Latent var only experiments

[zyaocoder]

Thanks for sharing the great work!

I was wondering how to try the Latent var (LV) only in Table 3: Decoder ablations of PGP paper. Is there any configuration?

Thanks in advance, Zhen

[nachiket92]

Sure, you can change the aggregator from PGP to Global Attention. You'll need to modify the training config file as follows:

Replace aggregator parameters with

# Aggregator parameters
aggregator_type: 'global_attention'
aggregator_args:
  target_agent_enc_size: 32
  context_enc_size: 32
  emb_size: 128
  num_heads: 32

Remove the behavior cloning loss from the config

losses: ['min_ade_k']
loss_weights: [1]
loss_args:
  - k: 10

Change the aggregation type to 'combined' in the decoder parameters

decoder_type: 'lvm'
decoder_args:
  num_samples: 1000
  op_len: 12
  hidden_size: 128
  encoding_size: 160
  agg_type: 'combined'
  lv_dim: 5
  num_clusters: 10

[zyaocoder]

• Thank you for your quick reply!