magma / configs /MAGMA_v2.yml
stellaathena's picture
This should work
bb5cd12
{
# image encoder settings
encoder_name: 'clip_resnet_large',
adapter_config: {"mlp": {"adapter_type": "normal", "downsample_factor": 8}, "attention": {"adapter_type": "normal", "downsample_factor": 8}},
freeze_img_encoder: false,
# train settings
batch_size: 256,
train_steps: 150000,
lr: 8.0e-4,
min_lr: 0.0,
lr_decay_iters: 300000,
image_enc_lr: 2.0e-6,
use_image_embed_layernorm: true,
image_embed_dropout_prob: 0.1,
image_size: 384,
gradient_accumulation_steps: 4,
zero_stage: 2,
gradient_clipping: 1.0,
# dataset / save / load settings
dataset_type: 'new',
train_dataset_dir: ['/mnt/localdisk/laion', '/mnt/brick/CC3M_converted', '/mnt/localdisk/localized_narratives', '/mnt/localdisk/visual_genome_converted', '/mnt/localdisk/hateful_memes_converted', '/mnt/localdisk/coco_converted', '/mnt/brick/wit_converted', '/mnt/localdisk/gqa_train_converted', '/mnt/localdisk/vqa_train_converted', '/mnt/localdisk/okvqa_train_converted'], #'/mnt/brick/wit_converted'
eval_dataset_dir: null, # if this is none, train dataset will be split
vqa_dir: "/mnt/localdisk/vqa_val_converted",
gqa_dir: "/mnt/localdisk/gqa_val_converted",
save: "/mnt/shared_vol/checkpoints/MAGMA_RN50x16",
load: "/mnt/shared_vol/checkpoints/MAGMA_RN50x16",
eval_every: 250,
wandb_project: "MAGMA_training",
name: "MAGMA_RN50x16_v1"
}