-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathnn_ds_dataset.py
402 lines (331 loc) · 17.4 KB
/
nn_ds_dataset.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
#!/usr/bin/env python3
from numpy import float64
__copyright__ = "Copyright 2018, Elphel, Inc."
__license__ = "GPL-3.0+"
__email__ = "andrey@elphel.com"
from PIL import Image
import os
import sys
import glob
import pack_tile as pile
import numpy as np
import itertools
import time
import matplotlib.pyplot as plt
#http://stackoverflow.com/questions/287871/print-in-terminal-with-colors-using-python
TIME_START = time.time()
TIME_LAST = TIME_START
DEBUG_LEVEL= 1
DISP_BATCH_BINS = 20 # Number of batch disparity bins
STR_BATCH_BINS = 10 # Number of batch strength bins
FILES_PER_SCENE = 5 # number of random offset files for the scene to select from (0 - use all available)
MIN_BATCH_CHOICES = 10 # minimal number of tiles in a file for each bin to select from
MAX_BATCH_FILES = 10 #maximal number of files to use in a batch
MAX_EPOCH = 500
LR = 3e-3 # learning rate
USE_CONFIDENCE = False
ABSOLUTE_DISPARITY = False
DEBUG_PLT_LOSS = True
FEATURES_PER_TILE = 324
EPOCHS_TO_RUN = 20 #0
RUN_TOT_AVG = 100 # last batches to average. Epoch is 307 training batches
#DEBUG_PACK_TILES = True
class bcolors:
HEADER = '\033[95m'
OKBLUE = '\033[94m'
OKGREEN = '\033[92m'
WARNING = '\033[38;5;214m'
FAIL = '\033[91m'
ENDC = '\033[0m'
BOLD = '\033[1m'
BOLDWHITE = '\033[1;37m'
UNDERLINE = '\033[4m'
def print_time(txt="",end="\n"):
global TIME_LAST
t = time.time()
if txt:
txt +=" "
print(("%s"+bcolors.BOLDWHITE+"at %.4fs (+%.4fs)"+bcolors.ENDC)%(txt,t-TIME_START,t-TIME_LAST), end = end)
TIME_LAST = t
#reading to memory (testing)
def readTFRewcordsEpoch(train_filename):
# filenames = [train_filename]
# dataset = tf.data.TFRecordDataset(filenames)
if not '.tfrecords' in train_filename:
train_filename += '.tfrecords'
record_iterator = tf.python_io.tf_record_iterator(path=train_filename)
corr2d_list=[]
target_disparity_list=[]
gt_ds_list = []
for string_record in record_iterator:
example = tf.train.Example()
example.ParseFromString(string_record)
corr2d_list.append(np.array(example.features.feature['corr2d'] .float_list .value))
target_disparity_list.append(np.array(example.features.feature['target_disparity'] .float_list .value[0]))
gt_ds_list.append(np.array(example.features.feature['gt_ds'] .float_list .value))
corr2d= np.array(corr2d_list)
target_disparity = np.array(target_disparity_list)
gt_ds = np.array(gt_ds_list)
return corr2d, target_disparity, gt_ds
#from http://warmspringwinds.github.io/tensorflow/tf-slim/2016/12/21/tfrecords-guide/
def read_and_decode(filename_queue):
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized_example,
# Defaults are not specified since both keys are required.
features={
'corr2d': tf.FixedLenFeature([324],tf.float32), #string),
'target_disparity': tf.FixedLenFeature([1], tf.float32), #.string),
'gt_ds': tf.FixedLenFeature([2], tf.float32) #.string)
})
corr2d = features['corr2d'] # tf.decode_raw(features['corr2d'], tf.float32)
target_disparity = features['target_disparity'] # tf.decode_raw(features['target_disparity'], tf.float32)
gt_ds = tf.cast(features['gt_ds'], tf.float32) # tf.decode_raw(features['gt_ds'], tf.float32)
in_features = tf.concat([corr2d,target_disparity],0)
# still some nan-s in correlation data?
# in_features_clean = tf.where(tf.is_nan(in_features), tf.zeros_like(in_features), in_features)
# corr2d_out, target_disparity_out, gt_ds_out = tf.train.shuffle_batch( [in_features_clean, target_disparity, gt_ds],
corr2d_out, target_disparity_out, gt_ds_out = tf.train.shuffle_batch( [in_features, target_disparity, gt_ds],
batch_size=1000, # 2,
capacity=30,
num_threads=2,
min_after_dequeue=10)
return corr2d_out, target_disparity_out, gt_ds_out
#http://adventuresinmachinelearning.com/introduction-tensorflow-queuing/
#Main code
try:
train_filenameTFR = sys.argv[1]
except IndexError:
train_filenameTFR = "/mnt/dde6f983-d149-435e-b4a2-88749245cc6c/home/eyesis/x3d_data/data_sets/tf_data/train.tfrecords"
#FILES_PER_SCENE
print_time("Importing TensorCrawl")
import tensorflow as tf
import tensorflow.contrib.slim as slim
print_time("TensorCrawl imported")
result_dir = './result/'
checkpoint_dir = './result/'
save_freq = 500
def lrelu(x):
return tf.maximum(x*0.2,x)
# return tf.nn.relu(x)
def network(input):
# fc1 = slim.fully_connected(input, 512, activation_fn=lrelu,scope='g_fc1')
# fc2 = slim.fully_connected(fc1, 512, activation_fn=lrelu,scope='g_fc2')
fc3 = slim.fully_connected(input, 256, activation_fn=lrelu,scope='g_fc3')
fc4 = slim.fully_connected(fc3, 128, activation_fn=lrelu,scope='g_fc4')
fc5 = slim.fully_connected(fc4, 64, activation_fn=lrelu,scope='g_fc5')
if USE_CONFIDENCE:
fc6 = slim.fully_connected(fc5, 2, activation_fn=lrelu,scope='g_fc6')
else:
fc6 = slim.fully_connected(fc5, 1, activation_fn=None,scope='g_fc6')
#If using residual disparity, split last layer into 2 or remove activation and add rectifier to confidence only
return fc6
def batchLoss(out_batch, # [batch_size,(1..2)] tf_result
target_disparity_batch, # [batch_size] tf placeholder
gt_ds_batch, # [batch_size,2] tf placeholder
absolute_disparity = True, #when false there should be no activation on disparity output !
use_confidence = True,
lambda_conf_avg = 0.01,
lambda_conf_pwr = 0.1,
conf_pwr = 2.0,
gt_conf_offset = 0.08,
gt_conf_pwr = 1.0):
"""
Here confidence should be after relU. Disparity - may be also if absolute, but no activation if output is residual disparity
"""
tf_lambda_conf_avg = tf.constant(lambda_conf_avg, dtype=tf.float32, name="tf_lambda_conf_avg")
tf_lambda_conf_pwr = tf.constant(lambda_conf_pwr, dtype=tf.float32, name="tf_lambda_conf_pwr")
tf_conf_pwr = tf.constant(conf_pwr, dtype=tf.float32, name="tf_conf_pwr")
tf_gt_conf_offset = tf.constant(gt_conf_offset, dtype=tf.float32, name="tf_gt_conf_offset")
tf_gt_conf_pwr = tf.constant(gt_conf_pwr, dtype=tf.float32, name="tf_gt_conf_pwr")
tf_num_tiles = tf.shape(gt_ds_batch)[0]
tf_0f = tf.constant(0.0, dtype=tf.float32, name="tf_0f")
tf_1f = tf.constant(1.0, dtype=tf.float32, name="tf_1f")
tf_maxw = tf.constant(1.0, dtype=tf.float32, name="tf_maxw")
if gt_conf_pwr == 0:
w = tf.ones((out_batch.shape[0]), dtype=tf.float32,name="w_ones")
else:
# w_slice = tf.slice(gt_ds_batch,[0,1],[-1,1], name = "w_gt_slice")
w_slice = tf.reshape(gt_ds_batch[:,1],[-1], name = "w_gt_slice")
w_sub = tf.subtract (w_slice, tf_gt_conf_offset, name = "w_sub")
# w_clip = tf.clip_by_value(w_sub, tf_0f,tf_maxw, name = "w_clip")
w_clip = tf.maximum(w_sub, tf_0f, name = "w_clip")
if gt_conf_pwr == 1.0:
w = w_clip
else:
w=tf.pow(w_clip, tf_gt_conf_pwr, name = "w")
if use_confidence:
tf_num_tilesf = tf.cast(tf_num_tiles, dtype=tf.float32, name="tf_num_tilesf")
# conf_slice = tf.slice(out_batch,[0,1],[-1,1], name = "conf_slice")
conf_slice = tf.reshape(out_batch[:,1],[-1], name = "conf_slice")
conf_sum = tf.reduce_sum(conf_slice, name = "conf_sum")
conf_avg = tf.divide(conf_sum, tf_num_tilesf, name = "conf_avg")
conf_avg1 = tf.subtract(conf_avg, tf_1f, name = "conf_avg1")
conf_avg2 = tf.square(conf_avg1, name = "conf_avg2")
cost2 = tf.multiply (conf_avg2, tf_lambda_conf_avg, name = "cost2")
iconf_avg = tf.divide(tf_1f, conf_avg, name = "iconf_avg")
nconf = tf.multiply (conf_slice, iconf_avg, name = "nconf") #normalized confidence
nconf_pwr = tf.pow(nconf, conf_pwr, name = "nconf_pwr")
nconf_pwr_sum = tf.reduce_sum(nconf_pwr, name = "nconf_pwr_sum")
nconf_pwr_offs = tf.subtract(nconf_pwr_sum, tf_1f, name = "nconf_pwr_offs")
cost3 = tf.multiply (conf_avg2, nconf_pwr_offs, name = "cost3")
w_all = tf.multiply (w, nconf, name = "w_all")
else:
w_all = w
cost2 = 0.0
cost3 = 0.0
# normalize weights
w_sum = tf.reduce_sum(w_all, name = "w_sum")
iw_sum = tf.divide(tf_1f, w_sum, name = "iw_sum")
w_norm = tf.multiply (w_all, iw_sum, name = "w_norm")
# disp_slice = tf.slice(out_batch,[0,0],[-1,1], name = "disp_slice")
# d_gt_slice = tf.slice(gt_ds_batch,[0,0],[-1,1], name = "d_gt_slice")
disp_slice = tf.reshape(out_batch[:,0],[-1], name = "disp_slice")
d_gt_slice = tf.reshape(gt_ds_batch[:,0],[-1], name = "d_gt_slice")
if absolute_disparity:
out_diff = tf.subtract(disp_slice, d_gt_slice, name = "out_diff")
else:
td_flat = tf.reshape(target_disparity_batch,[-1], name = "td_flat")
residual_disp = tf.subtract(d_gt_slice, td_flat, name = "residual_disp")
out_diff = tf.subtract(disp_slice, residual_disp, name = "out_diff")
out_diff2 = tf.square(out_diff, name = "out_diff2")
out_wdiff2 = tf.multiply (out_diff2, w_norm, name = "out_wdiff2")
cost1 = tf.reduce_sum(out_wdiff2, name = "cost1")
if use_confidence:
cost12 = tf.add(cost1, cost2, name = "cost12")
cost123 = tf.add(cost12, cost3, name = "cost123")
return cost123, disp_slice, d_gt_slice, out_diff,out_diff2, w_norm, out_wdiff2, cost1
else:
return cost1, disp_slice, d_gt_slice, out_diff,out_diff2, w_norm, out_wdiff2, cost1
filename_queue = tf.train.string_input_producer(
[train_filenameTFR], num_epochs = EPOCHS_TO_RUN) #0)
# Even when reading in multiple threads, share the filename
# queue.
corr2d325, target_disparity, gt_ds = read_and_decode(filename_queue)
# The op for initializing the variables.
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
#sess = tf.Session()
"""
in_tile = tf.placeholder(tf.float32,[None,9 * 9 * 4 + 1])
gt = tf.placeholder(tf.float32,[None,2])
target_d = tf.placeholder(tf.float32,[None])
out = network(in_tile)
"""
out = network(corr2d325)
#Try standard loss functions first
G_loss, _disp_slice, _d_gt_slice, _out_diff, _out_diff2, _w_norm, _out_wdiff2, _cost1 = batchLoss(out_batch = out, # [batch_size,(1..2)] tf_result
target_disparity_batch= target_disparity, ### target_d, # [batch_size] tf placeholder
gt_ds_batch = gt_ds, ### gt, # [batch_size,2] tf placeholder
absolute_disparity = ABSOLUTE_DISPARITY,
use_confidence = USE_CONFIDENCE, # True,
lambda_conf_avg = 0.01,
lambda_conf_pwr = 0.1,
conf_pwr = 2.0,
gt_conf_offset = 0.08,
gt_conf_pwr = 1.0)
t_vars=tf.trainable_variables()
lr=tf.placeholder(tf.float32)
G_opt=tf.train.AdamOptimizer(learning_rate=lr).minimize(G_loss)
saver=tf.train.Saver()
# ?!!!!!
#merged = tf.summary.merge_all()
#train_writer = tf.summary.FileWriter(result_dir + '/train', sess.graph)
#test_writer = tf.summary.FileWriter(result_dir + '/test')
#http://rtfcode.com/xref/tensorflow-1.4.1/tensorflow/docs_src/api_guides/python/reading_data.md
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
# sess.run(init_op) # Was reporting beta1 not initialized in Adam
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
writer = tf.summary.FileWriter('./attic/nn_ds_dataset_graph1', sess.graph)
writer.close()
# for i in range(1000):
loss_hist = np.zeros(RUN_TOT_AVG, dtype=np.float32)
i = 0
try:
while not coord.should_stop():
print_time("%d: Run "%(i), end = "")
_,G_current,output, disp_slice, d_gt_slice, out_diff, out_diff2, w_norm, out_wdiff2, out_cost1, corr2d325_out, target_disparity_out, gt_ds_out = sess.run(
[G_opt,G_loss,out,_disp_slice, _d_gt_slice, _out_diff, _out_diff2, _w_norm, _out_wdiff2, _cost1, corr2d325, target_disparity, gt_ds],
feed_dict={lr: LR})
# print_time("loss=%f, running average=%f"%(G_current,mean_loss))
loss_hist[i % RUN_TOT_AVG] = G_current
if (i < RUN_TOT_AVG):
loss_avg = np.average(loss_hist[:i])
else:
loss_avg = np.average(loss_hist)
print_time("loss=%f, running average=%f"%(G_current,loss_avg))
# print ("%d: corr2d_out.shape="%(i),corr2d325_out.shape)
## print ("target_disparity_out.shape=",target_disparity_out.shape)
## print ("gt_ds_out.shape=",gt_ds_out.shape)
i += 1
except tf.errors.OutOfRangeError:
print('Done training -- epoch limit reached')
finally:
# When done, ask the threads to stop.
coord.request_stop()
coord.join(threads)
#sess.close() ('whith' does that)
'''
ckpt=tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt:
print('loaded '+ckpt.model_checkpoint_path)
saver.restore(sess,ckpt.model_checkpoint_path)
allfolders = glob.glob('./result/*0')
lastepoch = 0
for folder in allfolders:
lastepoch = np.maximum(lastepoch, int(folder[-4:]))
recorded_loss = []
recorded_mean_loss = []
recorded_gt_d = []
recorded_gt_c = []
recorded_pr_d = []
recorded_pr_c = []
LR = 1e-3
print(bcolors.HEADER+"Last Epoch = "+str(lastepoch)+bcolors.ENDC)
if DEBUG_PLT_LOSS:
plt.ion() # something about plotting
plt.figure(1, figsize=(4,12))
pass
training_tiles = np.array([])
training_values = np.array([])
graph_saved = False
for epoch in range(20): #MAX_EPOCH):
print_time("epoch="+str(epoch))
train_seed_list = np.arange(len(ex_data.files_train))
np.random.shuffle(train_seed_list)
g_loss = np.zeros(len(train_seed_list))
for nscene, seed_index in enumerate(train_seed_list):
corr2d_batch, target_disparity_batch, gt_ds_batch = ex_data.prepareBatchData(seed_index)
num_tiles = corr2d_batch.shape[0] # 1000
num_tile_slices = corr2d_batch.shape[1] # 4
num_cell_in_slice = corr2d_batch.shape[2] # 81
in_data = np.empty((num_tiles, num_tile_slices*num_cell_in_slice + 1), dtype = np.float32)
in_data[...,0:num_tile_slices*num_cell_in_slice] = corr2d_batch.reshape((corr2d_batch.shape[0],corr2d_batch.shape[1]*corr2d_batch.shape[2]))
in_data[...,num_tile_slices*num_cell_in_slice] = target_disparity_batch
st=time.time()
#run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
#run_metadata = tf.RunMetadata()
#_,G_current,output = sess.run([G_opt,G_loss,out],feed_dict={in_tile:input_patch,gt:gt_patch,lr:LR},options=run_options,run_metadata=run_metadata)
print_time("%d:%d Run "%(epoch, nscene), end = "")
_,G_current,output, disp_slice, d_gt_slice, out_diff, out_diff2, w_norm = sess.run([G_opt,G_loss,out,_disp_slice, _d_gt_slice, _out_diff, _out_diff2, _w_norm],
feed_dict={in_tile: in_data,
gt: gt_ds_batch,
target_d: target_disparity_batch,
lr: LR})
if not graph_saved:
writer = tf.summary.FileWriter('./attic/nn_ds_single_graph1', sess.graph)
writer.close()
graph_saved = True
# exit(0)
g_loss[nscene]=G_current
mean_loss = np.mean(g_loss[np.where(g_loss)])
print_time("loss=%f, running average=%f"%(G_current,mean_loss))
pass
'''
#if wait_and_show: # wait and show images
# plt.show()
print_time("All done, exiting...")