Repository navigation
Overview
Full Changelog: v5.5.1...v5.5.2
Fixed
- Modifies the workflows and docs to support the new extras branch by @ClemensSchwarke in #242
- Keeps discounted reward history independent of caller storage by @Afloat16 in #239
- Resumes RND exploration schedules from PPO checkpoints by @Afloat16 in #240
- Clarifies VecEnv observation buffer lifetime by @Mikasa0503 in #235
- Detects infinite values in environment outputs by @watayu10 in #238
- Retains rollout remainders in mini-batch generators by @Afloat16 in #244
- Skips singleton PPO advantage normalization by @ClemensSchwarke in #251
- Steps CNN encoders shared by actor and critic once per update by @sxngt in #246
- Keeps the recurrent teacher's hidden state across distillation updates by @sxngt in #247
- Skips observation-normalizer updates for image-only CNN models by @ClemensSchwarke in #252
- Fixes logging writer lifecycle and duplicate uploads by @louislelay in #248
- Preserves environment dimension in recurrent PPO surrogate loss by @cyanseek in #249
- Cleans up tests and remove unused
verboseargument for ONNX classes by @ClemensSchwarke in #256
Breaking Changes
- The unused
verboseargument was removed fromMLPModel.as_onnx(),CNNModel.as_onnx(), andRNNModel.as_onnx(). Please usemodel.as_onnx()instead.OnPolicyRunner.export_policy_to_onnx(..., verbose=...)remains supported. - PPO optimizer parameters are now deduplicated when the actor and critic share a CNN encoder. Optimizer states from older checkpoints with shared parameters cannot be loaded directly. Model weights remain compatible; disable optimizer loading through
load_cfgand explicitly enable the other states you want to restore. - Rollout mini-batch generators now retain all samples, so mini-batch sizes can differ when the rollout is not evenly divisible. Invalid
num_mini_batchesvalues now raiseValueError: the count must be between 1 and the rollout sample count for feedforward models, or between 1 and the number of environments for recurrent models. - Environment output checks now reject infinite values in addition to NaN values when
check_for_nanis enabled. Train/mean_reward/timeandTrain/mean_episode_length/timeare no longer logged through Neptune or customLogWriterbackends. The iteration-based metrics remain available.Distillation.last_hidden_stateswas replaced bylast_student_hidden_state. Custom extensions accessing this attribute must adapt; the recurrent teacher now retains its hidden state across optimization updates.