[Fix] #1428: remove max_seq_len inconsistency (#1433)

open-mmlab · Oct 9, 2022 · 5fc9204 · 5fc9204
1 parent b422ded
commit 5fc9204
Show file tree

Hide file tree

Showing 13 changed files with 80 additions and 25 deletions.
diff --git a/configs/_base_/recog_models/abinet.py b/configs/_base_/recog_models/abinet.py
@@ -11,7 +11,7 @@
     with_unknown=False,
     with_padding=False,
     lower=True,
-)
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='ABINet',

diff --git a/configs/_base_/recog_models/master.py b/configs/_base_/recog_models/master.py
@@ -1,5 +1,10 @@
+max_seq_len = 30
+
 label_convertor = dict(
-    type='AttnConvertor', dict_type='DICT90', with_unknown=True)
+    type='AttnConvertor',
+    dict_type='DICT90',
+    with_unknown=True,
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='MASTER',
@@ -58,4 +63,4 @@
         feat_size=6 * 40),
     loss=dict(type='TFLoss', reduction='mean'),
     label_convertor=label_convertor,
-    max_seq_len=30)
+    max_seq_len=max_seq_len)
diff --git a/configs/_base_/recog_models/nrtr_modality_transform.py b/configs/_base_/recog_models/nrtr_modality_transform.py
@@ -1,5 +1,11 @@
+max_seq_len = 40
+
 label_convertor = dict(
-    type='AttnConvertor', dict_type='DICT36', with_unknown=True, lower=True)
+    type='AttnConvertor',
+    dict_type='DICT36',
+    with_unknown=True,
+    lower=True,
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='NRTR',
@@ -8,4 +14,4 @@
     decoder=dict(type='NRTRDecoder'),
     loss=dict(type='TFLoss'),
     label_convertor=label_convertor,
-    max_seq_len=40)
+    max_seq_len=max_seq_len)
diff --git a/configs/_base_/recog_models/robust_scanner.py b/configs/_base_/recog_models/robust_scanner.py
@@ -1,5 +1,10 @@
+max_seq_len = 30
+
 label_convertor = dict(
-    type='AttnConvertor', dict_type='DICT90', with_unknown=True)
+    type='AttnConvertor',
+    dict_type='DICT90',
+    with_unknown=True,
+    max_seq_len=max_seq_len)
 
 hybrid_decoder = dict(type='SequenceAttentionDecoder')
 
@@ -21,4 +26,4 @@
         position_decoder=position_decoder),
     loss=dict(type='SARLoss'),
     label_convertor=label_convertor,
-    max_seq_len=30)
+    max_seq_len=max_seq_len)
diff --git a/configs/_base_/recog_models/sar.py b/configs/_base_/recog_models/sar.py
@@ -1,5 +1,10 @@
+max_seq_len = 30
+
 label_convertor = dict(
-    type='AttnConvertor', dict_type='DICT90', with_unknown=True)
+    type='AttnConvertor',
+    dict_type='DICT90',
+    with_unknown=True,
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='SARNet',
@@ -21,4 +26,4 @@
         pred_concat=True),
     loss=dict(type='SARLoss'),
     label_convertor=label_convertor,
-    max_seq_len=30)
+    max_seq_len=max_seq_len)
diff --git a/configs/_base_/recog_models/satrn.py b/configs/_base_/recog_models/satrn.py
@@ -1,5 +1,11 @@
+max_seq_len = 40
+
 label_convertor = dict(
-    type='AttnConvertor', dict_type='DICT36', with_unknown=True, lower=True)
+    type='AttnConvertor',
+    dict_type='DICT36',
+    with_unknown=True,
+    lower=True,
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='SATRN',
@@ -8,4 +14,4 @@
     decoder=dict(type='TFDecoder'),
     loss=dict(type='TFLoss'),
     label_convertor=label_convertor,
-    max_seq_len=40)
+    max_seq_len=max_seq_len)
diff --git a/configs/textrecog/abinet/abinet_vision_only_academic.py b/configs/textrecog/abinet/abinet_vision_only_academic.py
@@ -21,7 +21,7 @@
     with_unknown=False,
     with_padding=False,
     lower=True,
-)
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='ABINet',

diff --git a/configs/textrecog/nrtr/nrtr_r31_1by16_1by8_academic.py b/configs/textrecog/nrtr/nrtr_r31_1by16_1by8_academic.py
@@ -12,8 +12,13 @@
 train_pipeline = {{_base_.train_pipeline}}
 test_pipeline = {{_base_.test_pipeline}}
 
+max_seq_len = 40
+
 label_convertor = dict(
-    type='AttnConvertor', dict_type='DICT90', with_unknown=True)
+    type='AttnConvertor',
+    dict_type='DICT90',
+    with_unknown=True,
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='NRTR',
@@ -27,7 +32,7 @@
     decoder=dict(type='NRTRDecoder'),
     loss=dict(type='TFLoss'),
     label_convertor=label_convertor,
-    max_seq_len=40)
+    max_seq_len=max_seq_len)
 
 data = dict(
     samples_per_gpu=128,

diff --git a/configs/textrecog/nrtr/nrtr_r31_1by8_1by4_academic.py b/configs/textrecog/nrtr/nrtr_r31_1by8_1by4_academic.py
@@ -12,8 +12,13 @@
 train_pipeline = {{_base_.train_pipeline}}
 test_pipeline = {{_base_.test_pipeline}}
 
+max_seq_len = 40
+
 label_convertor = dict(
-    type='AttnConvertor', dict_type='DICT90', with_unknown=True)
+    type='AttnConvertor',
+    dict_type='DICT90',
+    with_unknown=True,
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='NRTR',
@@ -27,7 +32,7 @@
     decoder=dict(type='NRTRDecoder'),
     loss=dict(type='TFLoss'),
     label_convertor=label_convertor,
-    max_seq_len=40)
+    max_seq_len=max_seq_len)
 
 data = dict(
     samples_per_gpu=64,

diff --git a/configs/textrecog/sar/sar_r31_parallel_decoder_chinese.py b/configs/textrecog/sar/sar_r31_parallel_decoder_chinese.py
@@ -2,10 +2,13 @@
     '../../_base_/default_runtime.py',
     '../../_base_/schedules/schedule_adam_step_5e.py'
 ]
-
+max_seq_len = 30
 dict_file = 'data/chineseocr/labels/dict_printed_chinese_english_digits.txt'
 label_convertor = dict(
-    type='AttnConvertor', dict_file=dict_file, with_unknown=True)
+    type='AttnConvertor',
+    dict_file=dict_file,
+    with_unknown=True,
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='SARNet',
@@ -27,7 +30,7 @@
         pred_concat=True),
     loss=dict(type='SARLoss'),
     label_convertor=label_convertor,
-    max_seq_len=30)
+    max_seq_len=max_seq_len)
 
 img_norm_cfg = dict(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
 train_pipeline = [

diff --git a/configs/textrecog/sar/sar_r31_sequential_decoder_academic.py b/configs/textrecog/sar/sar_r31_sequential_decoder_academic.py
@@ -12,8 +12,13 @@
 train_pipeline = {{_base_.train_pipeline}}
 test_pipeline = {{_base_.test_pipeline}}
 
+max_seq_len = 30
+
 label_convertor = dict(
-    type='AttnConvertor', dict_type='DICT90', with_unknown=True)
+    type='AttnConvertor',
+    dict_type='DICT90',
+    with_unknown=True,
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='SARNet',
@@ -35,7 +40,7 @@
         pred_concat=True),
     loss=dict(type='SARLoss'),
     label_convertor=label_convertor,
-    max_seq_len=30)
+    max_seq_len=max_seq_len)
 
 data = dict(
     samples_per_gpu=64,

diff --git a/configs/textrecog/satrn/satrn_academic.py b/configs/textrecog/satrn/satrn_academic.py
@@ -12,8 +12,13 @@
 train_pipeline = {{_base_.train_pipeline}}
 test_pipeline = {{_base_.test_pipeline}}
 
+max_seq_len = 25
+
 label_convertor = dict(
-    type='AttnConvertor', dict_type='DICT90', with_unknown=True)
+    type='AttnConvertor',
+    dict_type='DICT90',
+    with_unknown=True,
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='SATRN',
@@ -39,7 +44,7 @@
         d_v=512 // 8),
     loss=dict(type='TFLoss'),
     label_convertor=label_convertor,
-    max_seq_len=25)
+    max_seq_len=max_seq_len)
 
 # optimizer
 optimizer = dict(type='Adam', lr=3e-4)

diff --git a/configs/textrecog/satrn/satrn_small.py b/configs/textrecog/satrn/satrn_small.py
@@ -12,8 +12,13 @@
 train_pipeline = {{_base_.train_pipeline}}
 test_pipeline = {{_base_.test_pipeline}}
 
+max_seq_len = 25
+
 label_convertor = dict(
-    type='AttnConvertor', dict_type='DICT90', with_unknown=True)
+    type='AttnConvertor',
+    dict_type='DICT90',
+    with_unknown=True,
+    max_seq_len=max_seq_len)
 
 model = dict(
     type='SATRN',
@@ -39,7 +44,7 @@
         d_v=256 // 8),
     loss=dict(type='TFLoss'),
     label_convertor=label_convertor,
-    max_seq_len=25)
+    max_seq_len=max_seq_len)
 
 # optimizer
 optimizer = dict(type='Adam', lr=3e-4)