| stage1_params: | |
| name: "VSQ" # always keep this "VectorGPT" | |
| vector_decoder_model: "cnn" # "mlp" or "raster_conv" | |
| quantized_dim: 512 | |
| codebook_size: 4096 # will be ignored for FSQ | |
| image_loss: "pyramid" # "pyramid" or "mse" | |
| single_code_representation: true | |
| vq_method: "fsq" # "vqvae", "FSQ", "vqtorch" | |
| fsq_levels: [7,5,5,5,5] # will determine codebook_size, see Table 1 of FSQ paper - [7,5,5,5,5] for 4096 (e.g. StrokeNUWA), [8,5,5,5] for 1024 | |
| num_segments: 15 | |
| pred_color: true | |
| checkpoint_path: "/sc/projects/sci-aisc/marco.cipriano/results/svg/Grimoire/VSQ/TiledMNIST/checkpoints/last.ckpt" | |
| model_params: | |
| name: "VQ_SVG_Stage2" | |
| max_seq_len: 512 # 512 | |
| dim: 512 # 512 | |
| depth: 16 # 12 | |
| heads: 8 | |
| text_encoder_str: "bert-base-uncased" | |
| use_alibi_positional_bias : False | |
| data_params: | |
| dataset: "mnist" | |
| csv_path: "/sc/projects/sci-aisc/marco.cipriano/data/SVG/Grimoire/MNIST/8x8_randomcolor/split.csv" # root path for CausalSVGDataModule in dataset.py | |
| vq_token_npy_path: "/sc/projects/sci-aisc/marco.cipriano/data/SVG/Grimoire/MNIST/8x8_randomcolor/vsq_tokenized.npy" | |
| train_batch_size: 16 #32 | |
| val_batch_size: 16 # 32 | |
| num_workers: 8 | |
| patch_size: 34 | |
| width: 128 | |
| num_tiles_per_row: 3 | |
| min_context_length: 10 # all samples below this will be removed from the dataset | |
| fraction_of_strokenuwa_inputs: 0.0 | |
| fraction_of_class_only_inputs: 0.9 # fraction of samples that will only have the "class" entry of the dataframe as input | |
| fraction_of_blank_inputs: 0.1 # fraction of samples that will have empty text input | |
| fraction_of_iconshop_chatgpt_inputs: 0.0 | |
| shuffle_vq_order: False # whether to shuffle the order of the VQ tokens, its not really "shuffling", but more cutting the sequence into two parts and switching their order | |
| use_pre_computed_text_tokens_only: False # if True, the text tokens that were pre-computed will be used as input, if False, the text will be tokenized in the dataloader (according to the specified fractions). | |
| exp_params: | |
| lr: 0.0006 # 0.00002 | |
| weight_decay: 1.e-4 # specify positive float to enable, start experimenting with 1.e-4/1.e-3 | |
| scheduler_gamma: 0.96 # 0.95 is a good starting value | |
| train_log_interval: 0.05 # decides how often full generations are made and logged | |
| val_log_interval: 0.02 | |
| metric_log_interval: 0.4 | |
| manual_seed: 1265 | |
| post_process: False # whether to log with svg fixing | |
| trainer_params: | |
| devices: -1 # always keep at -1 as this takes all available GPUs specified through CUDA_VISIBLE_DEVICES | |
| max_epochs: 100 # dsnt matter too much, got early stopping implemented | |
| # accumulate_grad_batches: 2 | |
| logging_params: | |
| entity: "aiis-chair" # comment to use default wandb entity "mfeuer" | |
| project: "grimoire-2" # your wandb project name | |
| save_dir: "/sc/projects/sci-aisc/marco.cipriano/results/svg/Grimoire/ART" | |
| name: "Stage 2 - MNIST color" # name of the run in wandb | |
| version: 0 | |
| author: "Marco" # will be a tag in wandb | |
| id: null # id of wandb run to continue | |
| allow_val_change: False # allow changing values in this config w.r.t. the run that you're continuing (good for changing loss weightings mid-run) |