对比实验_eta.yaml

# 数据集相关配置
dataset_code: ml-1m # ml-1m beauty sports steam video xlong
min_rating: 0
min_uc: 5
min_sc: 5
split: leave_one_out
seed: 42
bert_max_len: 200 # 50-steam 1000-xlong 200-ml 50-other

# 数据加载器相关配置
train_batch_size: 64 # val/test = train_batch_size*2  32-xlong 64-other
val_batch_size: 128
test_batch_size: 128
num_workers: 0
sliding_window_size: 1
negative_sample_size: 100

# 训练器相关配置
device: cuda
num_epochs: 500
optimizer: AdamW
weight_decay: 0.01
adam_epsilon: 1.e-9
momentum: null
lr: 0.001
max_grad_norm: 5.0
enable_lr_schedule: False
decay_step: 10000
gamma: 0.1
enable_lr_warmup: False
warmup_steps: 100

# evaluation
val_strategy: epoch # 每隔多少个iter验证一次，而不是每一个epoch验证一次
val_iterations: 500 # 500-ml 2000-steam/xlong 1000-other
early_stopping: true
early_stopping_patience: 5
metric_ks:
  - 1
  - 5
  - 10
  - 50
best_metric: Recall@10
use_wandb: True

# 模型相关配置
model_code: ETA # 大写
K: 20 #search数量，din最近K个
hidden_size: 64 # 模型embedding的维度
gsu_embd_hidden_size: 10 # sim：gsu维度 ，eta：gsu中hash的bit数，设置成10比较合适，sdim：hash的bit数，din/twin中没用到
num_heads: 4 # head的个数，sdim是hash桶数
desc: "eta对比试验"