Upload mnist_color/art/config.yaml with huggingface_hub

mnist_color/art/config.yaml

@@ -0,0 +1,65 @@
+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)