Update README.md
Browse files
README.md
CHANGED
|
@@ -38,11 +38,15 @@ model='DiT-g/2' # model type
|
|
| 38 |
task_flag="lora_jade_ema_rank64" # task flag
|
| 39 |
resume=./ckpts/t2i/model/ # resume checkpoint
|
| 40 |
index_file=dataset/index_v2_json/jade.json # index file
|
|
|
|
| 41 |
batch_size=1 # training batch size
|
|
|
|
| 42 |
grad_accu_steps=2 # gradient accumulation steps
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
rank=64 # rank of lora
|
| 44 |
-
max_training_steps=2000 # max training steps
|
| 45 |
-
lr=0.0001 # learning rate
|
| 46 |
|
| 47 |
PYTHONPATH=./ deepspeed hydit/train_large_deepspeed.py \
|
| 48 |
--task-flag ${task_flag} \
|
|
@@ -58,20 +62,19 @@ PYTHONPATH=./ deepspeed hydit/train_large_deepspeed.py \
|
|
| 58 |
--uncond-p 0.44 \
|
| 59 |
--uncond-p-t5 0.44 \
|
| 60 |
--index-file ${index_file} \
|
| 61 |
-
--random-crop \
|
| 62 |
--random-flip \
|
| 63 |
--batch-size ${batch_size} \
|
| 64 |
-
--image-size
|
| 65 |
--global-seed 999 \
|
| 66 |
--grad-accu-steps ${grad_accu_steps} \
|
| 67 |
-
--warmup-num-steps
|
| 68 |
--use-flash-attn \
|
| 69 |
--use-fp16 \
|
| 70 |
--ema-dtype fp32 \
|
| 71 |
-
--results-dir
|
| 72 |
-
--ckpt-every
|
| 73 |
--max-training-steps ${max_training_steps}\
|
| 74 |
-
--ckpt-latest-every
|
| 75 |
--log-every 10 \
|
| 76 |
--deepspeed \
|
| 77 |
--deepspeed-optimizer \
|
|
|
|
| 38 |
task_flag="lora_jade_ema_rank64" # task flag
|
| 39 |
resume=./ckpts/t2i/model/ # resume checkpoint
|
| 40 |
index_file=dataset/index_v2_json/jade.json # index file
|
| 41 |
+
results_dir=./log_EXP # save root for results
|
| 42 |
batch_size=1 # training batch size
|
| 43 |
+
image_size=1024 # training image resolution
|
| 44 |
grad_accu_steps=2 # gradient accumulation steps
|
| 45 |
+
warmup_num_steps=0 # warm-up steps
|
| 46 |
+
lr=0.0001 # learning rate
|
| 47 |
+
ckpt_every=100 # create a ckpt every a few steps.
|
| 48 |
+
ckpt_latest_every=2000 # create a ckpt named `latest.pt` every a few steps.
|
| 49 |
rank=64 # rank of lora
|
|
|
|
|
|
|
| 50 |
|
| 51 |
PYTHONPATH=./ deepspeed hydit/train_large_deepspeed.py \
|
| 52 |
--task-flag ${task_flag} \
|
|
|
|
| 62 |
--uncond-p 0.44 \
|
| 63 |
--uncond-p-t5 0.44 \
|
| 64 |
--index-file ${index_file} \
|
|
|
|
| 65 |
--random-flip \
|
| 66 |
--batch-size ${batch_size} \
|
| 67 |
+
--image-size ${image_size} \
|
| 68 |
--global-seed 999 \
|
| 69 |
--grad-accu-steps ${grad_accu_steps} \
|
| 70 |
+
--warmup-num-steps ${warmup_num_steps} \
|
| 71 |
--use-flash-attn \
|
| 72 |
--use-fp16 \
|
| 73 |
--ema-dtype fp32 \
|
| 74 |
+
--results-dir ${results_dir} \
|
| 75 |
+
--ckpt-every ${ckpt_every} \
|
| 76 |
--max-training-steps ${max_training_steps}\
|
| 77 |
+
--ckpt-latest-every ${ckpt_latest_every} \
|
| 78 |
--log-every 10 \
|
| 79 |
--deepspeed \
|
| 80 |
--deepspeed-optimizer \
|