examples/vision/mae/pytorch/benchmarks.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
---
common_options: &common_options
  output:
    - [samples/sec, 'throughput']
    - [loss, 'loss']
  data:
    throughput:
      regexp: 'throughput:?[ \t]*(\d+)'
      skip: 1
    loss:
      reduction_type: 'final'
      regexp: 'loss:?[ \t]*([e\d\.\-\+]+)'
      skip: 1
  env:
    POPLAR_ENGINE_OPTIONS: '{
      "opt.enableMultiAccessCopies":"false",
      "target.hostSyncTimeout":"3000"
    }'
    PYTORCH_EXE_DIR: "/tmp/pt_cache/"

config_options: &config_options
  requirements_path: requirements.txt
  pre_run_commands: [cd remap, make clean, make, cd ..]

pytorch_mae_train_real_pod16:
  <<: [*common_options, *config_options]
  description: |
    Mae training with real data.
  cmd: >-
    python3 main_pretrain.py
      --config vit_base_pod16
      --data_path $DATASETS_DIR/imagenet-raw-dataset
      --epochs 1
      --warmup_epochs 0
      --wandb
      --wandb_run_name pytorch_mae_train_real_pod16

pytorch_mae_train_real_pod64:
  <<: [*common_options, *config_options]
  description: |
    Mae training performance test with real data.
  cmd: >-
    poprun
      -vv
      --host=$HOSTS
      --num-instances=4
      --num-replicas=16
      --ipus-per-replica=4
      --update-partition=yes
      --remove-partition=no
      --vipu-server-host=$VIPU_CLI_API_HOST
      --vipu-partition=$IPUOF_VIPU_API_PARTITION_ID
      --vipu-allocation=$VIPU_ALLOCATION_ID
      --vipu-server-timeout=3600
      --executable-cache-path=$PYTORCH_EXE_DIR
      --mpi-global-args="
        --mca oob_tcp_if_include $TCP_IF_INCLUDE
        --mca btl_tcp_if_include $TCP_IF_INCLUDE"
      --mpi-local-args="
        -x OPAL_PREFIX
        -x LD_LIBRARY_PATH
        -x PATH
        -x PYTHONPATH
        -x CPATH
        -x IPUOF_VIPU_API_TIMEOUT=3600
        -x POPLAR_LOG_LEVEL=WARN
        -x POPLAR_SDK_ENABLED
        -x POPLAR_ENGINE_OPTIONS"
    python3 main_pretrain.py
      --config vit_base_pod64
      --data_path $DATASETS_DIR/imagenet-raw-dataset
      --epochs 3
      --wandb
      --wandb_run_name pytorch_mae_train_real_pod64

pytorch_mae_train_real_pod64_conv:
  <<: [*common_options, *config_options]
  description: |
    Mae training convergence test with real data.
  cmd: >-
    poprun
      -vv
      --host=$HOSTS
      --num-instances=4
      --num-replicas=16
      --ipus-per-replica=4
      --update-partition=yes
      --remove-partition=no
      --vipu-server-host=$VIPU_CLI_API_HOST
      --vipu-partition=$IPUOF_VIPU_API_PARTITION_ID
      --vipu-allocation=$VIPU_ALLOCATION_ID
      --vipu-server-timeout=3600
      --executable-cache-path=$PYTORCH_EXE_DIR
      --mpi-global-args="
        --mca oob_tcp_if_include $TCP_IF_INCLUDE
        --mca btl_tcp_if_include $TCP_IF_INCLUDE"
      --mpi-local-args="
        -x OPAL_PREFIX
        -x LD_LIBRARY_PATH
        -x PATH
        -x PYTHONPATH
        -x CPATH
        -x IPUOF_VIPU_API_TIMEOUT=3600
        -x POPLAR_LOG_LEVEL=WARN
        -x POPLAR_SDK_ENABLED
        -x POPLAR_ENGINE_OPTIONS"
    python3 main_pretrain.py
      --config vit_base_pod64
      --data_path $DATASETS_DIR/imagenet-raw-dataset
      --wandb
      --wandb_run_name pytorch_mae_train_real_pod64_conv

pytorch_mae_finetune_real_pod16_conv:
  <<: [*common_options, *config_options]
  description: |
    Mae finetuning with real data.
  cmd:
    python3 main_finetune.py
      --finetune checkpoint_mae_vit_base_pod64/checkpoint.pth
      --data_path $DATASETS_DIR/imagenet-raw-dataset
      --wandb
      --wandb_run_name pytorch_mae_finetune_real_pod16_conv

pytorch_mae_validation_real_pod16:
  <<: [*common_options, *config_options]
  description: |
    Mae validation after pretraining.
  cmd:
    python3 finetune_validate.py
      --resume checkpoint_mae_vit_base_finetune/checkpoint.pth
      --batch_size 16
      --data_path $DATASETS_DIR/imagenet-raw-dataset
      --wandb
      --wandb_run_name pytorch_mae_validation_real_pod16