Download code/tests/test_batched.py from jisue/robottt-pi05-base-comparison: direct link, hf CLI and curl.
- Browser
- Download file 2.03 kB
-
https://huggingface.co/jisue/robottt-pi05-base-comparison/resolve/main/code/tests/test_batched.py
- Command line
-
hf download hf://jisue/robottt-pi05-base-comparison/code/tests/test_batched.py
-
curl -L -o test_batched.py https://huggingface.co/jisue/robottt-pi05-base-comparison/resolve/main/code/tests/test_batched.py
2.03 kB
| """The segment loop must preserve global scaling and truncated gradients.""" | |
| import unittest | |
| from unittest.mock import patch | |
| import torch | |
| from robottt_pi05.batched_trainer import run | |
| from robottt_pi05.sequence_trainer import SequenceTrainer,full_masks | |
| from robottt_pi05.config import ExperimentConfig | |
| from test_sequence import TinyRecurrent | |
| class TinySegment(TinyRecurrent): | |
| def velocity_segment(self,prepared,noisy,time,state,**kwargs): | |
| values=[] | |
| for t in range(len(noisy)): | |
| value,state=self.velocity(None,noisy[t:t+1],time[t:t+1],state) | |
| values.append(value) | |
| return torch.cat(values),state | |
| class BatchedLoopTests(unittest.TestCase): | |
| def test_float64_loss_gradient_state_equal_to_sequential(self): | |
| torch.manual_seed(31);a=torch.randn(1,32,10,32,dtype=torch.float64);e=torch.randn_like(a);t=torch.rand(1,32,1,1,dtype=torch.float64) | |
| masks=full_masks(1,32,10);masks.action_dimension=(torch.arange(32)<7)[None] | |
| config=ExperimentConfig(sequence_length=32,segment_length=8,observation_stride=1,loss_action_dim=7) | |
| sequential=TinyRecurrent();batched=TinySegment() | |
| expected=SequenceTrainer(sequential).run([None]*32,a,e,t,masks,segment_length=8,backward=lambda loss:loss.backward()) | |
| with patch('robottt_pi05.batched_trainer.batch_prepared',lambda x:x): | |
| actual=run(batched,config,[None]*32,a,e,t,masks,backward=lambda loss:loss.backward()) | |
| self.assertAlmostEqual(actual['loss'],expected['loss'],places=13) | |
| torch.testing.assert_close(actual['prediction'],expected['prediction'],rtol=0,atol=0) | |
| torch.testing.assert_close(batched.w.grad,sequential.w.grad,rtol=1e-13,atol=1e-13) | |
| torch.testing.assert_close(actual['state']['0']['weights'][0],expected['state']['0']['weights'][0],rtol=0,atol=0) | |
| self.assertEqual(actual['state']['0']['updates'].tolist(),[32]);self.assertEqual(actual['valid_scalar_count'],2240) | |
| self.assertEqual([r['end'] for r in actual['segments']],[8,16,24,32]) | |