readme

Author: dongzhuoyao

.gitignore

@@ -1,2 +1,3 @@
 static/gif/.DS_Store
 .DS_Store
+README copy.md

README.md

@@ -39,6 +39,21 @@ Please cite our paper:
 * **` Feb. 4th, 2025`**: Training code released.
 * **` Dec. 10th, 2024`**: Arxiv released.
 
+## 📦 Training
+
+
+#### COCO training(Deepspeed)
+
+```
+CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch  --num_processes 4 --num_machines 1 --main_process_ip 127.0.0.1 --main_process_port 8868    train_ds_vq.py   model=uvit_s2deep_it data=coco14_cond_indices dynamic=linear dynamic.mask_ce=1  input_tensor_type=bwh tokenizer=sd_vq_f8 optim.wd=0.00 "optim.betas=[0.9, 0.9]" data.train_steps=1_000_000 ckpt_every=20_000 data.sample_fid_every=100_000 data.sample_fid_n=20_000   data.batch_size=64 optim.name=adam optim.lr=2e-4 lrschedule.warmup_steps=5000 dstep_num=500  mixed_precision=bf16 accum=4
+```
+
+#### ImageNet training(accelerator,bs256)
+
+```
+CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch   --num_processes 4 --num_machines 1 --main_process_ip 127.0.0.1 --main_process_port 8868    train_acc_vq.py  model=uvit_h2_it dynamic=linear   input_tensor_type=bwh tokenizer=sd_vq_f8 data=imagenet256_cond_indices data.batch_size=64 data.sample_vis_n=16 data.sample_fid_every=50_000 ckpt_every=20_000 data.train_steps=1500_000  data.sample_fid_n=5_000 optim.name=adamw optim.lr=1e-4 optim.wd=0.0 lrschedule.warmup_steps=1     mixed_precision=bf16 accum=1
+```
+
 
 ## Trend