seppl/optical-regeneration
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor hypertraining.py to improve model layer handling and response plotting; adjust data settings for batch processing

Browse Source
This commit is contained in:
Joseph Hopfmüller
2024-11-29 15:48:27 +01:00
parent cfa08aae4e
commit b156b9ceaf
1 changed files with 12 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
src/single-core-regen/hypertraining/hypertraining.py
View File

@@ -245,18 +245,18 @@ class HyperTraining:
dtype = getattr(torch, dtype) dtype = getattr(torch, dtype)
afunc = trial.suggest_categorical_optional("model_activation_func", self.model_settings.model_activation_func) afunc = trial.suggest_categorical_optional("model_activation_func", self.model_settings.model_activation_func)
# T0 = trial.suggest_float_optional("T0", self.model_settings.satabsT0 , log=True)
layers = [] layers = []
last_dim = input_dim last_dim = input_dim
n_nodes = last_dim n_nodes = last_dim
for i in range(n_layers): for i in range(n_layers):
if hidden_dim_override := self.model_settings.overrides.get(f"n_hidden_nodes_{i}", False): if hidden_dim_override := self.model_settings.overrides.get(f"n_hidden_nodes_{i}", False):
hidden_dim = trial.suggest_int_optional(f"model_hidden_dim_{i}", hidden_dim_override, force=True) hidden_dim = trial.suggest_int_optional(f"model_hidden_dim_{i}", hidden_dim_override)
else: else:
hidden_dim = trial.suggest_int_optional( hidden_dim = trial.suggest_int_optional(
f"model_hidden_dim_{i}", f"model_hidden_dim_{i}",
self.model_settings.n_hidden_nodes, self.model_settings.n_hidden_nodes,
# step=2,
) )
layers.append(util.complexNN.SemiUnitaryLayer(last_dim, hidden_dim, dtype=dtype)) layers.append(util.complexNN.SemiUnitaryLayer(last_dim, hidden_dim, dtype=dtype))
last_dim = hidden_dim last_dim = hidden_dim
@@ -642,6 +642,7 @@ class HyperTraining:
if show: if show:
plt.show() plt.show()
return fig
def _plot_model_response_head(self, *signals, labels=None, sps=None, title_append="", subtitle="", show=True): def _plot_model_response_head(self, *signals, labels=None, sps=None, title_append="", subtitle="", show=True):
if not hasattr(labels, "__iter__") or isinstance(labels, (str, type(None))): if not hasattr(labels, "__iter__") or isinstance(labels, (str, type(None))):
@@ -684,7 +685,7 @@ class HyperTraining:
): ):
data_settings_backup = copy.deepcopy(self.data_settings) data_settings_backup = copy.deepcopy(self.data_settings)
pytorch_settings_backup = copy.deepcopy(self.pytorch_settings) pytorch_settings_backup = copy.deepcopy(self.pytorch_settings)
self.data_settings.drop_first = 100 self.data_settings.drop_first = 100*128
self.data_settings.shuffle = False self.data_settings.shuffle = False
self.data_settings.train_split = 1.0 self.data_settings.train_split = 1.0
self.pytorch_settings.batchsize = ( self.pytorch_settings.batchsize = (
@@ -739,11 +740,15 @@ class HyperTraining:
@staticmethod @staticmethod
def build_title(trial: optuna.trial.Trial): def build_title(trial: optuna.trial.Trial):
title_append = f"for trial {trial.number}" title_append = f"for trial {trial.number}"
model_n_layers = util.misc.multi_getattr((trial.params, trial.user_attrs), "model_n_layers", 0) model_n_hidden_layers = util.misc.multi_getattr((trial.params, trial.user_attrs), "model_n_hidden_layers", 0)
model_hidden_dims = [ input_dim = util.misc.multi_getattr((trial.params, trial.user_attrs), "model_input_dim", 0)
model_dims = [
util.misc.multi_getattr((trial.params, trial.user_attrs), f"model_hidden_dim_{i}", 0) util.misc.multi_getattr((trial.params, trial.user_attrs), f"model_hidden_dim_{i}", 0)
for i in range(model_n_layers) for i in range(model_n_hidden_layers)
] ]
model_dims.insert(0, input_dim)
model_dims.append(2)
model_dims = [str(dim) for dim in model_dims]
model_activation_func = util.misc.multi_getattr( model_activation_func = util.misc.multi_getattr(
(trial.params, trial.user_attrs), (trial.params, trial.user_attrs),
"model_activation_func", "model_activation_func",
@@ -752,7 +757,7 @@ class HyperTraining:
model_dtype = util.misc.multi_getattr((trial.params, trial.user_attrs), "model_dtype", "unknown dtype") model_dtype = util.misc.multi_getattr((trial.params, trial.user_attrs), "model_dtype", "unknown dtype")
subtitle = ( subtitle = (
f"{model_n_layers} layers à ({', '.join(model_hidden_dims)}) units, {model_activation_func}, {model_dtype}" f"{model_n_hidden_layers+2} layers à ({', '.join(model_dims)}) units, {model_activation_func}, {model_dtype}"
) )
return title_append, subtitle return title_append, subtitle