-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
283 lines (249 loc) · 15 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
import os
from losses.cable import CableBSplineLoss, CableSeqLoss
from models.separated_cnn_neural_predictor import SeparatedCNNNeuralPredictor
from models.separated_neural_predictor import SeparatedNeuralPredictor
from utils.bspline import BSpline
from utils.constants import BSplineConstants
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
from utils.dataset import _ds, prepare_dataset, whiten, mix_datasets
from utils.execution import ExperimentHandler
from models.basic_neural_predictor import BasicNeuralPredictor
np.random.seed(444)
# physical_devices = tf.config.experimental.list_physical_devices('GPU')
# assert len(physical_devices) > 0, "Not enough GPU hardware devices available"
# config = tf.config.experimental.set_memory_growth(physical_devices[0], True)
class args:
# batch_size = 128
#batch_size = 64
batch_size = 32
working_dir = './trainings'
# out_name = 'short_ds_bs16_lr5em5_l5x1024_scaledincms_regloss1em4'
# out_name = 'yz_bs16_lr5em5_l5x256_cm_regloss1em4_bs_keq_dsseparatedbutsorted_traindata10'
#out_name = 'xyzrpy_02_09__10_30_bs16_lr5em5_separated_l1x256_l3x256_diffs_mwhithened_regloss1em4_bs_keq'
# out_name = 'xyzrpy_02_09__10_30_bs16_lr5em5_l5x256_m_regloss1em4_bs_keq'
#out_name = 'xyzrpy_episodic_all2all_02_09__15_00_dxyzlg5cm_bs64_lr5em5_l5x256_m_regloss0em4_bs_keq_dsmixed'
#out_name = 'xyzrpy_episodic_all2all_02_09__15_00_dxyzlg5cm_bs64_lr5em5_separated_l1x256_l3x256_m_regloss0em4_bs_keq_len_yz_curv_yz'
#out_name = 'xyzrpy_episodic_all2all_02_10__09_15_bs32_lr5em5_separated_l1x256_l3x256_m_regloss0em4_bs_keq_dsnotmixed_length_loss_1em1'
#out_name = 'xyzrpy_episodic_all2all_02_10__10_00_bs32_lr5em5_separated_l1x256_l3x256_m_regloss0em4_bs_keq_dsnotmixed'
#out_name = 'xyzrpy_episodic_all2all_02_10__10_20_bs32_lr5em5_separated_l1x256_l3x256_m_regloss0em4_bs_keq_dsnotmixed'
#out_name = 'xyzrpy_episodic_all2all_02_10__14_00_bs32_lr5em5_separated_l1x256_l3x256_m_regloss0em4_bs_keq_dsnotmixed_absloss_withened_'
out_name = 'xyzrpy_episodic_all2all_02_10__14_00_bs32_lr5em5_separated_l1x256_l3x256_cable_diff_m_regloss0em4_bs_keq_dsnotmixed_absloss_withened'
#out_name = 'xyzrpy_episodic_all2all_02_10__14_00_bs32_lr5em5_separated_cablecnn_l1x128_l2x128_outputcnn_m_regloss0em4_bs_keq_dsnotmixed_absloss_withened'
#out_name = 'xyzrpy_all2all_02_10__14_00_bs32_lr5em5_separated_l1x256_l3x256_m_regloss0em4_bs_keq_dsmixed_absloss_withened'
#out_name = 'test'
log_interval = 100
learning_rate = 5e-5
l2reg = 0e-4
len_loss = 0
acc_loss = 0e-1
# dataset_path = "./data/prepared_datasets/xy_pts256/train.tsv"
# dataset_path = "./data/prepared_datasets/xy_bs/train.tsv"
# dataset_path = "./data/prepared_datasets/yz_big_pts256/train.tsv"
# dataset_path = "./data/prepared_datasets/yz_big_keq_sorted/train.tsv"
# dataset_path = "./data/prepared_datasets/xyzrpy/train.tsv"
#dataset_path = "./data/prepared_datasets/xyzrpy_02_09__10_30/train.tsv"
# dataset_path = "./data/prepared_datasets/xyzrpy_episodic_all2all_02_09__11_00/train.tsv"
#dataset_path = "./data/prepared_datasets/xyzrpy_episodic_all2all_02_09__13_30/train.tsv"
#dataset_path = "./data/prepared_datasets/xyzrpy_episodic_all2all_02_09__15_00_dxyzlg5cm/train.tsv"
#dataset_path = "./data/prepared_datasets/xyzrpy_episodic_all2all_02_10__09_00/train.tsv"
#dataset_path = "./data/prepared_datasets/xyzrpy_episodic_all2all_02_10__09_15/train.tsv"
#dataset_path = "./data/prepared_datasets/xyzrpy_episodic_all2all_02_10__10_00/train.tsv"
#dataset_path = "./data/prepared_datasets/xyzrpy_episodic_all2all_02_10__10_20/train.tsv"
#dataset_path = "./data/prepared_datasets/xyzrpy_episodic_all2all_02_10__10_40/train.tsv"
dataset_path = "./data/prepared_datasets/xyzrpy_episodic_all2all_02_10__14_00/train.tsv"
#dataset_path = "./data/prepared_datasets/xyzrpy_all2all_02_10__14_00/train.tsv"
# dataset_path = "./data/prepared_datasets/yz_big_keq/train.tsv"
# dataset_path = "./data/prepared_datasets/yz_big_keq_n1000/train.tsv"
# dataset_path = "./data/prepared_datasets/xy_bs_keq/train.tsv"
# dataset_path = "./data/prepared_datasets/short_ds/train.tsv"
train_ds, train_size, tX1, tX2, tX3, tY = prepare_dataset(args.dataset_path) # , n=10)
val_ds, val_size, vX1, vX2, vX3, vY = prepare_dataset(args.dataset_path.replace("train", "val"))
#tX1, tX2, tX3, tY, vX1, vX2, vX3, vY, train_size, val_size = mix_datasets(tX1, tX2, tX3, tY, vX1, vX2, vX3, vY)
tX1, vX1, m1, s1 = whiten(tX1, vX1)
tX2, vX2, m2, s2 = whiten(tX2, vX2)
tX3, vX3, m3, s3 = whiten(tX3, vX3)
#m3, s3 = 0., 1.
tY, vY, my, sy = whiten(tY, vY)
train_ds = tf.data.Dataset.from_tensor_slices({"x1": tX1, "x2": tX2, "x3": tX3, "y": tY})
val_ds = tf.data.Dataset.from_tensor_slices({"x1": vX1, "x2": vX2, "x3": vX3, "y": vY})
bsp = BSpline(25, 3)
bsp32 = BSpline(25, 3, num_T_pts=32)
# dttX = np.abs(tX[np.newaxis] - tX[:, np.newaxis])
# dtvX = np.abs(tX[np.newaxis] - vX[:, np.newaxis])
# nttX = np.linalg.norm(tX[np.newaxis] - tX[:, np.newaxis], axis=-1)
# ntvX = np.linalg.norm(tX[np.newaxis] - vX[:, np.newaxis], axis=-1)
#
# ncp = BSplineConstants.ncp
# vid = 0
# closest_idx = np.argmin(ntvX[vid])
# cpt = tX[closest_idx, 42:42 + ncp]
# cpt = cpt.reshape((BSplineConstants.n, BSplineConstants.dim))
# cpv = vX[vid, 42:42 + ncp]
# cpv = cpv.reshape((BSplineConstants.n, BSplineConstants.dim))
#
# plt.subplot(121)
# plt.plot(cpv[:, 1], cpv[:, 2], 'bo')
# plt.plot(cpt[:, 1], cpt[:, 2], 'g^')
# plt.subplot(122)
# v_t = (bsp.N[0] @ cpt)
# v_v = (bsp.N[0] @ cpv)
# plt.plot(v_t[:, 1], v_t[:, 2], label="train")
# plt.plot(v_v[:, 1], v_v[:, 2], label="val")
# plt.legend()
# plt.show()
#
#tX, tY, vX, vY, train_size, val_size = mix_datasets(tX, tY, vX, vY)
#train_ds = tf.data.Dataset.from_tensor_slices({"x": tX, "y": tY})
#val_ds = tf.data.Dataset.from_tensor_slices({"x": vX, "y": vY})
#tX1, tX2, tX3, tY, vX1, vX2, vX3, vY, train_size, val_size = mix_datasets(tX1, tX2, tX3, tY, vX1, vX2, vX3, vY)
#train_ds = tf.data.Dataset.from_tensor_slices({"x1": tX1, "x2": tX2, "x3": tX3, "y": tY})
#val_ds = tf.data.Dataset.from_tensor_slices({"x1": vX1, "x2": vX2, "x3": vX3, "y": vY})
opt = tf.keras.optimizers.Adam(args.learning_rate)
#loss = tf.keras.losses.mean_squared_error
#def loss(gt, pred, mul=1.):
# gt = mul * tf.reshape(gt, (-1, BSplineConstants.n, BSplineConstants.dim))
# pred = mul * tf.reshape(pred, (-1, BSplineConstants.n, BSplineConstants.dim))
# #plt.plot(pred[0, :, 0], pred[0, :, 1])
# #plt.plot(gt[0, :, 0], gt[0, :, 1])
# #plt.show()
# diff = tf.reduce_sum(tf.square(gt - pred), axis=-1)
# control_pts_loss = tf.reduce_mean(diff, axis=-1)
# vgt = bsp.N @ gt
# dvgt = bsp.dN @ gt
# ddvgt = bsp.ddN @ gt
# vpred = bsp.N @ pred
# dvpred = bsp.dN @ pred
# ddvpred = bsp.ddN @ pred
# len_gt = tf.reduce_sum(tf.sqrt(tf.reduce_sum(tf.square(vgt[:, 1:, 1:] - vgt[:, :-1, 1:]), axis=-1)), axis=-1)
# len_pred = tf.reduce_sum(tf.sqrt(tf.reduce_sum(tf.square(vpred[:, 1:, 1:] - vpred[:, :-1, 1:]), axis=-1)), axis=-1)
# length_loss = tf.square(len_gt - len_pred)
# curv_yz_gt = (dvgt[..., 1] * ddvgt[..., 2] - ddvgt[..., 1] * dvgt[..., 2]) / tf.sqrt(dvgt[..., 1] ** 2 + dvgt[..., 2] ** 2 + 1e-8)
# curv_yz_pred = (dvpred[..., 1] * ddvpred[..., 2] - ddvpred[..., 1] * dvpred[..., 2]) / tf.sqrt(dvpred[..., 1] ** 2 + dvpred[..., 2] ** 2 + 1e-8)
# dyz_gt = tf.sqrt(tf.reduce_sum(tf.square(vgt[:, 1:, 1:] - vgt[:, :-1, 1:]), axis=-1))
# dyz_pred = tf.sqrt(tf.reduce_sum(tf.square(vpred[:, 1:, 1:] - vpred[:, :-1, 1:]), axis=-1))
# acccurv_yz_gt = tf.reduce_sum(tf.abs(curv_yz_gt[:, :-1]) * dyz_gt, axis=-1)
# acccurv_yz_pred = tf.reduce_sum(tf.abs(curv_yz_pred[:, :-1]) * dyz_pred, axis=-1)
# acccurv_yz_loss = tf.square(acccurv_yz_gt - acccurv_yz_pred)
# return control_pts_loss, length_loss, acccurv_yz_loss
loss = CableBSplineLoss()
loss_pts = CableSeqLoss()
model = BasicNeuralPredictor()
#model = SeparatedNeuralPredictor()
#model = SeparatedCNNNeuralPredictor()
experiment_handler = ExperimentHandler(args.working_dir, args.out_name, args.log_interval, model, opt)
# experiment_handler.restore("./trainings/short_ds_bs16_lr5em5_run2/checkpoints/best-115")
# experiment_handler.restore("./trainings/short_ds_bs16_lr5em5_fixed_small_cable_changes/checkpoints/best-95")
# experiment_handler.restore("./trainings/short_ds_bs16_lr5em5_l5x1024_scaledincms_regloss1em4/checkpoints/last_n-9")
# experiment_handler.restore("./trained_models/last_n-274")
# experiment_handler.restore("./trainings/xy_bs8_lr5em5_l5x256_scaledincms_regloss1em4/checkpoints/last_n-34")
# experiment_handler.restore("./trainings/xy_bs8_lr5em5_l5x256_scaledincms_regloss1em4_fixedmul/checkpoints/last_n-162")
train_step = 0
val_step = 0
best_epoch_loss = 1e10
best_unscaled_epoch_loss = 1e10
mul = 1.
for epoch in range(30000):
# training
dataset_epoch = train_ds.shuffle(train_size)
dataset_epoch = dataset_epoch.batch(args.batch_size).prefetch(args.batch_size)
epoch_loss = []
prediction_losses = []
diff_ratios = []
pts_losses = []
experiment_handler.log_training()
for i, rotation, translation, cable, y_gt in _ds('Train', dataset_epoch, train_size, epoch, args.batch_size):
with tf.GradientTape(persistent=True) as tape:
y_base = cable * s3 + m3
y_pred = model(rotation, translation, cable, training=True)
y_pred = y_pred * sy + my
y_gt = y_gt * sy + my
prediction_loss, length_loss, accurv_yz_loss = loss(y_gt + y_base, y_pred + y_base)
gt_pts = bsp32.N @ tf.reshape(y_gt + y_base, (-1, BSplineConstants.n, BSplineConstants.dim))
pred_pts = bsp32.N @ tf.reshape(y_pred + y_base, (-1, BSplineConstants.n, BSplineConstants.dim))
pts_loss = loss_pts(gt_pts, pred_pts)
#prediction_loss_, length_loss = loss(y_gt + y_base, y_pred + y_base, 100.)
#prediction_loss = loss(mul * y_gt, mul * y_pred)
diff_ratio = tf.reduce_mean(tf.abs(y_gt - y_pred) / tf.abs(y_gt), axis=-1)
reg_loss = tf.add_n([tf.nn.l2_loss(v) for v in model.trainable_variables
if 'bias' not in v.name])
model_loss = prediction_loss + args.l2reg * reg_loss + args.len_loss * length_loss + args.acc_loss * accurv_yz_loss
grads = tape.gradient(model_loss, model.trainable_variables)
opt.apply_gradients(zip(grads, model.trainable_variables))
epoch_loss.append(model_loss)
prediction_losses.append(prediction_loss)
pts_losses.append(pts_loss)
diff_ratios.append(diff_ratio)
with tf.summary.record_if(train_step % args.log_interval == 0):
tf.summary.scalar('metrics/model_loss', tf.reduce_mean(model_loss), step=train_step)
tf.summary.scalar('metrics/pts_loss', tf.reduce_mean(pts_loss), step=train_step)
tf.summary.scalar('metrics/prediction_loss', tf.reduce_mean(prediction_loss), step=train_step)
tf.summary.scalar('metrics/length_loss', tf.reduce_mean(length_loss), step=train_step)
tf.summary.scalar('metrics/accurv_yz_loss', tf.reduce_mean(accurv_yz_loss), step=train_step)
tf.summary.scalar('metrics/diff_ratio', tf.reduce_mean(diff_ratio), step=train_step)
tf.summary.scalar('metrics/reg_loss', tf.reduce_mean(reg_loss), step=train_step)
train_step += 1
epoch_loss = tf.reduce_mean(tf.concat(epoch_loss, -1))
prediction_losses = tf.reduce_mean(tf.concat(prediction_losses, -1))
pts_losses = tf.reduce_mean(tf.concat(pts_losses, -1))
diff_ratios = tf.reduce_mean(tf.concat(diff_ratios, -1))
with tf.summary.record_if(True):
tf.summary.scalar('epoch/loss', epoch_loss, step=epoch)
tf.summary.scalar('epoch/prediction_loss', prediction_losses, step=epoch)
tf.summary.scalar('epoch/pts_loss', pts_losses, step=epoch)
tf.summary.scalar('epoch/diff_ratio', diff_ratios, step=epoch)
# validation
dataset_epoch = val_ds.shuffle(val_size)
dataset_epoch = dataset_epoch.batch(args.batch_size).prefetch(args.batch_size)
epoch_loss = []
prediction_losses = []
diff_ratios = []
pts_losses = []
experiment_handler.log_validation()
for i, rotation, translation, cable, y_gt in _ds('Val', dataset_epoch, val_size, epoch, args.batch_size):
y_base = cable * s3 + m3
y_pred = model(rotation, translation, cable, training=False)
y_gt = y_gt * sy + my
y_pred = y_pred * sy + my
prediction_loss, length_loss, accurv_yz_loss = loss(y_gt + y_base, y_pred + y_base)
gt_pts = bsp32.N @ tf.reshape(y_gt + y_base, (-1, BSplineConstants.n, BSplineConstants.dim))
pred_pts = bsp32.N @ tf.reshape(y_pred + y_base, (-1, BSplineConstants.n, BSplineConstants.dim))
pts_loss = loss_pts(gt_pts, pred_pts)
#prediction_loss_, length_loss = loss(y_gt + y_base, y_pred + y_base, 100.)
#prediction_loss, length_loss = loss(y_gt + y_base, y_pred + y_base)
# prediction_loss = loss(y_gt, y_pred)
#prediction_loss = loss(mul * y_gt, mul * y_pred)
diff_ratio = np.mean(np.abs(y_gt - y_pred) / np.abs(y_gt), axis=-1)
reg_loss = tf.add_n([tf.nn.l2_loss(v) for v in model.trainable_variables
if 'bias' not in v.name])
model_loss = prediction_loss + args.l2reg * reg_loss + args.len_loss * length_loss + args.acc_loss * accurv_yz_loss
epoch_loss.append(model_loss)
prediction_losses.append(prediction_loss)
pts_losses.append(pts_loss)
diff_ratios.append(diff_ratio)
with tf.summary.record_if(val_step % args.log_interval == 0):
tf.summary.scalar('metrics/model_loss', tf.reduce_mean(model_loss), step=val_step)
tf.summary.scalar('metrics/pts_loss', tf.reduce_mean(pts_loss), step=train_step)
tf.summary.scalar('metrics/prediction_loss', tf.reduce_mean(prediction_loss), step=val_step)
tf.summary.scalar('metrics/length_loss', tf.reduce_mean(length_loss), step=val_step)
tf.summary.scalar('metrics/accurv_yz_loss', tf.reduce_mean(accurv_yz_loss), step=val_step)
tf.summary.scalar('metrics/diff_ratio', tf.reduce_mean(diff_ratio), step=val_step)
tf.summary.scalar('metrics/reg_loss', tf.reduce_mean(reg_loss), step=val_step)
val_step += 1
epoch_loss = tf.reduce_mean(tf.concat(epoch_loss, -1))
prediction_losses = tf.reduce_mean(tf.concat(prediction_losses, -1))
pts_losses = tf.reduce_mean(tf.concat(pts_losses, -1))
diff_ratios = tf.reduce_mean(tf.concat(diff_ratios, -1))
with tf.summary.record_if(True):
tf.summary.scalar('epoch/loss', epoch_loss, step=epoch)
tf.summary.scalar('epoch/prediction_loss', prediction_losses, step=epoch)
tf.summary.scalar('epoch/pts_loss', pts_losses, step=epoch)
tf.summary.scalar('epoch/diff_ratio', diff_ratios, step=epoch)
w = 20
if epoch % w == w - 1:
experiment_handler.save_last()
if best_epoch_loss > epoch_loss:
best_epoch_loss = epoch_loss
experiment_handler.save_best()