1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
|
experiment_group: Sample Experiments
experiments:
- train_args:
batch_size: 64
max_epochs: 32
dataset:
type: EmnistLinesDataset
args:
subsample_fraction: 0.33
max_length: 34
min_overlap: 0
max_overlap: 0.33
num_samples: 10000
seed: 4711
blank: true
train_args:
num_workers: 6
train_fraction: 0.85
model: LineCTCModel
metrics: [cer, wer]
network:
type: LineRecurrentNetwork
args:
# encoder: ResidualNetworkEncoder
# encoder_args:
# in_channels: 1
# num_classes: 81
# depths: [2, 2]
# block_sizes: [64, 128]
# activation: SELU
# stn: false
encoder: WideResidualNetwork
encoder_args:
in_channels: 1
num_classes: 81
depth: 16
num_layers: 4
width_factor: 2
dropout_rate: 0.2
activation: selu
use_decoder: false
flatten: true
input_size: 256
hidden_size: 128
num_layers: 2
num_classes: 81
patch_size: [28, 14]
stride: [1, 5]
criterion:
type: CTCLoss
args:
blank: 80
optimizer:
type: AdamW
args:
lr: 1.e-02
betas: [0.9, 0.999]
eps: 1.e-08
weight_decay: 5.e-4
amsgrad: false
# lr_scheduler:
# type: OneCycleLR
# args:
# max_lr: 1.e-03
# epochs: null
# anneal_strategy: linear
lr_scheduler:
type: CosineAnnealingLR
args:
T_max: null
swa_args:
start: 4
lr: 5.e-2
callbacks: [Checkpoint, ProgressBar, WandbCallback, WandbImageLogger, SWA] # EarlyStopping, OneCycleLR]
callback_args:
Checkpoint:
monitor: val_loss
mode: min
ProgressBar:
epochs: null
log_batch_frequency: 100
# EarlyStopping:
# monitor: val_loss
# min_delta: 0.0
# patience: 5
# mode: min
WandbCallback:
log_batch_frequency: 10
WandbImageLogger:
num_examples: 6
# OneCycleLR:
# null
SWA:
null
verbosity: 1 # 0, 1, 2
resume_experiment: null
test: true
test_metric: test_cer
|
