-
Notifications
You must be signed in to change notification settings - Fork 31
/
Copy pathmibb.yaml
155 lines (147 loc) · 3.59 KB
/
mibb.yaml
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
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
rollout_type: discrete
## ---- Component search_space ----
# ---- Type cnn ----
search_space_type: cnn
search_space_cfg:
# Schedulable attributes:
num_cell_groups: 2
num_init_nodes: 2
num_layers: 8
cell_layout: null
reduce_cell_groups:
- 1
num_steps: 4
num_node_inputs: 2
shared_primitives:
- none
- max_pool_3x3
- avg_pool_3x3
- skip_connect
- sep_conv_3x3
- sep_conv_5x5
- dil_conv_3x3
- dil_conv_5x5
# - relu_conv_bn_3x3
cell_shared_primitives: null
# ---- End Type cnn ----
## ---- End Component search_space ----
## ---- Component dataset ----
# ---- Type cifar10 ----
dataset_type: cifar10
dataset_cfg:
# Schedulable attributes:
cutout: null
# ---- End Type cifar10 ----
## ---- End Component dataset ----
## ---- Component controller ----
# ---- Type rl ----
controller_type: rl
controller_cfg:
# Schedulable attributes:
independent_cell_group: true
mode: eval
controller_network_type: embed_lstm
controller_network_cfg:
# Schedulable attributes: softmax_temperature
attention_hid: 100
controller_hid: 100
num_lstm_layers: 1
op_tanh_reduce: null
schedule_cfg: null
softmax_temperature: 5.0
tanh_constant: null
rl_agent_type: pg
rl_agent_cfg:
# Schedulable attributes:
alpha: 0.99
batch_update: false
entropy_coeff: 0.01
gamma: 1.0
max_grad_norm: null
# ---- End Type rl ----
## ---- End Component controller ----
## ---- Component evaluator ----
# ---- Type mepa ----
evaluator_type: mepa
evaluator_cfg:
# Schedulable attributes: controller_surrogate_steps, mepa_surrogate_steps, mepa_samples
rollout_type: discrete
batch_size: 64
controller_surrogate_steps: 0
# derive_surrogate_steps: 0
mepa_surrogate_steps: 0
mepa_optimizer:
lr: 0.05
momentum: 0.9
type: SGD
weight_decay: 0.0001
mepa_scheduler:
eta_min: 0.0
T_max: 100
type: CosineAnnealingLR
surrogate_optimizer: null
surrogate_scheduler: null
mepa_samples: 1
data_portion:
- 0.0
- 0.8 # train evaluator weights
- 0.2 # for train controller
mepa_as_surrogate: false
bptt_steps: 35
schedule_cfg: null
# ---- End Type mepa ----
## ---- End Component evaluator ----
## ---- Component weights_manager ----
# ---- Type supernet ----
weights_manager_type: supernet
weights_manager_cfg:
# Schedulable attributes:
rollout_type: discrete
num_classes: 10
init_channels: 20
stem_multiplier: 3
max_grad_norm: 5.0
dropout_rate: 0.1
candidate_member_mask: true
candidate_cache_named_members: true
candidate_virtual_parameter_only: true
# ---- End Type supernet ----
## ---- End Component weights_manager ----
## ---- Component objective ----
# ---- Type fault_injection ----
objective_type: fault_injection
objective_cfg:
# Schedulable attributes:
inject_prob: 0.0001
inject_propto_flops: true
fault_reward_coeff: 0.5
fault_loss_coeff: 0.0
latency_reward_coeff: 0
as_evaluator_regularization: true
fault_modes: "fixed"
max_value_mode: false
activation_fixed_bitwidth: 8
# ---- End Type fault_injection ----
## ---- End Component objective ----
## ---- Component trainer ----
# ---- Type simple ----
trainer_type: simple
trainer_cfg:
# Schedulable attributes: controller_samples, derive_samples
rollout_type: discrete
epochs: 100
test_every: 10
controller_optimizer:
lr: 0.001
type: Adam
controller_scheduler: null
controller_samples: 1
derive_samples: 8
evaluator_steps: null
controller_steps: null
controller_train_every: 1
controller_train_begin: 1
interleave_controller_every: null
schedule_cfg: null
# ---- End Type simple ----
## ---- End Component trainer ----