FBAGSTM's picture
STM32 AI Experimentation Hub
747451d
general:
project_name: ''
output: ''
saved_models_dir: saved_models
display_figures: False
seed: 42
gpu_memory_limit: 3
workers: 4
log_interval: 50
recovery_interval: 0
checkpoint_hist: 10
save_images: False
amp: false
amp_dtype: "float16"
amp_impl: "native"
no_ddp_bb: false
synchronize_step: false
pin_mem: false
no_prefetcher: true
eval_metric: "top1"
tta: 0
local_rank: 0
use_multi_epochs_loader: false
log_wandb: false
log_tb: false
operation_mode: training
# quantization:
# quantizer: Onnx_quantizer
# quantization_type: PTQ
# quantization_input_type: uint8
# quantization_output_type: float
# export_dir: quantized_models
dataset:
dataset_name: "imagenet" # options "flowers102", "food101", "imagenet"
class_names: '' # how is class_names used, torch especially imagenet requires a dict called class_map
classes_file_path: ./datasets/deployment_labels_imagenet.txt
num_classes: 1000 # change according to your dataset
data_dir: 'local/datasets/' # provide root folder which cotnains imagenet folder and this can also be used for quantization as a fall back
#train_split: "train" # folder name under root (for imagenet) Optional for standard imagenet
#val_split: "validation" # folder name under root (for imagenet) Optional for standard imagenet
#test_path: '/local/datasets/ic_imagenet_2012/val/'
#quantization_path: '/local/datasets/ic_imagenet_2012/val/'
#prediction_path: '/local/datasets/ic_imagenet_2012/n01440764/'
preprocessing:
rescaling:
scale: 1/255.0 # TODO scale node is already present under data_augmentation
offset: 0
resizing:
interpolation: nearest # nearest 'Image resize interpolation type (overrides model)'
aspect_ratio: fit
color_mode: rgb
mean: [0.485, 0.456, 0.406] # 'Override mean pixel value of dataset'
std: [0.229, 0.224, 0.225] # 'Override std deviation of dataset'
data_augmentation:
no_aug: False
scale: [0.08, 1.0] # TODO scale node is already present under data_augmentation
ratio: [0.75, 1.33]
horizontal_flip: 0.5
vertical_flip: 0.0
hflip: 0.5
vflip: 0.0
color_jitter: 0.4
aa: null
aug_repeats: 0
aug_splits: 0
jsd_loss: False
bce_loss: False
bce_target_thresh: null
reprob: 0
remode: 'pixel'
recount: 1
resplit: False
mixup: 0.0
cutmix: 0.0
cutmix_minmax: null # Example: [0.3, 0.8]
mixup_prob: 1.0
mixup_switch_prob: 0.5
mixup_mode: "batch"
smoothing: 0.1
train_interpolation: "random"
drop: 0.0
drop_connect: null
drop_path: null
drop_block: null
model:
model_name: 'mobilenetv2_w035_pt'
pretrained: True
pretrained_dataset: "imagenet"
input_shape: [3, 224, 224]
training:
epochs: 2
batch_size: 256
validation_batch_size: null
optimizer:
opt: 'sgd'
opt-eps: null
opt-betas: null
momentum: 0.9
weight_decay: !!float 2e-5
clip_grad: null
clip_mode: 'norm'
layer_decay: null
lr_scheduler:
sched: 'cosine'
sched_on_updates: False
lr: null
lr_base: 0.1
lr_base_size: 256
lr_base_scale: ''
lr_noise: null
lr_noise_pct: 0.67
lr_noise_std: 1.0
lr_cycle_mul: 1.0
lr_cycle_decay: 0.5
lr_cycle_limit: 1
lr_k_decay: 1.0
warmup_lr: !!float 1e-5
min_lr: 0
epoch_repeats: 0
start_epoch: 0
decay_milestones: [90, 180, 270]
decay_epochs: 90
warmup_epochs: 5
warmup_prefix: False
cooldown_epochs: 0
patience_epochs: 10
decay_rate: 0.1
bn_momentum: null
bn_eps: null
sync_bn: false
dist_bn: "reduce"
split_bn: false
#distributed: True
model_ema: false
model_ema_force_cpu: false
model_ema_decay: 0.9998
worker_seeding: all
mlflow:
uri: ./pt/src/experiments_outputs/mlruns
hydra:
run:
dir: ./pt/src/experiments_outputs/${now:%Y_%m_%d_%H_%M_%S}