FangSen9000 commited on
Commit
dde130b
·
1 Parent(s): d6bdfb4

Runs normally, but training still has some issues

Browse files
.gitignore CHANGED
@@ -5,3 +5,6 @@ plugins/ASLLRP_utterances_results
5
  plugins/CorrNet_Plus_modified
6
  plugins/dtwSign-X
7
  plugins/ASLLRP_info
 
 
 
 
5
  plugins/CorrNet_Plus_modified
6
  plugins/dtwSign-X
7
  plugins/ASLLRP_info
8
+ plugins/YouTubeASL
9
+ plugins/SignX/smkd/work_dir
10
+ plugins/SignX/smkd/youtubeasl
env/{requirements.txt → dashboard.txt} RENAMED
File without changes
env/signx-slt.txt ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ absl-py==2.1.0
2
+ accelerate==1.0.1
3
+ Brotli @ file:///home/conda/feedstock_root/build_artifacts/brotli-split_1695989787169/work
4
+ cachetools==5.5.1
5
+ certifi @ file:///home/conda/feedstock_root/build_artifacts/certifi_1725278078093/work/certifi
6
+ cffi @ file:///home/conda/feedstock_root/build_artifacts/cffi_1723018376978/work
7
+ charset-normalizer @ file:///home/conda/feedstock_root/build_artifacts/charset-normalizer_1728479282467/work
8
+ click==8.1.8
9
+ contourpy==1.1.1
10
+ cycler==0.12.1
11
+ diffusers==0.32.2
12
+ fastdtw==0.3.4
13
+ filelock @ file:///home/conda/feedstock_root/build_artifacts/filelock_1726613473834/work
14
+ fonttools==4.55.4
15
+ fsspec==2024.12.0
16
+ gmpy2 @ file:///home/conda/feedstock_root/build_artifacts/gmpy2_1715527302982/work
17
+ google-auth==2.37.0
18
+ google-auth-oauthlib==1.0.0
19
+ GPUtil==1.4.0
20
+ graphviz==0.20.3
21
+ grpcio==1.69.0
22
+ h2 @ file:///home/conda/feedstock_root/build_artifacts/h2_1634280454336/work
23
+ h5py==3.11.0
24
+ hiddenlayer==0.3
25
+ hpack==4.0.0
26
+ huggingface-hub==0.27.1
27
+ hyperframe @ file:///home/conda/feedstock_root/build_artifacts/hyperframe_1619110129307/work
28
+ idna @ file:///home/conda/feedstock_root/build_artifacts/idna_1726459485162/work
29
+ importlib_metadata==8.5.0
30
+ importlib_resources==6.4.5
31
+ Jinja2 @ file:///home/conda/feedstock_root/build_artifacts/jinja2_1715127149914/work
32
+ joblib==1.4.2
33
+ kiwisolver==1.4.7
34
+ Markdown==3.7
35
+ MarkupSafe @ file:///home/conda/feedstock_root/build_artifacts/markupsafe_1706899923320/work
36
+ matplotlib==3.7.5
37
+ mpmath @ file:///home/conda/feedstock_root/build_artifacts/mpmath_1678228039184/work
38
+ netron==8.1.3
39
+ networkx @ file:///home/conda/feedstock_root/build_artifacts/networkx_1680692919326/work
40
+ nltk==3.9.1
41
+ numpy @ file:///home/conda/feedstock_root/build_artifacts/numpy_1687808301083/work
42
+ oauthlib==3.2.2
43
+ opencv-python==4.11.0.86
44
+ packaging==24.2
45
+ pandas==2.0.3
46
+ pillow @ file:///home/conda/feedstock_root/build_artifacts/pillow_1719903565503/work
47
+ protobuf==5.29.3
48
+ psutil==7.0.0
49
+ pyarrow==17.0.0
50
+ pyasn1==0.6.1
51
+ pyasn1_modules==0.4.1
52
+ pycparser @ file:///home/conda/feedstock_root/build_artifacts/pycparser_1711811537435/work
53
+ pyparsing==3.1.4
54
+ PySocks @ file:///home/conda/feedstock_root/build_artifacts/pysocks_1661604839144/work
55
+ python-dateutil==2.9.0.post0
56
+ pytz==2024.2
57
+ PyYAML @ file:///home/conda/feedstock_root/build_artifacts/pyyaml_1723018227672/work
58
+ regex==2024.11.6
59
+ requests @ file:///home/conda/feedstock_root/build_artifacts/requests_1717057054362/work
60
+ requests-oauthlib==2.0.0
61
+ rsa==4.9
62
+ safetensors==0.5.2
63
+ scikit-learn==1.3.2
64
+ scipy==1.10.1
65
+ sentencepiece==0.2.0
66
+ six==1.17.0
67
+ sympy @ file:///home/conda/feedstock_root/build_artifacts/sympy_1728484478345/work
68
+ tensorboard==2.14.0
69
+ tensorboard-data-server==0.7.2
70
+ threadpoolctl==3.5.0
71
+ timm==1.0.20
72
+ tokenizers==0.20.3
73
+ torch==2.4.1
74
+ torchaudio==2.4.1
75
+ torchinfo==1.8.0
76
+ torchview==0.2.6
77
+ torchvision==0.20.0
78
+ torchviz==0.0.3
79
+ tqdm==4.67.1
80
+ transformers==4.46.3
81
+ triton==3.0.0
82
+ typing_extensions @ file:///home/conda/feedstock_root/build_artifacts/typing_extensions_1717802530399/work
83
+ tzdata==2025.1
84
+ urllib3 @ file:///home/conda/feedstock_root/build_artifacts/urllib3_1726496430923/work
85
+ Werkzeug==3.0.6
86
+ zipp==3.20.2
87
+ zstandard==0.23.0
88
+
env/slt_tf1.txt ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ absl-py==2.1.0
2
+ astor==0.8.1
3
+ certifi @ file:///croot/certifi_1671487769961/work/certifi
4
+ cycler==0.11.0
5
+ fonttools==4.38.0
6
+ gast==0.2.2
7
+ google-pasta==0.2.0
8
+ grpcio==1.62.3
9
+ h5py==3.8.0
10
+ importlib-metadata==6.7.0
11
+ Keras-Applications==1.0.8
12
+ Keras-Preprocessing==1.1.2
13
+ kiwisolver==1.4.5
14
+ Markdown==3.4.4
15
+ MarkupSafe==2.1.5
16
+ matplotlib==3.5.3
17
+ numpy==1.21.6
18
+ nvidia-cublas-cu11==11.10.3.66
19
+ nvidia-cuda-nvrtc-cu11==11.7.99
20
+ nvidia-cuda-runtime-cu11==11.7.99
21
+ nvidia-cudnn-cu11==8.5.0.96
22
+ opt-einsum==3.3.0
23
+ packaging==24.0
24
+ Pillow==9.5.0
25
+ protobuf==4.24.4
26
+ pyparsing==3.1.4
27
+ python-dateutil==2.9.0.post0
28
+ PyYAML==6.0.1
29
+ six==1.17.0
30
+ tensorboard==1.15.0
31
+ tensorflow==1.15.0
32
+ tensorflow-estimator==1.15.1
33
+ termcolor==2.3.0
34
+ tqdm==4.67.1
35
+ typing_extensions==4.7.1
36
+ Werkzeug==2.2.3
37
+ wrapt==1.16.0
38
+ zipp==3.15.0
39
+
plugins/SignX/smkd/asllrp第七次训练全pose协助2000/gloss_dict.txt ADDED
@@ -0,0 +1,1243 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 1 13 "I
2
+ 2 34 "WHAT"
3
+ 3 3 "pause"
4
+ 4 4 "perplexed"
5
+ 5 7 "you
6
+ 6 10 #ALL
7
+ 7 2 #APT
8
+ 8 4 #ASL
9
+ 9 2 #BACK
10
+ 10 15 #BBQ
11
+ 11 2 #BURN
12
+ 12 3 #CAR
13
+ 13 4 #DO
14
+ 14 2 #DO-IT
15
+ 15 23 #DOG
16
+ 16 2 #ED
17
+ 17 2 #FRESH
18
+ 18 3 #FUCK
19
+ 19 4 #GAS
20
+ 20 3 #GO
21
+ 21 13 #HS
22
+ 22 173 #IF
23
+ 23 4 #JOB
24
+ 24 81 #NO
25
+ 25 3 #OFF
26
+ 26 13 #OK
27
+ 27 11 #OR
28
+ 28 2 #OWN
29
+ 29 13 #PO
30
+ 30 6 #QUIZ
31
+ 31 12 #SALE
32
+ 32 34 #SO
33
+ 33 6 #TV
34
+ 34 12 #VP
35
+ 35 4 #WHAT
36
+ 36 52 #WHEN
37
+ 37 7 (1h)5"perplexed"
38
+ 38 2 (1h)NOW+NIGHT
39
+ 39 2 (1h)TEACH+(1h)AGENT
40
+ 40 4 (25)WHY+NOT
41
+ 41 2 (2h)#NO+MORE
42
+ 42 4 (2h)ANY+(2h)THING
43
+ 43 7 (2h)KNOW+NEG
44
+ 44 3 (5)FIRST-IN-LIST
45
+ 45 12 (5)WEATHER
46
+ 46 19 (K)KITCHEN
47
+ 47 3 (L-X)INTERESTING
48
+ 48 2 1"no"
49
+ 49 4 5"dismissive
50
+ 50 10 5"exasperation"
51
+ 51 3 5"forget
52
+ 52 3 5"frustration"
53
+ 53 3 5"hands
54
+ 54 8 5"hesitation"
55
+ 55 3 5"hey"
56
+ 56 4 5"let's
57
+ 57 2 5"looking
58
+ 58 4 5"no
59
+ 59 5 5"not
60
+ 60 4 5"oh
61
+ 61 3 5"ok
62
+ 62 6 5"out
63
+ 63 2 5"panic"
64
+ 64 3 5"resignation"
65
+ 65 2 5"shocked"
66
+ 66 3 5"so
67
+ 67 5 5"surprised"
68
+ 68 2 5"that's
69
+ 69 2 5"trepidation"
70
+ 70 7 5"umm"
71
+ 71 87 5"wave
72
+ 72 2 5"well"
73
+ 73 13 5"whoa"
74
+ 74 19 5"wow"
75
+ 75 7 5"you
76
+ 76 3 A
77
+ 77 2 A-LEVEL-ABOVE
78
+ 78 3 A-LEVEL-BELOW
79
+ 79 20 A-LOT
80
+ 80 5 A-OK
81
+ 81 25 ABOUT
82
+ 82 5 ABOVE
83
+ 83 20 ACCEPT
84
+ 84 10 ACCIDENT
85
+ 85 2 ACCOMMODATE
86
+ 86 17 ACT
87
+ 87 17 ACTION
88
+ 88 4 ADD-TO
89
+ 89 2 ADDICTED
90
+ 90 6 ADULT-TALL
91
+ 91 3 ADVISE/INFLUENCE
92
+ 92 8 AFRAID
93
+ 93 11 AGAIN
94
+ 94 3 AGAINST
95
+ 95 2 AGE-TWENTY-ONE
96
+ 96 4 AGREE
97
+ 97 3 AIRPLANE
98
+ 98 4 ALARM
99
+ 99 7 ALL
100
+ 100 8 ALL-DAY
101
+ 101 12 ALL-NIGHT
102
+ 102 3 ALLERGY
103
+ 103 9 ALLOW
104
+ 104 6 ALMOST
105
+ 105 6 ALONE
106
+ 106 16 ALSO
107
+ 107 14 ALWAYS
108
+ 108 13 AMONG
109
+ 109 9 ANALYZE
110
+ 110 28 AND
111
+ 111 7 ANGRY
112
+ 112 2 ANIMAL
113
+ 113 4 ANSWER
114
+ 114 2 ANSWER+AGENT
115
+ 115 5 ANY
116
+ 116 3 ANY+ONE
117
+ 117 6 ANYWAY/NOT-MATTER
118
+ 118 7 APPLAUSE
119
+ 119 3 APPLE
120
+ 120 6 APPLY
121
+ 121 2 APPOINTMENT
122
+ 122 13 AREA
123
+ 123 8 AROUND
124
+ 124 80 ARRIVE
125
+ 125 3 ARTICLE
126
+ 126 16 ASK
127
+ 127 3 ATTEMPT
128
+ 128 8 ATTENTION-WAVE
129
+ 129 14 AUNT
130
+ 130 2 AVERAGE
131
+ 131 2 B
132
+ 132 10 BABY
133
+ 133 2 BACK
134
+ 134 1 BACKWARD_2
135
+ 135 11 BAD
136
+ 136 2 BAKING-SPRINKLES
137
+ 137 2 BALANCE
138
+ 138 7 BALD
139
+ 139 4 BASEBALL
140
+ 140 4 BASKETBALL
141
+ 141 5 BAWL-OUT
142
+ 142 2 BCL:C"grab
143
+ 143 2 BCL:crvd-5"holding
144
+ 144 2 BCL:crvd-5"swashbuckler
145
+ 145 3 BEAT
146
+ 146 13 BEAUTIFUL
147
+ 147 7 BECAUSE
148
+ 148 12 BECOME
149
+ 149 2 BED
150
+ 150 12 BEFORE
151
+ 151 5 BELIEVE
152
+ 152 22 BELONG
153
+ 153 8 BELOW_2
154
+ 154 11 BETTER
155
+ 155 9 BETWEEN/SHARE
156
+ 156 24 BICYCLE
157
+ 157 24 BIG
158
+ 158 2 BILL/OWE
159
+ 159 6 BIRTHDAY
160
+ 160 7 BIRTHDAY_3
161
+ 161 6 BLACK
162
+ 162 2 BLAME
163
+ 163 3 BLUE
164
+ 164 3 BLUE+BERRY
165
+ 165 4 BLUE+fs-BERRY
166
+ 166 2 BODY
167
+ 167 3 BOMB
168
+ 168 49 BOOK
169
+ 169 3 BOOK-PAGE
170
+ 170 13 BORE
171
+ 171 16 BORN
172
+ 172 7 BOSS
173
+ 173 2 BOTH
174
+ 174 3 BOTHER
175
+ 175 10 BOX/ROOM
176
+ 176 16 BOY
177
+ 177 2 BOY+SCOUT
178
+ 178 1 BPCL:1"going
179
+ 179 7 BPCL:1"voice
180
+ 180 1 BPCL:bent-B"going
181
+ 181 1 BPCL:bent-B"gradually
182
+ 182 4 BRACELET
183
+ 183 5 BRAVE/RECOVER
184
+ 184 34 BREAK
185
+ 185 30 BREAK-DOWN
186
+ 186 9 BRING/CARRY
187
+ 187 7 BROTHER
188
+ 188 2 BURP
189
+ 189 2 BUSY/BUSINESS
190
+ 190 42 BUT
191
+ 191 10 BUTTER
192
+ 192 60 BUY
193
+ 193 2 Bible"
194
+ 194 7 CALL-BY-PHONE
195
+ 195 13 CAMP
196
+ 196 42 CAN
197
+ 197 30 CANCEL/CRITICIZE
198
+ 198 33 CANNOT
199
+ 199 97 CAR
200
+ 200 7 CARE
201
+ 201 8 CAT
202
+ 202 4 CELEBRATE
203
+ 203 4 CHANGE
204
+ 204 11 CHARACTER
205
+ 205 3 CHASE
206
+ 206 19 CHAT
207
+ 207 2 CHECK
208
+ 208 20 CHEESE
209
+ 209 14 CHILD
210
+ 210 14 CHOCOLATE
211
+ 211 2 CHOOSE
212
+ 212 2 CHURCH
213
+ 213 6 CITY/COMMUNITY
214
+ 214 28 CLASS
215
+ 215 12 CLASS_2+BOX/ROOM
216
+ 216 5 CLEAR
217
+ 217 4 CLEVER
218
+ 218 8 CLOSE-DOOR
219
+ 219 2 COAT
220
+ 220 13 COLD
221
+ 221 2 COLLAR
222
+ 222 7 COLLEGE
223
+ 223 3 COLOR
224
+ 224 9 COME
225
+ 225 10 COMMUNICATE
226
+ 226 17 COMMUTE
227
+ 227 4 COMPARE
228
+ 228 5 COMPETITION
229
+ 229 3 COMPUTER
230
+ 230 2 COMPUTER_5
231
+ 231 6 CONCEPT
232
+ 232 2 CONFRONT
233
+ 233 3 CONTACT
234
+ 234 9 CONTINUE
235
+ 235 3 COOK/KITCHEN
236
+ 236 6 COOKIE
237
+ 237 13 COOPERATE/UNITE
238
+ 238 14 CORRECT
239
+ 239 8 COUGH
240
+ 240 2 COUNTRY
241
+ 241 11 COURT
242
+ 242 2 COW
243
+ 243 12 CRASH
244
+ 244 5 CROWDED
245
+ 245 31 CULTURE
246
+ 246 2 CURRENT-TOPIC
247
+ 247 9 DANCE
248
+ 248 2 DANGER
249
+ 249 2 DATE/DESSERT
250
+ 250 2 DCL"flat
251
+ 251 5 DCL"round
252
+ 252 4 DCL:1"point
253
+ 253 3 DCL:4"many
254
+ 254 2 DCL:4"people
255
+ 255 7 DCL:5"emit
256
+ 256 3 DCL:5"extremely
257
+ 257 6 DCL:5"hair
258
+ 258 2 DCL:5-C-L"nails
259
+ 259 3 DCL:A"tear
260
+ 260 3 DCL:B"pathway
261
+ 261 2 DCL:C"group
262
+ 262 3 DCL:F/9"peel
263
+ 263 2 DCL:L-X"tube
264
+ 264 8 DCL:M"humming"
265
+ 265 2 DCL:S"heavy
266
+ 266 2 DCL:S-5"throw
267
+ 267 2 DCL:U-L"horse
268
+ 268 3 DCL:baby-O"writing
269
+ 269 2 DCL:bent-B-L"close
270
+ 270 3 DCL:crvd-5"clouds
271
+ 271 9 DCL:crvd-5"group
272
+ 272 1 DCL:crvd-5"reach
273
+ 273 4 DCL:crvd-B"above
274
+ 274 2 DCL:crvd-B"curve
275
+ 275 7 DCL:crvd-B"small
276
+ 276 2 DCL:fanned-O"throat
277
+ 277 2 DCL:fanned-flat-O"throat"
278
+ 278 8 DCL:flat-O"rip
279
+ 279 81 DEAF
280
+ 280 23 DECIDE
281
+ 281 2 DECORATE
282
+ 282 5 DEEP
283
+ 283 3 DELETE
284
+ 284 3 DELIBERATE
285
+ 285 4 DELICIOUS
286
+ 286 50 DEPART
287
+ 287 4 DEVELOP
288
+ 288 51 DIFFERENT
289
+ 289 9 DINNER
290
+ 290 20 DIRECT
291
+ 291 17 DISCUSS
292
+ 292 4 DISGUST
293
+ 293 2 DISTRIBUTE
294
+ 294 2 DO
295
+ 295 41 DO-DO
296
+ 296 5 DOCTOR
297
+ 297 11 DOG
298
+ 298 7 DOLLAR
299
+ 299 2 DORM
300
+ 300 4 DOUBT
301
+ 301 17 DRESS/CLOTHES
302
+ 302 8 DRINK
303
+ 303 34 DRIVE
304
+ 304 4 DROP
305
+ 305 11 DRUNK
306
+ 306 7 DRY-CEREAL
307
+ 307 12 DURING/WHILE
308
+ 308 7 DYE
309
+ 309 3 EACH
310
+ 310 9 EACH+ONE
311
+ 311 5 EASY
312
+ 312 31 EAT
313
+ 313 11 EAT+NIGHT
314
+ 314 2 ECONOMIC
315
+ 315 2 EIGHT
316
+ 316 2 EMERGENCY
317
+ 317 9 EMPTY/EARLY
318
+ 318 2 EMPTY/EARLY+(1h)MOST
319
+ 319 2 END
320
+ 320 17 ENJOY
321
+ 321 3 ENOUGH
322
+ 322 21 ENTER
323
+ 323 2 ENVIRONMENT
324
+ 324 5 EQUAL
325
+ 325 2 EVERY+ONE
326
+ 326 5 EVERY-MONTH/RENT
327
+ 327 13 EVERY-YEAR
328
+ 328 10 EVERYDAY
329
+ 329 2 EXACT
330
+ 330 2 EXHAUST
331
+ 331 7 EXPENSIVE
332
+ 332 34 EXPERIENCE
333
+ 333 11 EXPERT
334
+ 334 5 EXPRESS
335
+ 335 14 FACE
336
+ 336 3 FACE+SAME
337
+ 337 9 FAIL
338
+ 338 32 FAMILY
339
+ 339 5 FAMOUS
340
+ 340 14 FARM
341
+ 341 8 FAST
342
+ 342 8 FAST_3
343
+ 343 60 FATHER
344
+ 344 20 FAVORITE/PREFER
345
+ 345 18 FEEL
346
+ 346 3 FIGHT
347
+ 347 6 FINALLY
348
+ 348 23 FIND/FIND-OUT
349
+ 349 2 FIND/FIND-OUT+#OUT
350
+ 350 37 FINE
351
+ 351 124 FINISH
352
+ 352 17 FIRE/BURN
353
+ 353 3 FIRED
354
+ 354 10 FIREWORKS
355
+ 355 5 FIRST
356
+ 356 3 FIRST-AND-SECOND-IN-LIST
357
+ 357 13 FIRST-IN-LIST
358
+ 358 6 FISH
359
+ 359 12 FISHING
360
+ 360 2 FLAT-TIRE
361
+ 361 2 FLAT/FLOOR
362
+ 362 2 FLAT/LEVEL
363
+ 363 2 FLIP-FLOP
364
+ 364 2 FLOW
365
+ 365 15 FLOWER
366
+ 366 3 FLY-BY-PLANE
367
+ 367 8 FOCUS/NARROW
368
+ 368 3 FOLLOW
369
+ 369 4 FOOTBALL
370
+ 370 73 FOR
371
+ 371 5 FOR-FOR
372
+ 372 13 FORGET-IT
373
+ 373 2 FORMERLY
374
+ 374 2 FREE
375
+ 375 2 FREQUENTLY
376
+ 376 2 FRESHMAN
377
+ 377 2 FRIDAY
378
+ 378 241 FRIEND
379
+ 379 14 FRISBEE
380
+ 380 24 FROM
381
+ 381 2 FROM-NOW-ON
382
+ 382 3 FULL
383
+ 383 9 FUNNY
384
+ 384 163 FUTURE
385
+ 385 14 GAME
386
+ 386 2 GARAGE/SUBMARINE
387
+ 387 3 GATE
388
+ 388 2 GENERAL
389
+ 389 5 GESTURE
390
+ 390 19 GET
391
+ 391 3 GET-TICKET
392
+ 392 3 GET-UP
393
+ 393 15 GIFT
394
+ 394 3 GIRL
395
+ 395 3 GIRL+CORRECT
396
+ 396 5 GIRL+FRIEND
397
+ 397 16 GIVE
398
+ 398 3 GLASSES_2
399
+ 399 60 GO
400
+ 400 141 GO-OUT
401
+ 401 2 GOLD/ns-CALIFORNIA
402
+ 402 7 GONE
403
+ 403 51 GOOD/THANK-YOU
404
+ 404 8 GRAB-CHANCE
405
+ 405 7 GRADUATE
406
+ 406 6 GRASS
407
+ 407 7 GREAT
408
+ 408 4 GRIEF
409
+ 409 2 GROUP
410
+ 410 71 GROUP/TOGETHER
411
+ 411 6 GROW
412
+ 412 11 GROW-UP
413
+ 413 21 GUARANTEE
414
+ 414 4 GUILTY
415
+ 415 2 GUITAR
416
+ 416 2 HABIT
417
+ 417 7 HAIR
418
+ 418 2 HAIR+DYE
419
+ 419 4 HAIR_2+DYE
420
+ 420 2 HAMBURGER
421
+ 421 3 HANDSOME/ns-HAWAII_2
422
+ 422 23 HAPPEN
423
+ 423 16 HAPPY
424
+ 424 7 HARD
425
+ 425 2 HATE
426
+ 426 134 HAVE
427
+ 427 6 HEADBAND
428
+ 428 5 HEADBAND_2
429
+ 429 22 HEAR
430
+ 430 2 HEAR+(1h)RUMBLE
431
+ 431 3 HEAR+NOISE
432
+ 432 14 HEAR/LISTEN
433
+ 433 18 HEARING
434
+ 434 4 HEAVY
435
+ 435 10 HELP
436
+ 436 21 HERE
437
+ 437 9 HIGH
438
+ 438 2 HISTORY
439
+ 439 7 HIT
440
+ 440 67 HOME
441
+ 441 5 HOME+WORK
442
+ 442 22 HOMEWORK
443
+ 443 6 HONEST
444
+ 444 8 HOPE
445
+ 445 19 HORSE
446
+ 446 7 HOSPITAL
447
+ 447 13 HOT
448
+ 448 16 HOUSE
449
+ 449 38 HOW
450
+ 450 9 HOW-MANY
451
+ 451 4 HUNGRY/WISH
452
+ 452 2 HURRICANE
453
+ 453 3 HURRY
454
+ 454 7 HUSBAND
455
+ 455 4 ICE-CREAM
456
+ 456 4 ICL:F/9"flip
457
+ 457 3 ICL:S"swinging
458
+ 458 2 ICL:X"spray
459
+ 459 11 IDEA
460
+ 460 15 IDENTIFY
461
+ 461 40 IF
462
+ 462 3 ILLEGAL
463
+ 463 2 IMAGINE
464
+ 464 6 IMPORTANT
465
+ 465 5 IMPOSSIBLE
466
+ 466 70 IN
467
+ 467 20 INCLUDE/INVOLVE
468
+ 468 2 INCREASE
469
+ 469 28 INFORM
470
+ 470 2 INFREQUENTLY
471
+ 471 9 INHALE
472
+ 472 2 INNOCENT
473
+ 473 4 INSTITUTE
474
+ 474 2 INSURANCE/INFECTION
475
+ 475 6 INTERCOURSE+fs-UP
476
+ 476 2 INTERNATIONAL
477
+ 477 15 INTERNET
478
+ 478 23 INTERPRET
479
+ 479 4 INTERRUPT
480
+ 480 3 INTERVIEW
481
+ 481 7 INTRODUCE
482
+ 482 10 INVITE/HIRE
483
+ 483 1313 IX
484
+ 484 812 IX-1p
485
+ 485 15 IX-1p-pl
486
+ 486 6 IX-1p-pl-2
487
+ 487 12 IX-honorific
488
+ 488 68 IX-pl
489
+ 489 31 IX-pl-2
490
+ 490 22 JOIN
491
+ 491 2 KEEP
492
+ 492 12 KEY
493
+ 493 2 KICK-OFF
494
+ 494 23 KID
495
+ 495 24 KIND
496
+ 496 7 KING
497
+ 497 39 KISS-FIST
498
+ 498 42 KNOW
499
+ 499 6 KNOW+(1h)NEG
500
+ 500 3 KNOW-NOTHING
501
+ 501 7 KNOW-THAT
502
+ 502 3 KOWTOW
503
+ 503 8 LANGUAGE
504
+ 504 7 LAPTOP
505
+ 505 11 LAST
506
+ 506 3 LAST-WEEK
507
+ 507 3 LAST-YEAR
508
+ 508 7 LATE
509
+ 509 6 LATER_2
510
+ 510 6 LAUGH
511
+ 511 3 LAUGH-HARD
512
+ 512 17 LEARN
513
+ 513 2 LEARN+AGENT
514
+ 514 6 LEAVE
515
+ 515 3 LEAVE-THERE
516
+ 516 6 LECTURE
517
+ 517 2 LEND
518
+ 518 3 LESS
519
+ 519 16 LETTER/MAIL
520
+ 520 11 LIBRARY
521
+ 521 2 LICENSE
522
+ 522 9 LIE
523
+ 523 7 LIE-DOWN
524
+ 524 3 LIGHT
525
+ 525 2 LIGHT-FLASH
526
+ 526 63 LIKE
527
+ 527 3 LINGUISTICS
528
+ 528 4 LIST
529
+ 529 8 LIST-ORDER
530
+ 530 4 LIST-ORDER-MIDDLE
531
+ 531 3 LITTLE-BIT
532
+ 532 23 LIVE
533
+ 533 3 LONG
534
+ 534 17 LONG-AGO
535
+ 535 7 LONG-LIST
536
+ 536 112 LOOK
537
+ 537 3 LOOK-BY-EYE
538
+ 538 2 LOSE
539
+ 539 9 LOUD
540
+ 540 9 LOUSY
541
+ 541 27 LOVE
542
+ 542 2 MAD
543
+ 543 5 MAGAZINE
544
+ 544 11 MAJOR
545
+ 545 69 MAKE
546
+ 546 2 MAKE+REALLY
547
+ 547 23 MAN
548
+ 548 17 MANAGE/CONTROL
549
+ 549 34 MANY
550
+ 550 4 MARRY
551
+ 551 10 MATCH
552
+ 552 29 MAYBE
553
+ 553 55 MEAN
554
+ 554 21 MEET
555
+ 555 3 MEETING
556
+ 556 2 MEMORIZE
557
+ 557 4 MESSED-UP
558
+ 558 6 MILK
559
+ 559 3 MIND
560
+ 560 2 MINOR
561
+ 561 4 MINUTE
562
+ 562 5 MISMATCH
563
+ 563 10 MISS/ASSUME
564
+ 564 2 MISUNDERSTAND
565
+ 565 3 MONDAY
566
+ 566 12 MONEY
567
+ 567 2 MONEY_2
568
+ 568 4 MONSTER
569
+ 569 22 MORE
570
+ 570 3 MORMON
571
+ 571 4 MORNING
572
+ 572 2 MOST
573
+ 573 105 MOTHER
574
+ 574 28 MOTHER+FATHER
575
+ 575 2 MOUSE/FICTION
576
+ 576 19 MOVE
577
+ 577 1 MOVE-AROUND
578
+ 578 12 MOVE-AWAY
579
+ 579 46 MOVIE
580
+ 580 3 MOVING-ON-TO-NEXT-TOPIC
581
+ 581 11 MOW
582
+ 582 5 MUSIC
583
+ 583 112 MUST
584
+ 584 2 MUTE
585
+ 585 2 Mary
586
+ 586 3 NAB
587
+ 587 25 NAME
588
+ 588 4 NEAR
589
+ 589 3 NEAT
590
+ 590 2 NEGATIVE
591
+ 591 2 NEGLECT
592
+ 592 15 NEVER
593
+ 593 18 NEW
594
+ 594 4 NEXT
595
+ 595 10 NEXT-TO
596
+ 596 2 NEXT-TOPIC
597
+ 597 8 NEXT-WEEK
598
+ 598 39 NICE/CLEAN
599
+ 599 24 NIGHT
600
+ 600 2 NINE
601
+ 601 2 NINETEEN
602
+ 602 44 NOISE
603
+ 603 3 NOISE+RUMBLE
604
+ 604 4 NONE
605
+ 605 83 NONE/NOTHING
606
+ 606 3 NONE/NOTHING+ONE
607
+ 607 283 NOT
608
+ 608 2 NOT+MUST
609
+ 609 6 NOT-MIND
610
+ 610 32 NOT-YET
611
+ 611 3 NOTHING
612
+ 612 2 NOTHING-AT-ALL
613
+ 613 5 NOTICE
614
+ 614 89 NOW
615
+ 615 2 NOW-WEEK
616
+ 616 2 NURSE
617
+ 617 4 NUT
618
+ 618 12 OF-COURSE
619
+ 619 21 OFTEN
620
+ 620 38 OH-I-SEE
621
+ 621 9 OLD
622
+ 622 24 ON
623
+ 623 2 ON-TIME
624
+ 624 24 ONE
625
+ 625 2 ONE+HUNDRED
626
+ 626 2 ONE-DOLLAR
627
+ 627 3 ONE-MONTH
628
+ 628 2 ONE-WEEK
629
+ 629 11 ONLY
630
+ 630 2 OPPORTUNITY
631
+ 631 2 OPPOSITE
632
+ 632 3 ORANGE
633
+ 633 7 ORDER
634
+ 634 20 OTHER
635
+ 635 7 OUT
636
+ 636 3 OVER/AFTER
637
+ 637 2 OVERSLEEP
638
+ 638 2 PAINT
639
+ 639 6 PAPER
640
+ 640 4 PAPER-CHECK/CARD
641
+ 641 55 PARTY
642
+ 642 6 PASS
643
+ 643 7 PASS-OUT
644
+ 644 12 PAST
645
+ 645 26 PAST+NIGHT
646
+ 646 17 PAY
647
+ 647 12 PAY/BUY
648
+ 648 16 PAY/SPEND
649
+ 649 1 PCL:crvd-5"people
650
+ 650 62 PEOPLE
651
+ 651 5 PERMIT
652
+ 652 28 PERSON
653
+ 653 3 PET-PEEVE
654
+ 654 2 PHONE
655
+ 655 7 PICK/CHOOSE
656
+ 656 12 PIE
657
+ 657 27 PINEAPPLE
658
+ 658 8 PING-PONG/TENNIS
659
+ 659 3 PISS-OFF
660
+ 660 20 PIZZA_4
661
+ 661 23 PLACE
662
+ 662 11 PLAID
663
+ 663 12 PLAN
664
+ 664 53 PLAY
665
+ 665 5 PLUS
666
+ 666 3 POINT
667
+ 667 5 POLITE
668
+ 668 3 POLITICS
669
+ 669 3 POP-UP
670
+ 670 120 POSS
671
+ 671 177 POSS-1p
672
+ 672 4 POSS-1p-pl
673
+ 673 8 POSS-pl
674
+ 674 2 POUND/WEIGH
675
+ 675 2 POWER
676
+ 676 2 PRACTICE
677
+ 677 14 PRESIDENT
678
+ 678 3 PRINCIPAL
679
+ 679 3 PROBLEM
680
+ 680 35 PROCEED
681
+ 681 3 PROFESSION
682
+ 682 8 PROGRESS
683
+ 683 5 PUT
684
+ 684 7 PUT-AWAY
685
+ 685 88 QMwg
686
+ 686 4 QUALITY
687
+ 687 2 QUESTION
688
+ 688 13 QUIET
689
+ 689 19 QUOTE/TOPIC
690
+ 690 65 RAIN
691
+ 691 9 RANGE
692
+ 692 54 READ
693
+ 693 10 READY
694
+ 694 68 REALLY
695
+ 695 26 REALLY+WORK
696
+ 696 6 REALLY+WORK/BUSINESS
697
+ 697 14 REASON
698
+ 698 34 RECENT-PAST
699
+ 699 2 RECORD
700
+ 700 4 RED
701
+ 701 25 REFUSE
702
+ 702 3 REGULAR
703
+ 703 4 RELATIONSHIP-OFF
704
+ 704 5 REMEMBER
705
+ 705 6 REMIND
706
+ 706 3 REMIND-2p
707
+ 707 2 REPLACE
708
+ 708 3 REPUBLICAN
709
+ 709 6 REQUIRE
710
+ 710 6 RESEARCH
711
+ 711 3 RESTAURANT
712
+ 712 2 RESUME
713
+ 713 4 ROAD
714
+ 714 3 ROLE
715
+ 715 6 ROOMMATE
716
+ 716 7 ROWING
717
+ 717 3 RULE
718
+ 718 7 RUN
719
+ 719 2 RUN-MACHINE
720
+ 720 4 S"cheering"
721
+ 721 5 SAFE
722
+ 722 45 SAME
723
+ 723 3 SAME-LIST
724
+ 724 3 SAME-OLD
725
+ 725 11 SAY
726
+ 726 53 SCHOOL
727
+ 727 2 SCL:1"John
728
+ 728 3 SCL:1"mentor
729
+ 729 2 SCL:1"person
730
+ 730 9 SCL:3"person
731
+ 731 2 SCL:5-C"big
732
+ 732 3 SCL:U-L"person
733
+ 733 2 SCL:V"person
734
+ 734 4 SCL:V/2"look
735
+ 735 3 SCL:V/2"person
736
+ 736 3 SCL:V/2"two
737
+ 737 11 SCL:bent-U-L"person
738
+ 738 9 SCL:crvd-5"group
739
+ 739 2 SCL:crvd-V"cat
740
+ 740 2 SCL:crvd-V"horse
741
+ 741 15 SCREAM
742
+ 742 2 SCREEN/FILTER
743
+ 743 4 SEARCH-FOR
744
+ 744 2 SECOND
745
+ 745 12 SECOND-IN-LIST
746
+ 746 29 SEE
747
+ 747 9 SEEM
748
+ 748 9 SELF
749
+ 749 7 SELF-1p
750
+ 750 25 SELL
751
+ 751 3 SEMESTER
752
+ 752 15 SEND
753
+ 753 6 SEND/MAIL
754
+ 754 3 SERIOUS
755
+ 755 6 SERVICE
756
+ 756 22 SET-UP
757
+ 757 6 SEVEN
758
+ 758 2 SHEET-CAKE
759
+ 759 3 SHH
760
+ 760 9 SHOE
761
+ 761 5 SHORT-HEIGHT
762
+ 762 10 SHORTS
763
+ 763 17 SHOULD
764
+ 764 25 SHOW
765
+ 765 2 SHUCKS
766
+ 766 27 SICK
767
+ 767 4 SICK-OF
768
+ 768 3 SICK/SILLY
769
+ 769 10 SIGN
770
+ 770 31 SISTER
771
+ 771 9 SIT
772
+ 772 7 SITUATION
773
+ 773 6 SIX
774
+ 774 7 SKI
775
+ 775 4 SKILL
776
+ 776 4 SLEEP
777
+ 777 6 SLOW
778
+ 778 15 SMALL
779
+ 779 2 SMALL-AMOUNT
780
+ 780 35 SNOW
781
+ 781 3 SO-SO
782
+ 782 10 SOCIAL/INTERACT
783
+ 783 15 SOME
784
+ 784 22 SOMETHING/ONE
785
+ 785 5 SOMETIMES
786
+ 786 4 SON
787
+ 787 14 SOON
788
+ 788 13 SORRY
789
+ 789 2 SOUP
790
+ 790 3 SOUTH
791
+ 791 3 SPAGHETTI
792
+ 792 7 SPECIAL/EXCEPT
793
+ 793 10 SPEECH/ORAL
794
+ 794 3 SPEND
795
+ 795 3 SPREAD
796
+ 796 2 SPRING
797
+ 797 3 STAND
798
+ 798 2 STAR
799
+ 799 3 STAR+BOMB
800
+ 800 37 START
801
+ 801 28 STAY
802
+ 802 2 STEAL
803
+ 803 15 STILL
804
+ 804 11 STOP
805
+ 805 16 STORY
806
+ 806 14 STRANGE
807
+ 807 3 STRIPES
808
+ 808 2 STRONG
809
+ 809 4 STRUGGLE
810
+ 810 2 STUCK
811
+ 811 22 STUDENT
812
+ 812 25 STUDY
813
+ 813 4 SUCCEED
814
+ 814 4 SUGGEST
815
+ 815 3 SUMMER
816
+ 816 2 SUMMON
817
+ 817 2 SUN
818
+ 818 6 SUNBATHE
819
+ 819 7 SUNSHINE
820
+ 820 4 SUPERIOR
821
+ 821 2 SUPPORT
822
+ 822 8 SURF-INTERNET
823
+ 823 26 SURPRISE
824
+ 824 2 SUSPECT
825
+ 825 5 SWEETHEART
826
+ 826 14 SWIM
827
+ 827 4 SYSTEM
828
+ 828 5 TABLE
829
+ 829 11 TAKE
830
+ 830 4 TAKE-BREAK
831
+ 831 3 TAKE-NOTES
832
+ 832 33 TAKE-UP
833
+ 833 49 TALK
834
+ 834 2 TALK-OVER
835
+ 835 4 TAN
836
+ 836 3 TAPE-RECORDING
837
+ 837 5 TASTE
838
+ 838 3 TEACH
839
+ 839 46 TEACH+AGENT
840
+ 840 2 TEAM
841
+ 841 26 TELL
842
+ 842 2 TEN
843
+ 843 51 TEND
844
+ 844 18 TEST
845
+ 845 5 TEXTING
846
+ 846 5 THAN
847
+ 847 101 THAT
848
+ 848 5 THEN
849
+ 849 4 THESE
850
+ 850 8 THING
851
+ 851 27 THINK
852
+ 852 1 THINK+HOPE
853
+ 853 2 THINK-OVER
854
+ 854 6 THIRD-IN-LIST
855
+ 855 2 THIRD-PLACE
856
+ 856 4 THIRTY
857
+ 857 9 THREE
858
+ 858 8 THRILL/WHAT'S-UP
859
+ 859 12 THROUGH
860
+ 860 43 TIME
861
+ 861 3 TIME-PERIOD
862
+ 862 7 TIRED
863
+ 863 52 TO/UNTIL
864
+ 864 17 TOGETHER/GO-STEADY
865
+ 865 16 TOILET
866
+ 866 38 TOMORROW
867
+ 867 7 TOOTH/GLASS
868
+ 868 4 TOTAL
869
+ 869 2 TRACK-IX
870
+ 870 2 TRACK-MIDDLE
871
+ 871 2 TRACK-RING
872
+ 872 2 TRACK-THUMB
873
+ 873 2 TRADITION
874
+ 874 3 TRAIN
875
+ 875 7 TRANSFER
876
+ 876 4 TREE
877
+ 877 19 TRIP
878
+ 878 3 TROPHY
879
+ 879 2 TROUBLE
880
+ 880 3 TRY
881
+ 881 2 TUBING
882
+ 882 5 TURKEY
883
+ 883 8 TWENTY
884
+ 884 2 TWENTY-EIGHT
885
+ 885 2 TWICE
886
+ 886 16 TWO
887
+ 887 7 TYPE
888
+ 888 6 UMBRELLA
889
+ 889 7 UNCLE
890
+ 890 12 UNDERSTAND
891
+ 891 2 UP
892
+ 892 10 UP-TO-NOW
893
+ 893 7 UPSET
894
+ 894 28 USE
895
+ 895 7 USE-SIGN-LANGUAGE
896
+ 896 15 VACATION
897
+ 897 9 VARY
898
+ 898 8 VEGETABLE
899
+ 899 3 VEGETABLE+AGENT
900
+ 900 4 VERY
901
+ 901 3 VERY-FAST
902
+ 902 15 VIDEO-GAME
903
+ 903 3 VIDEOTAPE
904
+ 904 3 VISIT
905
+ 905 26 VISUALIZE
906
+ 906 59 VOICE
907
+ 907 15 VOICE-RANGE
908
+ 908 4 VOICE_2
909
+ 909 16 VOLUNTEER/SHIRT
910
+ 910 23 VOMIT/HATE
911
+ 911 8 VOTE
912
+ 912 5 WAIT
913
+ 913 4 WALK
914
+ 914 62 WANT
915
+ 915 24 WANT+NEG
916
+ 916 12 WARM
917
+ 917 7 WARN
918
+ 918 4 WATCH-TV
919
+ 919 8 WATER
920
+ 920 2 WATER+DCL:crvd-L"water
921
+ 921 2 WATER+TUBING
922
+ 922 12 WEATHER
923
+ 923 11 WEDDING
924
+ 924 2 WEEKEND
925
+ 925 43 WHEN
926
+ 926 69 WHERE
927
+ 927 15 WHICH
928
+ 928 140 WHO
929
+ 929 191 WHY
930
+ 930 6 WIDE
931
+ 931 15 WIFE
932
+ 932 16 WIN
933
+ 933 8 WINDOW
934
+ 934 72 WITH
935
+ 935 6 WOMAN
936
+ 936 2 WONDER
937
+ 937 11 WORD
938
+ 938 88 WORK
939
+ 939 3 WORK-OUT
940
+ 940 8 WORLD
941
+ 941 41 WOW
942
+ 942 25 WOW/AWFUL
943
+ 943 2 WRESTLE
944
+ 944 12 WRITE
945
+ 945 14 WRONG
946
+ 946 13 YEAR
947
+ 947 11 YES
948
+ 948 20 YESTERDAY
949
+ 949 2 YOUNG
950
+ 950 2 ZOOM
951
+ 951 11 a
952
+ 952 9 all
953
+ 953 12 and
954
+ 954 3 approaching
955
+ 955 1 around
956
+ 956 3 around"
957
+ 957 2 at
958
+ 958 2 back
959
+ 959 1 below
960
+ 960 4 big
961
+ 961 2 book"
962
+ 962 4 bowl"
963
+ 963 1 bucket
964
+ 964 7 bucket"
965
+ 965 4 car"
966
+ 966 3 chalkboard"
967
+ 967 2 circle"
968
+ 968 2 coming
969
+ 969 3 coming"
970
+ 970 7 control"
971
+ 971 2 course"
972
+ 972 3 cup"
973
+ 973 4 deal"
974
+ 974 2 do"
975
+ 975 2 doll"
976
+ 976 13 don't
977
+ 977 1 down
978
+ 978 4 down"
979
+ 979 6 dyed
980
+ 980 2 each
981
+ 981 7 ear"
982
+ 982 3 embarrassed"
983
+ 983 5 envelope"
984
+ 984 3 fast"
985
+ 985 2 fish
986
+ 986 2 fish"
987
+ 987 6 flat"
988
+ 988 2 flip
989
+ 989 2 football"
990
+ 990 2 for
991
+ 991 2 forth"
992
+ 992 7 from
993
+ 993 5 fs-AT
994
+ 994 3 fs-AUDIO
995
+ 995 3 fs-AUDIOVOCAL
996
+ 996 4 fs-BAGEL
997
+ 997 2 fs-BE
998
+ 998 28 fs-BEACH
999
+ 999 4 fs-BONFIRE
1000
+ 1000 6 fs-BRIDE
1001
+ 1001 3 fs-BUBBLES
1002
+ 1002 2 fs-BURP
1003
+ 1003 4 fs-BY
1004
+ 1004 2 fs-CANCER
1005
+ 1005 3 fs-CARE
1006
+ 1006 3 fs-CAREER
1007
+ 1007 2 fs-CAUCUS
1008
+ 1008 5 fs-CEREAL
1009
+ 1009 2 fs-CLASS
1010
+ 1010 3 fs-CLUB
1011
+ 1011 4 fs-CODA
1012
+ 1012 3 fs-COLLIDED
1013
+ 1013 19 fs-CREAM
1014
+ 1014 7 fs-CREAM+CHEESE
1015
+ 1015 2 fs-CREAM+fs-CHEESE
1016
+ 1016 4 fs-DAY
1017
+ 1017 7 fs-DISORDER
1018
+ 1018 2 fs-DVD
1019
+ 1019 2 fs-ELECTION
1020
+ 1020 10 fs-EVEN
1021
+ 1021 2 fs-EXAM
1022
+ 1022 2 fs-EXCESS
1023
+ 1023 3 fs-FIELD
1024
+ 1024 3 fs-FUN
1025
+ 1025 2 fs-GLOBE
1026
+ 1026 5 fs-GROOM
1027
+ 1027 12 fs-GRUNT
1028
+ 1028 3 fs-HEGEMONY
1029
+ 1029 3 fs-HIT
1030
+ 1030 2 fs-HORROR
1031
+ 1031 2 fs-HUMAN
1032
+ 1032 2 fs-IPOD
1033
+ 1033 2 fs-KEY
1034
+ 1034 4 fs-LAKE
1035
+ 1035 36 fs-LATE
1036
+ 1036 2 fs-LINK
1037
+ 1037 3 fs-LIT
1038
+ 1038 1 fs-LOCKER+BOX/ROOM
1039
+ 1039 6 fs-LOTTERY
1040
+ 1040 8 fs-LOX
1041
+ 1041 3 fs-MALL
1042
+ 1042 2 fs-MARATHON
1043
+ 1043 4 fs-MENTOR
1044
+ 1044 2 fs-MODEL
1045
+ 1045 3 fs-MPH
1046
+ 1046 2 fs-MUTE
1047
+ 1047 4 fs-NASAL
1048
+ 1048 3 fs-NOISE
1049
+ 1049 2 fs-NOISY
1050
+ 1050 2 fs-OCEAN
1051
+ 1051 39 fs-OF
1052
+ 1052 2 fs-OH
1053
+ 1053 2 fs-OR
1054
+ 1054 14 fs-ORIENTATION
1055
+ 1055 2 fs-ORIENTED
1056
+ 1056 2 fs-OWN
1057
+ 1057 15 fs-PARK
1058
+ 1058 2 fs-PC
1059
+ 1059 2 fs-PISS
1060
+ 1060 3 fs-PITCHED
1061
+ 1061 3 fs-POP
1062
+ 1062 2 fs-PROCTOR
1063
+ 1063 2 fs-PUBLIC
1064
+ 1064 6 fs-PUPPIES
1065
+ 1065 2 fs-QUALITY
1066
+ 1066 9 fs-RANGE
1067
+ 1067 12 fs-RENT
1068
+ 1068 2 fs-SANCTUARY
1069
+ 1069 5 fs-SENSORY
1070
+ 1070 2 fs-SMASHED
1071
+ 1071 5 fs-SO
1072
+ 1072 3 fs-SOUND
1073
+ 1073 14 fs-SPEECH
1074
+ 1074 3 fs-SPINACH
1075
+ 1075 2 fs-STATUS
1076
+ 1076 2 fs-STORM
1077
+ 1077 3 fs-SUN
1078
+ 1078 2 fs-SUSHI
1079
+ 1079 2 fs-SYNTAX
1080
+ 1080 3 fs-TACTILE
1081
+ 1081 2 fs-TAX
1082
+ 1082 2 fs-TEMPO
1083
+ 1083 4 fs-TESTIMONIES
1084
+ 1084 3 fs-TOO
1085
+ 1085 2 fs-TRIP
1086
+ 1086 6 fs-TRUCK
1087
+ 1087 2 fs-TRY
1088
+ 1088 3 fs-VOCAL
1089
+ 1089 2 fs-WAS
1090
+ 1090 2 fs-WASTE
1091
+ 1091 5 galloping"
1092
+ 1092 4 gesture"
1093
+ 1093 24 go
1094
+ 1094 2 growing-up"
1095
+ 1095 1 grumble"
1096
+ 1096 2 hair"
1097
+ 1097 2 hey"
1098
+ 1098 2 house"
1099
+ 1099 5 in
1100
+ 1100 7 in"
1101
+ 1101 1 increasing
1102
+ 1102 1 into
1103
+ 1103 3 is"
1104
+ 1104 5 it
1105
+ 1105 4 it"
1106
+ 1106 2 know
1107
+ 1107 21 know"
1108
+ 1108 3 land"
1109
+ 1109 2 language"
1110
+ 1110 2 levitating"
1111
+ 1111 3 lifting"
1112
+ 1112 2 like"
1113
+ 1113 2 look
1114
+ 1114 7 loop
1115
+ 1115 4 lotion"
1116
+ 1116 2 loudly"
1117
+ 1117 2 man"
1118
+ 1118 2 me"
1119
+ 1119 4 more"
1120
+ 1120 7 mouth
1121
+ 1121 4 moving
1122
+ 1122 6 my
1123
+ 1123 2 natural
1124
+ 1124 2 next
1125
+ 1125 89 no"
1126
+ 1126 1 noises"
1127
+ 1127 2 ns-#AA
1128
+ 1128 4 ns-#FLA
1129
+ 1129 2 ns-#IBM
1130
+ 1130 3 ns-#RIT
1131
+ 1131 12 ns-#SD
1132
+ 1132 2 ns-#TENN
1133
+ 1133 30 ns-BOSTON
1134
+ 1134 2 ns-BROOKLINE
1135
+ 1135 5 ns-CHICAGO
1136
+ 1136 7 ns-FRANCE
1137
+ 1137 3 ns-GALLAUDET
1138
+ 1138 2 ns-INDIA
1139
+ 1139 2 ns-IRELAND
1140
+ 1140 3 ns-MCDONALD
1141
+ 1141 8 ns-NEW-YORK
1142
+ 1142 3 ns-PHILADELPHIA
1143
+ 1143 2 ns-PUERTO-RICO
1144
+ 1144 5 ns-ROCHESTER
1145
+ 1145 2 ns-fs-ALEC-BALDWIN
1146
+ 1146 2 ns-fs-AMY
1147
+ 1147 3 ns-fs-ASIA
1148
+ 1148 2 ns-fs-BARACK-OBAMA
1149
+ 1149 3 ns-fs-BAUMAN
1150
+ 1150 2 ns-fs-BETH
1151
+ 1151 7 ns-fs-BILL
1152
+ 1152 72 ns-fs-BOB
1153
+ 1153 2 ns-fs-BOSTON
1154
+ 1154 3 ns-fs-BU
1155
+ 1155 2 ns-fs-DEAN
1156
+ 1156 8 ns-fs-DONNA
1157
+ 1157 3 ns-fs-FEDEX
1158
+ 1158 2 ns-fs-HARRY
1159
+ 1159 13 ns-fs-JACK
1160
+ 1160 3 ns-fs-JAMES
1161
+ 1161 2 ns-fs-JEN
1162
+ 1162 5 ns-fs-JESSICA
1163
+ 1163 3 ns-fs-JIM
1164
+ 1164 3 ns-fs-JOE
1165
+ 1165 90 ns-fs-JOHN
1166
+ 1166 2 ns-fs-LIZ
1167
+ 1167 4 ns-fs-MAC
1168
+ 1168 80 ns-fs-MARY
1169
+ 1169 2 ns-fs-MITT-ROMNEY
1170
+ 1170 2 ns-fs-NETFLIX
1171
+ 1171 2 ns-fs-NEWTON
1172
+ 1172 11 ns-fs-OBAMA
1173
+ 1173 2 ns-fs-PAUL
1174
+ 1174 3 ns-fs-PETE
1175
+ 1175 3 ns-fs-PHIL
1176
+ 1176 3 ns-fs-RICK-PERRY
1177
+ 1177 2 ns-fs-SALLY
1178
+ 1178 2 ns-fs-SARAH-JESSICA-PARKER
1179
+ 1179 2 ns-fs-SARAH-PALIN
1180
+ 1180 2 ns-fs-SAT
1181
+ 1181 5 ns-fs-SUE
1182
+ 1182 2 ns-nat-AUSTRALIA
1183
+ 1183 26 of
1184
+ 1184 3 off
1185
+ 1185 4 off"
1186
+ 1186 15 on
1187
+ 1187 3 open
1188
+ 1188 2 other"
1189
+ 1189 2 out
1190
+ 1190 9 out"
1191
+ 1191 4 over
1192
+ 1192 7 over"
1193
+ 1193 7 pages"
1194
+ 1194 5 paper
1195
+ 1195 4 paper"
1196
+ 1196 451 part:indef
1197
+ 1197 7 passed
1198
+ 1198 2 pathway"
1199
+ 1199 7 people
1200
+ 1200 2 popcorn
1201
+ 1201 1 popcorn"
1202
+ 1202 7 puppies"
1203
+ 1203 8 range"
1204
+ 1204 3 raquet"
1205
+ 1205 2 rolling
1206
+ 1206 5 running"
1207
+ 1207 12 see"
1208
+ 1208 2 set"
1209
+ 1209 1 shoot"
1210
+ 1210 2 sign
1211
+ 1211 2 silly
1212
+ 1212 2 sit
1213
+ 1213 2 skipping
1214
+ 1214 3 slowly"
1215
+ 1215 2 staring"
1216
+ 1216 2 string"
1217
+ 1217 3 styrofoam
1218
+ 1218 2 table"
1219
+ 1219 2 tanning"
1220
+ 1220 3 tennis
1221
+ 1221 7 test"
1222
+ 1222 21 the
1223
+ 1223 2 them
1224
+ 1224 2 there"
1225
+ 1225 3 things"
1226
+ 1226 17 to
1227
+ 1227 25 together"
1228
+ 1228 3 toilet
1229
+ 1229 2 tract
1230
+ 1230 2 tubing"
1231
+ 1231 5 up
1232
+ 1232 6 voice
1233
+ 1233 1 voice"
1234
+ 1234 2 walk
1235
+ 1235 20 walking"
1236
+ 1236 3 waste
1237
+ 1237 5 waste"
1238
+ 1238 2 water"
1239
+ 1239 3 way
1240
+ 1240 2 well"
1241
+ 1241 2 what
1242
+ 1242 2 with
1243
+ 1243 2 words"
plugins/SignX/smkd/configs/youtubeasl.yaml ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ dataset_root: youtubeasl
2
+ dict_path: youtubeasl/gloss_dict.npy
3
+ evaluation_dir: ./evaluation/slr_eval
4
+ evaluation_prefix: youtubeasl-groundtruth
plugins/SignX/smkd/dataset/dataloader_video.py CHANGED
@@ -27,7 +27,9 @@ sys.path.append("..")
27
 
28
  class BaseFeeder(data.Dataset):
29
  def __init__(self, prefix, gloss_dict, drop_ratio=1, num_gloss=-1, mode="train", transform_mode=True,
30
- datatype="lmdb", skip_ids=None, c2d_type=None):
 
 
31
  self.mode = mode
32
  self.ng = num_gloss
33
  self.prefix = prefix
@@ -35,6 +37,13 @@ class BaseFeeder(data.Dataset):
35
  self.data_type = datatype
36
  self.transform_mode = "train" if transform_mode else "test"
37
  self.c2d_type = c2d_type # 用于确定归一化方式
 
 
 
 
 
 
 
38
  self.inputs_list = np.load(f"{prefix}/{mode}_info.npy", allow_pickle=True).item()
39
  self.skip_ids = skip_ids if isinstance(skip_ids, set) else set(skip_ids or [])
40
  self.sample_indices = self._build_sample_indices()
@@ -48,14 +57,37 @@ class BaseFeeder(data.Dataset):
48
  print("")
49
 
50
  def __getitem__(self, idx):
51
- real_idx = self.sample_indices[idx]
52
- if self.data_type == "video":
53
- input_data, label, fi = self.read_video(real_idx)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54
  input_data, label = self.normalize(input_data, label)
55
- # input_data, label = self.normalize(input_data, label, fi['fileid'])
 
56
  return input_data, torch.LongTensor(label), self.inputs_list[real_idx]['original_info']
57
- else:
58
- raise ValueError(f"Not supported: {self.data_type}")
 
 
 
 
59
 
60
  def read_video(self, index, num_glosses=-1):
61
  # load file info
@@ -86,6 +118,12 @@ class BaseFeeder(data.Dataset):
86
 
87
  # load next frame
88
  success, img = vidcap.read()
 
 
 
 
 
 
89
 
90
  for phase in fi['label'].split(" "):
91
  if phase == '':
@@ -94,14 +132,36 @@ class BaseFeeder(data.Dataset):
94
  label_list.append(self.dict[phase][0])
95
 
96
  if len(img_features) == 0:
 
 
97
  img_features = [np.zeros((256, 256, 3))]
98
  if len(label_list) == 0:
99
  label_list = [0]
100
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
101
  return img_features, label_list, fi
102
 
103
  def normalize(self, video, label, file_id=None):
104
  video, label = self.data_aug(video, label, file_id)
 
 
105
 
106
  # 根据backbone类型选择归一化方式
107
  if self.c2d_type and ('dino' in self.c2d_type.lower() or 'vit' in self.c2d_type.lower()):
@@ -116,15 +176,38 @@ class BaseFeeder(data.Dataset):
116
 
117
  return video, label
118
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
119
  def transform(self):
120
  if self.transform_mode == "train":
121
  print("Apply training transform.")
122
- return video_augmentation.Compose([
123
  video_augmentation.RandomCrop(224),
124
  video_augmentation.RandomHorizontalFlip(0.5),
125
  video_augmentation.ToTensor(),
126
- video_augmentation.TemporalRescale(0.2),
127
- ])
 
 
128
  else:
129
  print("Apply testing transform.")
130
  return video_augmentation.Compose([
 
27
 
28
  class BaseFeeder(data.Dataset):
29
  def __init__(self, prefix, gloss_dict, drop_ratio=1, num_gloss=-1, mode="train", transform_mode=True,
30
+ datatype="lmdb", skip_ids=None, c2d_type=None, max_frames=-1, ctc_length_guard=True,
31
+ skip_on_decode_failure=False, min_frames=0, skip_on_black_frames=False, temporal_rescale=0.2,
32
+ ctc_length_guard_post=True):
33
  self.mode = mode
34
  self.ng = num_gloss
35
  self.prefix = prefix
 
37
  self.data_type = datatype
38
  self.transform_mode = "train" if transform_mode else "test"
39
  self.c2d_type = c2d_type # 用于确定归一化方式
40
+ self.max_frames = max_frames
41
+ self.ctc_length_guard = ctc_length_guard
42
+ self.skip_on_decode_failure = skip_on_decode_failure
43
+ self.min_frames = min_frames
44
+ self.skip_on_black_frames = skip_on_black_frames
45
+ self.temporal_rescale = temporal_rescale
46
+ self.ctc_length_guard_post = ctc_length_guard_post
47
  self.inputs_list = np.load(f"{prefix}/{mode}_info.npy", allow_pickle=True).item()
48
  self.skip_ids = skip_ids if isinstance(skip_ids, set) else set(skip_ids or [])
49
  self.sample_indices = self._build_sample_indices()
 
57
  print("")
58
 
59
  def __getitem__(self, idx):
60
+ if self.data_type != "video":
61
+ raise ValueError(f"Not supported: {self.data_type}")
62
+
63
+ retries = 20 if self.skip_on_decode_failure else 1
64
+ last_err = None
65
+ skipped_ids = []
66
+ for attempt in range(retries):
67
+ if attempt == 0:
68
+ real_idx = self.sample_indices[idx % len(self.sample_indices)]
69
+ else:
70
+ real_idx = self.sample_indices[np.random.randint(0, len(self.sample_indices))]
71
+ try:
72
+ input_data, label, fi = self.read_video(real_idx)
73
+ except ValueError as e:
74
+ if self.skip_on_decode_failure and str(e).startswith("decode_failed:"):
75
+ last_err = e
76
+ sid = str(e).split("decode_failed:", 1)[-1]
77
+ if sid:
78
+ skipped_ids.append(sid)
79
+ continue
80
+ raise
81
  input_data, label = self.normalize(input_data, label)
82
+ if skipped_ids:
83
+ print(f"[BaseFeeder] Skipped {len(skipped_ids)} decode-failed sample(s): {', '.join(skipped_ids)}")
84
  return input_data, torch.LongTensor(label), self.inputs_list[real_idx]['original_info']
85
+
86
+ if last_err is not None:
87
+ if skipped_ids:
88
+ print(f"[BaseFeeder] Skipped {len(skipped_ids)} decode-failed sample(s) and exhausted retries: {', '.join(skipped_ids)}")
89
+ raise last_err
90
+ raise ValueError("Failed to load sample")
91
 
92
  def read_video(self, index, num_glosses=-1):
93
  # load file info
 
118
 
119
  # load next frame
120
  success, img = vidcap.read()
121
+ vidcap.release()
122
+
123
+ # Uniformly sample to a fixed maximum length for stability.
124
+ if self.max_frames is not None and self.max_frames > 0 and len(img_features) > self.max_frames:
125
+ idx = np.linspace(0, len(img_features) - 1, self.max_frames, dtype=int)
126
+ img_features = [img_features[i] for i in idx]
127
 
128
  for phase in fi['label'].split(" "):
129
  if phase == '':
 
132
  label_list.append(self.dict[phase][0])
133
 
134
  if len(img_features) == 0:
135
+ if self.skip_on_decode_failure:
136
+ raise ValueError(f"decode_failed:{fi.get('fileid','unknown')}")
137
  img_features = [np.zeros((256, 256, 3))]
138
  if len(label_list) == 0:
139
  label_list = [0]
140
 
141
+ # Skip extremely short clips if requested.
142
+ if self.min_frames and len(img_features) < self.min_frames:
143
+ if self.skip_on_decode_failure:
144
+ raise ValueError(f"decode_failed:{fi.get('fileid','unknown')}")
145
+
146
+ # Skip all-black clips if requested.
147
+ if self.skip_on_black_frames:
148
+ try:
149
+ frame_mean = float(np.mean(img_features[0]))
150
+ except Exception:
151
+ frame_mean = 0.0
152
+ if frame_mean < 1.0: # near-black frame
153
+ if self.skip_on_decode_failure:
154
+ raise ValueError(f"decode_failed:{fi.get('fileid','unknown')}")
155
+
156
+ if self.ctc_length_guard and not self.ctc_length_guard_post:
157
+ img_features = self._ensure_ctc_length(img_features, label_list)
158
+
159
  return img_features, label_list, fi
160
 
161
  def normalize(self, video, label, file_id=None):
162
  video, label = self.data_aug(video, label, file_id)
163
+ if self.ctc_length_guard and self.ctc_length_guard_post:
164
+ video = self._ensure_ctc_length(video, label)
165
 
166
  # 根据backbone类型选择归一化方式
167
  if self.c2d_type and ('dino' in self.c2d_type.lower() or 'vit' in self.c2d_type.lower()):
 
176
 
177
  return video, label
178
 
179
+ def _ensure_ctc_length(self, video, label):
180
+ # Ensure temporal length is long enough for CTC after TemporalConv (conv_type=2).
181
+ # Roughly require input_len >= 4 * label_len + 12.
182
+ if not label:
183
+ return video
184
+ min_required = 4 * len(label) + 12
185
+ if min_required <= 0:
186
+ return video
187
+ if isinstance(video, torch.Tensor):
188
+ t = video.shape[0]
189
+ if t == 0 or t >= min_required:
190
+ return video
191
+ idx = torch.linspace(0, t - 1, steps=min_required, device=video.device)
192
+ idx = torch.clamp(idx.round().long(), 0, t - 1)
193
+ return video.index_select(0, idx)
194
+ t = len(video)
195
+ if t == 0 or t >= min_required:
196
+ return video
197
+ idx = np.linspace(0, t - 1, min_required, dtype=int)
198
+ return [video[i] for i in idx]
199
+
200
  def transform(self):
201
  if self.transform_mode == "train":
202
  print("Apply training transform.")
203
+ transforms = [
204
  video_augmentation.RandomCrop(224),
205
  video_augmentation.RandomHorizontalFlip(0.5),
206
  video_augmentation.ToTensor(),
207
+ ]
208
+ if self.temporal_rescale and self.temporal_rescale > 0:
209
+ transforms.append(video_augmentation.TemporalRescale(self.temporal_rescale))
210
+ return video_augmentation.Compose(transforms)
211
  else:
212
  print("Apply testing transform.")
213
  return video_augmentation.Compose([
plugins/SignX/smkd/evaluation/slr_eval/wer_calculation.py CHANGED
@@ -1,4 +1,5 @@
1
  import os
 
2
  from .python_wer_evaluation import wer_calculation
3
 
4
 
@@ -36,10 +37,12 @@ def evaluate(prefix="./", mode="dev", evaluate_dir=None, dataset_dir=None, evalu
36
  f"{sclite_path} -h {prefix}out.{output_file} ctm"
37
  f" -r {prefix}tmp.stm stm -f 0 -o sgml sum rsum pra"
38
  )
39
- ret = os.popen(
40
  f"{sclite_path} -h {prefix}out.{output_file} ctm "
41
- f"-r {prefix}tmp.stm stm -f 0 -o dtl stdout |grep Error"
42
- ).readlines()[0]
 
 
43
  return float(ret.split("=")[1].split("%")[0])
44
 
45
 
 
1
  import os
2
+ import subprocess
3
  from .python_wer_evaluation import wer_calculation
4
 
5
 
 
37
  f"{sclite_path} -h {prefix}out.{output_file} ctm"
38
  f" -r {prefix}tmp.stm stm -f 0 -o sgml sum rsum pra"
39
  )
40
+ ret = subprocess.check_output(
41
  f"{sclite_path} -h {prefix}out.{output_file} ctm "
42
+ f"-r {prefix}tmp.stm stm -f 0 -o dtl stdout |grep Error",
43
+ shell=True,
44
+ text=True
45
+ ).splitlines()[0]
46
  return float(ret.split("=")[1].split("%")[0])
47
 
48
 
plugins/SignX/smkd/seq_scripts.py CHANGED
@@ -208,11 +208,11 @@ def seq_feature_generation(loader, model, device, mode, recoder, is_master=True)
208
 
209
 
210
  def write2file(path, info, output):
211
- filereader = open(path, "w")
212
- for sample_idx, sample in enumerate(output):
213
- for word_idx, word in enumerate(sample):
214
- filereader.writelines(
215
- "{} 1 {:.2f} {:.2f} {}\n".format(info[sample_idx],
216
- word_idx * 1.0 / 100,
217
- (word_idx + 1) * 1.0 / 100,
218
- word[0]))
 
208
 
209
 
210
  def write2file(path, info, output):
211
+ with open(path, "w") as filereader:
212
+ for sample_idx, sample in enumerate(output):
213
+ for word_idx, word in enumerate(sample):
214
+ filereader.writelines(
215
+ "{} 1 {:.2f} {:.2f} {}\n".format(info[sample_idx],
216
+ word_idx * 1.0 / 100,
217
+ (word_idx + 1) * 1.0 / 100,
218
+ word[0]))
plugins/SignX/training_yutubeasl/ENVIRONMENT_SETUP.md ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # 环境设置说明
2
+
3
+ ## 为什么需要两个环境?
4
+
5
+ SLTUNET 项目由**两个独立组件**组成,使用不同的深度学习框架:
6
+
7
+ 1. **SMKD** (步骤2-3) - 从视频提取特征
8
+ - 框架:**PyTorch**
9
+ - 文件:`smkd/` 目录下的代码
10
+ - 用途:训练 ResNet34 提取 1024 维特征
11
+
12
+ 2. **SLTUNET** (步骤4) - 翻译模型
13
+ - 框架:**TensorFlow 1.15**
14
+ - 文件:根目录下的 `main.py`, `sltunet.py` 等
15
+ - 用途:sign-to-gloss 翻译
16
+
17
+ **问题:** TensorFlow 1.15 和现代 PyTorch 的 CUDA 依赖冲突,无法在同一环境中共存。
18
+
19
+ ## 解决方案:两个 Conda 环境
20
+
21
+ ### 环境1:smkd_pytorch (用于步骤 2-3)
22
+
23
+ ```bash
24
+ # 创建环境
25
+ conda create -n smkd_pytorch python=3.8 -y
26
+ conda activate smkd_pytorch
27
+
28
+ # 安装 PyTorch (CUDA 11.8)
29
+ pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 \
30
+ --index-url https://download.pytorch.org/whl/cu118
31
+
32
+ # 安装其他依赖
33
+ pip install pyyaml tqdm h5py opencv-python pillow numpy
34
+ ```
35
+
36
+ ### 环境2:slt_tf1 (用于步骤 1, 4) - 已存在
37
+
38
+ ```bash
39
+ # 你已经创建了这个环境
40
+ conda activate slt_tf1
41
+ # 包含 TensorFlow 1.15
42
+ ```
43
+
44
+ ## 训练流程(使用正确的环境)
45
+
46
+ ### 步骤 1:数据预处理 (任意环境)
47
+
48
+ ```bash
49
+ conda activate slt_tf1 # 或 signx-slt
50
+ cd training_yutubeasl/scripts
51
+ python 1_preprocess.py
52
+ ```
53
+
54
+ ### 步骤 2:训练 SMKD (⚠️ 需要 PyTorch)
55
+
56
+ ```bash
57
+ conda activate signx-slt # ← 使用 PyTorch 环境
58
+ cd training_yutubeasl/scripts
59
+ bash 2_train_smkd.sh
60
+ ```
61
+
62
+ ### 步骤 3:提取特征 (⚠️ 需要 PyTorch)
63
+
64
+ ```bash
65
+ conda activate signx-slt # ← 保持 PyTorch 环境
66
+ bash 3_extract_features.sh
67
+ ```
68
+
69
+ ### 步骤 4:训练 SLTUNET (⚠️ 需要 TensorFlow)
70
+
71
+ ```bash
72
+ conda activate slt_tf1 # ← 切换到 TensorFlow 环境
73
+ bash 4_train_sltunet.sh
74
+ ```
75
+
76
+ ## 快速设置脚本
77
+
78
+ 运行这个脚本自动创建 PyTorch 环境:
79
+
80
+ ```bash
81
+ conda create -n smkd_pytorch python=3.8 -y && \
82
+ conda run -n smkd_pytorch pip install \
83
+ torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 \
84
+ --index-url https://download.pytorch.org/whl/cu118 && \
85
+ conda run -n smkd_pytorch pip install \
86
+ pyyaml tqdm h5py opencv-python pillow numpy
87
+
88
+ echo "✓ smkd_pytorch 环境创建完成"
89
+ ```
90
+
91
+ ## 为什么 README.md 没说需要 PyTorch?
92
+
93
+ README.md 只描述了 SLTUNET 主模型的要求(TensorFlow 1.15)。
94
+
95
+ SMKD 是一个**独立的 PyTorch 项目**,被集成到这个 repo 中用于特征提取。原作者可能假设用户会查看 `smkd/` 目录下的代码来了解其依赖。
96
+
97
+ ## 总结
98
+
99
+ | 步骤 | 环境 | 框架 | 用途 |
100
+ |------|------|------|------|
101
+ | 1. 预处理 | 任意 | - | 生成元数据 |
102
+ | 2. 训练SMKD | `smkd_pytorch` | PyTorch | 特征提取器 |
103
+ | 3. 提取特征 | `smkd_pytorch` | PyTorch | 生成.h5文件 |
104
+ | 4. 训练SLTUNET | `slt_tf1` | TensorFlow 1.15 | 翻译模型 |
105
+
106
+ **现在你需要创建 `smkd_pytorch` 环境,然后重新运行步骤2!**
plugins/SignX/training_yutubeasl/QUICK_START.md ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # 🚀 YouTubeASL 训练 - 快速开始
2
+
3
+ ## 一键运行(使用你现有的环境)
4
+
5
+ ### 步骤1:数据预处理 ✅ (已完成)
6
+
7
+ 已经生成:
8
+ - ✅ 1686 训练样本
9
+ - ✅ 210 验证样本
10
+ - ✅ 212 测试样本
11
+ - ✅ 1243 gloss 词汇
12
+
13
+ ### 步骤2:训练 SMKD (几小时)
14
+
15
+ ```bash
16
+ # 使用 signx-slt 环境(已有 PyTorch 2.4.1)
17
+ conda activate signx-slt
18
+
19
+ cd ~/code/Sign-X/output/huggingface_youtubeasl_repo/slt_new_new/training_yutubeasl/scripts
20
+
21
+ bash 2_train_smkd.sh
22
+ ```
23
+
24
+ ### 步骤3:提取特征 (1小时)
25
+
26
+ ```bash
27
+ # 保持在 signx-slt 环境
28
+ bash 3_extract_features.sh
29
+ ```
30
+
31
+ ### 步骤4:训练 SLTUNET (1天)
32
+
33
+ ```bash
34
+ # 切换到 TensorFlow 环境
35
+ conda activate slt_tf1
36
+
37
+ # 先准备文本数据(见下文)
38
+ # 然后运行
39
+ bash 4_train_sltunet.sh
40
+ ```
41
+
42
+ ## 为什么需要两个环境?
43
+
44
+ - **SMKD (步骤2-3)**: 使用 PyTorch → 用 `signx-slt` 环境
45
+ - **SLTUNET (步骤4)**: 使用 TensorFlow 1.15 → 用 `slt_tf1` 环境
46
+
47
+ 详细说明见 `ENVIRONMENT_SETUP.md`
48
+
49
+ ## 现在开始训练!
50
+
51
+ ```bash
52
+ conda activate signx-slt
53
+ cd ~/code/Sign-X/output/huggingface_youtubeasl_repo/slt_new_new/training_yutubeasl/scripts
54
+ bash 2_train_smkd.sh
55
+ ```
56
+
57
+ 训练大约需要几小时,请耐心等待。你可以用 `nvidia-smi` 监控 GPU 使用情况。
58
+
59
+ ## 需要帮助?
60
+
61
+ 查看 `docs/QUICKSTART.md` 获取详细说明。
plugins/SignX/training_yutubeasl/README.md ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # YouTubeASL Training Scripts and Documentation
2
+
3
+ This directory contains all scripts and documentation for training SLTUNET on the YouTubeASL dataset.
4
+
5
+ ## Directory Structure
6
+
7
+ ```
8
+ training_yutubeasl/
9
+ ├── README.md # This file
10
+ ├── scripts/
11
+ │ ├── 1_preprocess.py # Step 1: Data preprocessing
12
+ │ ├── 2_train_smkd.sh # Step 2: SMKD feature extraction training
13
+ │ ├── 3_extract_features.sh # Step 3: Extract features from trained SMKD
14
+ │ └── 4_train_sltunet.sh # Step 4: Train SLTUNET model
15
+ ├── configs/
16
+ │ ├── youtubeasl_smkd.yaml # SMKD training config
17
+ │ └── youtubeasl_dataset.yaml # Dataset config for SMKD
18
+ └── docs/
19
+ ├── QUICKSTART.md # Quick start guide (Chinese)
20
+ └── TUTORIAL.md # Detailed tutorial (English)
21
+ ```
22
+
23
+ ## Quick Start
24
+
25
+ ```bash
26
+ cd training_yutubeasl/scripts
27
+
28
+ # Step 1: Preprocess data
29
+ python 1_preprocess.py
30
+
31
+ # Step 2: Train SMKD
32
+ bash 2_train_smkd.sh
33
+
34
+ # Step 3: Extract features
35
+ bash 3_extract_features.sh
36
+
37
+ # Step 4: Train SLTUNET
38
+ bash 4_train_sltunet.sh
39
+ ```
40
+
41
+ See `docs/QUICKSTART.md` for detailed instructions.
plugins/SignX/training_yutubeasl/START_HERE.md ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # 🎯 从这里开始 - YouTubeASL训练
2
+
3
+ ## 📂 文件夹已整理完成!
4
+
5
+ 所有YouTubeASL训练相关的文件都已经整理到这个文件夹中。
6
+
7
+ ```
8
+ training_yutubeasl/ ← 你在这里
9
+ ├── START_HERE.md ← 📍 本文件(快速导航)
10
+ ├── run_all.sh ← 🚀 一键运行完整pipeline
11
+ ├── README.md ← 📖 目录说明
12
+ │
13
+ ├── scripts/ ← 📜 训练脚本(按顺序)
14
+ │ ├── 1_preprocess.py
15
+ │ ├── 2_train_smkd.sh
16
+ │ ├── 3_extract_features.sh
17
+ │ └── 4_train_sltunet.sh
18
+ │
19
+ ├── configs/ ← ⚙️ 配置文件
20
+ │ ├── youtubeasl_smkd.yaml
21
+ │ └── youtubeasl_dataset.yaml
22
+ │
23
+ └── docs/ ← 📚 文档
24
+ ├── QUICKSTART.md ← ⭐ 快速开始(中文)
25
+ └── TUTORIAL.md ← 📖 详细教程(英文)
26
+ ```
27
+
28
+ ## 🚀 三种使用方式
29
+
30
+ ### 方式1:一键运行(最简单)
31
+
32
+ ```bash
33
+ cd /research/cbim/vast/sf895/code/Sign-X/output/huggingface_youtubeasl_repo/slt_new_new/training_yutubeasl
34
+ bash run_all.sh
35
+ ```
36
+
37
+ ### 方式2:分步运行(推荐调试)
38
+
39
+ ```bash
40
+ cd scripts
41
+
42
+ # 步骤1
43
+ python 1_preprocess.py
44
+
45
+ # 步骤2
46
+ bash 2_train_smkd.sh
47
+
48
+ # 步骤3
49
+ bash 3_extract_features.sh
50
+
51
+ # 步骤4(需要先准备文本数据)
52
+ bash 4_train_sltunet.sh
53
+ ```
54
+
55
+ ### 方式3:手动运行(完全控制)
56
+
57
+ 查看 `docs/TUTORIAL.md` 获取每一步的详细命令。
58
+
59
+ ## 📖 下一步阅读什么?
60
+
61
+ 1. **新手?** → 先看 `docs/QUICKSTART.md`
62
+ 2. **想了解细节?** → 看 `docs/TUTORIAL.md`
63
+ 3. **遇到问题?** → `docs/QUICKSTART.md` 有故障排查
64
+ 4. **想调参数?** → 看 `configs/` 文件夹
65
+
66
+ ## ✅ 已经为你做的事情
67
+
68
+ - ✅ 整理了所有训练脚本到 `scripts/`
69
+ - ✅ 整理了所有配置文件到 `configs/`
70
+ - ✅ 整理了所有文档到 `docs/`
71
+ - ✅ 创建了主入口脚本 `run_all.sh`
72
+ - ✅ 更新了所有路径引用(无需修改)
73
+ - ✅ 添加了详细的中文说明
74
+
75
+ ## 🎓 训练流程概览
76
+
77
+ ```
78
+ 原始数据
79
+ ↓
80
+ [1_preprocess.py] → 生成训练元数据
81
+ ↓
82
+ [2_train_smkd.sh] → 训练特征提取器(几小时)
83
+ ↓
84
+ [3_extract_features.sh] → 提取1024维特征
85
+ ↓
86
+ [准备文本数据] → BPE分词(手动)
87
+ ↓
88
+ [4_train_sltunet.sh] → 训练翻译模型(1天)
89
+ ↓
90
+ 完成!
91
+ ```
92
+
93
+ ## ⚡ 快速测试
94
+
95
+ 想快速测试是否一切正常?
96
+
97
+ ```bash
98
+ cd scripts
99
+ python 1_preprocess.py --help # 应该显示帮助信息
100
+ ```
101
+
102
+ ## 🆘 需要帮助?
103
+
104
+ - 查看 `docs/QUICKSTART.md` 的 "常见问题" 部分
105
+ - 检查 `configs/` 中的参数设置
106
+ - 确保GPU可用:`nvidia-smi`
107
+
108
+ ## 🎉 准备好了吗?
109
+
110
+ ```bash
111
+ # 开始吧!
112
+ bash run_all.sh
113
+ ```
114
+
115
+ 祝训练顺利!🚀
plugins/SignX/training_yutubeasl/configs/youtubeasl_dataset.yaml ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ dataset_root: youtubeasl
2
+ dict_path: youtubeasl/gloss_dict.npy
3
+ evaluation_dir: ./evaluation/slr_eval
4
+ evaluation_prefix: youtubeasl-groundtruth
plugins/SignX/training_yutubeasl/configs/youtubeasl_smkd.yaml ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ feeder: dataset.dataloader_video.BaseFeeder
2
+ phase: train
3
+ dataset: youtubeasl
4
+ num_epoch: 120
5
+ work_dir: ./work_dir/youtubeasl_baseline/
6
+ batch_size: 2
7
+ random_seed: 0
8
+ test_batch_size: 8
9
+ num_worker: 4
10
+ device: 0
11
+ log_interval: 50
12
+ eval_interval: 1
13
+ save_interval: 1
14
+ evaluate_tool: sclite
15
+ num_feature_aug: 10
16
+ loss_weights:
17
+ SeqCTC: 1.0
18
+ ConvCTC: 1.0
19
+ Dist: 10.0
20
+
21
+ optimizer_args:
22
+ optimizer: Adam
23
+ base_lr: 0.0001
24
+ step: [ 20, 35]
25
+ learning_ratio: 1
26
+ weight_decay: 0.0001
27
+ start_epoch: 0
28
+ nesterov: False
29
+
30
+ feeder_args:
31
+ mode: 'train'
32
+ datatype: 'video'
33
+ num_gloss: -1
34
+ drop_ratio: 1.0
35
+ ctc_length_guard: true
36
+ ctc_length_guard_post: true
37
+ max_frames: -1
38
+ skip_on_decode_failure: true
39
+ min_frames: 5
40
+ skip_on_black_frames: true
41
+ temporal_rescale: 0.2
42
+
43
+ model: slr_network.SLRModel
44
+ decode_mode: max
45
+ model_args:
46
+ num_classes: 1296
47
+ c2d_type: resnet34
48
+ conv_type: 2
49
+ use_bn: 1
50
+ share_classifier: True
51
+ weight_norm: True
plugins/SignX/training_yutubeasl/docs/QUICKSTART.md ADDED
@@ -0,0 +1,203 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # YouTubeASL SLTUNET 训练快速开始指南
2
+
3
+ ## 概述
4
+
5
+ 本项目**完全按照 SLTUNET 原始方法**在 YouTubeASL 数据集上训练手语翻译模型。
6
+
7
+ **不修改任何模型代码**,只做数据预处理适配。
8
+
9
+ ## 训练流程
10
+
11
+ ### 第一步:数据预处理 (5分钟)
12
+
13
+ ```bash
14
+ cd /research/cbim/vast/sf895/code/Sign-X/output/huggingface_youtubeasl_repo/slt_new_new
15
+
16
+ python 1_preprocess.py
17
+ ```
18
+
19
+ 这将生成:
20
+ - `smkd/youtubeasl/train_info.npy` - 训练集元数据 (~1684 样本)
21
+ - `smkd/youtubeasl/dev_info.npy` - 验证集元数据 (~210 样本)
22
+ - `smkd/youtubeasl/test_info.npy` - 测试集元数据 (~214 样本)
23
+ - `smkd/youtubeasl/gloss_dict.npy` - Gloss 词典
24
+
25
+ ### 第二步:训练 SMKD 特征提取器 (几小时)
26
+
27
+ 从原始视频帧提取深度特征:
28
+
29
+ ```bash
30
+ cd smkd
31
+ python main.py --work-dir work_dir/youtubeasl_smkd --config youtubeasl_smkd.yaml --device 0
32
+ ```
33
+
34
+ 训练 ResNet34 模型提取 1024 维手语视频特征。
35
+
36
+ **可选参数:**
37
+ - `--device 0,1` - 使用多GPU
38
+ - 修改 `youtubeasl_smkd.yaml` 中的 `batch_size` 根据显存调整
39
+
40
+ ### 第三步:提取特征 (1小时)
41
+
42
+ 训练完成后,选择最好的 checkpoint 提取特征:
43
+
44
+ ```bash
45
+ # 平均 top-5 checkpoints
46
+ python ckpt_avg.py --path work_dir/youtubeasl_smkd --checkpoints 5 --output avg
47
+
48
+ # 提取特征
49
+ python main.py \
50
+ --load-weights avg/average.pt \
51
+ --phase features \
52
+ --device 0 \
53
+ --num-feature-aug 10 \
54
+ --work-dir work_dir/youtubeasl_smkd \
55
+ --config youtubeasl_smkd.yaml
56
+
57
+ # 合并训练特征
58
+ cd work_dir/youtubeasl_smkd
59
+ python ../../example/sign_feature_cmb.py train*.h5
60
+ cd ../..
61
+ ```
62
+
63
+ 生成 `train.h5`, `dev.h5`, `test.h5` 文件。
64
+
65
+ ### 第四步:准备文本数据 (手动)
66
+
67
+ **注意:这一步需要根据你的需求手动处理**
68
+
69
+ SLTUNET 原本是做 **手语→文本翻译**,但 YouTubeASL 只有 Gloss 标注。
70
+
71
+ 你需要决定:
72
+
73
+ **选项 A:只做 Gloss 识别(推荐开始)**
74
+ ```bash
75
+ mkdir -p preprocessed-youtubeasl
76
+ cd preprocessed-youtubeasl
77
+
78
+ # 从 smkd/youtubeasl/*_info.npy 提取 gloss 序列
79
+ # 创建 train.gloss, dev.gloss, test.gloss
80
+
81
+ # 学习 BPE
82
+ pip install subword-nmt
83
+ cat train.gloss | subword-nmt learn-bpe -s 1000 > youtubeasl.bpe
84
+
85
+ # 应用 BPE
86
+ for split in train dev test; do
87
+ subword-nmt apply-bpe -c youtubeasl.bpe < ${split}.gloss > ${split}.bpe.gloss
88
+ done
89
+
90
+ # 添加索引(第一列是特征索引)
91
+ awk 'BEGIN{i=0} /.*/{printf "%d %s\n",i,$0; i++}' < train.bpe.gloss > train.bpe.gloss.idx
92
+ awk 'BEGIN{i=0} /.*/{printf "%d %s\n",i,$0; i++}' < dev.bpe.gloss > dev.bpe.gloss.idx
93
+ awk 'BEGIN{i=0} /.*/{printf "%d %s\n",i,$0; i++}' < test.bpe.gloss > test.bpe.gloss.idx
94
+
95
+ # 打乱训练数据
96
+ python ../example/shuffle_corpus.py --corpus train.bpe.gloss.idx train.bpe.gloss.idx
97
+
98
+ # 创建词典
99
+ cut -d ' ' -f 2- < train.bpe.gloss.idx > train.bpe.gloss.noindex
100
+ python ../vocab.py train.bpe.gloss.noindex vocab.youtubeasl
101
+
102
+ cd ..
103
+ ```
104
+
105
+ **选项 B:添加英文翻译(需要额外标注)**
106
+ 如果你有英文翻译,可以做 Gloss→English 的翻译任务。
107
+
108
+ ### 第五步:训练 SLTUNET (1天)
109
+
110
+ ```bash
111
+ bash train_sltunet.sh
112
+ ```
113
+
114
+ 或者手动运行(可以调整参数):
115
+
116
+ ```bash
117
+ export CUDA_VISIBLE_DEVICES=0
118
+
119
+ python run.py --mode train --parameters=\
120
+ hidden_size=256,embed_size=256,filter_size=4096,\
121
+ sep_layer=0,num_encoder_layer=6,num_decoder_layer=6,\
122
+ ctc_enable=True,ctc_alpha=0.3,ctc_repeated=True,\
123
+ batch_size=40,eval_batch_size=16,\
124
+ max_len=256,max_img_len=512,\
125
+ img_feature_size=1024,img_aug_size=11,\
126
+ gpus=[0],\
127
+ img_train_file="smkd/work_dir/youtubeasl_smkd/train.h5",\
128
+ src_train_file="preprocessed-youtubeasl/train.bpe.gloss.shuf",\
129
+ tgt_train_file="preprocessed-youtubeasl/train.bpe.gloss.shuf",\
130
+ img_dev_file="smkd/work_dir/youtubeasl_smkd/dev.h5",\
131
+ src_dev_file="preprocessed-youtubeasl/dev.bpe.gloss.idx",\
132
+ tgt_dev_file="preprocessed-youtubeasl/dev.bpe.gloss.idx",\
133
+ output_dir="train_youtubeasl"
134
+ ```
135
+
136
+ ## 便捷脚本
137
+
138
+ ```bash
139
+ # 方式1:分步执行
140
+ bash train_youtubeasl.sh # 预处理 + SMKD训练
141
+ bash extract_features.sh # 提取特征
142
+ # 手动准备 BPE 数据...
143
+ bash train_sltunet.sh # SLTUNET训练
144
+
145
+ # 方式2:单独执行各步骤(见上文)
146
+ ```
147
+
148
+ ## 文件说明
149
+
150
+ | 文件 | 说明 |
151
+ |------|------|
152
+ | `1_preprocess.py` | YouTubeASL数据预处理脚本 |
153
+ | `train_youtubeasl.sh` | SMKD训练脚本 |
154
+ | `extract_features.sh` | 特征提取脚本 |
155
+ | `train_sltunet.sh` | SLTUNET训练脚本 |
156
+ | `smkd/youtubeasl_smkd.yaml` | SMKD配置文件 |
157
+ | `training_yutubeasl/configs/youtubeasl_dataset.yaml` | 数据集配置文件 |
158
+ | `README_YouTubeASL.md` | 详细文档 |
159
+
160
+ ## 预期结果
161
+
162
+ - **数据集大小:** ~2108 样本 (80% train / 10% dev / 10% test)
163
+ - **Gloss 词汇量:** ~几百个独特的 gloss tokens
164
+ - **特征维度:** 1024 (来自 ResNet34)
165
+ - **训练时间:**
166
+ - SMKD: ~10-20 epochs, 几小时
167
+ - SLTUNET: ~5000 steps, ~1天
168
+
169
+ ## 故障排查
170
+
171
+ ### 问题1:找不到视频��
172
+ 确保路径正确:
173
+ ```
174
+ YouTubeASL_utterances_results/
175
+ ├── 10006709/
176
+ │ └── crop_frame/
177
+ │ ├── 00000001.jpg
178
+ │ ├── 00000002.jpg
179
+ │ └── ...
180
+ ├── 10036884/
181
+ └── ...
182
+ ```
183
+
184
+ ### 问题2:CUDA out of memory
185
+ - 减小 `batch_size`(在 `youtubeasl_smkd.yaml` 或 `train_sltunet.sh`)
186
+ - 减少 `num_feature_aug`
187
+ - 使用多GPU:`--device 0,1`
188
+
189
+ ### 问题3:导入错误
190
+ 确保在正确的目录和环境:
191
+ ```bash
192
+ cd /research/cbim/vast/sf895/code/Sign-X/output/huggingface_youtubeasl_repo/slt_new_new
193
+ conda activate WaveGen # 或你的环境
194
+ ```
195
+
196
+ ## 下一步
197
+
198
+ 训练完成后:
199
+ 1. 评估模型性能(Gloss 识别准确率/WER)
200
+ 2. 如果效果不好,调整超参数
201
+ 3. 考虑添加更多数据或使用预训练模型
202
+
203
+ 详细信息请参考 `README_YouTubeASL.md`。
plugins/SignX/training_yutubeasl/docs/TUTORIAL.md ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # YouTubeASL Training Pipeline for SLTUNET
2
+
3
+ This guide describes how to train SLTUNET on the YouTubeASL dataset, following the exact same pipeline as Phoenix-2014T.
4
+
5
+ ## Prerequisites
6
+
7
+ 1. Python 3.8
8
+ 2. TensorFlow 1.15
9
+ 3. PyTorch (for SMKD)
10
+ 4. YouTubeASL data located at:
11
+ - `../../YouTubeASL_utterances_results/` (video frames)
12
+ - `../../YouTubeASL_utterances_mapping.txt` (gloss annotations)
13
+
14
+ ## Complete Training Pipeline
15
+
16
+ ### Step 1: Preprocess YouTubeASL Dataset
17
+
18
+ Generate `{train,dev,test}_info.npy` and `gloss_dict.npy` files:
19
+
20
+ ```bash
21
+ python 1_preprocess.py \
22
+ --data-root ../../YouTubeASL_utterances_results \
23
+ --mapping-file ../../YouTubeASL_utterances_mapping.txt \
24
+ --output-dir ./smkd/youtubeasl \
25
+ --train-ratio 0.8 \
26
+ --dev-ratio 0.1
27
+ ```
28
+
29
+ **Output:**
30
+ - `smkd/youtubeasl/train_info.npy` - Training set metadata
31
+ - `smkd/youtubeasl/dev_info.npy` - Dev set metadata
32
+ - `smkd/youtubeasl/test_info.npy` - Test set metadata
33
+ - `smkd/youtubeasl/gloss_dict.npy` - Gloss vocabulary
34
+ - `smkd/youtubeasl/youtubeasl-groundtruth-{train,dev,test}.stm` - Evaluation files
35
+
36
+ ### Step 2: Pretrain SMKD Embeddings
37
+
38
+ Train a ResNet34 model to extract sign video features:
39
+
40
+ ```bash
41
+ cd smkd
42
+ python main.py --work-dir work_dir/youtubeasl_smkd --config youtubeasl_smkd.yaml --device 0
43
+ ```
44
+
45
+ This will:
46
+ - Train SMKD model for 40 epochs
47
+ - Save checkpoints in `work_dir/youtubeasl_smkd/`
48
+ - Evaluate on dev set every epoch
49
+
50
+ **Training time:** ~Several hours depending on GPU
51
+
52
+ ### Step 3: Checkpoint Averaging
53
+
54
+ After training, select top-5 checkpoints based on dev set WER and average them:
55
+
56
+ ```bash
57
+ # Still in smkd directory
58
+ python ckpt_avg.py --path work_dir/youtubeasl_smkd --checkpoints 5 --output avg
59
+ ```
60
+
61
+ ### Step 4: Extract Sign Features
62
+
63
+ Extract 1024-dim features for all video frames:
64
+
65
+ ```bash
66
+ python main.py \
67
+ --load-weights avg/average.pt \
68
+ --phase features \
69
+ --device 0 \
70
+ --num-feature-aug 10 \
71
+ --work-dir work_dir/youtubeasl_smkd \
72
+ --config youtubeasl_smkd.yaml
73
+ ```
74
+
75
+ Then combine training features:
76
+
77
+ ```bash
78
+ cd work_dir/youtubeasl_smkd
79
+ python ../../example/sign_feature_cmb.py train*.h5
80
+ cd ../../..
81
+ ```
82
+
83
+ **Output:** `train.h5`, `dev.h5`, `test.h5` containing sign video features
84
+
85
+ ### Step 5: Prepare Text Data (BPE Tokenization)
86
+
87
+ Create `preprocessed-youtubeasl/` directory and prepare BPE files:
88
+
89
+ ```bash
90
+ mkdir -p preprocessed-youtubeasl
91
+ cd preprocessed-youtubeasl
92
+
93
+ # Extract glosses from train/dev/test splits
94
+ # Create train.gloss, dev.gloss, test.gloss files
95
+
96
+ # Learn BPE with subword-nmt (install if needed: pip install subword-nmt)
97
+ cat train.gloss | subword-nmt learn-bpe -s 1000 > youtubeasl.bpe
98
+
99
+ # Apply BPE
100
+ for split in train dev test; do
101
+ subword-nmt apply-bpe -c youtubeasl.bpe < ${split}.gloss > ${split}.bpe.gloss
102
+ done
103
+
104
+ # Shuffle training data
105
+ python ../example/shuffle_corpus.py --corpus train.bpe.gloss train.bpe.gloss
106
+
107
+ # Create vocabulary
108
+ python ../vocab.py train.bpe.gloss vocab.youtubeasl
109
+
110
+ cd ..
111
+ ```
112
+
113
+ ### Step 6: Train SLTUNET Model
114
+
115
+ ```bash
116
+ bash train_sltunet.sh
117
+ ```
118
+
119
+ This will train the full SLTUNET model with:
120
+ - Multi-task learning (gloss recognition + translation)
121
+ - CTC alignment
122
+ - BPE dropout for regularization
123
+
124
+ **Output:** Model checkpoints in `train_youtubeasl/` directory
125
+
126
+ ### Step 7: Evaluation
127
+
128
+ After training, average top-10 checkpoints and evaluate:
129
+
130
+ ```bash
131
+ python models/checkpoint_averaging.py \
132
+ --path train_youtubeasl \
133
+ --checkpoints 10 \
134
+ --output avg \
135
+ --gpu 0
136
+ ```
137
+
138
+ Then decode on test set (see `example/test.sh` for reference).
139
+
140
+ ## Quick Start Scripts
141
+
142
+ We provide three convenience scripts:
143
+
144
+ 1. **`train_youtubeasl.sh`** - Run Steps 1-2 (preprocessing + SMKD training)
145
+ 2. **`extract_features.sh`** - Run Steps 3-4 (checkpoint averaging + feature extraction)
146
+ 3. **`train_sltunet.sh`** - Run Step 6 (SLTUNET training)
147
+
148
+ Run them sequentially:
149
+
150
+ ```bash
151
+ bash 2_train_smkd.sh # Wait for completion
152
+ bash 3_extract_features.sh # Wait for completion
153
+ # Prepare BPE data (Step 5)
154
+ bash 4_train_sltunet.sh # Final training
155
+ ```
156
+
157
+ ## File Structure
158
+
159
+ ```
160
+ slt_new_new/
161
+ ├── 1_preprocess.py # Data preprocessing script
162
+ ├── 2_train_smkd.sh # SMKD training script
163
+ ├── 3_extract_features.sh # Feature extraction script
164
+ ├── 4_train_sltunet.sh # SLTUNET training script
165
+ ├── smkd/
166
+ │ ├── youtubeasl_smkd.yaml # SMKD config for YouTubeASL
167
+ │ ├── configs/
168
+ │ │ └── youtubeasl_dataset.yaml # Dataset config
169
+ │ ├── youtubeasl/ # Preprocessed data (generated)
170
+ │ │ ├── train_info.npy
171
+ │ │ ├── dev_info.npy
172
+ │ │ ├── test_info.npy
173
+ │ │ └── gloss_dict.npy
174
+ │ └── work_dir/
175
+ │ └── youtubeasl_smkd/ # SMKD checkpoints & features
176
+ │ ├── train.h5
177
+ │ ├── dev.h5
178
+ │ └── test.h5
179
+ └── preprocessed-youtubeasl/ # BPE tokenized data (manual)
180
+ ├── youtubeasl.bpe
181
+ ├── train.bpe.gloss.shuf
182
+ ├── dev.bpe.gloss
183
+ ├── test.bpe.gloss
184
+ └── vocab.youtubeasl
185
+ ```
186
+
187
+ ## Notes
188
+
189
+ - The pipeline exactly follows Phoenix-2014T methodology
190
+ - No model modifications - only data preprocessing
191
+ - SMKD extracts 1024-dim features from video frames (256x256 → 224x224 crop)
192
+ - SLTUNET uses these features for sign-to-gloss translation
193
+ - Total training time: ~1-2 days on single GPU (SMKD) + ~1 day (SLTUNET)
plugins/SignX/training_yutubeasl/run_all.sh ADDED
@@ -0,0 +1,247 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ # Complete YouTubeASL training pipeline with automatic environment switching
4
+ # This will automatically use the correct environment for each step
5
+
6
+ set -e
7
+
8
+ echo "========================================"
9
+ echo " YouTubeASL SLTUNET Training Pipeline "
10
+ echo "========================================"
11
+ echo ""
12
+ echo "This will run the complete pipeline:"
13
+ echo " 1. Data preprocessing"
14
+ echo " 2. SMKD feature extractor training (PyTorch)"
15
+ echo " 3. Feature extraction (PyTorch)"
16
+ echo " 4. SLTUNET training (TensorFlow)"
17
+ echo ""
18
+ echo "Environments will be switched automatically:"
19
+ echo " - Steps 2-3: signx-slt (PyTorch)"
20
+ echo " - Step 4: slt_tf1 (TensorFlow)"
21
+ echo ""
22
+ echo "========================================"
23
+ echo ""
24
+
25
+ # Get conda base
26
+ CONDA_BASE=$(conda info --base)
27
+ source "${CONDA_BASE}/etc/profile.d/conda.sh"
28
+
29
+ cd scripts
30
+ SCRIPTS_DIR=$(pwd)
31
+
32
+ # Step 1: Preprocess (use signx-slt to avoid numpy pickle incompatibility)
33
+ echo ">>> Step 1/4: Preprocessing YouTubeASL dataset..."
34
+ echo "Switching to signx-slt environment for preprocessing..."
35
+ conda activate signx-slt
36
+ python 1_preprocess.py \
37
+ --video-root /common/users/sf895/SignerX-inference/plugins/YouTubeASL \
38
+ --frames-root /common/users/sf895/SignerX-inference/plugins/YouTubeASL/video1000size_frames \
39
+ --manifest-path /common/users/sf895/SignerX-inference/plugins/YouTubeASL/video1000size/manifest_top1000.tsv
40
+ echo "✓ Preprocessing complete"
41
+ echo ""
42
+
43
+ # Step 2: Train SMKD (need PyTorch)
44
+ echo ">>> Step 2/4: Training SMKD feature extractor..."
45
+ echo "Using signx-slt environment (PyTorch)..."
46
+ SMKD_DIR=../../smkd
47
+ SMKD_WORK_DIR="${SMKD_DIR}/work_dir/youtubeasl_smkd"
48
+ BEST_SMKD_CKPT="${SMKD_WORK_DIR}/best_model.pt"
49
+
50
+ # Decide GPU usage (interactive unless SMKD_NUM_GPUS already set)
51
+ if [[ -z "${SMKD_NUM_GPUS:-}" ]]; then
52
+ TOTAL_GPUS=$(nvidia-smi -L 2>/dev/null | wc -l | tr -d ' ')
53
+ if [[ -z "$TOTAL_GPUS" || "$TOTAL_GPUS" -lt 1 ]]; then
54
+ TOTAL_GPUS=1
55
+ fi
56
+ DEFAULT_GPU_COUNT=$TOTAL_GPUS
57
+ read -p "How many GPUs should SMKD training use? (1-${TOTAL_GPUS}, default ${DEFAULT_GPU_COUNT}): " -r USER_GPU_COUNT
58
+ if [[ -n "$USER_GPU_COUNT" ]]; then
59
+ SMKD_NUM_GPUS=$USER_GPU_COUNT
60
+ else
61
+ SMKD_NUM_GPUS=$DEFAULT_GPU_COUNT
62
+ fi
63
+ fi
64
+ export SMKD_NUM_GPUS
65
+ echo "SMKD training will run on ${SMKD_NUM_GPUS} GPU(s). Override by exporting SMKD_NUM_GPUS before running this script."
66
+
67
+ # Choose training mode
68
+ SMKD_MODE="train"
69
+ RESUME_CHECKPOINT=""
70
+ if [[ -f "${BEST_SMKD_CKPT}" ]]; then
71
+ echo ""
72
+ echo "Detected existing SMKD checkpoint at ${BEST_SMKD_CKPT}."
73
+ echo "Choose training mode:"
74
+ echo " 1) Train from scratch (will overwrite existing checkpoints)"
75
+ echo " 2) Continue training from existing checkpoint (resume)"
76
+ echo " 3) Skip training and use existing checkpoint"
77
+ echo ""
78
+ read -p "Enter choice (1/2/3, default 3): " -r SMKD_CHOICE
79
+
80
+ case "${SMKD_CHOICE}" in
81
+ 1)
82
+ SMKD_MODE="train"
83
+ echo "Will train from scratch (existing checkpoints will be overwritten)."
84
+ ;;
85
+ 2)
86
+ SMKD_MODE="resume"
87
+ echo ""
88
+ echo "Available checkpoints in ${SMKD_WORK_DIR}:"
89
+ ls -lh "${SMKD_WORK_DIR}"/*.pt 2>/dev/null | awk '{print " " $9 " (" $5 ")"}'
90
+ echo ""
91
+ read -p "Enter checkpoint path (default: ${BEST_SMKD_CKPT}): " -r USER_CHECKPOINT
92
+ if [[ -n "${USER_CHECKPOINT}" ]]; then
93
+ # Handle relative paths
94
+ if [[ "${USER_CHECKPOINT}" != /* ]]; then
95
+ RESUME_CHECKPOINT="${SMKD_WORK_DIR}/${USER_CHECKPOINT}"
96
+ else
97
+ RESUME_CHECKPOINT="${USER_CHECKPOINT}"
98
+ fi
99
+ else
100
+ RESUME_CHECKPOINT="${BEST_SMKD_CKPT}"
101
+ fi
102
+
103
+ # Convert to absolute path
104
+ RESUME_CHECKPOINT=$(cd "$(dirname "${RESUME_CHECKPOINT}")" && pwd)/$(basename "${RESUME_CHECKPOINT}")
105
+
106
+ if [[ ! -f "${RESUME_CHECKPOINT}" ]]; then
107
+ echo "ERROR: Checkpoint not found: ${RESUME_CHECKPOINT}"
108
+ exit 1
109
+ fi
110
+ echo "Will continue training from ${RESUME_CHECKPOINT}."
111
+ ;;
112
+ 3|"")
113
+ SMKD_MODE="skip"
114
+ echo "Skipping SMKD training, will reuse existing checkpoint."
115
+ ;;
116
+ *)
117
+ echo "Invalid choice. Defaulting to skip."
118
+ SMKD_MODE="skip"
119
+ ;;
120
+ esac
121
+ else
122
+ echo "No existing checkpoint found. Will train from scratch."
123
+ SMKD_MODE="train"
124
+ fi
125
+
126
+ # Check PyTorch
127
+ python -c "import torch; print('Using PyTorch', torch.__version__)" || {
128
+ echo "ERROR: PyTorch not found in signx-slt!"
129
+ exit 1
130
+ }
131
+
132
+ cd "${SMKD_DIR}"
133
+ if [[ "${SMKD_MODE}" != "skip" ]]; then
134
+ echo "Running SMKD training (this may take several hours)..."
135
+ cd "$SCRIPTS_DIR"
136
+ if [[ "${SMKD_MODE}" == "resume" ]]; then
137
+ export SMKD_RESUME_CHECKPOINT="${RESUME_CHECKPOINT}"
138
+ fi
139
+ bash 2_train_smkd.sh
140
+ cd "${SMKD_DIR}"
141
+ else
142
+ echo "SMKD training skipped."
143
+ fi
144
+ echo ""
145
+
146
+ # Step 3: Extract features (stay in PyTorch)
147
+ echo ">>> Step 3/4: Extracting features..."
148
+ echo "Staying in signx-slt environment..."
149
+
150
+ FEATURE_DIR=../smkd/features
151
+ WORK_FEATURE_DIR=../smkd/work_dir/youtubeasl_smkd
152
+
153
+ SKIP_FEATURES="n"
154
+ if [[ -f "${WORK_FEATURE_DIR}/train.h5" && -f "${WORK_FEATURE_DIR}/dev.h5" && -f "${WORK_FEATURE_DIR}/test.h5" ]]; then
155
+ echo "Detected existing merged feature files in ${WORK_FEATURE_DIR}."
156
+ read -p "Skip feature extraction and reuse existing .h5 files? (Y/n) " -n 1 -r
157
+ echo
158
+ if [[ -z $REPLY || $REPLY =~ ^[Yy]$ ]]; then
159
+ SKIP_FEATURES="y"
160
+ echo "Skipping feature extraction; existing files will be used."
161
+ fi
162
+ fi
163
+
164
+ if [[ "${SKIP_FEATURES}" != "y" ]]; then
165
+ # Checkpoint averaging
166
+ echo "Averaging top-5 checkpoints..."
167
+ python ckpt_avg.py \
168
+ --path work_dir/youtubeasl_smkd \
169
+ --checkpoints 5 \
170
+ --output avg
171
+
172
+ # Extract features
173
+ echo "Extracting sign features..."
174
+ python main.py \
175
+ --load-weights avg/average.pt \
176
+ --phase features \
177
+ --device 0 \
178
+ --num-feature-aug 10 \
179
+ --work-dir work_dir/youtubeasl_smkd \
180
+ --config ../training_yutubeasl/configs/youtubeasl_smkd.yaml
181
+
182
+ # Combine training features
183
+ echo "Combining training features..."
184
+ cd features
185
+ python ../training_phoenix/sign_feature_cmb.py train_*.h5
186
+ echo "Copying merged features to work_dir..."
187
+ cp train.h5 ../work_dir/youtubeasl_smkd/train.h5
188
+ cp dev.h5 ../work_dir/youtubeasl_smkd/dev.h5
189
+ cp test.h5 ../work_dir/youtubeasl_smkd/test.h5
190
+ cd ..
191
+ cd "$SCRIPTS_DIR"
192
+
193
+ echo "✓ Feature extraction complete"
194
+ else
195
+ echo "Feature extraction skipped."
196
+ fi
197
+ echo ""
198
+
199
+ # Step 4: Train SLTUNET (need TensorFlow)
200
+ echo ">>> Step 4/4: SLTUNET training"
201
+ echo ""
202
+ echo "⚠️ IMPORTANT: Before running SLTUNET, you need to prepare text data:"
203
+ echo " 1. Create preprocessed-youtubeasl/ directory"
204
+ echo " 2. Extract gloss sequences from *_info.npy files"
205
+ echo " 3. Apply BPE tokenization"
206
+ echo " 4. Create vocabulary file"
207
+ echo ""
208
+ echo "See training_yutubeasl/docs/QUICKSTART.md for detailed instructions"
209
+ echo ""
210
+ read -p "Have you prepared the text data? (y/n) " -n 1 -r
211
+ echo
212
+ if [[ $REPLY =~ ^[Yy]$ ]]
213
+ then
214
+ echo "Switching to slt_tf1 environment (TensorFlow)..."
215
+ conda activate slt_tf1
216
+ export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
217
+
218
+ # Check TensorFlow
219
+ python -c "import tensorflow as tf; print('Using TensorFlow', tf.__version__)" || {
220
+ echo "ERROR: TensorFlow not found in slt_tf1!"
221
+ exit 1
222
+ }
223
+
224
+ echo "Starting SLTUNET training..."
225
+ cd "$SCRIPTS_DIR"
226
+ bash 4_train_sltunet.sh
227
+ echo "✓ SLTUNET training complete"
228
+ else
229
+ echo ""
230
+ echo "Please prepare text data first."
231
+ echo "Steps:"
232
+ echo " 1. cd ../../"
233
+ echo " 2. mkdir -p preprocessed-youtubeasl"
234
+ echo " 3. Follow instructions in training_yutubeasl/docs/QUICKSTART.md"
235
+ echo " 4. Then run: conda activate slt_tf1 && bash training_yutubeasl/scripts/4_train_sltunet.sh"
236
+ exit 0
237
+ fi
238
+
239
+ echo ""
240
+ echo "========================================"
241
+ echo " Training Pipeline Complete! "
242
+ echo "========================================"
243
+ echo ""
244
+ echo "Output files:"
245
+ echo " - SMKD checkpoints: smkd/work_dir/youtubeasl_smkd/"
246
+ echo " - Features: smkd/work_dir/youtubeasl_smkd/{train,dev,test}.h5"
247
+ echo " - SLTUNET model: train_youtubeasl/"
plugins/SignX/training_yutubeasl/scripts/.nfs000000004234454e00002a55 ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ # Train SLTUNET model on ASLLRP dataset
4
+ # Run this after extract_features.sh completes
5
+
6
+ set -e
7
+
8
+ SCRIPT_DIR=$(cd "$(dirname "$0")" && pwd)
9
+ REPO_ROOT=$(cd "$SCRIPT_DIR/../.." && pwd)
10
+
11
+ # Auto-select available GPUs (skip GPUs with >20% memory usage)
12
+ echo "========================================="
13
+ echo "Auto-selecting available GPUs..."
14
+ echo "========================================="
15
+ eval $(bash "$SCRIPT_DIR/select_gpus.sh")
16
+
17
+ export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
18
+
19
+ # Paths
20
+ data=preprocessed-asllrp/
21
+ feature=smkd/work_dir/asllrp_smkd/
22
+
23
+ echo "========================================="
24
+ echo "Training SLTUNET on ASLLRP Dataset"
25
+ echo "========================================="
26
+
27
+ # First, you need to prepare the text data (glosses)
28
+ # Create preprocessed-asllrp directory and prepare BPE tokenization
29
+
30
+ echo ""
31
+ echo "Make sure you have:"
32
+ echo "1. Created $data directory"
33
+ echo "2. Prepared BPE tokenized gloss files"
34
+ echo "3. Extracted SMKD features in $feature"
35
+ echo ""
36
+
37
+ # Train SLTUNET model
38
+ # Create log directory
39
+ LOG_DIR="$REPO_ROOT/checkpoints_asllrp/logs"
40
+ mkdir -p "$LOG_DIR"
41
+ TIMESTAMP=$(date +%Y%m%d_%H%M%S)
42
+ LOG_FILE="$LOG_DIR/training_${TIMESTAMP}.log"
43
+
44
+ echo "Training logs will be saved to: $LOG_FILE"
45
+ echo "You can monitor the log with: tail -f $LOG_FILE"
46
+ echo ""
47
+
48
+ (cd "$REPO_ROOT" && python run.py --mode train --parameters=\
49
+ hidden_size=256,embed_size=256,filter_size=4096,\
50
+ sep_layer=0,num_encoder_layer=6,num_decoder_layer=6,\
51
+ ctc_enable=True,ctc_alpha=0.3,ctc_repeated=True,\
52
+ src_bpe_dropout=0.2,tgt_bpe_dropout=0.2,bpe_dropout_stochastic_rate=0.6,\
53
+ initializer="uniform_unit_scaling",initializer_gain=0.5,\
54
+ dropout=0.3,label_smooth=0.1,attention_dropout=0.3,relu_dropout=0.5,residual_dropout=0.4,\
55
+ max_len=256,max_img_len=512,batch_size=240,eval_batch_size=80,\
56
+ token_size=4800,batch_or_token='token',beam_size=8,remove_bpe=True,decode_alpha=1.0,\
57
+ scope_name="transformer",buffer_size=50000,data_leak_ratio=0.1,\
58
+ img_feature_size=1024,img_aug_size=10,\
59
+ clip_grad_norm=5.0,\
60
+ num_heads=4,\
61
+ process_num=16,\
62
+ lrate=1.0,\
63
+ estop_patience=100,\
64
+ warmup_steps=500,\
65
+ epoches=5000,\
66
+ update_cycle=16,\
67
+ gpus=$GPU_LIST,\
68
+ disp_freq=1,\
69
+ eval_freq=10,\
70
+ save_freq=10,\
71
+ sample_freq=10,\
72
+ checkpoints=1,\
73
+ best_checkpoints=1,\
74
+ max_training_steps=30000,\
75
+ nthreads=8,\
76
+ beta1=0.9,\
77
+ beta2=0.998,\
78
+ random_seed=1234,\
79
+ train_continue=True,\
80
+ src_codes="$data/asllrp.bpe",tgt_codes="$data/asllrp.bpe",\
81
+ src_vocab_file="$data/vocab.asllrp",\
82
+ tgt_vocab_file="$data/vocab.asllrp",\
83
+ img_train_file="$feature/train.h5",\
84
+ src_train_file="$data/train.bpe.gloss.shuf",\
85
+ tgt_train_file="$data/train.bpe.gloss.shuf",\
86
+ img_dev_file="$feature/dev.h5",\
87
+ src_dev_file="$data/dev.bpe.gloss",\
88
+ tgt_dev_file="$data/dev.bpe.gloss",\
89
+ img_test_file="$feature/test.h5",\
90
+ src_test_file="$data/test.bpe.gloss",\
91
+ tgt_test_file="$data/test.bpe.gloss",\
92
+ output_dir="checkpoints_asllrp",\
93
+ test_output="",\
94
+ shared_source_target_embedding=True,\
95
+ collect_attention_weights=False) 2>&1 | tee "$LOG_FILE"
plugins/SignX/training_yutubeasl/scripts/1_preprocess.py ADDED
@@ -0,0 +1,303 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ YouTubeASL Dataset Preprocessing for SMKD
3
+ Generates {mode}_info.npy and gloss_dict.npy in the same format as Phoenix/CSL-Daily.
4
+ """
5
+
6
+ import argparse
7
+ import csv
8
+ import os
9
+ import random
10
+ import re
11
+ from collections import Counter
12
+
13
+ import numpy as np
14
+
15
+
16
+ REQUIRED_COLUMNS = {"VIDEO_NAME", "CLIP_INDEX", "TEXT", "CLIP_PATH"}
17
+
18
+
19
+ def _read_manifest(manifest_path, video_root=None):
20
+ samples = []
21
+ skipped = {"missing": 0, "empty": 0, "no_text": 0}
22
+
23
+ with open(manifest_path, "r", encoding="utf-8") as f:
24
+ reader = csv.DictReader(f, delimiter="\t")
25
+ missing = REQUIRED_COLUMNS - set(reader.fieldnames or [])
26
+ if missing:
27
+ raise ValueError(f"Manifest missing required columns: {sorted(missing)}")
28
+
29
+ for row in reader:
30
+ clip_path = (row.get("CLIP_PATH") or "").strip()
31
+ text = (row.get("TEXT") or "").strip()
32
+ if not text:
33
+ skipped["no_text"] += 1
34
+ continue
35
+
36
+ if video_root and clip_path and not os.path.isabs(clip_path):
37
+ clip_path = os.path.join(video_root, clip_path)
38
+
39
+ if not clip_path or not os.path.exists(clip_path):
40
+ skipped["missing"] += 1
41
+ continue
42
+ if os.path.getsize(clip_path) == 0:
43
+ skipped["empty"] += 1
44
+ continue
45
+
46
+ video_name = (row.get("VIDEO_NAME") or "").strip()
47
+ clip_index_raw = (row.get("CLIP_INDEX") or "").strip()
48
+ try:
49
+ clip_index = int(clip_index_raw)
50
+ file_id = f"{video_name}_{clip_index:04d}"
51
+ except ValueError:
52
+ file_id = f"{video_name}_{clip_index_raw}"
53
+
54
+ normalized_text = text.lower()
55
+ normalized_text = re.sub(r"[^\w']+", " ", normalized_text)
56
+ normalized_text = " ".join(normalized_text.split())
57
+ sample = {
58
+ "id": file_id,
59
+ "video_name": video_name,
60
+ "clip_index": clip_index_raw,
61
+ "text": normalized_text,
62
+ "clip_path": clip_path,
63
+ "start": (row.get("START") or "").strip(),
64
+ "end": (row.get("END") or "").strip(),
65
+ }
66
+ samples.append(sample)
67
+
68
+ return samples, skipped
69
+
70
+
71
+ def _split_by_video(samples, seed, val_size, test_size):
72
+ groups = {}
73
+ for sample in samples:
74
+ groups.setdefault(sample["video_name"], []).append(sample)
75
+
76
+ rng = random.Random(seed)
77
+ video_names = list(groups.keys())
78
+ rng.shuffle(video_names)
79
+
80
+ train, val, test = [], [], []
81
+ for video_name in video_names:
82
+ group = groups[video_name]
83
+ if len(test) < test_size:
84
+ test.extend(group)
85
+ continue
86
+ if len(val) < val_size:
87
+ val.extend(group)
88
+ continue
89
+ train.extend(group)
90
+
91
+ return train, val, test
92
+
93
+
94
+ def preprocess_youtubeasl(
95
+ manifest_path,
96
+ output_dir,
97
+ prefix_for_smkd=None,
98
+ seed=1234,
99
+ val_size=300,
100
+ test_size=100,
101
+ video_root=None,
102
+ frames_root=None,
103
+ min_frames=0,
104
+ max_frames=0,
105
+ evaluation_prefix="youtubeasl-groundtruth",
106
+ ):
107
+ print(f"Loading manifest: {manifest_path}")
108
+ samples, skipped = _read_manifest(manifest_path, video_root=video_root)
109
+
110
+ print(f"Total samples: {len(samples)}")
111
+ if any(skipped.values()):
112
+ print(
113
+ "Skipped samples: "
114
+ + ", ".join(f"{k}={v}" for k, v in skipped.items() if v)
115
+ )
116
+
117
+ train_samples, val_samples, test_samples = _split_by_video(
118
+ samples, seed=seed, val_size=val_size, test_size=test_size
119
+ )
120
+
121
+ if len(val_samples) < val_size or len(test_samples) < test_size:
122
+ print(
123
+ "Warning: requested split sizes not reached. "
124
+ f"val={len(val_samples)} (target {val_size}), "
125
+ f"test={len(test_samples)} (target {test_size})."
126
+ )
127
+
128
+ if len(val_samples) > val_size or len(test_samples) > test_size:
129
+ print(
130
+ "Note: split sizes exceeded targets due to grouping by VIDEO_NAME. "
131
+ f"val={len(val_samples)} (target {val_size}), "
132
+ f"test={len(test_samples)} (target {test_size})."
133
+ )
134
+
135
+ splits = {
136
+ "train": train_samples,
137
+ "dev": val_samples,
138
+ "test": test_samples,
139
+ }
140
+
141
+ print(
142
+ f"Split - Train: {len(train_samples)}, Dev: {len(val_samples)}, Test: {len(test_samples)}"
143
+ )
144
+
145
+ os.makedirs(output_dir, exist_ok=True)
146
+
147
+ if frames_root:
148
+ frames_root = os.path.abspath(frames_root)
149
+ prefix_for_smkd = frames_root
150
+
151
+ kept_by_split = {}
152
+ for mode, samples_list in splits.items():
153
+ info_dict = {"prefix": prefix_for_smkd}
154
+ kept = 0
155
+ kept_samples = []
156
+
157
+ for sample in samples_list:
158
+ frame_count = -1
159
+ if frames_root:
160
+ clip_name = os.path.splitext(os.path.basename(sample["clip_path"]))[0]
161
+ folder_path = os.path.join(sample["video_name"], clip_name, "crop_frame", "*.jpg")
162
+ abs_dir = os.path.join(frames_root, sample["video_name"], clip_name, "crop_frame")
163
+ if not os.path.isdir(abs_dir):
164
+ continue
165
+ frame_count = len([p for p in os.listdir(abs_dir) if p.lower().endswith(".jpg")])
166
+ if min_frames and frame_count < min_frames:
167
+ continue
168
+ if max_frames and frame_count > max_frames:
169
+ continue
170
+ else:
171
+ folder_path = sample["clip_path"]
172
+
173
+ info_dict[kept] = {
174
+ "fileid": sample["id"],
175
+ "folder": folder_path,
176
+ "signer": "Unknown",
177
+ "label": sample["text"],
178
+ "num_frames": frame_count,
179
+ "original_info": (sample, kept, sample["id"]),
180
+ }
181
+ kept += 1
182
+ kept_samples.append(sample)
183
+
184
+ info_path = os.path.join(output_dir, f"{mode}_info.npy")
185
+ np.save(info_path, info_dict)
186
+ print(f"Saved {mode}_info.npy with {kept} samples")
187
+
188
+ kept_by_split[mode] = kept_samples
189
+
190
+ gloss_counter = Counter()
191
+ for kept_samples in kept_by_split.values():
192
+ for sample in kept_samples:
193
+ gloss_counter.update(sample["text"].split())
194
+
195
+ gloss_dict = {}
196
+ for idx, (gloss, count) in enumerate(sorted(gloss_counter.items()), start=1):
197
+ gloss_dict[gloss] = [idx, count]
198
+
199
+ gloss_dict_path = os.path.join(output_dir, "gloss_dict.npy")
200
+ np.save(gloss_dict_path, gloss_dict)
201
+ print(f"Saved gloss_dict.npy with {len(gloss_dict)} glosses")
202
+
203
+ for mode, kept_samples in kept_by_split.items():
204
+ stm_path = os.path.join(output_dir, f"{evaluation_prefix}-{mode}.stm")
205
+ with open(stm_path, "w", encoding="utf-8") as f:
206
+ for sample in kept_samples:
207
+ utt_id = sample["id"]
208
+ gloss = sample["text"]
209
+ signer = "Unknown"
210
+ f.write(f"{utt_id} 1 {signer} 0.0 1.79769e+308 {gloss}\n")
211
+ print(f"Saved {stm_path}")
212
+
213
+ print("\nPreprocessing complete!")
214
+ print(f"Output directory: {output_dir}")
215
+ print("Files generated:")
216
+ print(" - gloss_dict.npy")
217
+ print(" - train_info.npy, dev_info.npy, test_info.npy")
218
+ print(
219
+ f" - {evaluation_prefix}-train.stm, {evaluation_prefix}-dev.stm, {evaluation_prefix}-test.stm"
220
+ )
221
+
222
+
223
+ if __name__ == "__main__":
224
+ parser = argparse.ArgumentParser(
225
+ description="Preprocess YouTubeASL dataset for SMKD (video inputs)."
226
+ )
227
+ parser.add_argument(
228
+ "--manifest-path",
229
+ type=str,
230
+ default="../../../YouTubeASL/video1000size/manifest_top1000.tsv",
231
+ help="Path to manifest_top1000.tsv (relative to scripts/)",
232
+ )
233
+ parser.add_argument(
234
+ "--video-root",
235
+ type=str,
236
+ default="../../../YouTubeASL",
237
+ help="Root for relative CLIP_PATH entries (relative to scripts/)",
238
+ )
239
+ parser.add_argument(
240
+ "--prefix-for-smkd",
241
+ type=str,
242
+ default="",
243
+ help="Prefix stored in info.npy; leave empty to use absolute clip paths",
244
+ )
245
+ parser.add_argument(
246
+ "--output-dir",
247
+ type=str,
248
+ default="../../smkd/youtubeasl",
249
+ help="Output directory for preprocessed files",
250
+ )
251
+ parser.add_argument("--seed", type=int, default=1234, help="Random seed")
252
+ parser.add_argument(
253
+ "--val-size",
254
+ type=int,
255
+ default=300,
256
+ help="Validation split size (by clips)",
257
+ )
258
+ parser.add_argument(
259
+ "--test-size",
260
+ type=int,
261
+ default=100,
262
+ help="Test split size (by clips)",
263
+ )
264
+ parser.add_argument(
265
+ "--frames-root",
266
+ type=str,
267
+ default="../../../YouTubeASL/video1000size_frames",
268
+ help="Root directory containing extracted frames (relative to scripts/).",
269
+ )
270
+ parser.add_argument(
271
+ "--min-frames",
272
+ type=int,
273
+ default=5,
274
+ help="Minimum number of frames required when using extracted frames.",
275
+ )
276
+ parser.add_argument(
277
+ "--evaluation-prefix",
278
+ type=str,
279
+ default="youtubeasl-groundtruth",
280
+ help="Prefix for generated STM files",
281
+ )
282
+
283
+ args = parser.parse_args()
284
+
285
+ prefix_for_smkd = args.prefix_for_smkd if args.prefix_for_smkd else None
286
+ video_root = args.video_root if args.video_root else None
287
+ frames_root = args.frames_root if args.frames_root else None
288
+ min_frames = args.min_frames if args.min_frames else 0
289
+ if frames_root and not prefix_for_smkd:
290
+ prefix_for_smkd = frames_root
291
+
292
+ preprocess_youtubeasl(
293
+ manifest_path=args.manifest_path,
294
+ output_dir=args.output_dir,
295
+ prefix_for_smkd=prefix_for_smkd,
296
+ seed=args.seed,
297
+ val_size=args.val_size,
298
+ test_size=args.test_size,
299
+ video_root=video_root,
300
+ frames_root=frames_root,
301
+ min_frames=min_frames,
302
+ evaluation_prefix=args.evaluation_prefix,
303
+ )
plugins/SignX/training_yutubeasl/scripts/2_train_smkd.sh ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ # Train SMKD feature extractor on YouTubeASL dataset
4
+ # Auto-switches to signx-slt environment (PyTorch)
5
+
6
+ set -e
7
+
8
+ echo "========================================="
9
+ echo "Training SMKD Feature Extractor"
10
+ echo "========================================="
11
+
12
+ SCRIPT_DIR="$(cd -- "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
13
+ REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../.." && pwd)"
14
+ SMKD_DIR="${REPO_ROOT}/smkd"
15
+
16
+ # Get conda base and activate signx-slt
17
+ CONDA_BASE=$(conda info --base)
18
+ source "${CONDA_BASE}/etc/profile.d/conda.sh"
19
+
20
+ echo "Switching to signx-slt environment (PyTorch)..."
21
+ conda activate signx-slt
22
+
23
+ # Check PyTorch
24
+ python -c "import torch; print('Using PyTorch', torch.__version__)" || {
25
+ echo ""
26
+ echo "ERROR: PyTorch not found in signx-slt environment!"
27
+ echo "Please ensure signx-slt environment has PyTorch installed."
28
+ exit 1
29
+ }
30
+
31
+ echo ""
32
+ echo "✓ Environment ready"
33
+ CFG_PATH="$REPO_ROOT/training_yutubeasl/configs/youtubeasl_smkd.yaml"
34
+ NUM_EPOCHS=$(python - "$CFG_PATH" <<'PY'
35
+ import re, sys
36
+ path = sys.argv[1]
37
+ try:
38
+ text = open(path, 'r', encoding='utf-8').read()
39
+ except Exception:
40
+ print("unknown")
41
+ sys.exit(0)
42
+ m = re.search(r'^\s*num_epoch\s*:\s*(\d+)\s*$', text, re.M)
43
+ print(m.group(1) if m else "unknown")
44
+ PY
45
+ )
46
+ echo "Training time: ~${NUM_EPOCHS} epochs, several hours"
47
+ echo "GPU: Will use device 0"
48
+ echo ""
49
+
50
+ # Navigate to smkd directory
51
+ cd "${SMKD_DIR}"
52
+
53
+ # Decide which GPUs to use (defaults to all visible GPUs on the node)
54
+ TOTAL_GPUS=$(nvidia-smi -L 2>/dev/null | wc -l | tr -d ' ')
55
+ if [ -z "$TOTAL_GPUS" ] || [ "$TOTAL_GPUS" -lt 1 ]; then
56
+ TOTAL_GPUS=1
57
+ fi
58
+ GPU_LIST=${SMKD_GPU_LIST:-$(seq -s, 0 $((TOTAL_GPUS-1)))}
59
+ export CUDA_VISIBLE_DEVICES=${GPU_LIST}
60
+ VISIBLE_GPUS=$(echo "${GPU_LIST}" | awk -F',' '{print NF}')
61
+ NUM_GPUS=${SMKD_NUM_GPUS:-$VISIBLE_GPUS}
62
+ if [ -z "$NUM_GPUS" ] || [ "$NUM_GPUS" -lt 1 ]; then
63
+ NUM_GPUS=1
64
+ fi
65
+ echo "Using CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}"
66
+ export NCCL_ASYNC_ERROR_HANDLING=1
67
+ export TORCH_NCCL_BLOCKING_WAIT=1
68
+ export NCCL_BLOCKING_WAIT=1
69
+
70
+ COMMON_ARGS="--work-dir work_dir/youtubeasl_smkd --config ../training_yutubeasl/configs/youtubeasl_smkd.yaml --device 0"
71
+
72
+ # Check if resuming from checkpoint
73
+ if [[ -n "${SMKD_RESUME_CHECKPOINT:-}" ]]; then
74
+ echo "Resuming training from checkpoint: ${SMKD_RESUME_CHECKPOINT}"
75
+ COMMON_ARGS="${COMMON_ARGS} --load-checkpoints ${SMKD_RESUME_CHECKPOINT}"
76
+ else
77
+ echo "Training from scratch (no checkpoint)"
78
+ fi
79
+
80
+ if [ "$NUM_GPUS" -gt 1 ]; then
81
+ echo "Launching distributed training on ${NUM_GPUS} GPUs..."
82
+ torchrun --standalone --nproc_per_node=${NUM_GPUS} main.py ${COMMON_ARGS}
83
+ else
84
+ echo "Launching single-GPU training..."
85
+ python main.py ${COMMON_ARGS}
86
+ fi
87
+
88
+ echo ""
89
+ echo "========================================="
90
+ echo "SMKD Training Complete!"
91
+ echo "========================================="
92
+ echo ""
93
+ echo "Checkpoints saved in: smkd/work_dir/youtubeasl_smkd/"
94
+ echo ""
95
+ echo "Next step: bash 3_extract_features.sh"
plugins/SignX/training_yutubeasl/scripts/3_extract_features.sh ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ # Extract SMKD features from trained model
4
+ # Auto-switches to signx-slt environment (PyTorch)
5
+
6
+ set -e
7
+
8
+ echo "========================================="
9
+ echo "Extracting SMKD Features for YouTubeASL"
10
+ echo "========================================="
11
+
12
+ # Get conda base and activate signx-slt
13
+ CONDA_BASE=$(conda info --base)
14
+ source "${CONDA_BASE}/etc/profile.d/conda.sh"
15
+ conda activate signx-slt
16
+
17
+ # Check PyTorch
18
+ python -c "import torch; print('Using PyTorch', torch.__version__)"
19
+
20
+ # Navigate to smkd directory
21
+ cd ../../smkd
22
+
23
+ # Step 1: Checkpoint averaging
24
+ echo ""
25
+ echo "Step 1: Averaging top-5 checkpoints..."
26
+ python ckpt_avg.py \
27
+ --path work_dir/youtubeasl_smkd \
28
+ --checkpoints 5 \
29
+ --output avg
30
+
31
+ # Step 2: Extract sign features
32
+ echo ""
33
+ echo "Step 2: Extracting sign features..."
34
+ python main.py \
35
+ --load-weights avg/average.pt \
36
+ --phase features \
37
+ --device 0 \
38
+ --num-feature-aug 10 \
39
+ --work-dir work_dir/youtubeasl_smkd \
40
+ --config ../training_yutubeasl/configs/youtubeasl_smkd.yaml
41
+
42
+ # Step 3: Combine training features
43
+ echo ""
44
+ echo "Step 3: Combining training features..."
45
+ cd work_dir/youtubeasl_smkd
46
+ python ../../example/sign_feature_cmb.py train*.h5
47
+
48
+ echo ""
49
+ echo "========================================="
50
+ echo "Feature Extraction Complete!"
51
+ echo "========================================="
52
+ echo ""
53
+ echo "Output files:"
54
+ echo " - train.h5, dev.h5, test.h5"
55
+ echo ""
56
+ echo "Next: Prepare text data for SLTUNET (see docs/QUICKSTART.md)"
plugins/SignX/training_yutubeasl/scripts/4_train_sltunet.sh ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ # Train SLTUNET model on YouTubeASL dataset
4
+ # Run this after extract_features.sh completes
5
+
6
+ set -e
7
+
8
+ SCRIPT_DIR=$(cd "$(dirname "$0")" && pwd)
9
+ REPO_ROOT=$(cd "$SCRIPT_DIR/../.." && pwd)
10
+
11
+ # Auto-select a single available GPU (skip GPUs with >20% memory usage)
12
+ echo "========================================="
13
+ echo "Auto-selecting an available GPU..."
14
+ echo "========================================="
15
+ eval $(bash "$SCRIPT_DIR/select_gpus.sh")
16
+
17
+ export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
18
+
19
+ # Paths
20
+ data=preprocessed-youtubeasl/
21
+ feature=smkd/work_dir/youtubeasl_smkd/
22
+
23
+ echo "========================================="
24
+ echo "Training SLTUNET on YouTubeASL Dataset"
25
+ echo "========================================="
26
+
27
+ # First, you need to prepare the text data (glosses)
28
+ # Create preprocessed-youtubeasl directory and prepare BPE tokenization
29
+
30
+ echo ""
31
+ echo "Make sure you have:"
32
+ echo "1. Created $data directory"
33
+ echo "2. Prepared BPE tokenized gloss files"
34
+ echo "3. Extracted SMKD features in $feature"
35
+ echo ""
36
+
37
+ # Train SLTUNET model
38
+ # Create log directory
39
+ LOG_DIR="$REPO_ROOT/checkpoints_youtubeasl/logs"
40
+ mkdir -p "$LOG_DIR"
41
+ TIMESTAMP=$(date +%Y%m%d_%H%M%S)
42
+ LOG_FILE="$LOG_DIR/training_${TIMESTAMP}.log"
43
+
44
+ echo "Training logs will be saved to: $LOG_FILE"
45
+ echo "You can monitor the log with: tail -f $LOG_FILE"
46
+ echo ""
47
+
48
+ (cd "$REPO_ROOT" && python run.py --mode train --parameters=\
49
+ hidden_size=256,embed_size=256,filter_size=4096,\
50
+ sep_layer=0,num_encoder_layer=6,num_decoder_layer=6,\
51
+ ctc_enable=True,ctc_alpha=0.3,ctc_repeated=True,\
52
+ src_bpe_dropout=0.2,tgt_bpe_dropout=0.2,bpe_dropout_stochastic_rate=0.6,\
53
+ initializer="uniform_unit_scaling",initializer_gain=0.5,\
54
+ dropout=0.3,label_smooth=0.1,attention_dropout=0.3,relu_dropout=0.5,residual_dropout=0.4,\
55
+ max_len=256,max_img_len=512,batch_size=80,eval_batch_size=32,\
56
+ token_size=1600,batch_or_token='token',beam_size=8,remove_bpe=True,decode_alpha=1.0,\
57
+ scope_name="transformer",buffer_size=50000,data_leak_ratio=0.1,\
58
+ img_feature_size=1024,img_aug_size=10,\
59
+ clip_grad_norm=0.0,\
60
+ num_heads=4,\
61
+ process_num=2,\
62
+ lrate=1.0,\
63
+ estop_patience=100,\
64
+ warmup_steps=4000,\
65
+ epoches=5000,\
66
+ update_cycle=16,\
67
+ gpus=$GPU_LIST,\
68
+ disp_freq=1,\
69
+ eval_freq=100,\
70
+ save_freq=10,\
71
+ sample_freq=10,\
72
+ checkpoints=1,\
73
+ best_checkpoints=1,\
74
+ max_training_steps=30000,\
75
+ nthreads=8,\
76
+ beta1=0.9,\
77
+ beta2=0.998,\
78
+ random_seed=1234,\
79
+ train_continue=True,\
80
+ src_codes="$data/youtubeasl.bpe",tgt_codes="$data/youtubeasl.bpe",\
81
+ src_vocab_file="$data/vocab.youtubeasl",\
82
+ tgt_vocab_file="$data/vocab.youtubeasl",\
83
+ img_train_file="$feature/train.h5",\
84
+ src_train_file="$data/train.bpe.gloss.shuf",\
85
+ tgt_train_file="$data/train.bpe.gloss.shuf",\
86
+ img_dev_file="$feature/dev.h5",\
87
+ src_dev_file="$data/dev.bpe.gloss",\
88
+ tgt_dev_file="$data/dev.bpe.gloss",\
89
+ img_test_file="$feature/test.h5",\
90
+ src_test_file="$data/test.bpe.gloss",\
91
+ tgt_test_file="$data/test.bpe.gloss",\
92
+ output_dir="checkpoints_youtubeasl",\
93
+ test_output="",\
94
+ shared_source_target_embedding=True,\
95
+ collect_attention_weights=False) 2>&1 | tee "$LOG_FILE"
plugins/SignX/training_yutubeasl/scripts/prepare_youtubeasl_text.py ADDED
@@ -0,0 +1,185 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """
3
+ Prepare YouTubeASL gloss text data for SLTUNET training.
4
+
5
+ Steps:
6
+ 1. Load *_info.npy files produced during preprocessing to collect gloss sequences.
7
+ 2. Learn BPE merge rules on the training glosses.
8
+ 3. Apply BPE to train/dev/test splits.
9
+ 4. Add feature indices, shuffle the training split, and build vocabulary files.
10
+ """
11
+
12
+ import argparse
13
+ import collections
14
+ import io
15
+ import os
16
+ import random
17
+ import sys
18
+ from pathlib import Path
19
+
20
+ import numpy as np
21
+
22
+ # Ensure project root is on sys.path so we can import helper utilities.
23
+ REPO_ROOT = Path(__file__).resolve().parents[2]
24
+ sys.path.append(str(REPO_ROOT))
25
+
26
+ from utils.apply_bpe import BPE # noqa: E402
27
+ from models.vocab import Vocab # noqa: E402
28
+
29
+
30
+ def load_gloss_sequences(info_path):
31
+ """Return gloss strings in dataset order from *_info.npy."""
32
+ info = np.load(info_path, allow_pickle=True).item()
33
+ glosses = []
34
+ ordered_keys = sorted(
35
+ [k for k in info.keys() if k != "prefix"],
36
+ key=lambda x: int(x) if isinstance(x, (int, np.integer)) else int(x)
37
+ )
38
+ for key in ordered_keys:
39
+ entry = info[key]
40
+ gloss = ' '.join(entry.get('label', '').strip().split())
41
+ if not gloss:
42
+ gloss = "<blank>"
43
+ glosses.append(gloss)
44
+ return glosses
45
+
46
+
47
+ def build_initial_vocab(lines):
48
+ """Create initial word->frequency mapping for BPE."""
49
+ vocab = collections.Counter()
50
+ for line in lines:
51
+ for word in line.strip().split():
52
+ symbols = tuple(list(word) + ['</w>'])
53
+ vocab[symbols] += 1
54
+ return vocab
55
+
56
+
57
+ def get_pair_stats(vocab):
58
+ """Count frequency of symbol pairs in the vocabulary."""
59
+ pairs = collections.Counter()
60
+ for word, freq in vocab.items():
61
+ symbols = list(word)
62
+ for i in range(len(symbols) - 1):
63
+ pairs[(symbols[i], symbols[i + 1])] += freq
64
+ return pairs
65
+
66
+
67
+ def merge_vocab(pair, vocab):
68
+ """Merge all occurrences of the provided pair in the vocabulary."""
69
+ new_vocab = {}
70
+ first, second = pair
71
+ pair_str = first + second
72
+ for word, freq in vocab.items():
73
+ symbols = []
74
+ i = 0
75
+ while i < len(word):
76
+ if (i < len(word) - 1) and word[i] == first and word[i + 1] == second:
77
+ symbols.append(pair_str)
78
+ i += 2
79
+ else:
80
+ symbols.append(word[i])
81
+ i += 1
82
+ new_vocab[tuple(symbols)] = freq
83
+ return new_vocab
84
+
85
+
86
+ def learn_bpe_merges(train_lines, num_merges):
87
+ vocab = build_initial_vocab(train_lines)
88
+ merges = []
89
+ for _ in range(num_merges):
90
+ stats = get_pair_stats(vocab)
91
+ if not stats:
92
+ break
93
+ best = stats.most_common(1)[0][0]
94
+ vocab = merge_vocab(best, vocab)
95
+ merges.append(best)
96
+ return merges
97
+
98
+
99
+ def write_bpe_codes(merges, output_path):
100
+ with open(output_path, 'w', encoding='utf-8') as writer:
101
+ writer.write("#version: 0.2\n")
102
+ for left, right in merges:
103
+ writer.write(f"{left} {right}\n")
104
+
105
+
106
+ def apply_bpe_to_lines(lines, codes_path):
107
+ with open(codes_path, 'r', encoding='utf-8') as reader:
108
+ bpe = BPE(reader, merges=-1, separator='@@', vocab=None, glossaries=None)
109
+ return [bpe.process_line(line.strip()) for line in lines]
110
+
111
+
112
+ def write_indexed_file(lines, path, include_index=True):
113
+ """Write lines to file, optionally with line indices."""
114
+ with open(path, 'w', encoding='utf-8') as writer:
115
+ for idx, line in enumerate(lines):
116
+ if include_index:
117
+ writer.write(f"{idx} {line.strip()}\n")
118
+ else:
119
+ writer.write(f"{line.strip()}\n")
120
+
121
+
122
+ def main():
123
+ parser = argparse.ArgumentParser(description="Prepare YouTubeASL gloss text/BPE data.")
124
+ parser.add_argument("--num-merges", type=int, default=1000,
125
+ help="Number of BPE merges to learn (default: 1000)")
126
+ parser.add_argument("--seed", type=int, default=1234, help="Random seed for shuffling.")
127
+ parser.add_argument("--output-dir", type=str, default="preprocessed-youtubeasl",
128
+ help="Output directory for BPE/tokenized files (relative to repo root).")
129
+ parser.add_argument("--info-dir", type=str, default="smkd/youtubeasl",
130
+ help="Directory containing *_info.npy files (relative to repo root).")
131
+ parser.add_argument("--dataset-name", type=str, default="youtubeasl",
132
+ help="Dataset name used for BPE/vocab filenames.")
133
+ args = parser.parse_args()
134
+
135
+ output_dir = REPO_ROOT / args.output_dir
136
+ output_dir.mkdir(parents=True, exist_ok=True)
137
+
138
+ info_dir = REPO_ROOT / args.info_dir
139
+
140
+ splits = {}
141
+ for split in ["train", "dev", "test"]:
142
+ glosses = load_gloss_sequences(info_dir / f"{split}_info.npy")
143
+ splits[split] = glosses
144
+ with open(output_dir / f"{split}.gloss", 'w', encoding='utf-8') as writer:
145
+ writer.write("\n".join(glosses) + "\n")
146
+
147
+ merges = learn_bpe_merges(splits["train"], args.num_merges)
148
+ codes_path = output_dir / f"{args.dataset_name}.bpe"
149
+ write_bpe_codes(merges, codes_path)
150
+
151
+ bpe_lines = {split: apply_bpe_to_lines(glosses, codes_path)
152
+ for split, glosses in splits.items()}
153
+
154
+ # Write BPE files: ALL files WITH indices for uniform data loading
155
+ # Indices will be stripped during BLEU evaluation (not in the data files)
156
+ for split, lines in bpe_lines.items():
157
+ bpe_path = output_dir / f"{split}.bpe.gloss"
158
+ write_indexed_file(lines, bpe_path, include_index=True)
159
+ print(f"Generated {split}.bpe.gloss (with indices)")
160
+
161
+ # Shuffle training split (train.bpe.gloss has indices)
162
+ with open(output_dir / "train.bpe.gloss", encoding='utf-8') as reader:
163
+ train_lines = [line.strip() for line in reader if line.strip()]
164
+ rnd = random.Random(args.seed)
165
+ rnd.shuffle(train_lines)
166
+ with open(output_dir / "train.bpe.gloss.shuf", 'w', encoding='utf-8') as writer:
167
+ writer.write("\n".join(train_lines) + "\n")
168
+ print(f"Generated train.bpe.gloss.shuf (shuffled, with indices)")
169
+
170
+ # Build vocabulary from training tokens (skip the first index column)
171
+ vocab_builder = Vocab()
172
+ with open(output_dir / "train.bpe.gloss", 'r', encoding='utf-8') as reader:
173
+ for line in reader:
174
+ parts = line.strip().split()
175
+ tokens = parts[1:] if len(parts) > 1 else [] # Skip index
176
+ for token in tokens:
177
+ vocab_builder.insert(token)
178
+ vocab_builder.sort_vocab()
179
+ vocab_builder.save_vocab(output_dir / f"vocab.{args.dataset_name}")
180
+
181
+ print("Prepared text data in:", output_dir)
182
+
183
+
184
+ if __name__ == "__main__":
185
+ main()
plugins/SignX/training_yutubeasl/scripts/select_gpus.sh ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+ # Auto-select available GPUs based on memory usage
3
+ # Usage: source select_gpus.sh
4
+ # Returns: CUDA_VISIBLE_DEVICES and GPU_LIST for training
5
+
6
+ MEMORY_THRESHOLD=20 # Skip GPUs with >20% memory usage
7
+ MAX_GPUS=1 # Maximum number of GPUs to use
8
+
9
+ # Get GPU info: index, memory_used, memory_total
10
+ gpu_info=$(nvidia-smi --query-gpu=index,memory.used,memory.total --format=csv,noheader,nounits)
11
+
12
+ available_gpus=()
13
+ gpu_list=()
14
+
15
+ while IFS=, read -r idx mem_used mem_total; do
16
+ # Trim whitespace
17
+ idx=$(echo $idx | xargs)
18
+ mem_used=$(echo $mem_used | xargs)
19
+ mem_total=$(echo $mem_total | xargs)
20
+
21
+ # Calculate memory usage percentage
22
+ mem_percent=$(awk "BEGIN {printf \"%.2f\", ($mem_used/$mem_total)*100}")
23
+
24
+ # Check if GPU is available (memory usage < threshold)
25
+ if (( $(awk "BEGIN {print ($mem_percent < $MEMORY_THRESHOLD)}") )); then
26
+ available_gpus+=($idx)
27
+ gpu_list+=($idx)
28
+ echo "GPU $idx: ${mem_percent}% memory used - AVAILABLE" >&2
29
+ else
30
+ echo "GPU $idx: ${mem_percent}% memory used - BUSY (skipped)" >&2
31
+ fi
32
+
33
+ # Stop if we have enough GPUs
34
+ if [ ${#available_gpus[@]} -ge $MAX_GPUS ]; then
35
+ break
36
+ fi
37
+ done <<< "$gpu_info"
38
+
39
+ # Export CUDA_VISIBLE_DEVICES
40
+ if [ ${#available_gpus[@]} -eq 0 ]; then
41
+ echo "ERROR: No available GPUs found!" >&2
42
+ exit 1
43
+ fi
44
+
45
+ export CUDA_VISIBLE_DEVICES=$(IFS=,; echo "${available_gpus[*]}")
46
+ export GPU_LIST="[$(IFS=,; echo "${gpu_list[*]}")]"
47
+ export NUM_GPUS=${#available_gpus[@]}
48
+
49
+ echo "" >&2
50
+ echo "Selected GPUs: $CUDA_VISIBLE_DEVICES" >&2
51
+ echo "GPU List for training: $GPU_LIST" >&2
52
+ echo "Number of GPUs: $NUM_GPUS" >&2
53
+ echo "" >&2
54
+
55
+ # Print variables for sourcing
56
+ echo "export CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES"
57
+ echo "export GPU_LIST='$GPU_LIST'"
58
+ echo "export NUM_GPUS=$NUM_GPUS"