| import numpy as np |
| import torch |
| from synapse.config import SynapseConfig |
| from synapse.agent import IntegratedAgent, AcquiringMushroomBody |
|
|
| def test_acquisition_from_neutral_and_reversal(): |
| torch.manual_seed(1) |
| mb=AcquiringMushroomBody(4) |
| cue=torch.tensor([1.,0.,0.,1.]) |
| mb(cue); mb.dopaminergic_update(1.) |
| assert mb(cue)[0]>0 |
| for _ in range(8): mb(cue); mb.dopaminergic_update(-1.) |
| assert mb(cue)[0]<0 |
|
|
| def test_restore_and_inference_freeze(): |
| a=IntegratedAgent(SynapseConfig(nodes_per_source=16,cognition_interval=1)) |
| obs=np.linspace(0,1,155,dtype=np.float32) |
| before={k:v.clone() for k,v in a.editable_tensors().items()} |
| state=a.export_state(); first=a.act(obs); a.import_state(state); second=a.act(obs) |
| assert np.array_equal(first['action'],second['action']) |
| assert first['features'].shape==(64,) |
| assert np.isfinite(first['action']).all() |
| assert all(torch.equal(v,a.editable_tensors()[k]) for k,v in before.items()) |
|
|
| def test_bidirectional_coupling_and_bridge_preparation(): |
| a=IntegratedAgent(SynapseConfig(nodes_per_source=16,cognition_interval=1)) |
| obs=np.linspace(0,1,155,dtype=np.float32) |
| initial=a.export_state(); base=a.act(obs) |
| a.import_state(initial) |
| with torch.no_grad(): a.donor_to_graph.weight.zero_() |
| lesion=a.act(obs) |
| assert not np.array_equal(base['features'][:32],lesion['features'][:32]) |
| a.import_state(initial) |
| a.brain.r.fill_(.8) |
| changed=a.act(obs) |
| assert not np.array_equal(base['features'][32:48],changed['features'][32:48]) |
| a.import_state(initial) |
| report=a.prepare_observations(np.stack([obs,obs[::-1].copy()]),np.ones((2,5),np.float32)*.3) |
| for prefix in ('encoder.','donor_to_graph.','graph_to_resource.','motor.'): |
| assert any(v>0 for k,v in report['gradient_norms'].items() if k.startswith(prefix)) |
|
|
