-
Notifications
You must be signed in to change notification settings - Fork 0
/
model_base_spad_lit.py
381 lines (321 loc) · 13.8 KB
/
model_base_spad_lit.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
'''
Base lightning model that contains the share loss functions and training parameters across all models
'''
#### Standard Library Imports
import os
import logging
#### Library imports
import numpy as np
import matplotlib.pyplot as plt
import torch
from torch.autograd import Variable
import pytorch_lightning as pl
import torchvision
from IPython.core import debugger
breakpoint = debugger.set_trace
#### Local imports
from losses import criterion_L2, criterion_L1, criterion_KL, criterion_RMSE, criterion_TV
from research_utils.shared_constants import TWOPI, EPSILON
from research_utils.np_utils import calc_mean_percentile_errors
import tof_utils
def make_zeromean_normalized_bins(bins):
return (2*bins) - 1
def normdepth2phasor(d):
phase = TWOPI*d
cos = torch.cos(phase)
sin = torch.sin(phase)
return torch.cat((cos, sin), dim=1)
def hist2rec_soft_argmax(hist_img):
smax = torch.nn.Softmax2d()
weights = Variable(torch.linspace(0, 1, steps=hist_img.size()[1]).unsqueeze(1).unsqueeze(1)).to(hist_img.device)
weighted_smax = weights * smax(hist_img)
soft_argmax = weighted_smax.sum(1).unsqueeze(1)
return soft_argmax
class LITBaseSPADModel(pl.LightningModule):
def __init__(self,
backbone_net,
init_lr = 1e-4,
p_tv = 1e-5,
lr_decay_gamma = 0.9,
data_loss_id = 'kldiv',
):
super(LITBaseSPADModel, self).__init__()
self.lsmx = torch.nn.LogSoftmax(dim=-3)
self.smx = torch.nn.Softmax(dim=-3)
self.print_logger = logging.getLogger(__name__)
# Train hyperparams
self.init_lr = init_lr
self.lr_decay_gamma = lr_decay_gamma
self.p_tv = p_tv
self.data_loss_id = data_loss_id
## Only for Compressive Histogram models. All other models are not affected by this flag
self.emulate_int8_quantization = False
self.backbone_net = backbone_net
self.example_input_array = torch.randn([1, 1, 1024, 32, 32])
self.test_rmse_all = []
self.save_hyperparameters(ignore=['backbone_net'])
def forward(self, x):
# use forward for inference/predictions
out = self.backbone_net(x)
return out
def configure_optimizers(self):
optimizer = torch.optim.Adam(self.parameters(), self.init_lr)
lr_scheduler = {
'scheduler': torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=self.lr_decay_gamma, verbose=False)
, 'name': 'epoch/Adam-lr' # Name for logging in tensorboard (used by lr_monitor callback)
}
return [optimizer], [lr_scheduler]
def get_input_data(self, sample):
return sample["spad"]
def forward_wrapper(self, sample):
# Input the correct type of data to the model which should output the recovered histogram and depths
input_data = self.get_input_data(sample)
M_mea_re, rec = self(input_data)
return M_mea_re, rec
def rec2depth(self, rec):
'''
For some models the reconstructed signal are not depths
'''
return rec
def compute_losses(self, sample, M_mea_re, rec):
# load data and compute losses
M_gt = sample["rates"]
dep = sample["bins"]
# Get recovered depths normalized between 0,1
dep_re = self.rec2depth(rec)
# Normalize
M_mea_re_lsmx = self.lsmx(M_mea_re).unsqueeze(1)
# Compute metrics
loss_kl = criterion_KL(M_mea_re_lsmx, M_gt)
loss_tv = criterion_TV(dep_re)
loss_l1 = criterion_L1(dep_re, dep)
loss_l2 = criterion_L2(dep_re, dep)
loss_phasorl1 = criterion_L1(normdepth2phasor(dep_re), normdepth2phasor(dep))
loss_emd = criterion_L1(torch.cumsum(M_gt, dim=-3), torch.cumsum(M_mea_re.unsqueeze(1), dim=-3))
loss_rmse = criterion_RMSE(dep_re, dep)
if(self.data_loss_id == 'kldiv'):
# For the first 100 steps only consider the data loss term
if(self.global_step < 500): loss = loss_kl
else: loss = loss_kl + self.p_tv*loss_tv
# loss = loss_kl + self.p_tv*loss_tv
elif(self.data_loss_id == 'L1'):
loss = loss_l1 + self.p_tv*loss_tv
elif(self.data_loss_id == 'L2'):
loss = loss_l2 + self.p_tv*loss_tv
elif(self.data_loss_id == 'EMD'):
loss = loss_emd + self.p_tv*loss_tv
elif(self.data_loss_id == 'PhasorL1'):
loss = loss_phasorl1 + self.p_tv*loss_tv
elif(self.data_loss_id == 'DirectPhasorL1'):
loss_directphasorl1 = criterion_L1(rec, normdepth2phasor(dep))
loss = loss_directphasorl1 + self.p_tv*loss_tv
elif(self.data_loss_id == 'FFTHistL1'):
fft_gt_hist = torch.fft.rfft(M_gt.squeeze(1), dim=-3)
loss_real_l1 = criterion_L1(fft_gt_hist.real, rec[:, 0::2, :])
loss_imag_l1 = criterion_L1(fft_gt_hist.imag, rec[:, 1::2, :])
loss = loss_real_l1 + loss_imag_l1 + self.p_tv*loss_tv
elif(self.data_loss_id == 'DispL1'):
loss_displ1 = criterion_L1(rec, 1 / (dep + EPSILON))
loss = loss_displ1 + self.p_tv*loss_tv
else:
assert(False), "Invalid loss id"
return {'loss': loss,
'loss_kl': loss_kl,
'loss_l1': loss_l1,
'loss_l2': loss_l2,
'loss_emd': loss_emd,
'loss_tv': loss_tv,
'rmse': loss_rmse}
def training_step(self, sample, batch_idx):
# Forward pass
M_mea_re, rec = self.forward_wrapper(sample)
# Compute Losses
losses = self.compute_losses(sample, M_mea_re, rec)
loss = losses['loss']
loss_l1 = losses['loss_l1']
loss_l2 = losses['loss_l2']
loss_kl = losses['loss_kl']
loss_tv = losses['loss_tv']
rmse = losses['rmse']
# Log to logger (i.e., tensorboard), if you want it to be displayed at progress bar, use prog_bar=True
self.log_dict(
{
"loss/train": loss
, "rmse/train": rmse
, "train_losses/kldiv": loss_kl
, "train_losses/l1": loss_l1
, "train_losses/l2": loss_l2
, "train_losses/tv": self.p_tv*loss_tv
}
# , prog_bar=True
)
# # Log some images every 500 training steps
# if((batch_idx % 20 == 0)):
# self.log_depth_img(gt_dep=sample["bins"], rec_dep=dep_re, img_title_prefix='Train - ')
if((batch_idx % 250) == 0):
self.print_logger.info("Train - Epoch {} - Batch {} - Step {}".format(self.current_epoch, batch_idx, self.global_step))
self.print_logger.info(" RMSE: {:.4f}, L1: {:.4f}, KL: {:.4f}, TV: {:.4f}".format(rmse, loss_l1, loss_kl, self.p_tv*loss_tv))
return {'loss': loss}
def validation_step(self, sample, batch_idx):
# Forward pass
M_mea_re, rec = self.forward_wrapper(sample)
dep_re = self.rec2depth(rec)
# Compute Losses
val_losses = self.compute_losses(sample, M_mea_re, rec)
val_loss = val_losses['loss']
val_rmse = val_losses['rmse']
val_loss_l1 = val_losses['loss_l1']
val_loss_kl = val_losses['loss_kl']
val_loss_tv = val_losses['loss_tv']
# Log the losses
self.log("rmse/avg_val", val_rmse, prog_bar=True)
# Important NOTE: Newer version of lightning accumulate the val_loss for each batch and then take the mean at the end of the epoch
self.log_dict(
{
"loss/avg_val": val_loss
}
)
# Return depths
dep = sample["bins"]
if((batch_idx % 300) == 0):
self.print_logger.info("Validation - Epoch {} - Batch {} - Step {}".format(self.current_epoch, batch_idx, self.global_step))
self.print_logger.info(" RMSE: {:.4f}, L1: {:.4f}, KL: {:.4f}, TV: {:.4f}".format(val_rmse, val_loss_l1, val_loss_kl, self.p_tv*val_loss_tv))
return {'dep': dep, 'dep_re': dep_re}
def get_sample_spad_data_ids(self, sample):
'''
Get the sample ids. This function was created mainly to be able to change the ids depending on the processing mode
'''
dataloader_idx = 0 # If multiple dataloaders are available you need to change this using the input args to the test_step function
curr_dataloader = self.trainer.test_dataloaders[dataloader_idx]
spad_data_ids = []
for idx in sample['idx']:
spad_data_ids.append(curr_dataloader.dataset.get_spad_data_sample_id(idx))
if(self.emulate_int8_quantization):
for i in range(len(spad_data_ids)):
spad_data_ids[i] = 'quantized_{}'.format(spad_data_ids[i])
return spad_data_ids
def test_step(self, sample, batch_idx):
# Forward pass
M_mea_re, rec = self.forward_wrapper(sample)
# crop the recovered histogram image and compute rec again
# we do this because if we test with inputs that are not divisible by the encoding kernel dimensions then they were padded and here we want to remove the padding so that dimensions will match the original input dimensions
# NOTE: right now this is not done during training and validation but eventually it should be done
M_gt = sample["rates"] # use M_gt to figure out the original input dimensions
M_mea_re = M_mea_re[...,0:M_gt.shape[-3],0:M_gt.shape[-2],0:M_gt.shape[-1]]
# recompute rec
rec = hist2rec_soft_argmax(M_mea_re)
# Compute depths
dep_re = self.rec2depth(rec)
# Compute Losses
test_losses = self.compute_losses(sample, M_mea_re, rec)
test_loss = test_losses['loss']
test_rmse = test_losses['rmse']
## Save some model outputs
# Access dataloader to get some metadata for computation
dataloader_idx = 0 # If multiple dataloaders are available you need to change this using the input args to the test_step function
curr_dataloader = self.trainer.test_dataloaders[dataloader_idx]
# Get tof params to compute depths
tres = curr_dataloader.dataset.tres_ps*1e-12
nt = M_mea_re.shape[-3]
tau = nt*tres
### Save model outputs in a folder with the dataset name and with a filename equal to the train data filename
# First get dataloader to generate the data ids
spad_data_ids = self.get_sample_spad_data_ids(sample)
out_rel_dirpath = os.path.dirname(spad_data_ids[0])
if(not os.path.exists(out_rel_dirpath)):
os.makedirs(out_rel_dirpath, exist_ok=True)
for i in range(dep_re.shape[0]):
out_data_fpath = spad_data_ids[i]
np.savez(out_data_fpath, dep_re=dep_re[i,:].cpu().numpy())
# Load GT depths
dep = sample["bins"]
# Compute depths and RMSE on depths
rec_depths = tof_utils.bin2depth(dep_re*nt, num_bins=nt, tau=tau)
gt_depths = tof_utils.bin2depth(dep*nt, num_bins=nt, tau=tau)
# the following two lines give the same result
# depths_rmse = torch.sqrt(torch.mean((rec_depths - gt_depths)**2))
depths_mse = criterion_L2(rec_depths, gt_depths)
depths_rmse = criterion_RMSE(rec_depths, gt_depths)
depths_mae = criterion_L1(rec_depths, gt_depths)
percentiles = [0.5, 0.75, 0.95, 0.99]
(mean_percentile_errs, _) = calc_mean_percentile_errors(np.abs(rec_depths.cpu().numpy()-gt_depths.cpu().numpy()), percentiles=percentiles)
# Important NOTE: Newer version of lightning accumulate the test_loss for each batch and then take the mean at the end of the epoch
# Log results
self.log_dict(
{
"loss/avg_test": test_loss
, "rmse/avg_test": test_rmse
, "depths/test_rmse": depths_rmse
, "depths/test_mae": depths_mae
, "depths/test_mse": depths_mse
, "depths/test_mean_abs_perc{:.2f}".format(percentiles[0]): mean_percentile_errs[0]
, "depths/test_mean_abs_perc{:.2f}".format(percentiles[1]): mean_percentile_errs[1]
, "depths/test_mean_abs_perc{:.2f}".format(percentiles[2]): mean_percentile_errs[2]
, "depths/test_mean_abs_perc{:.2f}".format(percentiles[3]): mean_percentile_errs[3]
}
)
return {'dep': dep, 'dep_re': dep_re}
def log_depth_img(self, gt_dep, rec_dep, img_title_prefix):
n_img_per_row = gt_dep.shape[0]
# NOTE: By setting it to global step, we will log more images inside tensorboard, which may require more space
# If we set global_step to a constant, we will keep overwriting the images.
grid1 = torchvision.utils.make_grid(gt_dep, nrow=n_img_per_row, value_range=(0,1))
self.logger.experiment.add_image(img_title_prefix + 'GT Depths', grid1, global_step=self.global_step)
grid2 = torchvision.utils.make_grid(rec_dep, nrow=n_img_per_row, value_range=(0,1))
self.logger.experiment.add_image(img_title_prefix + 'Rec. Depths', grid2, global_step=self.global_step)
## NOTE: The following pause statement helps avoid a random segfault that happens when logging the images inside
# training step
plt.pause(0.1)
def validation_epoch_end(self, outputs):
'''
Important NOTE: In newer lightning versions, for single value metrix like val_loss, we can just add them to the log_dict at val_step
and lightning will aggregate them correctly.
'''
# Stack some of the images from the outputs
dep = outputs[-1]['dep']
dep_re = outputs[-1]['dep_re']
n_samples = min(3, len(outputs))
dep_all = torch.zeros((n_samples, 1, dep.shape[-2], dep.shape[-1])).type(dep.dtype)
dep_re_all = torch.zeros((n_samples, 1, dep_re.shape[-2], dep_re.shape[-1])).type(dep_re.dtype)
for i in range(n_samples):
dep_all[i,:] = outputs[i]['dep'][0,:] # Grab first img in batch
dep_re_all[i,:] = outputs[i]['dep_re'][0,:]
self.log_depth_img(dep_all, dep_re_all, img_title_prefix='')
# # NOTE: By setting it to global step, we will log more images inside tensorboard, which may require more space
# # If we set global_step to a constant, we will keep overwriting the images.
# grid = torchvision.utils.make_grid(dep_all, nrow=n_samples, value_range=(0,1))
# self.logger.experiment.add_image('GT Depths', grid, global_step=self.global_step)
# grid = torchvision.utils.make_grid(dep_re_all, nrow=n_samples, value_range=(0,1))
# self.logger.experiment.add_image('Rec. Depths', grid, global_step=self.global_step)
def on_train_epoch_end(self) -> None:
print("")
return super().on_train_epoch_start()
def on_validation_epoch_end(self) -> None:
print("")
return super().on_validation_epoch_end()
def on_train_start(self):
# Proper logging of hyperparams and metrics in TB
# self.logger.log_hyperparams(self.hparams, {"loss/train": 0, "loss/avg_val": 0, "rmse/train": 0, "rmse/avg_val": 0})
self.logger.log_hyperparams(self.hparams)
def enable_quantization_emulation(self):
'''
Do nothing if not implemented
'''
self.emulate_int8_quantization = True
return
class LITL1LossBaseSpadModel(LITBaseSPADModel):
'''
Same as BaseSpadModel, but we use L1 instead of KLDiv loss
'''
def __init__(self,
backbone_net,
init_lr = 1e-4,
p_tv = 1e-5,
lr_decay_gamma = 0.9):
super(LITL1LossBaseSpadModel, self).__init__(
backbone_net=backbone_net,
init_lr = init_lr,
p_tv = p_tv,
lr_decay_gamma = lr_decay_gamma,
data_loss_id = 'L1'
)