AlexeyGHT/decoder_start

args = Namespace(

output_dir='decoder_lora',

gradient_accumulation_steps=1,

mixed_precision='bf16',

seed=3434,

pretrained_prior_model_name_or_path="kandinsky-community/kandinsky-2-2-prior",

pretrained_decoder_model_name_or_path="kandinsky-community/kandinsky-2-2-decoder",

rank=32,

lr=1e-4,

use_8bit_adam=True,

lr_scheduler='constant',

lr_warmup_steps=0,

num_train_epochs=5,

max_train_steps=None,

checkpointing_steps=438,

train_batch_size=16,

resume_from_checkpoint=None,

max_grad_norm=1.0,

size=268

)

t=1,5h