modurl 0.1.0

Modular reinforcement learning algorithms and utilities built on Candle
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
//! Advantage Actor-Critic implemented as a constrained PPO configuration.
//!
//! A2C uses the PPO training implementation with policy and value clipping
//! disabled, one optimization epoch, and the entire rollout as one minibatch.
//! The public wrappers in this module keep PPO-specific types out of A2C
//! training programs.

use std::fmt::Debug;

use bon::bon;
use candle_core::Tensor;
use candle_nn::Optimizer;

pub use super::ppo::{
    PPOCollectionLogEntry as A2CCollectionLogEntry, PPOEpisodeLogEntry as A2CEpisodeLogEntry,
    PPOLogEntry as A2CLogEntry, PPOLogger as A2CLogger,
};
use super::{
    Agent,
    ppo::{
        FakeOptimizer, PPOAgent, PPOConfigurationError, PPOError, PPONetworkInfo,
        SeparatePPONetwork, SharedPPONetwork,
    },
};
use crate::{
    gym::MultiGym, models::probabilistic_model::ProbabilisticPolicy,
    parameter_schedule::ParameterSchedule, spaces::Space,
};

#[derive(Debug, thiserror::Error)]
pub enum A2CError<AE, GE, SE>
where
    AE: Debug,
    GE: Debug,
    SE: Debug,
{
    #[error("policy failed: {0}")]
    PolicyError(#[source] AE),
    #[error("gym failed: {0}")]
    GymError(#[source] GE),
    #[error("A2C tensor operation failed: {0}")]
    TensorError(#[source] candle_core::Error),
    #[error("space operation failed: {0}")]
    SpaceError(#[source] SE),
    #[error("invalid A2C configuration: {0}")]
    ConfigurationError(#[source] PPOConfigurationError),
    #[error("termination batches differ in length: {dones} dones and {truncateds} truncations")]
    MismatchedTerminationBatch { dones: usize, truncateds: usize },
    #[error("no prepared rollout is available")]
    MissingPreparedRollout,
    #[error("old value estimates are missing")]
    MissingOldValues,
}

impl<AE, GE, SE> From<PPOError<AE, GE, SE>> for A2CError<AE, GE, SE>
where
    AE: Debug,
    GE: Debug,
    SE: Debug,
{
    fn from(error: PPOError<AE, GE, SE>) -> Self {
        match error {
            PPOError::PolicyError(error) => Self::PolicyError(error),
            PPOError::GymError(error) => Self::GymError(error),
            PPOError::TensorError(error) => Self::TensorError(error),
            PPOError::SpaceError(error) => Self::SpaceError(error),
            PPOError::ConfigurationError(error) => Self::ConfigurationError(error),
            PPOError::MismatchedTerminationBatch { dones, truncateds } => {
                Self::MismatchedTerminationBatch { dones, truncateds }
            }
            PPOError::MissingPreparedRollout => Self::MissingPreparedRollout,
            PPOError::MissingOldValues => Self::MissingOldValues,
        }
    }
}

/// Shared actor-critic network configuration for A2C.
pub struct SharedA2CNetwork<O, E> {
    inner: SharedPPONetwork<O, E>,
}

#[bon]
impl<O, E> SharedA2CNetwork<O, E>
where
    O: Optimizer,
    E: Debug,
{
    #[builder]
    pub fn new(
        optimizer: O,
        #[builder(with = |network: impl candle_core::Module + 'static| Box::new(network))]
        shared_network: Box<dyn candle_core::Module>,
        #[builder(with = |network: impl candle_core::Module + 'static| Box::new(network))]
        critic_head: Box<dyn candle_core::Module>,
        #[builder(with = |policy: impl ProbabilisticPolicy<Error = E> + 'static| Box::new(policy))]
        actor_head: Box<dyn ProbabilisticPolicy<Error = E>>,
        #[builder(with = |schedule: impl ParameterSchedule + 'static| Box::new(schedule))]
        lr_scheduler: Option<Box<dyn ParameterSchedule>>,
    ) -> Self {
        Self {
            inner: SharedPPONetwork::from_boxed(
                optimizer,
                shared_network,
                critic_head,
                actor_head,
                lr_scheduler,
            ),
        }
    }
}

/// Separate actor and critic network configuration for A2C.
pub struct SeparateA2CNetwork<O1, O2, E> {
    inner: SeparatePPONetwork<O1, O2, E>,
}

#[bon]
impl<O1, O2, E> SeparateA2CNetwork<O1, O2, E>
where
    O1: Optimizer,
    O2: Optimizer,
    E: Debug,
{
    #[builder]
    pub fn new(
        actor_optimizer: O1,
        critic_optimizer: O2,
        #[builder(with = |policy: impl ProbabilisticPolicy<Error = E> + 'static| Box::new(policy))]
        actor_network: Box<dyn ProbabilisticPolicy<Error = E>>,
        #[builder(with = |network: impl candle_core::Module + 'static| Box::new(network))]
        critic_network: Box<dyn candle_core::Module>,
        #[builder(with = |schedule: impl ParameterSchedule + 'static| Box::new(schedule))]
        actor_lr_scheduler: Option<Box<dyn ParameterSchedule>>,
        #[builder(with = |schedule: impl ParameterSchedule + 'static| Box::new(schedule))]
        critic_lr_scheduler: Option<Box<dyn ParameterSchedule>>,
        #[builder(default = false)] combined_loss: bool,
    ) -> Self {
        Self {
            inner: SeparatePPONetwork::from_boxed(
                actor_optimizer,
                critic_optimizer,
                actor_network,
                critic_network,
                actor_lr_scheduler,
                critic_lr_scheduler,
                combined_loss,
            ),
        }
    }
}

/// An A2C-typed shared or separate network configuration.
pub struct A2CNetworkInfo<O1, E, O2 = FakeOptimizer> {
    inner: PPONetworkInfo<O1, E, O2>,
}

impl<O, E> A2CNetworkInfo<O, E, FakeOptimizer> {
    pub fn shared(network: SharedA2CNetwork<O, E>) -> Self {
        Self {
            inner: PPONetworkInfo::Shared(network.inner),
        }
    }
}

impl<O1, E, O2> A2CNetworkInfo<O1, E, O2> {
    pub fn separate(network: SeparateA2CNetwork<O1, O2, E>) -> Self {
        Self {
            inner: PPONetworkInfo::Separate(network.inner),
        }
    }
}

impl<O, E> From<SharedA2CNetwork<O, E>> for A2CNetworkInfo<O, E, FakeOptimizer> {
    fn from(network: SharedA2CNetwork<O, E>) -> Self {
        Self::shared(network)
    }
}

impl<O1, O2, E> From<SeparateA2CNetwork<O1, O2, E>> for A2CNetworkInfo<O1, E, O2> {
    fn from(network: SeparateA2CNetwork<O1, O2, E>) -> Self {
        Self::separate(network)
    }
}

/// Advantage Actor-Critic agent backed by a constrained PPO agent.
pub struct A2CAgent<'a, O1, O2, AE, GE, SE, I = ()>
where
    O1: Optimizer,
    O2: Optimizer,
    AE: Debug,
    GE: Debug,
    SE: Debug,
{
    inner: PPOAgent<'a, O1, O2, AE, GE, SE, I>,
}

#[bon]
impl<'a, O1, O2, AE, GE, SE, I> A2CAgent<'a, O1, O2, AE, GE, SE, I>
where
    O1: Optimizer,
    O2: Optimizer,
    AE: Debug,
    GE: Debug,
    SE: Debug,
{
    /// Builds A2C as unclipped PPO with one full-rollout optimization epoch.
    ///
    /// `batch_size` should be divisible by the number of vectorized
    /// environments so the collected rollout contains exactly `batch_size`
    /// transitions.
    #[builder]
    pub fn new(
        action_space: Box<dyn Space<Error = SE>>,
        network_info: A2CNetworkInfo<O1, AE, O2>,
        #[builder(default = 0.99)] gamma: f32,
        #[builder(default = 1.0)] gae_lambda: f32,
        #[builder(default = false)] normalize_advantage: bool,
        #[builder(default = false)] normalize_returns: bool,
        #[builder(default = 0.5)] vf_coef: f32,
        #[builder(default = 0.0)] ent_coef: f32,
        #[builder(default = 1024)] batch_size: usize,
        #[builder(default = 0.5)] gradient_clip: f32,
        training_horizon: usize,
        logging_info: Option<&'a mut dyn A2CLogger<I>>,
        device: candle_core::Device,
        #[builder(default = candle_core::DType::F32)] dtype: candle_core::DType,
    ) -> Result<Self, PPOConfigurationError> {
        let inner = PPOAgent::builder()
            .action_space(action_space)
            .network_info(network_info.inner)
            .clipped(false)
            .clip_value_loss(false)
            .gamma(gamma)
            .gae_lambda(gae_lambda)
            .normalize_advantage(normalize_advantage)
            .normalize_returns(normalize_returns)
            .vf_coef(vf_coef)
            .ent_coef(ent_coef)
            .batch_size(batch_size)
            .mini_batch_size(batch_size)
            .num_epochs(1)
            .gradient_clip(gradient_clip)
            .training_horizon(training_horizon)
            .maybe_logging_info(logging_info)
            .device(device)
            .dtype(dtype)
            .build()?;

        Ok(Self { inner })
    }

    pub fn set_learning_rate(&mut self, lr: f64) {
        self.inner.set_learning_rate(lr);
    }

    /// Selects deterministic environment actions for observations shaped
    /// `[batch, ...observation_shape]`.
    pub fn act_deterministic(
        &mut self,
        observation: &Tensor,
    ) -> Result<Tensor, A2CError<AE, GE, SE>> {
        self.inner
            .act_deterministic(observation)
            .map_err(A2CError::from)
    }

    pub fn reset_current_states(&mut self) {
        self.inner.reset_current_states();
    }
}

impl<'a, O1, O2, AE, GE, SE, I> Agent<I> for A2CAgent<'a, O1, O2, AE, GE, SE, I>
where
    O1: Optimizer,
    O2: Optimizer,
    AE: Debug,
    GE: Debug,
    SE: Debug,
{
    type Error = A2CError<AE, GE, SE>;
    type GymError = GE;
    type SpaceError = SE;

    /// Selects stochastic environment actions for observations shaped
    /// `[batch, ...observation_shape]`.
    fn act(&mut self, observation: &Tensor) -> Result<Tensor, Self::Error> {
        self.inner.act(observation).map_err(A2CError::from)
    }

    #[cfg_attr(
        feature = "tracing",
        tracing::instrument(
            name = "a2c.learn",
            target = "modurl::performance",
            skip_all,
            fields(num_timesteps)
        )
    )]
    fn learn(
        &mut self,
        env: &mut dyn MultiGym<I, Error = GE, SpaceError = SE>,
        num_timesteps: usize,
    ) -> Result<(), Self::Error> {
        self.inner.learn(env, num_timesteps).map_err(A2CError::from)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::{
        agents::test_support::{CountingOptimizer, FixedEnv},
        distributions::{CategoricalDistribution, CategoricalDistributionError},
        gym::VectorizedGymWrapper,
        models::{
            MLP,
            probabilistic_model::{ProbabilisticPolicyModel, ProbabilisticPolicyModelError},
        },
        spaces::Discrete,
    };
    use candle_core::{DType, Device};
    use candle_nn::{VarBuilder, VarMap};

    #[derive(Default)]
    struct RecordingLogger {
        updates: usize,
        collections: usize,
        completed_episodes: usize,
        advantages: Vec<Vec<f32>>,
    }

    impl A2CLogger for RecordingLogger {
        fn log(&mut self, entry: &A2CLogEntry) {
            self.updates += 1;
            assert_eq!(entry.epoch, 0);
            assert_eq!(entry.advantages.dims(), &[2]);
            self.advantages
                .push(entry.advantages.to_vec1::<f32>().unwrap());
        }

        fn log_collection(&mut self, entry: &A2CCollectionLogEntry) {
            self.collections += 1;
            self.completed_episodes += entry.completed_episodes.len();
        }
    }

    fn separate_network(
        device: &Device,
    ) -> A2CNetworkInfo<
        CountingOptimizer,
        ProbabilisticPolicyModelError<CategoricalDistributionError>,
        CountingOptimizer,
    > {
        let actor_vars = VarMap::new();
        let critic_vars = VarMap::new();
        let actor = MLP::builder()
            .input_size(4)
            .output_size(2)
            .vb(VarBuilder::from_varmap(&actor_vars, DType::F32, device))
            .hidden_layer_sizes(vec![2])
            .build()
            .unwrap();
        let critic = MLP::builder()
            .input_size(4)
            .output_size(1)
            .vb(VarBuilder::from_varmap(&critic_vars, DType::F32, device))
            .hidden_layer_sizes(vec![2])
            .build()
            .unwrap();

        SeparateA2CNetwork::builder()
            .actor_optimizer(CountingOptimizer::with_learning_rate(1e-3))
            .critic_optimizer(CountingOptimizer::with_learning_rate(1e-3))
            .actor_network(ProbabilisticPolicyModel::<CategoricalDistribution>::new(
                actor,
            ))
            .critic_network(critic)
            .build()
            .into()
    }

    #[test]
    fn public_agent_runs_one_full_rollout_update_and_forwards_actions() {
        let device = Device::Cpu;
        let mut env: VectorizedGymWrapper<FixedEnv> =
            vec![FixedEnv::new(device.clone()), FixedEnv::new(device.clone())].into();
        let mut logger = RecordingLogger::default();
        let mut agent = A2CAgent::builder()
            .action_space(Box::new(Discrete::new(2)))
            .network_info(separate_network(&device))
            .batch_size(2)
            .training_horizon(4)
            .logging_info(&mut logger)
            .device(device.clone())
            .build()
            .unwrap();

        agent.learn(&mut env, 2).unwrap();
        let observations = Tensor::zeros((2, 4), DType::F32, &device).unwrap();
        assert_eq!(agent.act(&observations).unwrap().dims(), &[2]);
        assert_eq!(agent.act_deterministic(&observations).unwrap().dims(), &[2]);
        agent.set_learning_rate(0.25);
        agent.reset_current_states();

        assert_eq!(logger.updates, 1);
        assert_eq!(logger.collections, 1);
        assert_eq!(logger.completed_episodes, 0);
        assert!(
            logger.advantages[0]
                .iter()
                .any(|advantage| advantage.abs() > 1e-6),
            "A2C should not normalize advantages by default"
        );
    }

    #[test]
    fn network_configuration_is_a2c_typed_until_it_is_wrapped() {
        let network = separate_network(&Device::Cpu);
        assert!(matches!(network.inner, PPONetworkInfo::Separate(_)));
    }
}