examples/multimodal/mini_dalle/pytorch/configs.yml at master · graphcore/examples · GitHub

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
155
156
157
158
159
160
161
162
#----------------------------------------------------------------------------------
defaults: &defaults
  dataloader_workers: 4
  loss_img_weight: 7
  text_seq_len: 80
  truncate_captions: True
  lr_scheduler: "ReduceLROnPlateau"
  checkpoint_output_dir: "./output/ckpt"
  wandb: False
#----------------------------------------------------------------------------------

#----------------------------------------------------------------------------------
L16_CLIP_vocab:
  <<: *defaults

  # Execution
  batch_size: 1
  epochs: 200
  device_iterations: 1
  replication_factor: 1
  gradient_accumulation: 15
  stochastic_rounding: True
  embedding_serialization_factor: 8
  enable_half_partials: True
  ipus_per_replica: 4
  layers_per_ipu: [0,7,7,2]
  matmul_proportion: 0.2
  fp16: True

  # Optimizer
  optimizer: "Adam"
  learning_rate: 3e-4
  enable_half_first_order_momentum: True
  loss_scaling: 16384

  # Model
  hidden_size: 512
  num_hidden_layers: 16
  num_attention_heads: 16
  dim_head: 64
  ff_dropout: 0.0
  attn_dropout: 0.0
  sandwich_norm: True
  attn_types: "axial_row,axial_row,axial_col,axial_row,axial_row,axial_row,axial_col,axial_row,axial_row,axial_row,axial_col,full,axial_row,axial_row,axial_col,full"
  checkpoint_save_steps: 5000

  # Dataset
  input_folder: "./data/COCO"

  # Misc
  wandb_project_name: "miniDALL-E_CLIP_vocab"
#----------------------------------------------------------------------------------

#----------------------------------------------------------------------------------
L16: &L16
  <<: *defaults

  # Execution
  batch_size: 2
  epochs: 800
  device_iterations: 1
  replication_factor: 1
  gradient_accumulation: 8192
  stochastic_rounding: True
  embedding_serialization_factor: 4
  enable_half_partials: True
  ipus_per_replica: 4
  layers_per_ipu: [0,6,6,4]
  fp16: True

  # Optimizer
  optimizer: "Adam"
  learning_rate: 5e-3
  lr_scheduler: "multi_step"
  enable_half_first_order_momentum: True
  loss_scaling: 32768

  # Model
  hidden_size: 512
  num_hidden_layers: 16
  num_attention_heads: 16
  dim_head: 64
  ff_dropout: 0.0
  attn_dropout: 0.0
  attn_types: "axial_row,axial_row,axial_col,axial_row,axial_row,axial_row,axial_col,axial_row,axial_row,axial_row,axial_col,full,axial_row,axial_row,axial_col,full"
  bpe_path: ./models/bpe/bpe_yttm_vocab.txt
  checkpoint_save_steps: 5000

  # Dataset
  input_folder: "./data/COCO"
  byteio: False

  # Misc
  wandb_project_name: "miniDALL-E"
#----------------------------------------------------------------------------------

#----------------------------------------------------------------------------------
L16_POD16:
  <<: *L16

  # Execution
  epochs: 800
  replication_factor: 4
  gradient_accumulation: 2048
  enable_rts: True

  # Optimizer
  learning_rate: 5e-3
  lr_scheduler: "multi_step"
  loss_scaling: 32768

  # Misc
  dataloader_workers: 64
  wandb_project_name: "miniDALL-E_POD16"
#----------------------------------------------------------------------------------

#----------------------------------------------------------------------------------
L16_POD64:
  <<: *L16

  # Execution
  epochs: 800
  replication_factor: 16
  gradient_accumulation: 512
  enable_rts: True

  # Optimizer
  learning_rate: 5e-3
  lr_scheduler: "multi_step"
  loss_scaling: 16384

  # Misc
  dataloader_workers: 64
  wandb_project_name: "miniDALL-E_POD64"
#----------------------------------------------------------------------------------

#----------------------------------------------------------------------------------
unit_test:
  <<: *defaults

  # Execution
  batch_size: 1
  epochs: 1
  device_iterations: 1
  replication_factor: 1
  gradient_accumulation: 2

  # Model
  hidden_size: 64
  num_hidden_layers: 1
  num_attention_heads: 1
  dim_head: 64
  ff_dropout: 0.0
  attn_dropout: 0.0
  enable_half_partials: False
  attn_types: "axial_row"
  bpe_path: ./models/bpe/bpe_yttm_vocab.txt

  # Optimizer
  learning_rate: 3e-4
  loss_scaling: 1
#----------------------------------------------------------------------------------