Hello，
I have a problem that can be reproduced.
When I resume a training run that has been running for several hours,the unity editor has a long frozen in a few minutes(only one time and lasts for almost half an hour).
But python still working.When the editor returns to normal, all behavior got a big reward.Most indicators are abnormal on the Tensorboard.I don’t know where I made a mistake.
Here is my config file:
default:
trainer: sac
batch_size: 128
buffer_size: 50000
buffer_init_steps: 0
hidden_units: 128
init_entcoef: 1.0
learning_rate: 3.0e-4
learning_rate_schedule: constant
max_steps: 5.0e5
memory_size: 128
normalize: false
steps_per_update: 10
num_layers: 2
time_horizon: 64
sequence_length: 64
summary_freq: 5000
tau: 0.005
use_recurrent: false
vis_encode_type: simple
reward_signals:
extrinsic:
strength: 1.0
gamma: 0.99
Cha 1:
init_entcoef: 0.1
reward_signals:
extrinsic:
strength: 1.0
gamma: 0.99
curiosity:
strength: 0.003
gamma: 0.99
encoding_size: 256
Cha 1 c1:
init_entcoef: 0.1
reward_signals:
extrinsic:
strength: 1.0
gamma: 0.99
curiosity:
strength: 0.003
gamma: 0.99
encoding_size: 256
Cha 1 r1:
init_entcoef: 0.1
reward_signals:
extrinsic:
strength: 1.0
gamma: 0.99
curiosity:
strength: 0.003
gamma: 0.99
encoding_size: 256
Cha 1 b1:
init_entcoef: 0.1
reward_signals:
extrinsic:
strength: 1.0
gamma: 0.99
curiosity:
strength: 0.003
gamma: 0.99
encoding_size: 256
Cha 1 d1:
init_entcoef: 0.1
reward_signals:
extrinsic:
strength: 1.0
gamma: 0.99
curiosity:
strength: 0.003
gamma: 0.99
encoding_size: 256
If you haven’t yet, could you please submit a bug report (making sure to include a repro project), so that we can investigate further? Thanks!