caption_model: transformer
noamopt: true
noamopt_warmup: 20000
label_smoothing: 0.0
input_json: data/cocotalk.json
input_label_h5: data/cocotalk_label.h5
input_fc_dir: data/cocotalk_clip_RN50_fc
input_att_dir: data/cocotalk_clip_RN50_att
input_clipscore_vis_dir: data/cocotalk_clipscore_vis
seq_per_img: 5
batch_size: 160
learning_rate: 0.0005

checkpoint_path: save/clipRN50_cider_clips/clipRN50_cider_clips

# Notice: because I'm to lazy, I reuse the option name for RNNs to set the hyperparameters for transformer:
# N=num_layers
# d_model=input_encoding_size
# d_ff=rnn_size

# will be ignored
num_layers: 6
input_encoding_size: 512
rnn_size: 2048

# Transformer config
N_enc: 6
N_dec: 6
d_model: 512
d_ff: 2048
num_att_heads: 8
dropout: 0.1


learning_rate_decay_start: 0
scheduled_sampling_start: -1 
save_checkpoint_every: 3000
language_eval: 1
val_images_use: 5000
max_epochs: 15
train_sample_n: 5

REFORWARD: false

# _BASE_: transformer.yml
reduce_on_plateau: false
noamopt: false
learning_rate: 0.000005
learning_rate_decay_start: -1

self_critical_after: 15
max_epochs: 40

verbose: false
precision: 32

use_clipscore: true
clipscore_reward_weight: 2.0
clipscore_mode: clip_s

use_multi_rewards: true