Download code/tests/test_protocol.py from jisue/robottt-pi05-base-comparison: direct link, hf CLI and curl.
- Browser
- Download file 3.23 kB
-
https://huggingface.co/jisue/robottt-pi05-base-comparison/resolve/main/code/tests/test_protocol.py
- Command line
-
hf download hf://jisue/robottt-pi05-base-comparison/code/tests/test_protocol.py
-
curl -L -o test_protocol.py https://huggingface.co/jisue/robottt-pi05-base-comparison/resolve/main/code/tests/test_protocol.py
3.23 kB
| import json,tempfile,unittest | |
| from pathlib import Path | |
| import torch | |
| from robottt_pi05.config import ExperimentConfig | |
| from robottt_pi05.prepare import prepare_protocol | |
| from robottt_pi05.assets import SavedSchedule | |
| from robottt_pi05.eval_io import sha | |
| from robottt_pi05.sequence_trainer import SequenceTrainer,full_masks | |
| from test_sequence import TinyRecurrent | |
| class ProtocolTests(unittest.TestCase): | |
| def test_t32_s8_masks_scaling_and_detach(self): | |
| config=ExperimentConfig(sequence_length=32,segment_length=8,observation_stride=1,learn_inner_lr=True,loss_action_dim=7) | |
| torch.manual_seed(42) | |
| actions=torch.randn(1,32,10,32,dtype=torch.float64);epsilon=torch.randn_like(actions) | |
| times=torch.full((1,32,1,1),.5,dtype=torch.float64) | |
| outputs=[] | |
| for segment in (8,32): | |
| adapter=TinyRecurrent();masks=full_masks(1,32,10) | |
| masks.action_dimension=(torch.arange(32)<7)[None] | |
| out=SequenceTrainer(adapter).run([None]*32,actions,epsilon,times,masks,segment_length=segment,backward=lambda x:x.backward()) | |
| self.assertEqual(out['valid_scalar_count'],32*10*7) | |
| expected=(out['prediction'][...,:7]-(epsilon-actions)[...,:7]).square().mean() | |
| self.assertAlmostEqual(out['loss'],float(expected),places=12) | |
| self.assertEqual(out['state']['0']['updates'].tolist(),[32]) | |
| self.assertTrue(all(r['denominator']==2240 for r in out['segments'])) | |
| outputs.append((out,adapter.w.grad.clone())) | |
| torch.testing.assert_close(outputs[0][0]['prediction'],outputs[1][0]['prediction'],rtol=0,atol=0) | |
| self.assertNotEqual(float(outputs[0][1]),float(outputs[1][1])) | |
| self.assertEqual(config.expected_trainables,(561540,41)) | |
| def test_new_protocol_roundtrip_and_seed(self): | |
| with tempfile.TemporaryDirectory() as tmp: | |
| d=Path(tmp);cp=d/'config.json';index=d/'index.json' | |
| cp.write_text(json.dumps(dict(sequence_length=32,segment_length=8,observation_stride=1,learn_inner_lr=True,loss_action_dim=7))) | |
| index.write_text(json.dumps(dict(train_ids=[0],episodes={'0':{'manifest':{'rows':list(range(45))}}}))) | |
| rng=torch.get_rng_state().clone() | |
| for name in ('a','b'): | |
| prepare_protocol(config_path=cp,asset_index=index,output=d/name,updates=3,seed=1000) | |
| bank=SavedSchedule(d/name/'schedule.json',d/name/'bank.pt',schedule_sha256=sha(d/name/'schedule.json'),bank_sha256=sha(d/name/'bank.pt'),train_ids=[0],sequence_length=32,stride=1) | |
| self.assertEqual(len(bank),3) | |
| self.assertEqual(bank[0][1].shape,(1,32,10,32)) | |
| self.assertTrue(bool((bank[0][2]>=.001).all() & (bank[0][2]<=1).all())) | |
| self.assertTrue(torch.equal(rng,torch.get_rng_state())) | |
| a=torch.load(d/'a/bank.pt',weights_only=True);b=torch.load(d/'b/bank.pt',weights_only=True) | |
| for k in a:torch.testing.assert_close(a[k],b[k],rtol=0,atol=0) | |
| self.assertEqual((d/'a/schedule.json').read_text(),(d/'b/schedule.json').read_text()) | |
| with self.assertRaises(FileExistsError):prepare_protocol(config_path=cp,asset_index=index,output=d/'a',updates=3) | |