['step_frac_misIDas_endpoint'],step_counter)
writer_train.add_scalar("Step/lr", PARAMS['CURRENT_LR'],step_counter)
log_stats_dict_step_train = make_log_stat_dict('step_')
if PARAMS['VALIDATION_TRACKIDX_LOGINTERVAL'] !=-1 and step_counter%PARAMS['VALIDATION_TRACKIDX_LOGINTERVAL'] == 0:
print("Logging Val Step",step_counter)
if not PARAMS['TWOWRITERS']:
log_stats_dict_step_val = make_log_stat_dict('step_val_')
else:
log_stats_dict_step_val = make_log_stat_dict('step_')
run_validation_pass(PARAMS, model, ComplexReformattedDataLoader_Val, loss_function_next_step, loss_function_endpoint, writer_val, log_stats_dict_step_val, step_counter, is_epoch=False)
if step_counter%PARAMS["STOP_AFTER_NTRACKS"] == 0:
break
####### DO VALIDATION PASS
if PARAMS['DO_TENSORLOG'] and epoch%PARAMS['VALIDATION_EPOCH_LOGINTERVAL']==0:
print("Logging Val Epoch", epoch)
if not PARAMS['TWOWRITERS']:
log_stats_dict_epoch_val = make_log_stat_dict('epoch_val_')
else:
log_stats_dict_epoch_val = make_log_stat_dict('epoch_')
run_validation_pass(PARAMS, model, ComplexReformattedDataLoader_Val, loss_function_next_step, loss_function_endpoint, writer_val, log_stats_dict_epoch_val, epoch, is_epoch=True)
# if epoch%50 ==0:
# print("Training Epoch Averaged")
# print("Exact Accuracy Endpoints:")
# print(log_stats_dict_epoch_train['epoch_train_acc_endpoint'])
# print("Fraction misID as Endpoints:")
# print(log_stats_dict_epoch_train['epoch_train_frac_misIDas_endpoint'])
# print("Exact Accuracy Trackpoints:")
# print(log_stats_dict_epoch_train['epoch_train_acc_exact'])
# print("Within 2 Accuracy:")
# print(log_stats_dict_epoch_train['epoch_train_acc_2dist'])
# print("Within 5 Accuracy:")
# print(log_stats_dict_epoch_train['epoch_train_acc_5dist'])
# print("Within 10 Accuracy:")
# print(log_stats_dict_epoch_train['epoch_train_acc_10dist'])
# print("Average Distance Off:")
# print(log_stats_dict_epoch_train['epoch_train_average_distance_off'])
# print("Loss:")
# print(log_stats_dict_epoch_train['epoch_train_loss_average'])
# print("/////////////////////////////")
# print()
if PARAMS['DO_TENSORLOG'] and epoch%PARAMS['TRAIN_EPOCH_LOGINTERVAL']==0:
print("Logging Train Epoch", epoch)
if not PARAMS['TWOWRITERS']:
writer_train.add_scalar('Epoch/train_loss_total', log_stats_dict_epoch_train['epoch_train_loss_average'], epoch)
writer_train.add_scalar('Epoch/train_loss_endpointnet', log_stats_dict_epoch_train['epoch_train_loss_endptnet'], epoch)
writer_train.add_scalar('Epoch/train_loss_stepnet', log_stats_dict_epoch_train['epoch_train_loss_stepnet'], epoch)
writer_train.add_scalar('Epoch/train_acc_endpoint', log_stats_dict_epoch_train['epoch_train_acc_endpoint'], epoch)
writer_train.add_scalar('Epoch/train_acc_exact', log_stats_dict_epoch_train['epoch_train_acc_exact'], epoch)
writer_train.add_scalar('Epoch/train_acc_2dist', log_stats_dict_epoch_train['epoch_train_acc_2dist'], epoch)
writer_train.add_scalar('Epoch/train_acc_5dist', log_stats_dict_epoch_train['epoch_train_acc_5dist'], epoch)
writer_train.add_scalar('Epoch/train_acc_10dist', log_stats_dict_epoch_train['epoch_train_acc_10dist'], epoch)
writer_train.add_scalar('Epoch/train_num_correct_exact', log_stats_dict_epoch_train['epoch_train_num_correct_exact'], epoch)
writer_train.add_scalar("Epoch/train_average_off_distance", log_stats_dict_epoch_train['epoch_train_average_distance_off'],epoch)
writer_train.add_scalar("Epoch/train_frac_misIDas_endpoint", log_stats_dict_epoch_train['epoch_train_frac_misIDas_endpoint'],epoch)
writer_train.add_scalar("Epoch/train_lr", PARAMS['CURRENT_LR'],epoch)
log_stats_dict_epoch_train = make_log_stat_dict('epoch_train_')
else:
writer_train.add_scalar('Epoch/loss_total', log_stats_dict_epoch_train['epoch_loss_average'], epoch)
writer_train.add_scalar('Epoch/loss_endpointnet', log_stats_dict_epoch_train['epoch_loss_endptnet'], epoch)
writer_train.add_scalar('Epoch/loss_stepnet', log_stats_dict_epoch_train['epoch_loss_stepnet'], epoch)
writer_train.add_scalar('Epoch/acc_endpoint', log_stats_dict_epoch_train['epoch_acc_endpoint'], epoch)
writer_train.add_scalar('Epoch/acc_exact', log_stats_dict_epoch_train['epoch_acc_exact'], epoch)
writer_train.add_scalar('Epoch/acc_2dist', log_stats_dict_epoch_train['epoch_acc_2dist'], epoch)
writer_train.add_scalar('Epoch/acc_5dist', log_stats_dict_epoch_train['epoch_acc_5dist'], epoch)
writer_train.add_scalar('Epoch/acc_10dist', log_stats_dict_epoch_train['epoch_acc_10dist'], epoch)
writer_train.add_scalar('Epoch/num_correct_exact', log_stats_dict_epoch_train['epoch_num_correct_exact'], epoch)
writer_train.add_scalar("Epoch/average_off_distance", log_stats_dict_epoch_train['epoch_average_distance_off'],epoch)
writer_train.add_scalar("Epoch/frac_misIDas_endpoint", log_stats_dict_epoch_train['epoch_frac_misIDas_endpoint'],epoch)
writer_train.add_scalar("Epoch/lr", PARAMS['CURRENT_LR'],epoch)
log_stats_dict_epoch_train = make_log_stat_dict('epoch_')
if step_counter%PARAMS["STOP_AFTER_NTRACKS"] == 0:
break
print()
print("End of Training")
print()
if PARAMS['DO_TENSORLOG']:
writer_train.close()
writer_val.close()
# See what the scores are after training
if PARAMS['SAVE_MODEL']:
print("CANT SAVE NEED TO SPECIFY SUBFOLDER")
torch.save(model.state_dict(), writer_dir+"TrackerCheckPoint_"+str(PARAMS['EPOCHS'])+"_Fin.pt")
# with torch.no_grad():
# train_idx = -1
# for step_images, next_steps in training_data:
# train_idx += 1
# print()
# print("Event:", event_ids[train_idx])
# print("Track Idx:",train_idx)
# step_images_in = prepare_sequence_steps(step_images).to(torch.device(PARAMS['DEVICE']))
# targets = prepare_sequence_steps(next_steps,long=is_long).to(torch.device(PARAMS['DEVICE']))
# np_targ = targets.cpu().detach().numpy()
# next_steps_pred_scores, endpoint_scores = model(step_images_in)
#
# np_pred = None
# if PARAMS['CLASSIFIER_NOT_DISTANCESHIFTER']:
# np_pred = np.argmax(next_steps_pred_scores.cpu().detach().numpy(),axis=1)
# else:
# np_pred = np.rint(next_steps_pred_scores.cpu().detach().numpy()) # Rounded to integers
#
# num_correct_exact = 0
# for ix in range(np_pred.shape[0]):
# if PARAMS['CLASSIFIER_NOT_DISTANCESHIFTER']:
# if np_pred[ix] == np_targ[ix]:
# num_correct_exact = num_correct_exact + 1
# else:
# if np.array_equal(np_pred[ix], np_targ[ix]):
# num_correct_exact += 1
# print("Accuracy",float(num_correct_exact)/float(np_pred.shape[0]))
# print("Points:",float(np_pred.shape[0]))
#
# np_targ = targets.cpu().detach().numpy()
# if not PARAMS['CLASSIFIER_NOT_DISTANCESHIFTER']:
# print("Predictions Raw")
# print(next_steps_pred_scores.cpu().detach().numpy())
#
# # make_steps_images(step_images_in.cpu().detach().numpy(),"images/PredStep_Final_"+str(train_idx).zfill(2)+"_",PADDING*2+1,pred=np_pred,targ=np_targ)
canv = ROOT.TCanvas('canv','canv',1000,800)
ROOT.gStyle.SetOptStat(0)
# pred_h.SetMaximum(500.0)
pred_h.DrawNormalized("COLZ")
# canv.SaveAs('pred_h.png')
# targ_h.SetMaximum(500.0)
targ_h.DrawNormalized("COLZ")
# canv.SaveAs('targ_h.png')
print("End of Main")
return 0
if __name__ == '__main__':
main()