git-test

第一步：制作数据集

1. 生成文本描述 python create_ascii_captions.py
   --dataset /data/way/segment_generate/datasets/level_windows_with_pacing.json
   --tileset /data/way/segment_generate/datasets/smb.json
   --output /data/way/segment_generate/datasets/level_windows_with_captions.json
2. 生成tokennizer python tokenizer.py save
   --json_file /data/way/segment_generate/datasets/level_windows_with_captions.json
   --pkl_file /data/way/segment_generate/datasets/level_windows_tokenizer.pkl
3. 随机测试caption python create_random_test_captions.py
   --dataset /data/way/segment_generate/datasets/level_windows_with_captions.json
   --output /data/way/segment_generate/datasets/random_test_captions.json
4. 划分训练集测试集验证集 python split_data.py
   --input /data/way/segment_generate/datasets/level_windows_with_captions.json
   --train /data/way/segment_generate/datasets/level_windows_train.json
   --val /data/way/segment_generate/datasets/level_windows_val.json
   --test /data/way/segment_generate/datasets/level_windows_test.json

第二步：训练文本编码器

1. 训练文本编码器: python train_mlm.py --epochs 300 --save_checkpoints --json /data/way/segment_generate/datasets/level_windows_with_captions.json --val_json /data/way/segment_generate/datasets/level_windows_with_captions-validate.json --test_json /data/way/segment_generate/datasets/level_windows_with_captions-test.json --pkl /data/way/segment_generate/datasets/level_windows_tokenizer.pkl --output_dir /data/way/segment_generate/Mar1and2-MLM-regular0 --seed 0
2. 测试文本编码器训练效果： python evaluate_masked_token_prediction.py --model_path Mar1and2-MLM-regular0 --json datasets\level_windows_with_captions-train.json

第三步：训练文本扩散模型 python train_diffusion.py --save_image_epochs 20 --augment --text_conditional --output_dir Mar1and2-conditional-regular0 --num_epochs 500 --json datasets\level_windows_with_captions-train.json --val_json datasets\level_windows_with_captions-validate.json --pkl datasets\level_windows_tokenizer.pkl --mlm_model_dir Mar1and2-MLM-regular0 --plot_validation_caption_score --seed 0