def test_training_only_cnn_weights(self):
with initialize(config_path=os.path.join("..", "fixtures", "conf")):
config = compose(
config_name="config",
overrides=[
"runner/model=resnet18fusionpolicy",
"runner/dataset=filetextbertforwardsatosapolicy",
"runner/val_dataset=filetextbertforwardsatosapolicy",
"runner.runner_name=SOCTextForwardPolicyRunner"
])
config.runner.dataset.dataset_path = _RAW_TEXT_BERT_DATASET_PATH
config.runner.val_dataset.dataset_path = _RAW_TEXT_BERT_DATASET_PATH
config.runner.train_cnn = True
config.runner.train_fusion = False
config.runner.train_heads = True
config.trainer.default_root_dir = self.folder
config.trainer.fast_dev_run = False
# We rely on seeds to copy the init weights
seed_everything(config['runner']['seed'])
r_copy = make_runner(config['runner'])
r_copy.setup('fit')
seed_everything(config['runner']['seed'])
runner = make_runner(config['runner'])
trainer = Trainer(**config['trainer'], deterministic=True)
trainer.fit(runner)
zipped_params = zip(r_copy.model.fusion.parameters(),
runner.model.fusion.parameters())
for param_copy, param in zipped_params:
assert torch.all(torch.eq(param_copy, param))
zipped_params = zip(r_copy.model.spatial_state_head.parameters(),
runner.model.spatial_state_head.parameters())
for param_copy, param in zipped_params:
assert not torch.all(torch.eq(param_copy, param))
zipped_params = zip(r_copy.model.linear_state_head.parameters(),
runner.model.linear_state_head.parameters())
for param_copy, param in zipped_params:
assert not torch.all(torch.eq(param_copy, param))
# zipped_params = zip(
# r_copy.model.policy_head.parameters(), runner.model.policy_head.parameters()
# )
# for param_copy, param in zipped_params:
# assert not torch.all(torch.eq(param_copy, param))
zipped_params = zip(r_copy.model.cnn.parameters(),
runner.model.cnn.parameters())
for param_copy, param in zipped_params:
assert not torch.all(torch.eq(param_copy, param))
break # Not all layers are learnable so we check only the first one
def test_training_soc_preprocessed_seq_convlstmpolicy(self):
with initialize(config_path=os.path.join(".", "fixtures", "conf")):
config = compose(config_name="config",
overrides=[
"runner/model=convlstmpolicy",
"runner/dataset=preprocessedseqsatosapolicy",
"runner.runner_name=SOCSeqPolicyRunner"
])
config.runner.dataset.dataset_path = _DATASET_PATH
config.trainer.default_root_dir = self.folder
seed_everything(config['runner']['seed'])
runner = make_runner(config['runner'])
trainer = Trainer(**config['trainer'], deterministic=True)
trainer.fit(runner)
def test_training_soc_psql_seq_sas_conv3d(self):
with initialize(config_path=os.path.join(".", "fixtures", "conf")):
config = compose(config_name="config",
overrides=[
"runner/model=conv3d",
"runner/dataset=psqlseqsatos",
"runner.runner_name=SOCSupervisedSeqRunner"
])
config.trainer.default_root_dir = self.folder
seed_everything(config['runner']['seed'])
runner = make_runner(config['runner'])
runner.setup_dataset = self.setup_dataset
trainer = Trainer(**config['trainer'], deterministic=True)
trainer.fit(runner)
def test_training_soc_preprocessed_forward_resnet(self):
with initialize(config_path=os.path.join(".", "fixtures", "conf")):
config = compose(
config_name="config",
overrides=[
"runner/model=resnet18",
"runner/dataset=preprocessedforwardsatosa",
"runner.runner_name=SOCSupervisedForwardRunner"
])
config.runner.dataset.dataset_path = _DATASET_PATH
config.trainer.default_root_dir = self.folder
seed_everything(config['runner']['seed'])
runner = make_runner(config['runner'])
trainer = Trainer(**config['trainer'], deterministic=True)
trainer.fit(runner)
def test_training_soc_file_humantrade_forward_resnetmeanffpolicy(self):
with initialize(config_path=os.path.join(".", "fixtures", "conf")):
config = compose(
config_name="config",
overrides=[
"runner/model=resnet18meanffpolicy",
"runner/dataset=filetextberthumantradeforwardsatosapolicy",
"runner.runner_name=SOCTextForwardPolicyRunner"
])
config.runner.dataset.dataset_path = _RAW_TEXT_BERT_DATASET_PATH
config.trainer.default_root_dir = self.folder
seed_everything(config['runner']['seed'])
runner = make_runner(config['runner'])
runner.num_workers = 1
trainer = Trainer(**config['trainer'], deterministic=True)
trainer.fit(runner)
def test_training_soc_psql_forward_resnetmeanconcatpolicy(self):
with initialize(config_path=os.path.join(".", "fixtures", "conf")):
config = compose(
config_name="config",
overrides=[
"runner/model=resnet18meanconcatpolicy",
"runner/dataset=psqltextbertforwardsatosapolicy",
"runner.runner_name=SOCTextForwardPolicyRunner"
])
config.trainer.default_root_dir = self.folder
seed_everything(config['runner']['seed'])
runner = make_runner(config['runner'])
runner.setup_dataset = self.setup_text_dataset
runner.num_workers = 1
trainer = Trainer(**config['trainer'], deterministic=True)
trainer.fit(runner)
'hyper_parameters']['dataset']['use_pooler_features']
ckpt['hyper_parameters']['val_dataset']['set_empty_text_to_zero'] = ckpt[
'hyper_parameters']['dataset']['set_empty_text_to_zero']
ckpt['hyper_parameters']['model'] = DictConfig(
ckpt['hyper_parameters']['model'])
ckpt['hyper_parameters']['dataset'][
'dataset_path'] = _RAW_SOC5_TEXT_BERT_DATASET_PATH
ckpt['hyper_parameters']['val_dataset'][
'dataset_path'] = _RAW_SOC5_TEXT_BERT_DATASET_PATH
ckpt['hyper_parameters']['dataset']['shuffle'] = False
ckpt['hyper_parameters']['batch_size'] = 1
pl.seed_everything(ckpt['hyper_parameters']['seed'])
runner = make_runner(ckpt['hyper_parameters'])
runner.setup('fit')
runner.load_state_dict(ckpt['state_dict'])
runner.eval()
pprint.pprint(runner.hparams)
def format_res_tensor(res_tensor):
return res_tensor.view(4, 6).detach().numpy().tolist()
def format_res_pred(res_pred):
return format_res_tensor(ds_utils.unnormalize_playersresources(res_pred))