zhuoyang20 Skhaki commited on
Commit
2735865
1 Parent(s): 95655d2

Add HRNet low-resolution checkpoint (#1)

Browse files

- Add HRNet low-resolution checkpoint (e4fc1c64bb7978bcfabe57558e909ba4cec1f1fe)


Co-authored-by: Samir Khaki <[email protected]>

fcn-hr48-4xb2-512x1024-80k/fcn-hr48-4xb2-512x1024-80k.py ADDED
@@ -0,0 +1,316 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ crop_size = (
2
+ 256,
3
+ 512,
4
+ )
5
+ data_preprocessor = dict(
6
+ bgr_to_rgb=True,
7
+ mean=[
8
+ 123.675,
9
+ 116.28,
10
+ 103.53,
11
+ ],
12
+ pad_val=0,
13
+ seg_pad_val=255,
14
+ size=(
15
+ 256,
16
+ 512,
17
+ ),
18
+ std=[
19
+ 58.395,
20
+ 57.12,
21
+ 57.375,
22
+ ],
23
+ type='SegDataPreProcessor')
24
+ data_root = '/dataset/cityscapes/'
25
+ dataset_type = 'CityscapesDataset'
26
+ default_hooks = dict(
27
+ checkpoint=dict(by_epoch=False, interval=4000, type='CheckpointHook'),
28
+ logger=dict(interval=50, log_metric_by_epoch=False, type='LoggerHook'),
29
+ param_scheduler=dict(type='ParamSchedulerHook'),
30
+ sampler_seed=dict(type='DistSamplerSeedHook'),
31
+ timer=dict(type='IterTimerHook'),
32
+ visualization=dict(type='SegVisualizationHook'))
33
+ default_scope = 'mmseg'
34
+ env_cfg = dict(
35
+ cudnn_benchmark=True,
36
+ dist_cfg=dict(backend='nccl'),
37
+ mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0))
38
+ img_ratios = [
39
+ 0.5,
40
+ 0.75,
41
+ 1.0,
42
+ 1.25,
43
+ 1.5,
44
+ 1.75,
45
+ ]
46
+ launcher = 'pytorch'
47
+ load_from = 'work_dirs/fcn-hr48-4xb2-512x1024-80k/fcn-hr48-4xb2-512x1024-80k_ckpt.pth'
48
+ log_level = 'INFO'
49
+ log_processor = dict(by_epoch=False)
50
+ model = dict(
51
+ backbone=dict(
52
+ extra=dict(
53
+ stage1=dict(
54
+ block='BOTTLENECK',
55
+ num_blocks=(4, ),
56
+ num_branches=1,
57
+ num_channels=(64, ),
58
+ num_modules=1),
59
+ stage2=dict(
60
+ block='BASIC',
61
+ num_blocks=(
62
+ 4,
63
+ 4,
64
+ ),
65
+ num_branches=2,
66
+ num_channels=(
67
+ 48,
68
+ 96,
69
+ ),
70
+ num_modules=1),
71
+ stage3=dict(
72
+ block='BASIC',
73
+ num_blocks=(
74
+ 4,
75
+ 4,
76
+ 4,
77
+ ),
78
+ num_branches=3,
79
+ num_channels=(
80
+ 48,
81
+ 96,
82
+ 192,
83
+ ),
84
+ num_modules=4),
85
+ stage4=dict(
86
+ block='BASIC',
87
+ num_blocks=(
88
+ 4,
89
+ 4,
90
+ 4,
91
+ 4,
92
+ ),
93
+ num_branches=4,
94
+ num_channels=(
95
+ 48,
96
+ 96,
97
+ 192,
98
+ 384,
99
+ ),
100
+ num_modules=3)),
101
+ norm_cfg=dict(requires_grad=True, type='SyncBN'),
102
+ norm_eval=False,
103
+ type='HRNet'),
104
+ data_preprocessor=dict(
105
+ bgr_to_rgb=True,
106
+ mean=[
107
+ 123.675,
108
+ 116.28,
109
+ 103.53,
110
+ ],
111
+ pad_val=0,
112
+ seg_pad_val=255,
113
+ size=(
114
+ 256,
115
+ 512,
116
+ ),
117
+ std=[
118
+ 58.395,
119
+ 57.12,
120
+ 57.375,
121
+ ],
122
+ type='SegDataPreProcessor'),
123
+ decode_head=dict(
124
+ align_corners=False,
125
+ channels=720,
126
+ concat_input=False,
127
+ dropout_ratio=-1,
128
+ in_channels=[
129
+ 48,
130
+ 96,
131
+ 192,
132
+ 384,
133
+ ],
134
+ in_index=(
135
+ 0,
136
+ 1,
137
+ 2,
138
+ 3,
139
+ ),
140
+ input_transform='resize_concat',
141
+ kernel_size=1,
142
+ loss_decode=dict(
143
+ loss_weight=1.0, type='CrossEntropyLoss', use_sigmoid=False),
144
+ norm_cfg=dict(requires_grad=True, type='SyncBN'),
145
+ num_classes=19,
146
+ num_convs=1,
147
+ type='FCNHead'),
148
+ pretrained='open-mmlab://msra/hrnetv2_w48',
149
+ test_cfg=dict(mode='whole'),
150
+ train_cfg=dict(),
151
+ type='EncoderDecoder')
152
+ norm_cfg = dict(requires_grad=True, type='SyncBN')
153
+ optim_wrapper = dict(
154
+ clip_grad=None,
155
+ optimizer=dict(lr=0.01, momentum=0.9, type='SGD', weight_decay=0.0005),
156
+ type='OptimWrapper')
157
+ optimizer = dict(lr=0.01, momentum=0.9, type='SGD', weight_decay=0.0005)
158
+ param_scheduler = [
159
+ dict(
160
+ begin=0,
161
+ by_epoch=False,
162
+ end=80000,
163
+ eta_min=0.0001,
164
+ power=0.9,
165
+ type='PolyLR'),
166
+ ]
167
+ resume = False
168
+ test_cfg = dict(type='TestLoop')
169
+ test_dataloader = dict(
170
+ batch_size=1,
171
+ dataset=dict(
172
+ data_prefix=dict(
173
+ img_path='leftImg8bit/val', seg_map_path='gtFine/val'),
174
+ data_root='/dataset/cityscapes/',
175
+ pipeline=[
176
+ dict(type='LoadImageFromFile'),
177
+ dict(keep_ratio=True, scale=(
178
+ 2048,
179
+ 1024,
180
+ ), type='Resize'),
181
+ dict(type='LoadAnnotations'),
182
+ dict(type='PackSegInputs'),
183
+ ],
184
+ type='CityscapesDataset'),
185
+ num_workers=4,
186
+ persistent_workers=True,
187
+ sampler=dict(shuffle=False, type='DefaultSampler'))
188
+ test_evaluator = dict(
189
+ iou_metrics=[
190
+ 'mIoU',
191
+ ], type='IoUMetric')
192
+ test_pipeline = [
193
+ dict(type='LoadImageFromFile'),
194
+ dict(keep_ratio=True, scale=(
195
+ 2048,
196
+ 1024,
197
+ ), type='Resize'),
198
+ dict(type='LoadAnnotations'),
199
+ dict(type='PackSegInputs'),
200
+ ]
201
+ train_cfg = dict(max_iters=80000, type='IterBasedTrainLoop', val_interval=8000)
202
+ train_dataloader = dict(
203
+ batch_size=2,
204
+ dataset=dict(
205
+ data_prefix=dict(
206
+ img_path='leftImg8bit/train', seg_map_path='gtFine/train'),
207
+ data_root='/dataset/cityscapes/',
208
+ pipeline=[
209
+ dict(type='LoadImageFromFile'),
210
+ dict(type='LoadAnnotations'),
211
+ dict(
212
+ keep_ratio=True,
213
+ ratio_range=(
214
+ 0.5,
215
+ 2.0,
216
+ ),
217
+ scale=(
218
+ 2048,
219
+ 1024,
220
+ ),
221
+ type='RandomResize'),
222
+ dict(
223
+ cat_max_ratio=0.75, crop_size=(
224
+ 256,
225
+ 512,
226
+ ), type='RandomCrop'),
227
+ dict(prob=0.5, type='RandomFlip'),
228
+ dict(type='PhotoMetricDistortion'),
229
+ dict(type='PackSegInputs'),
230
+ ],
231
+ type='CityscapesDataset'),
232
+ num_workers=2,
233
+ persistent_workers=True,
234
+ sampler=dict(shuffle=True, type='InfiniteSampler'))
235
+ train_pipeline = [
236
+ dict(type='LoadImageFromFile'),
237
+ dict(type='LoadAnnotations'),
238
+ dict(
239
+ keep_ratio=True,
240
+ ratio_range=(
241
+ 0.5,
242
+ 2.0,
243
+ ),
244
+ scale=(
245
+ 2048,
246
+ 1024,
247
+ ),
248
+ type='RandomResize'),
249
+ dict(cat_max_ratio=0.75, crop_size=(
250
+ 256,
251
+ 512,
252
+ ), type='RandomCrop'),
253
+ dict(prob=0.5, type='RandomFlip'),
254
+ dict(type='PhotoMetricDistortion'),
255
+ dict(type='PackSegInputs'),
256
+ ]
257
+ tta_model = dict(type='SegTTAModel')
258
+ tta_pipeline = [
259
+ dict(backend_args=None, type='LoadImageFromFile'),
260
+ dict(
261
+ transforms=[
262
+ [
263
+ dict(keep_ratio=True, scale_factor=0.5, type='Resize'),
264
+ dict(keep_ratio=True, scale_factor=0.75, type='Resize'),
265
+ dict(keep_ratio=True, scale_factor=1.0, type='Resize'),
266
+ dict(keep_ratio=True, scale_factor=1.25, type='Resize'),
267
+ dict(keep_ratio=True, scale_factor=1.5, type='Resize'),
268
+ dict(keep_ratio=True, scale_factor=1.75, type='Resize'),
269
+ ],
270
+ [
271
+ dict(direction='horizontal', prob=0.0, type='RandomFlip'),
272
+ dict(direction='horizontal', prob=1.0, type='RandomFlip'),
273
+ ],
274
+ [
275
+ dict(type='LoadAnnotations'),
276
+ ],
277
+ [
278
+ dict(type='PackSegInputs'),
279
+ ],
280
+ ],
281
+ type='TestTimeAug'),
282
+ ]
283
+ val_cfg = dict(type='ValLoop')
284
+ val_dataloader = dict(
285
+ batch_size=1,
286
+ dataset=dict(
287
+ data_prefix=dict(
288
+ img_path='leftImg8bit/val', seg_map_path='gtFine/val'),
289
+ data_root='/dataset/cityscapes/',
290
+ pipeline=[
291
+ dict(type='LoadImageFromFile'),
292
+ dict(keep_ratio=True, scale=(
293
+ 2048,
294
+ 1024,
295
+ ), type='Resize'),
296
+ dict(type='LoadAnnotations'),
297
+ dict(type='PackSegInputs'),
298
+ ],
299
+ type='CityscapesDataset'),
300
+ num_workers=4,
301
+ persistent_workers=True,
302
+ sampler=dict(shuffle=False, type='DefaultSampler'))
303
+ val_evaluator = dict(
304
+ iou_metrics=[
305
+ 'mIoU',
306
+ ], type='IoUMetric')
307
+ vis_backends = [
308
+ dict(type='LocalVisBackend'),
309
+ ]
310
+ visualizer = dict(
311
+ name='visualizer',
312
+ type='SegLocalVisualizer',
313
+ vis_backends=[
314
+ dict(type='LocalVisBackend'),
315
+ ])
316
+ work_dir = './work_dirs/fcn-hr48-4xb2-512x1024-80k'
fcn-hr48-4xb2-512x1024-80k/fcn-hr48-4xb2-512x1024-80k_ckpt.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f016b92b0c7acf787eb7ad44c726f47bb90098b61d551b50a5dbf999f7bf90d1
3
+ size 536540243