-
Notifications
You must be signed in to change notification settings - Fork 3
/
imagenet_multifold_train_and_evaluate_loop_icm.sh
executable file
·203 lines (183 loc) · 7.84 KB
/
imagenet_multifold_train_and_evaluate_loop_icm.sh
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
#! /usr/bin/env bash
eval_device=0
train_device=0
cur_dir=$PWD
experiment_name="icm_10fold_k8_init"
train_split="multifea_k8_init"
gt_split="alexfea300_revisit_val1"
train_config_template="configs/mil/imagenet/inception_resnet/agnostic_model/agnostic_box_multi_fea/pairwise_loop/templates/k2_301.config"
eval_config_template="configs/mil/imagenet/inception_resnet/agnostic_model/agnostic_box_multi_fea/pairwise_loop/templates/k8_transferred_301.config"
eval_config_template_icm="configs/mil/imagenet/inception_resnet/agnostic_model/agnostic_box_multi_fea/pairwise_loop/templates/k2_icm_301.config"
root="mil/inception_resnet/agnostic_model/agnostic_box_multi_fea/icm_10folds_k8_init"
configs_root="configs/${root}"
logs_root="logs/${root}"
eval_ncobj_proposals=500
aggregate_ncobj_proposals=500
num_loop_iters=2
train_iters=80000
num_eval_examples=1000
num_aggregate_examples=2500
# 65000/10 = 6500 (folds might not be even use 8000 instead)
# icm uses k2 so divide by 2 (8000/2 = 4000)
num_icm_examples=4000
# use smaller number if you want to compute energy
icm_save_eval_freq=8000
ds_root="../feature_extractor/logs/imagenet/revisit_agnostic_box_multi_feas"
start_iter=1 #1
start_fold=0 #0
num_folds=10
evaluate_trws(){
train_dir=$1
eval_config_file=$2
evaldir=$3
method=$4
CUDA_VISIBLE_DEVICES="${eval_device}" python eval.py --logtostderr \
--checkpoint_dir="${train_dir}/calib_train" \
--pipeline_config_path="${eval_config_file}" \
--eval_dir="${evaldir}" \
--calibration=true \
--add_mrf=true \
--add_unaries=true \
--unary_scale=2.0 \
--mrf_type="${method}" >> "${train_dir}/eval_${method}.log" 2>&1
}
evaluate_greedy(){
train_dir=$1
eval_config_file=$2
evaldir=$3
CUDA_VISIBLE_DEVICES="${eval_device}" python eval.py --logtostderr \
--checkpoint_dir="${train_dir}/calib_train" \
--pipeline_config_path="${eval_config_file}" \
--eval_dir="${evaldir}" \
--calibration=true >> "${train_dir}/eval.log" 2>&1
}
train_and_eval () {
train_config_file=$1
train_dir=$2
checkpoint=$3 #last checkpoint file
eval_config_file=$4
evaldir="${train_dir}/calib_eval"
echo $checkpoint
cp $train_config_file $train_dir
cp $eval_config_file "${train_dir}/calib_eval.config"
#run training in background
CUDA_VISIBLE_DEVICES="${train_device}" TF_ENABLE_WINOGRAD_NONFUSED=1 python train.py \
--logtostderr \
--pipeline_config_path="${train_config_file}" \
--train_dir="${train_dir}/train/" \
--num_clones=1 > "${train_dir}/train.log" 2>&1 &
while :
do
# NOTE: Uncomment (next 3 commented lines)
# if you want to evaluate during training
#evaluate_greedy $train_dir $eval_config_file $evaldir
sleep 20
if [ -f $checkpoint ]; then
echo "FOUND LAST ITERATION"
#evaluate_greedy $train_dir $eval_config_file $evaldir
#rm dataflow-pipe*
break
fi
done
}
mkdir -p ${configs_root}
#loop
for iter in $(seq $start_iter $num_loop_iters);
do
##some complex logic for enabling continue from broken run
starting_fold=0
if [ $iter = $start_iter ];
then
if [ $start_fold -ge 1 ];
then
starting_fold=$start_fold
last_fold=$((start_fold-1))
train_split="${experiment_name}_${iter}_${last_fold}_ICM"
fi
if [ $iter -ge 2 -a $start_fold = 0 ];
then
last_iter=$((iter-1))
train_split="${experiment_name}_${last_iter}_$((num_folds-1))_ICM"
fi
fi
##
if [ ! $iter = $start_iter -o ! $start_fold -ge 1 ];
then
echo "Iteration ${iter}, ASSIGNING FOLDS..."
python assign_folds.py --ds_root ${ds_root} --train_split ${train_split} --gt_split ${gt_split} --num_folds ${num_folds}
fi
for fold in $(seq $starting_fold $((num_folds-1)));
do
echo "Fold ${fold}"
#create train config
echo "CREATING TRAIN CONFIG"
train_config_file="${configs_root}/${experiment_name}_${iter}_${fold}.config"
echo $train_config_file
python config_utils.py --config_template_path ${train_config_template} \
--is_training true --split ${train_split} \
--train_num_steps ${train_iters} \
--write_path ${train_config_file} \
--eval_fold ${fold} --num_folds ${num_folds}
#create train and calib eval folders
echo "CREATING TRAIN AND CALIB EVAL FOLDERS"
train_dir="${logs_root}/${experiment_name}_${iter}_${fold}"
mkdir -p "${train_dir}/train"
cd ${train_dir} && ln -s train calib_train
cd $cur_dir
#create eval config
echo "CREATING EVAL CONFIG"
eval_config_file="${configs_root}/${experiment_name}_${iter}_${fold}_eval.config"
echo $eval_config_file
python config_utils.py --config_template_path ${eval_config_template} \
--is_training false --eval_num_examples ${num_eval_examples} \
--aggregate false --write_path ${eval_config_file} \
--ncobj_proposals ${eval_ncobj_proposals} \
--eval_fold ${fold} --num_folds ${num_folds}
#create eval folder
echo "CREATING EVAL FOLDER"
cd ${logs_root} && ln -s "${experiment_name}_${iter}_${fold}" "${experiment_name}_${iter}_${fold}_eval"
cd $cur_dir
#perform train and eval
echo "PERFORMING TRAINING AND EVALUATION"
last_checkpoint_file="${train_dir}/train/model.ckpt-${train_iters}.index"
train_log_file="${train_dir}/train.log"
#cp ${init_checkpoint_folder}/* ${train_dir}/train/
train_and_eval $train_config_file ${train_dir} $last_checkpoint_file $eval_config_file
#create aggregate config (create save_split)
echo "CREATING AGGREGATE CONFIG"
save_split="${experiment_name}_${iter}_${fold}"
update_split=${train_split}
python config_utils.py --config_template_path ${eval_config_template} \
--is_training false --eval_num_examples ${num_aggregate_examples} \
--aggregate true --aggregate_save_split ${save_split} \
--aggregate_update_split ${update_split} \
--write_path ${eval_config_file} \
--ncobj_proposals ${aggregate_ncobj_proposals} \
--eval_fold ${fold} --num_folds ${num_folds}
#aggregate
echo "PERFORMING INITIALIZATION AGGREGATION"
evaluate_greedy $train_dir $eval_config_file "${train_dir}/aggregate"
#use save split to create bcd labelling info
bcd_info="${save_split}_labelling_info"
echo "CREATING LABELLING INFO: ${bcd_info}"
python name2fea.py --split ${save_split} --save_name ${bcd_info} --ds_root ${ds_root} --eval_fold ${fold} --doublefeas true
bcd_init_file="${ds_root}/ImageSet/${bcd_info}.pkl"
old_save_split=${save_split}
save_split="${save_split}_ICM"
#create icm config with bcd_init
python config_utils.py --config_template_path ${eval_config_template_icm} \
--is_training false --eval_num_examples ${num_icm_examples} \
--aggregate true --aggregate_save_split ${save_split} \
--aggregate_update_split ${old_save_split} \
--write_path ${eval_config_file} \
--ncobj_proposals ${aggregate_ncobj_proposals} \
--eval_fold ${fold} --num_folds ${num_folds} \
--bcd_init ${bcd_init_file} --k_shot 1 \
--save_eval_freq ${icm_save_eval_freq}
#perform ICM
echo "Performing ICM"
evaluate_trws $train_dir $eval_config_file "${train_dir}/icm" "bcd_dense_trws"
train_split=$save_split
echo "ITERATION ${iter} DONE"
done
done