deviandice/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Update BYOL docs

Browse Source
This commit is contained in:
James Betker 2020-12-24 09:35:03 -07:00
parent 29db7c7a02
commit a947f064cc
2 changed files with 34 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
recipes/byol/README.md
View File

@ -30,8 +30,8 @@ Run the trainer by:
`python train.py -opt train_div2k_byol.yml`
BYOL is data hungry, as most unsupervised training methods are. You'll definitely want to provide
your own dataset - DIV2K is here as an example only.
BYOL is data hungry, as most unsupervised training methods are. If you're providing your own dataset, make sure it is
the hundreds of K-images or more!
## Using your own model

55
recipes/byol/train_div2k_byol.yml
View File

@ -1,55 +1,59 @@
#### general settings
name: train_div2k_byol
name: train_imageset_byol
use_tb_logger: true
model: extensibletrainer
scale: 1
gpu_ids: [0]
fp16: false
start_step: 0
checkpointing_enabled: true # <-- Highly recommended for single-GPU training. Will not work with DDP.
checkpointing_enabled: true # <-- Highly recommended for single-GPU training. May not work in distributed settings.
wandb: false
datasets:
train:
n_workers: 4
batch_size: 32
batch_size: 256 # <-- BYOL trains on very large batch sizes. 256 was the smallest batch size possible before a
# severe drop off in performance. Other parameters here are set to enable this to train on a
# single 10GB GPU.
mode: byol_dataset
crop_size: 256
crop_size: 224
normalize: true
key1: hq
key2: hq
dataset:
mode: imagefolder
paths: /content/div2k # <-- Put your path here. Note: full images.
target_size: 256
paths: /content/imagenet # <-- Put your path here. Directory should be filled with square images.
target_size: 224
scale: 1
skip_lq: true
networks:
generator:
type: generator
which_model_G: byol
image_size: 256
subnet: # <-- Specify your own network to pretrain here.
which_model_G: spinenet
arch: 49
use_input_norm: true
hidden_layer: endpoint_convs.4.conv # <-- Specify a hidden layer from your network here.
subnet:
which_model_G: resnet52 # <-- Specify your own network to pretrain here.
pretrained: false
hidden_layer: avgpool # <-- Specify a hidden layer from your network here.
#### path
path:
#pretrain_model_generator: <insert pretrained model path if desired>
strict_load: true
#resume_state: ../experiments/train_div2k_byol/training_state/0.state # <-- Set this to resume from a previous training state.
#resume_state: ../experiments/train_imageset_byol/training_state/0.state # <-- Set this to resume from a previous training state.
steps:
generator:
training: generator
optimizer: lars
optimizer_params:
# Optimizer params
lr: !!float 3e-4
weight_decay: 0
beta1: 0.9
beta2: 0.99
# All parameters from appendix J of BYOL.
lr: .2 # From BYOL paper: LR=.2*<batch_size>/256
weight_decay: !!float 1.5e-6
lars_coefficient: .001
momentum: .9
injectors:
gen_inj:
@ -67,13 +71,18 @@ steps:
train:
niter: 500000
warmup_iter: -1
mega_batch_factor: 1 # <-- Gradient accumulation factor. If you are running OOM, increase this to [2,4,8].
mega_batch_factor: 4 # <-- Gradient accumulation factor. If you are running OOM, increase this to [8].
# Likewise, if you are running on a 24GB GPU, decrease this to [1] to improve batch stats.
val_freq: 2000
# Default LR scheduler options
default_lr_scheme: MultiStepLR
gen_lr_steps: [50000, 100000, 150000, 200000]
lr_gamma: 0.5
default_lr_scheme: CosineAnnealingLR_Restart
T_period: [120000, 120000, 120000]
warmup: 10000
eta_min: .01 # Unspecified by the paper..
restarts: [140000, 280000] # Paper specifies a different, longer schedule that is not practical for anyone not using
# 4x V100s+. Modify these parameters if you are.
restart_weights: [.5, .25]
eval:
output_state: loss
@ -81,5 +90,5 @@ eval:
logger:
print_freq: 30
save_checkpoint_freq: 1000
visuals: [hq, lq, aug1, aug2]
visuals: [hq, aug1, aug2]
visual_debug_rate: 100