Reza2kn commited on
Commit
204b15a
·
verified ·
1 Parent(s): 6d05289

Preserve aligned169k-step12000-gated-native-v4

Browse files
.gitattributes CHANGED
@@ -42,3 +42,4 @@ checkpoints/eval-00006/best_accuracy.pdparams filter=lfs diff=lfs merge=lfs -tex
42
  checkpoints/eval-00009/best_accuracy.pdparams filter=lfs diff=lfs merge=lfs -text
43
  Bina-0.2-RizehPizeh.pdparams filter=lfs diff=lfs merge=lfs -text
44
  inference/inference.pdiparams filter=lfs diff=lfs merge=lfs -text
 
 
42
  checkpoints/eval-00009/best_accuracy.pdparams filter=lfs diff=lfs merge=lfs -text
43
  Bina-0.2-RizehPizeh.pdparams filter=lfs diff=lfs merge=lfs -text
44
  inference/inference.pdiparams filter=lfs diff=lfs merge=lfs -text
45
+ checkpoints/aligned169k-step12000-gated-native-v4/best_accuracy.pdparams filter=lfs diff=lfs merge=lfs -text
checkpoints/aligned169k-step12000-gated-native-v4/BEST_METRIC.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ [2026/07/25 19:13:54] ppocr INFO: best metric, worst_norm_edit_dis: 0.935414074851704, is_float16: False, acc: 0.6499023421633244, norm_edit_dis: 0.9447693838339172, handwriting_acc: 0.6782226529383659, handwriting_norm_edit_dis: 0.9541246930858112, printed_acc: 0.6215820282149315, printed_norm_edit_dis: 0.935414074851704, fps: 99.02456991783002, best_epoch: 3
checkpoints/aligned169k-step12000-gated-native-v4/SNAPSHOT.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"schema":"bina02.gated-checkpoint-snapshot.v1","source_run":"/home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/runs/aligned169k-fallback-blocks5-ctc-lr1e5-bs32-v4","metric_count":12,"label":"aligned169k-step12000-gated-native-v4","exact_train_rows":169000,"eval_handwriting":2048,"eval_printed":2048,"trainable_parameters":7584099,"reading_order_pipeline":"spread_right_first","benchmark_gate_status":"not_yet_below_10"}
checkpoints/aligned169k-step12000-gated-native-v4/best_accuracy.pdparams ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cb7fda82737fc2a0d5b7e620c04d1b5fe57c08def29cf7527225590c7c55e76a
3
+ size 142364527
checkpoints/aligned169k-step12000-gated-native-v4/best_accuracy.states ADDED
Binary file (345 Bytes). View file
 
checkpoints/aligned169k-step12000-gated-native-v4/config.yml ADDED
@@ -0,0 +1,164 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Global:
2
+ model_name: PP-OCRv6_medium_rec
3
+ debug: false
4
+ use_gpu: true
5
+ epoch_num: 4
6
+ log_smooth_window: 20
7
+ print_batch_step: 10
8
+ save_model_dir: /home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/runs/aligned169k-fallback-blocks5-ctc-lr1e5-bs32-v4
9
+ save_epoch_step: 1
10
+ eval_batch_step:
11
+ - 0
12
+ - 1000
13
+ cal_metric_during_train: false
14
+ pretrained_model: /home/rezo/bina02-ppocr/clean-restart/packages/Bina-0.2-RizehPizeh/Bina-0.2-RizehPizeh
15
+ checkpoints: null
16
+ save_inference_dir: null
17
+ use_visualdl: false
18
+ infer_img: doc/imgs_words/ch/word_1.jpg
19
+ character_dict_path: /home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/datasets/aligned169k-v1/persian_arabic_bina02_dict.txt
20
+ max_text_length: 192
21
+ infer_mode: false
22
+ use_space_char: true
23
+ distributed: false
24
+ save_res_path: /home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/runs/aligned169k-fallback-blocks5-ctc-lr1e5-bs32-v4/predicts.txt
25
+ d2s_train_image_shape:
26
+ - 3
27
+ - 48
28
+ - 1536
29
+ freeze_batch_norm: true
30
+ trainable_prefixes:
31
+ - backbone.blocks5
32
+ - head.ctc_encoder
33
+ - head.ctc_head
34
+ eval_handwriting_rows: 2048
35
+ eval_printed_rows: 2048
36
+ use_amp: false
37
+ Optimizer:
38
+ name: Adam
39
+ beta1: 0.9
40
+ beta2: 0.999
41
+ lr:
42
+ name: Cosine
43
+ learning_rate: 1.0e-05
44
+ warmup_epoch: 0
45
+ regularizer:
46
+ name: L2
47
+ factor: 3.0e-05
48
+ Architecture:
49
+ model_type: rec
50
+ algorithm: SVTR_LCNet
51
+ Transform: null
52
+ Backbone:
53
+ name: PPLCNetV4
54
+ model_size: medium
55
+ train_pool_width: 192
56
+ force_pool_width: true
57
+ Head:
58
+ name: MultiHead
59
+ head_list:
60
+ - CTCHead:
61
+ Neck:
62
+ name: lightsvtr
63
+ dims: 192
64
+ depth: 2
65
+ mlp_ratio: 4.0
66
+ local_kernel: 7
67
+ use_guide: false
68
+ Head:
69
+ fc_decay: 1.0e-05
70
+ - NRTRHead:
71
+ nrtr_dim: 512
72
+ max_text_length: 192
73
+ Loss:
74
+ name: MultiLoss
75
+ weight_1: 1.0
76
+ weight_2: 0.0
77
+ loss_config_list:
78
+ - CTCLoss: null
79
+ - NRTRLoss: null
80
+ PostProcess:
81
+ name: CTCLabelDecode
82
+ Metric:
83
+ name: RecMetric
84
+ main_indicator: worst_norm_edit_dis
85
+ ignore_space: false
86
+ Train:
87
+ dataset:
88
+ name: MultiScaleDataSet
89
+ ds_width: false
90
+ data_dir: /
91
+ ext_op_transform_idx: 1
92
+ label_file_list:
93
+ - /home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/datasets/aligned169k-v1/train_ctc.txt
94
+ ratio_list:
95
+ - 1.0
96
+ transforms:
97
+ - DecodeImage:
98
+ img_mode: BGR
99
+ channel_first: false
100
+ - RecConAug:
101
+ prob: 0.0
102
+ ext_data_num: 0
103
+ image_shape:
104
+ - 48
105
+ - 1536
106
+ - 3
107
+ max_text_length: 192
108
+ - RecAug: null
109
+ - MultiLabelEncode:
110
+ gtc_encode: NRTRLabelEncode
111
+ - KeepKeys:
112
+ keep_keys:
113
+ - image
114
+ - label_ctc
115
+ - label_gtc
116
+ - length
117
+ - valid_ratio
118
+ sampler:
119
+ name: MultiScaleSampler
120
+ scales:
121
+ - - 1536
122
+ - 48
123
+ first_bs: 32
124
+ fix_bs: true
125
+ divided_factor:
126
+ - 8
127
+ - 16
128
+ is_training: true
129
+ loader:
130
+ shuffle: true
131
+ batch_size_per_card: 32
132
+ drop_last: false
133
+ num_workers: 8
134
+ Eval:
135
+ dataset:
136
+ name: SimpleDataSet
137
+ data_dir: /
138
+ label_file_list:
139
+ - /home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/datasets/aligned169k-v1/eval_ctc.txt
140
+ transforms:
141
+ - DecodeImage:
142
+ img_mode: BGR
143
+ channel_first: false
144
+ - MultiLabelEncode:
145
+ gtc_encode: NRTRLabelEncode
146
+ - RecResizeImg:
147
+ image_shape:
148
+ - 3
149
+ - 48
150
+ - 1536
151
+ padding: true
152
+ - KeepKeys:
153
+ keep_keys:
154
+ - image
155
+ - label_ctc
156
+ - label_gtc
157
+ - length
158
+ - valid_ratio
159
+ loader:
160
+ shuffle: false
161
+ drop_last: false
162
+ batch_size_per_card: 32
163
+ num_workers: 4
164
+ profiler_options: null
checkpoints/aligned169k-step12000-gated-native-v4/input_config.yml ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Global:
2
+ model_name: PP-OCRv6_medium_rec
3
+ debug: false
4
+ use_gpu: true
5
+ epoch_num: 4
6
+ log_smooth_window: 20
7
+ print_batch_step: 10
8
+ save_model_dir: /home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/runs/aligned169k-fallback-blocks5-ctc-lr1e5-bs32-v4
9
+ save_epoch_step: 1
10
+ eval_batch_step:
11
+ - 0
12
+ - 1000
13
+ cal_metric_during_train: false
14
+ pretrained_model: /home/rezo/bina02-ppocr/clean-restart/packages/Bina-0.2-RizehPizeh/Bina-0.2-RizehPizeh
15
+ checkpoints: null
16
+ save_inference_dir: null
17
+ use_visualdl: false
18
+ infer_img: doc/imgs_words/ch/word_1.jpg
19
+ character_dict_path: /home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/datasets/aligned169k-v1/persian_arabic_bina02_dict.txt
20
+ max_text_length: 192
21
+ infer_mode: false
22
+ use_space_char: true
23
+ distributed: false
24
+ save_res_path: /home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/runs/aligned169k-fallback-blocks5-ctc-lr1e5-bs32-v4/predicts.txt
25
+ d2s_train_image_shape:
26
+ - 3
27
+ - 48
28
+ - 1536
29
+ freeze_batch_norm: true
30
+ trainable_prefixes:
31
+ - backbone.blocks5
32
+ - head.ctc_encoder
33
+ - head.ctc_head
34
+ eval_handwriting_rows: 2048
35
+ eval_printed_rows: 2048
36
+ use_amp: false
37
+ Optimizer:
38
+ name: Adam
39
+ beta1: 0.9
40
+ beta2: 0.999
41
+ lr:
42
+ name: Cosine
43
+ learning_rate: 1.0e-05
44
+ warmup_epoch: 0
45
+ regularizer:
46
+ name: L2
47
+ factor: 3.0e-05
48
+ Architecture:
49
+ model_type: rec
50
+ algorithm: SVTR_LCNet
51
+ Transform: null
52
+ Backbone:
53
+ name: PPLCNetV4
54
+ model_size: medium
55
+ train_pool_width: 192
56
+ force_pool_width: true
57
+ Head:
58
+ name: MultiHead
59
+ head_list:
60
+ - CTCHead:
61
+ Neck:
62
+ name: lightsvtr
63
+ dims: 192
64
+ depth: 2
65
+ mlp_ratio: 4.0
66
+ local_kernel: 7
67
+ use_guide: false
68
+ Head:
69
+ fc_decay: 1.0e-05
70
+ - NRTRHead:
71
+ nrtr_dim: 512
72
+ max_text_length: 192
73
+ Loss:
74
+ name: MultiLoss
75
+ weight_1: 1.0
76
+ weight_2: 0.0
77
+ loss_config_list:
78
+ - CTCLoss: null
79
+ - NRTRLoss: null
80
+ PostProcess:
81
+ name: CTCLabelDecode
82
+ Metric:
83
+ name: RecMetric
84
+ main_indicator: worst_norm_edit_dis
85
+ ignore_space: false
86
+ Train:
87
+ dataset:
88
+ name: MultiScaleDataSet
89
+ ds_width: false
90
+ data_dir: /
91
+ ext_op_transform_idx: 1
92
+ label_file_list:
93
+ - /home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/datasets/aligned169k-v1/train_ctc.txt
94
+ ratio_list:
95
+ - 1.0
96
+ transforms:
97
+ - DecodeImage:
98
+ img_mode: BGR
99
+ channel_first: false
100
+ - RecConAug:
101
+ prob: 0.0
102
+ ext_data_num: 0
103
+ image_shape:
104
+ - 48
105
+ - 1536
106
+ - 3
107
+ max_text_length: 192
108
+ - RecAug: null
109
+ - MultiLabelEncode:
110
+ gtc_encode: NRTRLabelEncode
111
+ - KeepKeys:
112
+ keep_keys:
113
+ - image
114
+ - label_ctc
115
+ - label_gtc
116
+ - length
117
+ - valid_ratio
118
+ sampler:
119
+ name: MultiScaleSampler
120
+ scales:
121
+ - - 1536
122
+ - 48
123
+ first_bs: 32
124
+ fix_bs: true
125
+ divided_factor:
126
+ - 8
127
+ - 16
128
+ is_training: true
129
+ loader:
130
+ shuffle: true
131
+ batch_size_per_card: 32
132
+ drop_last: false
133
+ num_workers: 8
134
+ Eval:
135
+ dataset:
136
+ name: SimpleDataSet
137
+ data_dir: /
138
+ label_file_list:
139
+ - /home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/datasets/aligned169k-v1/eval_ctc.txt
140
+ transforms:
141
+ - DecodeImage:
142
+ img_mode: BGR
143
+ channel_first: false
144
+ - MultiLabelEncode:
145
+ gtc_encode: NRTRLabelEncode
146
+ - RecResizeImg:
147
+ image_shape:
148
+ - 3
149
+ - 48
150
+ - 1536
151
+ padding: true
152
+ - KeepKeys:
153
+ keep_keys:
154
+ - image
155
+ - label_ctc
156
+ - label_gtc
157
+ - length
158
+ - valid_ratio
159
+ loader:
160
+ shuffle: false
161
+ drop_last: false
162
+ batch_size_per_card: 32
163
+ num_workers: 4
checkpoints/aligned169k-step12000-gated-native-v4/run.json ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "bina02.ppocrv6.native-training-run.v1",
3
+ "created_at": "2026-07-26T00:16:36.287774+00:00",
4
+ "dataset_contract": "/home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/datasets/aligned169k-v1/contract.json",
5
+ "train_rows": 169000,
6
+ "eval_rows": 4096,
7
+ "input_shape": [
8
+ 3,
9
+ 48,
10
+ 1536
11
+ ],
12
+ "aspect_ratio_padding": true,
13
+ "ctc_timesteps": 192,
14
+ "maximum_required_timesteps": 95,
15
+ "train_inference_pooling_parity": true,
16
+ "trainable_prefixes": [
17
+ "backbone.blocks5",
18
+ "head.ctc_encoder",
19
+ "head.ctc_head"
20
+ ],
21
+ "freeze_batch_norm": true,
22
+ "devices": [
23
+ "0"
24
+ ],
25
+ "world_size": 1,
26
+ "batch_size_per_card": 32,
27
+ "global_batch_size": 32,
28
+ "steps_per_epoch": 5282,
29
+ "eval_every_steps": 1000,
30
+ "learning_rate": 1e-05,
31
+ "checkpoint_indicator": "worst_norm_edit_dis",
32
+ "command": [
33
+ "/home/rezo/bina02-ppocr/.venv/bin/python",
34
+ "-u",
35
+ "/home/rezo/bina02-ppocr/recovery-native-20260725/train_bina02_native_stage.py",
36
+ "-c",
37
+ "/home/rezo/bina02-ppocr/recovery-native-20260725/workspace/native/runs/aligned169k-fallback-blocks5-ctc-lr1e5-bs32-v4/input_config.yml"
38
+ ]
39
+ }