# =============================================================================
# configs/base.yaml — single canonical config for training and inference.
# -----------------------------------------------------------------------------
# Every value here mirrors the IEEE notebook (cell 6 hyperparams + cell 21
# layer wiring) so behaviour is identical to the published research. Override
# any field on the CLI or via env var (CAPTIONING__TRAIN__BATCH_SIZE=32) — see
# src/captioning/config/schema.py for the full validated schema.
# =============================================================================

data:
  # Local path; scripts/prepare_data.py downloads COCO into this directory.
  base_path: data/coco2017
  annotations_filename: captions_train2017.json
  images_subdir: train2017
  sample_size: 120000           # Notebook: captions.sample(120000)
  train_val_split: 0.8          # Notebook cell 11: int(len(img_keys) * 0.8)

model:
  embedding_dim: 512            # Notebook: EMBEDDING_DIM = 512
  units: 512                    # Notebook: UNITS = 512
  max_length: 40                # Notebook: MAX_LENGTH = 40
  vocabulary_size: 15000        # Notebook: VOCABULARY_SIZE = 15000
  encoder_num_heads: 1          # Notebook cell 21: TransformerEncoderLayer(EMBEDDING_DIM, 1)
  decoder_num_heads: 8          # Notebook cell 21: TransformerDecoderLayer(..., 8)
  decoder_dropout_inner: 0.3    # Notebook cell 19: dropout_1 = Dropout(0.3)
  decoder_dropout_outer: 0.5    # Notebook cell 19: dropout_2 = Dropout(0.5)
  decoder_attention_dropout: 0.1  # Notebook cell 19: MultiHeadAttention(dropout=0.1)

train:
  epochs: 10                    # Notebook: EPOCHS = 10
  batch_size: 64                # Notebook: BATCH_SIZE = 64
  buffer_size: 1000             # Notebook: BUFFER_SIZE = 1000
  early_stopping_patience: 3    # Notebook cell 22: EarlyStopping(patience=3, ...)
  seed: 42                      # NEW: pin RNGs (notebook didn't seed; results varied)
  learning_rate: 0.001          # Keras Adam default — what the notebook uses implicitly
  weights_filename: model.h5    # Notebook cell 30: caption_model.save_weights('model.h5')

serve:
  max_upload_bytes: 10485760    # 10 MB — guard at the API edge
  decode_strategy: greedy       # Phase 1b: "beam"
  beam_width: 3
  cors_allowed_origins:
    - http://localhost:3000
    - http://localhost:5173
    - http://localhost:5174
    - http://127.0.0.1:5173
    - http://127.0.0.1:5174