File size: 1,980 Bytes
002bd9b |
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 |
#!/bin/bash
# Check if the wandb name argument is provided
if [ "$#" -ne 1 ]; then
echo "Usage: $0 <wandb_name>"
exit 1
fi
WANDB_NAME="$1"
# Generate timestamp
TIMESTAMP=$(python -c "from datetime import datetime; print(datetime.now().strftime('%Y%m%d_%H%M%S'))")
# Create directories
mkdir -p ./exp/checkpoints/$TIMESTAMP
mkdir -p ./exp/logs/$TIMESTAMP
HYDRA_FULL_ERROR=1 \
accelerate launch --config_file amlt_configs/accelerate_deepspeed_config.local.yaml \
-m src.train \
train_data='[m3d_2d]' eval_data='[m3d_2d]' \
+model=base_sca_multitask_v2 \
training.do_train=True \
training.do_eval=True \
training.do_inference=True \
+data.streaming=False \
training.max_eval_samples=800 \
training.max_steps=100000 \
training.fp16=True \
training.output_dir=./exp/checkpoints/$TIMESTAMP \
training.output_log_dir=./exp/logs/$TIMESTAMP \
model.cache_dir=/mnt/blob/weights/.model.cache/ \
training.save_strategy=steps \
training.save_steps=5000 \
training.save_total_limit=3 \
training.optim=adamw_torch \
training.evaluate_before_train=True \
training.per_device_train_batch_size=1 \
training.evaluation_strategy=steps \
training.eval_steps=5000 \
training.logging_steps=500 \
training.logging_first_step=True \
training.dataloader_num_workers=4 \
training.num_masks_per_sample=16 \
wandb.project=train_script \
wandb.name=$WANDB_NAME \
model.num_caption_tokens=8 \
model.additional_num_hidden_layers=12 \
model.num_task_tokens=6 \
training.lr_scheduler_type=cosine \
model.lm_head_model_name_or_path=StanfordAIMI/RadLLaMA-7b \
# model.lm_head_model_name_or_path=gpt2-large \
training.learning_rate=1e-4 \
training.weight_decay=1e-4 \
training.warmup_steps=200 \
training.warmup_ratio=0.33333333 \
training.compute_metrics=True \
+data_transforms=lsj-1_0-1_0 \
model.sam_model_name_or_path=facebook/sam-vit-base |