forked from bhuvi3/camera_trap_animal_classification
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathhelper_commands.py
414 lines (267 loc) · 49 KB
/
helper_commands.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
Capstone Project Commands
# VM:
bhuvanvm: ssh [email protected]
cpumachine-1: ssh [email protected]
cpumachine-2: ssh [email protected]
gpumachine-1: ssh [email protected]
gpumachine-2: ssh [email protected]
# Mounting the disk:
dmesg | grep SCSI
sudo fdisk /dev/<>
n
p
q
sudo mkdir /datadrive
sudo mount /dev/<>1 /datadrive
sudo -i blkid
sudo vi /etc/fstab
UUID=<> /datadrive ext4 defaults,nofail 1 2
sudo reboot
ssh-keygen -t rsa -b 4096 -C "[email protected]"
cat .ssh/id_rsa.pub
# Connecting to Jupyter
with_ssh_tunnel_for_Jupyter: ssh -N -f -L localhost:8889:localhost:8889 [email protected]
stop connection: sudo netstat -lpn |grep :8889 # and kill the shown PID.
starting jupyter: jupyter notebook --no-browser --port=8889 --allow-root &> jupyter.log
cat jupyter.log
# Docker.
docker pull tensorflow/tensorflow:latest-gpu-py3-jupyter [one time] # tensorflow/tensorflow:2.0.0-gpu-py3-jupyter
docker run -itd --gpus all -p 8889:8889 -e USER_HOME=$HOME/vm -v /datadrive:/datadrive tensorflow/tensorflow:latest-gpu-py3-jupyter bash
docker exec -it 4e267b28e70f /bin/bash # Get the container_id: b619db070bb6, using docker ps.
cd ../datadrive/camera_trap_animal_classification/code
docker exec -it 3734e31aa5a6 /bin/bash # gpumachine-2
# Things to install on docker
pip install pandas
pip install scikit-learn
pip install wget
apt-get install less
# Image Resizing:
from joblib import Parallel, delayed
from skimage import io
from skimage.transform import resize
import os
import time
from tqdm import tqdm
def resize_images_dir(src_dir, dest_dir, output_shape, njobs=1):
def _resize_img(src_image_file):
img = io.imread(os.path.join(src_dir, src_image_file))
img_resized = resize(img, output_shape, anti_aliasing=True)
img_resized = img_resized * 255 # The original resize function returns normalized image.
img_resized = img_resized.astype('uint8')
io.imsave(os.path.join(dest_dir, src_image_file), img_resized)
start_time = time.time()
os.makedirs(dest_dir)
src_images = os.listdir(src_dir)
num_images = len(src_images)
print("Total number of images: %s" % num_images)
Parallel(n_jobs=njobs)(delayed(_resize_img)(src_image_file) for src_image_file in tqdm(src_images))
end_time = time.time()
time_taken = end_time - start_time
print("Time taken to resize %s images: %s seconds." % (num_images, time_taken))
resize_images_dir("/datadrive2/dataset/images", "/datadrive2/dataset/images-resized-224", (224,224) , njobs=16)
resize_images_dir("/datadrive2/dataset/images", "/datadrive2/dataset/images-resized", (512,512) , njobs=16) # TODO Rename to: images-resized-512
#resize_images_dir("/datadrive2/dataset/images_test", "/datadrive2/dataset/images_test-resized", (224,224), njobs=16)
# Model training
The basic training code: https://towardsdatascience.com/easy-image-classification-with-tensorflow-2-0-f734fee52d13
tf image summary issues: https://github.com/tensorflow/tensorflow/issues/28868
# Baseline modeling: mode_flat_all
# class imbalance:
{0: 10738, 1: 51426} -> {0:1, 1: 0.21}
### Test baseline training model.
python train_baseline_model.py --train-meta-file ../data/final_dataset_train-trial.csv --val-meta-file ../data/final_dataset_val-trial.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline-trial --batch-size 32 --epochs 10 --learning-rate 0.001 --image-size 224 > ../logs/baseline-trial.log 2>&1
python train_pipeline.py --train-meta-file ../data/final_dataset_train-trial.csv --val-meta-file ../data/final_dataset_val-trial.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline-trial-1 --model-arch vgg16_batchnorm --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 2 --learning-rate 0.001 --image-size 224
python train_pipeline.py --train-meta-file ../data/final_dataset_train-trial.csv --val-meta-file ../data/final_dataset_val-trial.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline-trial-2 --model-arch vgg16_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 2 --learning-rate 0.001 --image-size 224
python train_pipeline.py --train-meta-file ../data/final_dataset_train-trial.csv --val-meta-file ../data/final_dataset_val-trial.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline-trial-4 --model-arch resnet50_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 64 --epochs 1 --learning-rate 0.001 --image-size 224
python train_pipeline.py --train-meta-file ../data/final_dataset_train-trial.csv --val-meta-file ../data/final_dataset_val-trial.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline-trial-5 --model-arch inceptionresnetv2_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 64 --epochs 1 --learning-rate 0.001 --image-size 224
### Baseline model training.
python train_baseline_model.py --train-meta-file ../data/final_dataset_train.csv --val-meta-file ../data/final_dataset_val.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_1 --batch-size 32 --epochs 10 --learning-rate 0.001 --image-size 224 > ../logs/baseline_1.log 2>&1 & [INCOMPLETE]
python train_baseline_model.py --train-meta-file ../data/final_dataset_train.csv --val-meta-file ../data/final_dataset_val.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_2 --batch-size 32 --epochs 10 --learning-rate 0.001 --image-size 224 > ../logs/baseline_2.log 2>&1 & [INCOMPLETE]
python train_pipeline.py --train-meta-file ../data/final_dataset_train.csv --val-meta-file ../data/final_dataset_val.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_3 --model-arch vgg16_batchnorm --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 10 --learning-rate 0.001 --image-size 224 > ../logs/baseline_3.log 2>&1 & [gpumachine-1, COMPLETED]
python train_pipeline.py --train-meta-file ../data/final_dataset_train.csv --val-meta-file ../data/final_dataset_val.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_4 --model-arch vgg16_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 10 --learning-rate 0.001 --image-size 224 > ../logs/baseline_4.log 2>&1 & [gpumachine-2, COMPLETED]
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_3_balanced --model-arch vgg16_batchnorm --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 10 --learning-rate 0.001 --image-size 224 > ../logs/baseline_3_balanced.log 2>&1 & [gpumachine-2, COMPLETED]
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_4_balanced --model-arch vgg16_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 10 --learning-rate 0.0001 --image-size 224 > ../logs/baseline_4_balanced.log 2>&1 & [gpumachine-2, COMPLETED]
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_5_balanced --model-arch resnet50_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 64 --epochs 10 --learning-rate 0.0001 --image-size 224 > ../logs/baseline_5_balanced.log 2>&1 & [gpumachine-3, COMPLETED]
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_6_balanced --model-arch inceptionresnetv2_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 64 --epochs 10 --learning-rate 0.0001 --image-size 224 > ../logs/baseline_6_balanced.log 2>&1 & [gpumachine-4, COMPLETED]
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_7_balanced --model-arch resnet101_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 10 --learning-rate 0.0001 --image-size 224 > ../logs/baseline_7_balanced.log 2>&1 & [gpumachine-2, COMPLETED]
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_8_balanced --model-arch resnet152_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 10 --learning-rate 0.0001 --image-size 224 > ../logs/baseline_8_balanced.log 2>&1 & [gpumachine-3, COMPLETED]
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_9_balanced --model-arch resnet152v2_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 10 --learning-rate 0.0001 --image-size 224 > ../logs/baseline_9_balanced.log 2>&1 & [gpumachine-3, COMPLETED]
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_8_updated --model-arch resnet152_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/baseline_8_updated.log 2>&1 & [gpumachine-3, COMPLETED]
# -------------------
# Darshan:
# Model: mask_background_mog2_single_balanced
python train_pipeline.py --train-meta-file ../data/final_dataset_train-trial.csv --val-meta-file ../data/final_dataset_val-trial.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/mask-background-trial-1 --model-arch resnet152_mask_pretrained_imagenet --data-pipeline-mode mode_mask_mog2_single --batch-size 16 --epochs 1 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/mask_MOG2_4channel_1 --model-arch resnet152_mask_pretrained_imagenet --data-pipeline-mode mode_mask_mog2_single --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/mask_MOG2_4channel_1.log 2>&1 & [gpumachine-4, IN PROCESS(132)]
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/mask_MOG2_4channel_1/val_auc --batch-size 32 --trained-model-arch resnet152_mask_pretrained_imagenet --data-pipeline-mode mode_mask_mog2_single --trained-checkpoint-dir ../trained_models/mask_MOG2_4channel_1/best_model_dir-auc.ckpt --image-size 224 > ../logs/inference-mask_MOG2_4channel_1-val_auc.log 2>&1 &
python compute_roc.py --preds-labels-file ../inference_outputs/mask_MOG2_4channel_1/val_auc/pred_labels-individual.pickle --out-file ../inference_outputs/mask_MOG2_4channel_1/val_auc/evaluation/individual-roc
# Model: mask_background_mog2_multichannel
python train_pipeline.py --train-meta-file ../data/final_dataset_train-trial.csv --val-meta-file ../data/final_dataset_val-trial.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/mask-background-trial-2 --model-arch resnet152_mask_mog2_10channel_pretrained_imagenet --data-pipeline-mode mode_mask_mog2_multichannel --batch-size 16 --epochs 1 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/mask_MOG2_10channel_1 --model-arch resnet152_mask_mog2_10channel_pretrained_imagenet --data-pipeline-mode mode_mask_mog2_multichannel --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/mask_MOG2_10channel_1.log 2>&1 & [gpumachine-4, IN PROCESS(1283)]
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/mask_MOG2_10channel_1/val_auc --batch-size 32 --trained-model-arch resnet152_mask_mog2_10channel_pretrained_imagenet --data-pipeline-mode mode_mask_mog2_multichannel --trained-checkpoint-dir ../trained_models/mask_MOG2_10channel_1/best_model_dir-auc.ckpt --image-size 224 > ../logs/inference-mask_MOG2_10channel_1-val_auc.log 2>&1 &
python compute_roc.py --preds-labels-file ../inference_outputs/mask_MOG2_4channel_1/val_auc/pred_labels-individual.pickle --out-file ../inference_outputs/mask_MOG2_4channel_1/val_auc/evaluation/individual-roc
# Train Optical Flow based models.
# Trial
python train_pipeline.py --train-meta-file ../data/final_dataset_train-trial.csv --val-meta-file ../data/final_dataset_val-trial.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/opticalflow_6channel_1-trial --model-arch resnet152_6channel_opticalflow --data-pipeline-mode mode_opticalflow_single --num-channels 6 --batch-size 8 --epochs 1 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005
python train_pipeline.py --train-meta-file ../data/final_dataset_train-trial.csv --val-meta-file ../data/final_dataset_val-trial.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/opticalflow_6channel_allpretrained_1-trial --model-arch resnet152_6channel_allpretrained_opticalflow --data-pipeline-mode mode_opticalflow_single --num-channels 6 --batch-size 8 --epochs 1 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005
# Model: opticalflow_6channel_1
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/opticalflow_6channel_1 --model-arch resnet152_6channel_opticalflow --data-pipeline-mode mode_opticalflow_single --num-channels 6 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/opticalflow_6channel_1.log 2>&1 & [gpumachine-2, COMPLETED]
# Model: opticalflow_6channel_allpretrained_1
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/opticalflow_6channel_allpretrained_1 --model-arch resnet152_6channel_allpretrained_opticalflow --data-pipeline-mode mode_opticalflow_single --num-channels 6 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/opticalflow_6channel_allpretrained_1.log 2>&1 & [gpumachine-3, COMPLETED]
# Model: opticalflow_15channel_1
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/opticalflow_15channel_1 --model-arch resnet152_15channel_opticalflow --data-pipeline-mode mode_opticalflow_multichannel --num-channels 15 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/opticalflow_15channel_1.log 2>&1 & [gpumachine-2, COMPLETED]
# Model: opticalflow_15channel_allpretrained_1
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/opticalflow_15channel_allpretrained_1 --model-arch resnet152_15channel_allpretrained_opticalflow --data-pipeline-mode mode_opticalflow_multichannel --num-channels 15 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/opticalflow_15channel_allpretrained_1.log 2>&1 & [gpumachine-3, COMPLETED]
# Train Hybrid Models.
# Model: hybrid_13channel_allpretrained_1
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/hybrid_13channel_allpretrained_1 --model-arch resnet152_13channel_allpretrained_hybrid --data-pipeline-mode mode_hybrid_13channel --num-channels 13 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/hybrid_13channel_allpretrained_1.log 2>&1 & [gpumachine-2, COMPLETED]
# Model: hybrid_16channel_allpretrained_1 (Very huge and runs out of the memory with shuffle buffer size of 10000, trying again, but if it fails, then reduce this to say 5000 and rerun).
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/hybrid_16channel_allpretrained_1 --model-arch resnet152_16channel_allpretrained_hybrid --data-pipeline-mode mode_hybrid_16channel --num-channels 16 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/hybrid_16channel_allpretrained_1.log 2>&1 & [gpumachine-3, COMPLETED]
# Train opticalflow or mask only models.
# Model: opticalflowonly_6channel_allpretrained_1
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/opticalflowonly_6channel_allpretrained_1 --model-arch resnet152_6channel_allpretrained_opticalflow --data-pipeline-mode mode_opticalflowonly_6channel --num-channels 6 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/opticalflowonly_6channel_allpretrained_1.log 2>&1 & [gpumachine-2, COMPLETED]
# Model: maskopticalflowonly_7channel_allpretrained_1
python train_pipeline.py --train-meta-file ../data/final_dataset_train_balanced.csv --val-meta-file ../data/final_dataset_val_balanced.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/maskopticalflowonly_7channel_allpretrained_1 --model-arch resnet152_7channel_allpretrained_maskopticalflow --data-pipeline-mode mode_maskopticalflowonly_7channel --num-channels 7 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/maskopticalflowonly_7channel_allpretrained_1.log 2>&1 & [gpumachine-4, COMPLETED]
#-----------------------------------------------------------------------------------------------------------------------------------------------------------
# Running inference pipeline. (specify --is-sequence-model flag for sequence models.)
# Model: baseline_3
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_3/val_acc --batch-size 32 --trained-model-arch vgg16_batchnorm --trained-checkpoint-dir ../trained_models/baseline_3/best_model_dir-acc.ckpt --extract-layers "block1_conv1-Relu,block3_conv3-Relu,block5_conv3-Relu" --image-size 224 --data-pipeline-mode mode_flat_all > ../logs/inference-baseline_3-val_acc.log 2>&1 &
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_3/val_loss --batch-size 32 --trained-model-arch vgg16_batchnorm --trained-checkpoint-dir ../trained_models/baseline_3/best_model_dir-loss.ckpt --extract-layers "block1_conv1-Relu,block3_conv3-Relu,block5_conv3-Relu" --image-size 224 > ../logs/inference-baseline_3-val_loss.log 2>&1 &
# Model: baseline_4
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_4/val_acc --batch-size 32 --trained-model-arch vgg16_pretrained_imagenet --trained-checkpoint-dir ../trained_models/baseline_4/best_model_dir-acc.ckpt --image-size 224 > ../logs/inference-baseline_4-val_acc.log 2>&1 &
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_4/val_loss --batch-size 32 --trained-model-arch vgg16_pretrained_imagenet --trained-checkpoint-dir ../trained_models/baseline_4/best_model_dir-loss.ckpt --image-size 224 > ../logs/inference-baseline_4-val_loss.log 2>&1 &
#---------
# Model: baseline_3_balanced
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_3_balanced/val_acc --batch-size 32 --trained-model-arch vgg16_batchnorm --trained-checkpoint-dir ../trained_models/baseline_3_balanced/best_model_dir-acc.ckpt --image-size 224 --data-pipeline-mode mode_flat_all > ../logs/inference-baseline_3_balanced-val_acc.log 2>&1
python compute_roc.py --preds-labels-file ../inference_outputs/baseline_3_balanced/val_acc/pred_labels-individual.pickle --out-file ../inference_outputs/baseline_3_balanced/val_acc/evaluation/individual-roc
# Model: baseline_4_balanced
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_4_balanced/val_acc --batch-size 32 --trained-model-arch vgg16_pretrained_imagenet --trained-checkpoint-dir ../trained_models/baseline_4_balanced/best_model_dir-acc.ckpt --image-size 224 --data-pipeline-mode mode_flat_all > ../logs/inference-baseline_4_balanced-val_acc.log 2>&1
python compute_roc.py --preds-labels-file ../inference_outputs/baseline_4_balanced/val_acc/pred_labels-individual.pickle --out-file ../inference_outputs/baseline_4_balanced/val_acc/evaluation/individual-roc
# Model: baseline_5_balanced
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_5_balanced/val_auc --batch-size 32 --trained-model-arch resnet50_pretrained_imagenet --trained-checkpoint-dir ../trained_models/baseline_5_balanced/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_flat_all > ../logs/inference-baseline_5_balanced-val_auc.log 2>&1
python compute_roc.py --preds-labels-file ../inference_outputs/baseline_5_balanced/val_auc/pred_labels-individual.pickle --out-file ../inference_outputs/baseline_5_balanced/val_auc/evaluation/individual-roc
# Model: baseline_6_balanced
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_6_balanced/val_auc --batch-size 32 --trained-model-arch inceptionresnetv2_pretrained_imagenet --trained-checkpoint-dir ../trained_models/baseline_6_balanced/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_flat_all > ../logs/inference-baseline_6_balanced-val_auc.log 2>&1
python compute_roc.py --preds-labels-file ../inference_outputs/baseline_6_balanced/val_auc/pred_labels-individual.pickle --out-file ../inference_outputs/baseline_6_balanced/val_auc/evaluation/individual-roc
# Model: baseline_7_balanced
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_7_balanced/val_auc --batch-size 32 --trained-model-arch resnet101_pretrained_imagenet --trained-checkpoint-dir ../trained_models/baseline_7_balanced/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_flat_all > ../logs/inference-baseline_7_balanced-val_auc.log 2>&1
python compute_roc.py --preds-labels-file ../inference_outputs/baseline_7_balanced/val_auc/pred_labels-individual.pickle --out-file ../inference_outputs/baseline_7_balanced/val_auc/evaluation/individual-roc
# Model: baseline_8_balanced
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_8_balanced/val_auc --batch-size 32 --trained-model-arch resnet152_pretrained_imagenet --trained-checkpoint-dir ../trained_models/baseline_8_balanced/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_flat_all > ../logs/inference-baseline_8_balanced-val_auc.log 2>&1
python compute_roc.py --preds-labels-file ../inference_outputs/baseline_8_balanced/val_auc/pred_labels-individual.pickle --out-file ../inference_outputs/baseline_8_balanced/val_auc/evaluation/individual-roc
# Model: baseline_9_balanced
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_9_balanced/val_auc --batch-size 32 --trained-model-arch resnet152v2_pretrained_imagenet --trained-checkpoint-dir ../trained_models/baseline_9_balanced/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_flat_all > ../logs/inference-baseline_9_balanced-val_auc.log 2>&1
python compute_roc.py --preds-labels-file ../inference_outputs/baseline_9_balanced/val_auc/pred_labels-individual.pickle --out-file ../inference_outputs/baseline_9_balanced/val_auc/evaluation/individual-roc
# Model: baseline_8_updated
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_8_updated/val_auc --batch-size 32 --trained-model-arch resnet152_pretrained_imagenet --trained-checkpoint-dir ../trained_models/baseline_8_updated/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_flat_all > ../logs/inference-baseline_8_updated-val_auc.log 2>&1
python compute_roc.py --preds-labels-file ../inference_outputs/baseline_8_updated/val_auc/pred_labels-individual.pickle --out-file ../inference_outputs/baseline_8_updated/val_auc/evaluation/individual-roc
#------
# Model: mask_MOG2_4channel_1
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/mask_MOG2_4channel_1/val_auc --batch-size 16 --trained-model-arch resnet152_mask_pretrained_imagenet --trained-checkpoint-dir ../trained_models/mask_MOG2_4channel_1/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_mask_mog2_single --num-channels 4 > ../logs/inference-mask_MOG2_4channel_1-val_auc.log 2>&1
# Model: mask_MOG2_10channel_1
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/mask_MOG2_10channel_1/val_auc --batch-size 16 --trained-model-arch resnet152_mask_mog2_10channel_pretrained_imagenet --trained-checkpoint-dir ../trained_models/mask_MOG2_10channel_1/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_mask_mog2_multichannel --num-channels 10 > ../logs/inference-mask_MOG2_10channel_1-val_auc.log 2>&1
#------
# Model: opticalflow_6channel_1
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/opticalflow_6channel_1/val_auc --batch-size 16 --trained-model-arch resnet152_6channel_opticalflow --trained-checkpoint-dir ../trained_models/opticalflow_6channel_1/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_opticalflow_single --num-channels 6 > ../logs/inference-opticalflow_6channel_1-val_auc.log 2>&1
# Model: opticalflow_6channel_allpretrained_1
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/opticalflow_6channel_allpretrained_1/val_auc --batch-size 16 --trained-model-arch resnet152_6channel_allpretrained_opticalflow --trained-checkpoint-dir ../trained_models/opticalflow_6channel_allpretrained_1/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_opticalflow_single --num-channels 6 > ../logs/inference-opticalflow_6channel_allpretrained_1-val_auc.log 2>&1
# Model: opticalflow_15channel_1
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/opticalflow_15channel_1/val_auc --batch-size 16 --trained-model-arch resnet152_15channel_opticalflow --trained-checkpoint-dir ../trained_models/opticalflow_15channel_1/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_opticalflow_multichannel --num-channels 15 > ../logs/inference-opticalflow_15channel_1-val_auc.log 2>&1
# Model: opticalflow_15channel_allpretrained_1
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/opticalflow_15channel_allpretrained_1/val_auc --batch-size 16 --trained-model-arch resnet152_15channel_allpretrained_opticalflow --trained-checkpoint-dir ../trained_models/opticalflow_15channel_allpretrained_1/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_opticalflow_multichannel --num-channels 15 > ../logs/inference-opticalflow_15channel_allpretrained_1-val_auc.log 2>&1
#------
# Model: hybrid_13channel_allpretrained_1
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/hybrid_13channel_allpretrained_1/val_auc --batch-size 16 --trained-model-arch resnet152_13channel_allpretrained_hybrid --trained-checkpoint-dir ../trained_models/hybrid_13channel_allpretrained_1/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_hybrid_13channel --num-channels 13 > ../logs/inference-hybrid_13channel_allpretrained_1-val_auc.log 2>&1
# Model: hybrid_16channel_allpretrained_1
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/hybrid_16channel_allpretrained_1/val_auc --batch-size 16 --trained-model-arch resnet152_16channel_allpretrained_hybrid --trained-checkpoint-dir ../trained_models/hybrid_16channel_allpretrained_1/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_hybrid_16channel --num-channels 16 > ../logs/inference-hybrid_16channel_allpretrained_1-val_auc.log 2>&1
#-----
# Model: opticalflowonly_6channel_allpretrained_1
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/opticalflowonly_6channel_allpretrained_1/val_auc --batch-size 16 --trained-model-arch resnet152_6channel_allpretrained_opticalflow --trained-checkpoint-dir ../trained_models/opticalflowonly_6channel_allpretrained_1/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_opticalflowonly_6channel --num-channels 6 > ../logs/inference-opticalflowonly_6channel_allpretrained_1-val_auc.log 2>&1
# Model: maskopticalflowonly_7channel_allpretrained_1
python inference_pipeline.py --test-meta-file ../data/final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/maskopticalflowonly_7channel_allpretrained_1/val_auc --batch-size 16 --trained-model-arch resnet152_7channel_allpretrained_maskopticalflow --trained-checkpoint-dir ../trained_models/maskopticalflowonly_7channel_allpretrained_1/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_maskopticalflowonly_7channel --num-channels 7 > ../logs/inference-maskopticalflowonly_7channel_allpretrained_1-val_auc.log 2>&1
### Optical Flow.
# Generate optical flow images [gpumachine-1]
python generate_optical_flow_images.py --metadata-file ../data/final_dataset_train.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../../wellington_data/images-resized-224-opticalflow/ & [DONE]
python generate_optical_flow_images.py --metadata-file ../data/final_dataset_val.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../../wellington_data/images-resized-224-opticalflow/ & [DONE]
python generate_optical_flow_images.py --metadata-file ../data/final_dataset_test.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../../wellington_data/images-resized-224-opticalflow/ & [DONE]
#----------------------------------------
### Experiment with train-test splits based on the camera-sites.
# train set (up-sampled): loc_final_dataset_train_balanced-shuffled.csv (89777) | unique: 53856 | unique sites: 111
# val set (down-sampled): loc_final_dataset_val_balanced-shuffled.csv (5969) | unique: 16412 | unique sites: 36
# test set (down-sampled): loc_final_dataset_test_balanced-shuffled.csv (6453) | unique: 18450 | unieuq sites: 35
### Train and Infererence the models:
# Model: baseline_8_balanced_loc
- training
python train_pipeline.py --train-meta-file ../data/loc_final_dataset_train_balanced-shuffled.csv --val-meta-file ../data/loc_final_dataset_val_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/baseline_8_balanced_loc --model-arch resnet152_pretrained_imagenet --data-pipeline-mode mode_flat_all --batch-size 32 --epochs 10 --learning-rate 0.0001 --image-size 224 > ../logs/baseline_8_balanced_loc.log 2>&1 & [gpumachine-2, COMPLETED]
- inference
python inference_pipeline.py --test-meta-file ../data/loc_final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/baseline_8_balanced_loc/val_auc --batch-size 32 --trained-model-arch resnet152_pretrained_imagenet --trained-checkpoint-dir ../trained_models/baseline_8_balanced_loc/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_flat_all > ../logs/inference-baseline_8_balanced_loc-val_auc.log 2>&1
# Model: resnet50_lstm_avg_pool_loc
- training
python train_pipeline.py --train-meta-file ../data/loc_final_dataset_train_balanced-shuffled.csv --val-meta-file ../data/loc_final_dataset_val_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/resnet50_lstm_avg_pool_loc --model-arch resnet50_pretrained_imagenet_lstm_avg_pool --data-pipeline-mode mode_sequence --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/resnet50_lstm_avg_pool_loc.log 2>&1 & [gpumachine-4, COMPLETED]
- inference
python inference_pipeline.py --test-meta-file ../data/loc_final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/resnet50_lstm_avg_pool_loc/val_auc --batch-size 16 --trained-model-arch resnet50_pretrained_imagenet_lstm_avg_pool --trained-checkpoint-dir ../trained_models/resnet50_lstm_avg_pool_loc/ --filetype .h5 --image-size 224 --data-pipeline-mode mode_sequence > ../logs/inference-resnet50_lstm_avg_pool_loc-val_auc.log 2>&1
# Model: mask_MOG2_4channel_1_loc
- training
python train_pipeline.py --train-meta-file ../data/loc_final_dataset_train_balanced-shuffled.csv --val-meta-file ../data/loc_final_dataset_val_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/mask_MOG2_4channel_1_loc --model-arch resnet152_mask_pretrained_imagenet --data-pipeline-mode mode_mask_mog2_single --num-channels 4 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/mask_MOG2_4channel_1_loc.log 2>&1 & [gpumachine-2, COMPLETED]
- inference
python inference_pipeline.py --test-meta-file ../data/loc_final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/mask_MOG2_4channel_1_loc/val_auc --batch-size 32 --trained-model-arch resnet152_mask_pretrained_imagenet --data-pipeline-mode mode_mask_mog2_single --trained-checkpoint-dir ../trained_models/mask_MOG2_4channel_1_loc/best_model_dir-auc.ckpt --image-size 224 --num-channels 4 > ../logs/inference-mask_MOG2_4channel_1_loc-val_auc.log 2>&1
# Model: mask_MOG2_10channel_1_loc
- training
python train_pipeline.py --train-meta-file ../data/loc_final_dataset_train_balanced-shuffled.csv --val-meta-file ../data/loc_final_dataset_val_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/mask_MOG2_10channel_1_loc --model-arch resnet152_mask_mog2_10channel_pretrained_imagenet --data-pipeline-mode mode_mask_mog2_multichannel --num-channels 10 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/mask_MOG2_10channel_1_loc.log 2>&1 & [gpumachine-3, COMPLETED]
- inference
python inference_pipeline.py --test-meta-file ../data/loc_final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/mask_MOG2_10channel_1_loc/val_auc --batch-size 32 --trained-model-arch resnet152_mask_mog2_10channel_pretrained_imagenet --data-pipeline-mode mode_mask_mog2_multichannel --trained-checkpoint-dir ../trained_models/mask_MOG2_10channel_1_loc/best_model_dir-auc.ckpt --image-size 224 --num-channels 10 > ../logs/inference-mask_MOG2_10channel_1_loc-val_auc.log 2>&1
# Model: opticalflow_6channel_allpretrained_1_loc
- training
python train_pipeline.py --train-meta-file ../data/loc_final_dataset_train_balanced-shuffled.csv --val-meta-file ../data/loc_final_dataset_val_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/opticalflow_6channel_allpretrained_1_loc --model-arch resnet152_6channel_allpretrained_opticalflow --data-pipeline-mode mode_opticalflow_single --num-channels 6 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/opticalflow_6channel_allpretrained_1_loc.log 2>&1 & [gpumachine-3, COMPLETED]
- inference
python inference_pipeline.py --test-meta-file ../data/loc_final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/opticalflow_6channel_allpretrained_1_loc/val_auc --batch-size 16 --trained-model-arch resnet152_6channel_allpretrained_opticalflow --trained-checkpoint-dir ../trained_models/opticalflow_6channel_allpretrained_1_loc/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_opticalflow_single --num-channels 6 > ../logs/inference-opticalflow_6channel_allpretrained_1_loc-val_auc.log 2>&1
# Model: opticalflow_15channel_allpretrained_1_loc
- training
python train_pipeline.py --train-meta-file ../data/loc_final_dataset_train_balanced-shuffled.csv --val-meta-file ../data/loc_final_dataset_val_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/opticalflow_15channel_allpretrained_1_loc --model-arch resnet152_15channel_allpretrained_opticalflow --data-pipeline-mode mode_opticalflow_multichannel --num-channels 15 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/opticalflow_15channel_allpretrained_1_loc.log 2>&1 & [gpumachine-4, COMPLETED]
- inference
python inference_pipeline.py --test-meta-file ../data/loc_final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/opticalflow_15channel_allpretrained_1_loc/val_auc --batch-size 16 --trained-model-arch resnet152_15channel_allpretrained_opticalflow --trained-checkpoint-dir ../trained_models/opticalflow_15channel_allpretrained_1_loc/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_opticalflow_multichannel --num-channels 15 > ../logs/inference-opticalflow_15channel_allpretrained_1_loc-val_auc.log 2>&1
# Model: hybrid_13channel_allpretrained_1_loc
- training
python train_pipeline.py --train-meta-file ../data/loc_final_dataset_train_balanced-shuffled.csv --val-meta-file ../data/loc_final_dataset_val_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/hybrid_13channel_allpretrained_1_loc --model-arch resnet152_13channel_allpretrained_hybrid --data-pipeline-mode mode_hybrid_13channel --num-channels 13 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/hybrid_13channel_allpretrained_1_loc.log 2>&1 & [gpumachine-4, COMPLETED]
- inference
python inference_pipeline.py --test-meta-file ../data/loc_final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/hybrid_13channel_allpretrained_1_loc/val_auc --batch-size 16 --trained-model-arch resnet152_13channel_allpretrained_hybrid --trained-checkpoint-dir ../trained_models/hybrid_13channel_allpretrained_1_loc/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_hybrid_13channel --num-channels 13 > ../logs/inference-hybrid_13channel_allpretrained_1_loc-val_auc.log 2>&1
# Model: opticalflowonly_6channel_allpretrained_1_loc
- training
python train_pipeline.py --train-meta-file ../data/loc_final_dataset_train_balanced-shuffled.csv --val-meta-file ../data/loc_final_dataset_val_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/opticalflowonly_6channel_allpretrained_1_loc --model-arch resnet152_6channel_allpretrained_opticalflow --data-pipeline-mode mode_opticalflowonly_6channel --num-channels 6 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/opticalflowonly_6channel_allpretrained_1_loc.log 2>&1 & [gpumachine-3, COMPLETED]
- inference
python inference_pipeline.py --test-meta-file ../data/loc_final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/opticalflowonly_6channel_allpretrained_1_loc/val_auc --batch-size 16 --trained-model-arch resnet152_6channel_allpretrained_opticalflow --trained-checkpoint-dir ../trained_models/opticalflowonly_6channel_allpretrained_1_loc/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_opticalflowonly_6channel --num-channels 6 > ../logs/inference-opticalflowonly_6channel_allpretrained_1_loc-val_auc.log 2>&1
# Model: maskopticalflowonly_7channel_allpretrained_1_loc
- training
python train_pipeline.py --train-meta-file ../data/loc_final_dataset_train_balanced-shuffled.csv --val-meta-file ../data/loc_final_dataset_val_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../trained_models/maskopticalflowonly_7channel_allpretrained_1_loc --model-arch resnet152_7channel_allpretrained_maskopticalflow --data-pipeline-mode mode_maskopticalflowonly_7channel --num-channels 7 --batch-size 16 --epochs 10 --learning-rate 0.0001 --image-size 224 --patience 3 --min-delta-auc 0.005 > ../logs/maskopticalflowonly_7channel_allpretrained_1_loc.log 2>&1 & [gpumachine-4, COMPLETED]
- inference
python inference_pipeline.py --test-meta-file ../data/loc_final_dataset_test_balanced-shuffled.csv --images-dir ../../wellington_data/images-resized-224/ --out-dir ../inference_outputs/maskopticalflowonly_7channel_allpretrained_1_loc/val_auc --batch-size 16 --trained-model-arch resnet152_7channel_allpretrained_maskopticalflow --trained-checkpoint-dir ../trained_models/maskopticalflowonly_7channel_allpretrained_1_loc/best_model_dir-auc.ckpt --image-size 224 --data-pipeline-mode mode_maskopticalflowonly_7channel --num-channels 7 > ../logs/inference-maskopticalflowonly_7channel_allpretrained_1_loc-val_auc.log 2>&1
# ---------------------------------------
# Paper details:
### Sequence ids presented in sequence_example image.
left-top (animal, full daylight): 29125 | info: 29125,"image1","110116151818047bs131.jpg","BIRD","047b","1/11/2016 15:18","s13"
right-top (animal, night time): 40156 | info: 40156,"image1","060216224444038bs163.jpg","HEDGEHOG","038b","2/6/2016 22:44","s16"
left-bottom (animal, low daylight): 17068 | info: 17068,"image1","160216060344039bs021.jpg","BIRD","039b","2/16/2016 6:03","s02"
right-bottom (empty, full daylight): 49305 | info: 49305,"image1","010416125916018c5171.jpg","NOTHINGHERE","018c","4/1/2016 12:59","517"
# Data upload:
The data, trained models and inferene outputs uploaded to capstone_project container in storage account (resourcegroupcapstonedia).
Public Blob read-access SAS URL till 2021: https://resourcegroupcapstonedia.blob.core.windows.net/?sv=2019-10-10&ss=bfqt&srt=sco&sp=rl&se=2021-04-29T13:21:03Z&st=2020-04-29T05:21:03Z&spr=https&sig=rk8VhQ%2BTKhRmfSLAEgeUjmCFNBTksuSPkKYPBefbmrE%3D
# ArXiv details:
Title: Sequence Information Channel Concatenation for Improving Camera Trap Image Burst Classification
Authors: Bhuvan Malladihalli Shashidhara, Darshan Mehta, Yash Kale, Dan Morris, Megan Hazen
Abstract:
Camera Traps are extensively used to observe wildlife in their natural habitat without disturbing the ecosystem. This could help in the early detection of natural or human threats to animals, and help towards ecological conservation. Currently, a massive number of such camera traps have been deployed at various ecological conservation areas around the world, collecting data for decades, thereby requiring automation to detect images containing animals. Existing systems perform classification to detect if images contain animals by considering a single image. However, due to challenging scenes with animals camouflaged in their natural habitat, it sometimes becomes difficult to identify the presence of animals from merely a single image. We hypothesize that a short burst of images instead of a single image, assuming that the animal moves, makes it much easier for a human as well as a machine to detect the presence of animals. In this work, we explore a variety of approaches, and measure the impact of using short image sequences (burst of 3 images) on improving the camera trap image classification. We show that concatenating masks containing sequence information and the images from the 3-image-burst across channels, improves the ROC AUC by 20% on a test-set from unseen camera-sites, as compared to an equivalent model that learns from a single image.
Comments: 9 pages, 4 figures, 2 tables. Git repository can be found at: https://github.com/bhuvi3/camera_trap_animal_classification.
ACM-class: I.4.9; I.4.10; I.2.10
# ---------------------------------------
# XXX: Keep pushing this section towards the end. This summary might help later in future.
### Learnings (include mistakes and self-realizations):
- Adding channels helps in Improvements.
- Weak label models can be still trained, but consider deep models with residual layers, like ResNet152. Also, not much differnce from V1 and V2 (ResNet), maybe V1 is better due to pre-activation while batch-norm. For weak label models, Average Pool may help better than dense layers.
- Optical flow Gunner Faurnebacks algorithm works great and similar to MOG2 background subtraction based motion detection for short sequences with objects moving in a relatively stationary environment.
- Pretraining almost always helps, and can be loaded by altering the intial layer with just a few lines of code. Additional channels can be initialized with random weights (or maybe with pretrained weights as well as these initial kernels learn very simple features which could be meaningful for any kind of mask or channel information).
- Be careful about 'training' argument in layers, and make sure that the inference is not run in training mode, as that leads to inconsistent predictions.
- Never ever, Never ever ever use "[[]] * x" syntax to initialize. Python will take you down unconsciously with its reference copies. Use list comprehension instead of ensure reference copy is avoided.
- Read TF and Keras documentation in detail, in all corners and crevices, they dont have good documentation. Also, consider reading their guides and examples before using.
### Improvements to consider for future projects.
# Code.
- Models could be maintained in configs. Or the model functions could take kwargs to put similar model generation together.
# Methods.
- Each method needs to get its optimal hyperparams like patience, but that takes a lot of time.