jonathang commited on
Commit
759e5fd
1 Parent(s): f45a4ca

Upload 7 files

Browse files
.gitattributes CHANGED
@@ -32,3 +32,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
32
  *.zip filter=lfs diff=lfs merge=lfs -text
33
  *.zst filter=lfs diff=lfs merge=lfs -text
34
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
32
  *.zip filter=lfs diff=lfs merge=lfs -text
33
  *.zst filter=lfs diff=lfs merge=lfs -text
34
  *tfevents* filter=lfs diff=lfs merge=lfs -text
35
+ checkpoint/ckpt-0.data-00000-of-00001 filter=lfs diff=lfs merge=lfs -text
checkpoint/checkpoint ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ model_checkpoint_path: "ckpt-0"
2
+ all_model_checkpoint_paths: "ckpt-0"
3
+ all_model_checkpoint_timestamps: 1594350057.1342435
4
+ last_preserved_timestamp: 1594350055.0480783
checkpoint/ckpt-0.data-00000-of-00001 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8c5ff5d40d12e72808a8267ffd95b3a1422b1e423f7bdf6335f280e3f21f9a3b
3
+ size 131768450
checkpoint/ckpt-0.index ADDED
Binary file (7.52 kB). View file
 
context_rcnn_camera_trap.config ADDED
@@ -0,0 +1,161 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ # Context R-CNN unit test configuration using sequence example sample data with
3
+ # context_features created from the Snapshot Serengeti Dataset and stored in
4
+ # snapshot_serengeti_sequence_examples.record.
5
+ # This model uses attention into contextual features within the Faster R-CNN
6
+ # object detection framework to improve object detection performance.
7
+ # See https://arxiv.org/abs/1912.03538 for more information.
8
+
9
+ model {
10
+ faster_rcnn {
11
+ num_classes: 48
12
+ image_resizer {
13
+ fixed_shape_resizer {
14
+ height: 640
15
+ width: 640
16
+ }
17
+ }
18
+ feature_extractor {
19
+ type: "faster_rcnn_resnet101"
20
+ first_stage_features_stride: 16
21
+ batch_norm_trainable: true
22
+ }
23
+ first_stage_anchor_generator {
24
+ grid_anchor_generator {
25
+ height_stride: 16
26
+ width_stride: 16
27
+ scales: 0.25
28
+ scales: 0.5
29
+ scales: 1.0
30
+ scales: 2.0
31
+ aspect_ratios: 0.5
32
+ aspect_ratios: 1.0
33
+ aspect_ratios: 2.0
34
+ }
35
+ }
36
+ first_stage_box_predictor_conv_hyperparams {
37
+ op: CONV
38
+ regularizer {
39
+ l2_regularizer {
40
+ weight: 0.0
41
+ }
42
+ }
43
+ initializer {
44
+ truncated_normal_initializer {
45
+ stddev: 0.00999999977648
46
+ }
47
+ }
48
+ }
49
+ first_stage_nms_score_threshold: 0.0
50
+ first_stage_nms_iou_threshold: 0.699999988079
51
+ first_stage_max_proposals: 300
52
+ first_stage_localization_loss_weight: 2.0
53
+ first_stage_objectness_loss_weight: 1.0
54
+ initial_crop_size: 14
55
+ maxpool_kernel_size: 2
56
+ maxpool_stride: 2
57
+ second_stage_box_predictor {
58
+ mask_rcnn_box_predictor {
59
+ fc_hyperparams {
60
+ op: FC
61
+ regularizer {
62
+ l2_regularizer {
63
+ weight: 0.0
64
+ }
65
+ }
66
+ initializer {
67
+ variance_scaling_initializer {
68
+ factor: 1.0
69
+ uniform: true
70
+ mode: FAN_AVG
71
+ }
72
+ }
73
+ }
74
+ use_dropout: false
75
+ dropout_keep_probability: 1.0
76
+ share_box_across_classes: true
77
+ }
78
+ }
79
+ second_stage_post_processing {
80
+ batch_non_max_suppression {
81
+ score_threshold: 0.0
82
+ iou_threshold: 0.600000023842
83
+ max_detections_per_class: 100
84
+ max_total_detections: 300
85
+ }
86
+ score_converter: SOFTMAX
87
+ }
88
+ second_stage_localization_loss_weight: 2.0
89
+ second_stage_classification_loss_weight: 1.0
90
+ use_matmul_crop_and_resize: true
91
+ clip_anchors_to_image: true
92
+ use_matmul_gather_in_matcher: true
93
+ use_static_balanced_label_sampler: true
94
+ use_static_shapes: true
95
+ context_config {
96
+ max_num_context_features: 2000
97
+ context_feature_length: 2057
98
+ }
99
+ }
100
+ }
101
+ train_config {
102
+ batch_size: 1
103
+ data_augmentation_options {
104
+ random_horizontal_flip {
105
+ }
106
+ }
107
+ sync_replicas: true
108
+ optimizer {
109
+ momentum_optimizer {
110
+ learning_rate {
111
+ manual_step_learning_rate {
112
+ initial_learning_rate: 0.0
113
+ schedule {
114
+ step: 2000
115
+ learning_rate: 0.00200000009499
116
+ }
117
+ schedule {
118
+ step: 200000
119
+ learning_rate: 0.000199999994948
120
+ }
121
+ schedule {
122
+ step: 300000
123
+ learning_rate: 1.99999994948e-05
124
+ }
125
+ warmup: true
126
+ }
127
+ }
128
+ momentum_optimizer_value: 0.899999976158
129
+ }
130
+ use_moving_average: false
131
+ }
132
+ gradient_clipping_by_norm: 10.0
133
+ num_steps: 500000
134
+ replicas_to_aggregate: 8
135
+ max_number_of_boxes: 100
136
+ unpad_groundtruth_tensors: false
137
+ }
138
+ train_input_reader {
139
+ label_map_path: ""
140
+ tf_record_input_reader {
141
+ input_path: ""
142
+ }
143
+ load_context_features: true
144
+ input_type: TF_SEQUENCE_EXAMPLE
145
+ }
146
+ eval_config {
147
+ max_evals: 50
148
+ metrics_set: "coco_detection_metrics"
149
+ use_moving_averages: false
150
+ batch_size: 1
151
+ }
152
+ eval_input_reader {
153
+ label_map_path: ""
154
+ shuffle: false
155
+ num_epochs: 1
156
+ tf_record_input_reader {
157
+ input_path: ""
158
+ }
159
+ load_context_features: true
160
+ input_type: TF_SEQUENCE_EXAMPLE
161
+ }
pets_examples.record ADDED
Binary file (874 kB). View file
 
snapshot_serengeti_sequence_examples.record ADDED
Binary file (821 kB). View file
 
ssd_mobilenet_v1_fpp.config ADDED
@@ -0,0 +1,251 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model {
2
+ ssd {
3
+ num_classes: 2
4
+ box_coder {
5
+ keypoint_box_coder {
6
+ num_keypoints: 23
7
+ y_scale: 10.0
8
+ x_scale: 10.0
9
+ height_scale: 5.0
10
+ width_scale: 5.0
11
+ }
12
+ }
13
+ matcher {
14
+ argmax_matcher {
15
+ matched_threshold: 0.5
16
+ unmatched_threshold: 0.5
17
+ ignore_thresholds: false
18
+ negatives_lower_than_unmatched: true
19
+ force_match_for_each_row: true
20
+ }
21
+ }
22
+ similarity_calculator {
23
+ iou_similarity {
24
+ }
25
+ }
26
+ anchor_generator {
27
+ ssd_anchor_generator {
28
+ num_layers: 6
29
+ min_scale: 0.2
30
+ max_scale: 0.95
31
+ aspect_ratios: 1.0
32
+ aspect_ratios: 2.0
33
+ aspect_ratios: 0.5
34
+ aspect_ratios: 3.0
35
+ aspect_ratios: 0.3333
36
+ height_stride: 16
37
+ height_stride: 32
38
+ height_stride: 64
39
+ height_stride: 128
40
+ height_stride: 256
41
+ height_stride: 512
42
+ width_stride: 16
43
+ width_stride: 32
44
+ width_stride: 64
45
+ width_stride: 128
46
+ width_stride: 256
47
+ width_stride: 512
48
+ height_offset: 0
49
+ height_offset: 0
50
+ height_offset: 0
51
+ height_offset: 0
52
+ height_offset: 0
53
+ height_offset: 0
54
+ width_offset: 0
55
+ width_offset: 0
56
+ width_offset: 0
57
+ width_offset: 0
58
+ width_offset: 0
59
+ width_offset: 0
60
+ }
61
+ }
62
+ image_resizer {
63
+ keep_aspect_ratio_resizer {
64
+ min_dimension: 320
65
+ max_dimension: 640
66
+ convert_to_grayscale: true
67
+ }
68
+ }
69
+ box_predictor {
70
+ convolutional_box_predictor {
71
+ min_depth: 0
72
+ max_depth: 0
73
+ num_layers_before_predictor: 0
74
+ use_dropout: false
75
+ kernel_size: 3
76
+ box_code_size: 50
77
+ apply_sigmoid_to_scores: false
78
+ conv_hyperparams {
79
+ activation: RELU_6
80
+ regularizer {
81
+ l2_regularizer {
82
+ weight: 0.00004
83
+ }
84
+ }
85
+ initializer {
86
+ truncated_normal_initializer {
87
+ stddev: 0.03
88
+ mean: 0.0
89
+ }
90
+ }
91
+ batch_norm {
92
+ train: true
93
+ scale: true
94
+ center: true
95
+ decay: 0.9997
96
+ epsilon: 0.001
97
+ }
98
+ }
99
+ }
100
+ }
101
+ feature_extractor {
102
+ type: "ssd_mobilenet_v1"
103
+ min_depth: 16
104
+ depth_multiplier: 0.25
105
+ use_explicit_padding: true
106
+ conv_hyperparams {
107
+ activation: RELU_6
108
+ regularizer {
109
+ l2_regularizer {
110
+ weight: 0.00004
111
+ }
112
+ }
113
+ initializer {
114
+ truncated_normal_initializer {
115
+ stddev: 0.03
116
+ mean: 0.0
117
+ }
118
+ }
119
+ batch_norm {
120
+ train: true
121
+ scale: true
122
+ center: true
123
+ decay: 0.9997
124
+ epsilon: 0.001
125
+ }
126
+ }
127
+ }
128
+ loss {
129
+ classification_loss {
130
+ weighted_sigmoid {
131
+ }
132
+ }
133
+ localization_loss {
134
+ weighted_smooth_l1 {
135
+ }
136
+ }
137
+ hard_example_miner {
138
+ num_hard_examples: 3000
139
+ iou_threshold: 0.99
140
+ loss_type: CLASSIFICATION
141
+ max_negatives_per_positive: 3
142
+ min_negatives_per_image: 10
143
+ }
144
+ classification_weight: 1.0
145
+ localization_weight: 1.0
146
+ }
147
+ normalize_loss_by_num_matches: true
148
+ post_processing {
149
+ batch_non_max_suppression {
150
+ score_threshold: 1e-8
151
+ iou_threshold: 0.5
152
+ max_detections_per_class: 100
153
+ max_total_detections: 100
154
+ }
155
+ score_converter: SIGMOID
156
+ }
157
+ }
158
+ }
159
+
160
+ train_config: {
161
+ fine_tune_checkpoint: ""
162
+ num_steps: 10000000
163
+ batch_size: 32
164
+ data_augmentation_options {
165
+ random_horizontal_flip {
166
+ keypoint_flip_permutation: 1
167
+ keypoint_flip_permutation: 0
168
+ keypoint_flip_permutation: 2
169
+ keypoint_flip_permutation: 3
170
+ keypoint_flip_permutation: 5
171
+ keypoint_flip_permutation: 4
172
+ keypoint_flip_permutation: 6
173
+ keypoint_flip_permutation: 8
174
+ keypoint_flip_permutation: 7
175
+ keypoint_flip_permutation: 10
176
+ keypoint_flip_permutation: 9
177
+ keypoint_flip_permutation: 12
178
+ keypoint_flip_permutation: 11
179
+ keypoint_flip_permutation: 14
180
+ keypoint_flip_permutation: 13
181
+ keypoint_flip_permutation: 16
182
+ keypoint_flip_permutation: 15
183
+ keypoint_flip_permutation: 18
184
+ keypoint_flip_permutation: 17
185
+ keypoint_flip_permutation: 20
186
+ keypoint_flip_permutation: 19
187
+ keypoint_flip_permutation: 22
188
+ keypoint_flip_permutation: 21
189
+ }
190
+ }
191
+ data_augmentation_options {
192
+ ssd_random_crop_fixed_aspect_ratio {
193
+ }
194
+ }
195
+ optimizer {
196
+ rms_prop_optimizer {
197
+ learning_rate {
198
+ exponential_decay_learning_rate {
199
+ initial_learning_rate: 0.0004
200
+ decay_steps: 800720
201
+ decay_factor: 0.95
202
+ }
203
+ }
204
+ momentum_optimizer_value: 0.9
205
+ decay: 0.9
206
+ epsilon: 1.0
207
+ }
208
+ }
209
+ }
210
+
211
+ train_input_reader {
212
+ label_map_path: "PATH_TO_BE_CONFIGURED/face_person_with_keypoints_label_map.pbtxt"
213
+ tf_record_input_reader {
214
+ input_path: "PATH_TO_BE_CONFIGURED/faces_train.record-?????-of-00010"
215
+ }
216
+ num_keypoints: 23
217
+ }
218
+
219
+ eval_config {
220
+ num_visualizations: 10
221
+ metrics_set: "coco_detection_metrics"
222
+ use_moving_averages: true
223
+ parameterized_metric {
224
+ coco_keypoint_metrics {
225
+ class_label: "face"
226
+ }
227
+ }
228
+ parameterized_metric {
229
+ coco_keypoint_metrics {
230
+ class_label: "PERSON"
231
+ }
232
+ }
233
+ }
234
+
235
+ eval_input_reader {
236
+ label_map_path: "PATH_TO_BE_CONFIGURED/face_person_with_keypoints_label_map.pbtxt"
237
+ shuffle: true
238
+ num_epochs: 1
239
+ tf_record_input_reader {
240
+ input_path: "PATH_TO_BE_CONFIGURED/faces_val.record-?????-of-00010"
241
+ }
242
+ num_keypoints: 23
243
+ }
244
+
245
+ graph_rewriter {
246
+ quantization {
247
+ delay: 2000000
248
+ activation_bits: 8
249
+ weight_bits: 8
250
+ }
251
+ }