SLM Lab
Search…
v4.2.3
🖥Setup
🚀Using SLM Lab
📈Analyzing Results
🥇Benchmark Results
📖Publications and Talks
Meta Spec: High Level Specifications
📂The Meta Spec
📂
In this tutorial we look at how to adjust the checkpointing frequency.
The meta spec is used to specify higher level configuration that don't fit within specific trial or session, such as how the experiments and trials should behave. It is specified using the meta key in a spec file with the following format:
1
{
2
"{spec_name}": {
3
"agent": [{...}],
4
"env": [{...}],
5
"meta": {
6
// Parameter for Trial.run(), to enable network sharing in memory among Sessions.
7
// - false: disable Hogwild!
8
// - "shared": enable Hogwild!, share parameters all the time
9
// - "synced": enabled Hogwild!, sync parameters only after training step
10
"distributed": str|bool,
11
12
// Whether to use rigorous eval.
13
// Note that this will be slower since it spawns vector environments separately to run evaluations, but it is more rigorous.
14
// This is useful when the agent and environment respond differently when in LAB_MODE=eval
15
// However, for all SLM-Lab-supported environments, this is optional since it can infer the total eval-rewards from training environments (see env.wrapper.TrackReward)
16
// - int: If > 0, spawn a number of vector environments to run eval separately
17
// - int|null: If 0 or null, infer eval scores from training checkpoints
18
// - defaults to null for performance
19
"rigorous_eval": int|null,
20
21
// Frequency to checkpoint on evaluation environment: logging and saving.
22
// This uses env.clock, which counts the total timesteps across vector environments by summing them up.
23
"eval_frequency": int,
24
25
// Frequency to checkpoint on training environment: logging and saving.
26
// This uses env.clock, which counts the total timesteps across vector environments by summing them up.
27
"log_frequency": int,
28
29
// The maximum number of Sessions to spawn per Trial, indexed 0 to (max_session - 1)
30
"max_session": int,
31
32
// The maximum number of Trials to spawn per Experiment, indexed 0 to (max_session - 1)
33
"max_trial": int,
34
35
// The number of CPUs to allocate for Ray.tune per Trial.
36
// If null, the trial is given {max_session} CPUs
37
"num_cpus": int|null,
38
39
// The number of GPUs to allocate for Ray.tune per Trial
40
// If null, the trial is given {max_session} GPUs
41
"num_gpus": int|null,
42
}
43
}
44
}
Copied!
We have already encountered some of the meta spec hyperparameters. Hogwild can be enabled using "distributed". We can also adjust the checkpoint mode and frequency. The max session and trials (when running an experiment) are specified here. Finally, although an experiment will automatically assign resources for each trial, advanced users can also configure the resource allocation for Ray.tune through "num_cpus" and "num_gpus".
✍ Meta Spec for Fast Checkpointing in Atari
✍
Atari games usually have multiple lives per episode. During training, we split it up and treat each life as an episode to encourage the agent to appreciate all its lives. This split also implies that total rewards tracked during training need to be summed up over the lives to yield the true episodic rewards for evaluation.
One method for evaluation is to spawn a new vector environments, say using "rigorous_eval": 8 and run them in evaluation mode so they do not split up an episode. However, this can be quite slow and expensive to run.
Thankfully, SLM Lab has a preprocessor which tracks the true episodic rewards for Atari environments at env.wrapper.TrackReward. This means we can use the true episodic rewards from the training environment and run a trial much faster. To do so, disable the rigorous eval, and set the evaluation frequency to a desired value, as shown below.
1
{
2
"agent_pong": {
3
"agent": [{...}],
4
"env": [{
5
"name": "PongNoFrameskip-v4",
6
"frame_op": "concat",
7
"frame_op_len": 4,
8
"reward_scale": "sign",
9
"num_envs": 16,
10
"max_t": null,
11
"max_frame": 1e7
12
}],
13
...
14
"meta": {
15
"distributed": false,
16
"rigorous_eval": 0,
17
"eval_frequency": 10000,
18
"log_frequency": 5000,
19
"max_session": 4,
20
"max_trial": 1
21
}
22
}
23
}
Copied!
Note that "eval_frequency" (evaluation checkpointing) and "log_frequency" (training checkpointing) are independent. Since evaluation needs multiple lives to tally the episodic rewards, it is usually set higher than "log_frequency".
Copy link