def test_logger(test_str):
logger.critical(test_str)
logger.debug(test_str)
logger.error(test_str)
logger.exception(test_str)
logger.info(test_str)
logger.warning(test_str)
def test_logger(test_multiline_str):
logger.critical(test_multiline_str)
logger.debug(test_multiline_str)
logger.error(test_multiline_str)
logger.exception(test_multiline_str)
logger.info(test_multiline_str)
logger.warn(test_multiline_str)
def check_fn(*args, **kwargs):
if not to_check_training_step():
return fn(*args, **kwargs)
net = args[0] # first arg self
# get pre-update parameters to compare
pre_params = [param.clone() for param in net.parameters()]
# run training_step, get loss
loss = fn(*args, **kwargs)
# get post-update parameters to compare
post_params = [param.clone() for param in net.parameters()]
if loss == 0.0:
# if loss is 0, there should be no updates
# TODO if without momentum, parameters should not change too
for p_name, param in net.named_parameters():
assert param.grad.norm() == 0
else:
# check parameter updates
try:
assert not all(
torch.equal(w1, w2)
for w1, w2 in zip(pre_params, post_params)
), f'Model parameter is not updated in training_step(), check if your tensor is detached from graph. Loss: {loss:g}'
logger.info(
f'Model parameter is updated in training_step(). Loss: {loss: g}'
)
except Exception as e:
logger.error(e)
if os.environ.get('PY_ENV') == 'test':
# raise error if in unit test
raise (e)
# check grad norms
min_norm, max_norm = 0.0, 1e5
for p_name, param in net.named_parameters():
try:
grad_norm = param.grad.norm()
assert min_norm < grad_norm < max_norm, f'Gradient norm for {p_name} is {grad_norm:g}, fails the extreme value check {min_norm} < grad_norm < {max_norm}. Loss: {loss:g}. Check your network and loss computation.'
logger.info(
f'Gradient norm for {p_name} is {grad_norm:g}; passes value check.'
)
except Exception as e:
logger.warn(e)
logger.debug('Passed network parameter update check.')
# store grad norms for debugging
net.store_grad_norms()
return loss
def save_image(figure, filepath=None):
if os.environ.get('PY_ENV') == 'test':
return
if filepath is None:
filepath = f'{PLOT_FILEDIR}/{_.get(figure, "layout.title")}.png'
try:
plotly.tools.set_credentials_file(
username=_.get(config, 'plotly.username'),
api_key=_.get(config, 'plotly.api_key'))
plotly.tools.set_config_file(world_readable=True, sharing='public')
return plotly.plotly.image.save_as(figure, filepath)
except Exception:
logger.error(
'Plotly server unreachable, but you can save the plots later via retro-analysis.'
)
return None
def update_nets(self):
# NOTE: Once polyak updating for multi-headed networks is supported via updates to flatten_params and load_params then this can be removed
space_clock = util.s_get(self, 'aeb_space.clock')
t = space_clock.get('t')
if self.update_type == 'replace':
if t % self.update_frequency == 0:
logger.debug('Updating target_net by replacing')
self.target_net = deepcopy(self.net)
self.online_net = self.target_net
self.eval_net = self.target_net
elif self.update_type == 'polyak':
logger.error(
'"polyak" updating not supported yet for MultiHeadDQN, please use "replace" instead. Exiting.')
sys.exit()
else:
logger.error(
'Unknown net.update_type. Should be "replace" or "polyak". Exiting.')
sys.exit()
def update_nets(self):
space_clock = util.s_get(self, 'aeb_space.clock')
t = space_clock.get('t')
if self.update_type == 'replace':
if t % self.update_frequency == 0:
logger.debug('Updating target_net by replacing')
self.target_net = deepcopy(self.net)
self.online_net = self.target_net
self.eval_net = self.target_net
elif self.update_type == 'polyak':
logger.debug('Updating net by averaging')
avg_params = self.polyak_weight * net_util.flatten_params(self.target_net) + \
(1 - self.polyak_weight) * net_util.flatten_params(self.net)
self.target_net = net_util.load_params(self.target_net, avg_params)
self.online_net = self.target_net
self.eval_net = self.target_net
else:
logger.error(
'Unknown net.update_type. Should be "replace" or "polyak". Exiting.')
sys.exit()
