def __init__(self,
atoms,
states,
invd_index=None,
angle_index=None,
dihed_index=None,
nn_size=100,
depth=3,
activ="selu",
use_reg_activ=None,
use_reg_weight=None,
use_reg_bias=None,
use_dropout=False,
dropout=0.01,
normalization_mode=1,
precomputed_features=False,
model_module="mlp_nac2",
**kwargs):
"""Initialize a NACModel with hyperparameters.
Args:
hyper (dict): Hyperparamters.
**kwargs (dict): Additional keras.model parameters.
Returns:
tf.keras.model.
"""
super(NACModel2, self).__init__(**kwargs)
self.model_module = model_module
self.in_invd_index = invd_index
self.in_angle_index = angle_index
self.in_dihed_index = dihed_index
self.nn_size = nn_size
self.depth = depth
self.activ = activ
self.use_reg_activ = use_reg_activ
self.use_reg_weight = use_reg_weight
self.use_reg_bias = use_reg_bias
self.use_dropout = use_dropout
self.dropout = dropout
self.normalization_mode = normalization_mode
self.y_atoms = int(atoms)
self.in_states = int(states)
out_dim = int(states * (states - 1) / 2)
indim = int(atoms)
# Allow for all distances, backward compatible
if isinstance(invd_index, bool):
if invd_index:
invd_index = [[i, j] for i in range(0, int(atoms))
for j in range(0, i)]
use_invd_index = len(invd_index) > 0 if isinstance(
invd_index, list) or isinstance(invd_index, np.ndarray) else False
use_angle_index = len(angle_index) > 0 if isinstance(
angle_index, list) or isinstance(angle_index,
np.ndarray) else False
use_dihed_index = len(dihed_index) > 0 if isinstance(
dihed_index, list) or isinstance(dihed_index,
np.ndarray) else False
invd_index = np.array(invd_index,
dtype=np.int64) if use_invd_index else None
angle_index = np.array(angle_index,
dtype=np.int64) if use_angle_index else None
dihed_index = np.array(dihed_index,
dtype=np.int64) if use_dihed_index else None
invd_shape = invd_index.shape if use_invd_index else None
angle_shape = angle_index.shape if use_angle_index else None
dihed_shape = dihed_index.shape if use_dihed_index else None
in_model_dim = 0
if use_invd_index:
in_model_dim += len(invd_index)
if use_angle_index:
in_model_dim += len(angle_index)
if use_dihed_index:
in_model_dim += len(dihed_index)
self.feat_layer = FeatureGeometric(invd_shape=invd_shape,
angle_shape=angle_shape,
dihed_shape=dihed_shape,
name="feat_geo")
self.feat_layer.set_mol_index(invd_index, angle_index, dihed_index)
if normalization_mode == 1:
self.std_layer = tf.keras.layers.BatchNormalization(
name='feat_std')
elif normalization_mode == 2:
self.std_layer = tf.keras.layers.LayerNormalization(
name='feat_std')
else:
self.std_layer = DummyLayer()
self.mlp_layer = MLP(nn_size,
dense_depth=depth,
dense_bias=True,
dense_bias_last=False,
dense_activ=activ,
dense_activ_last=activ,
dense_activity_regularizer=use_reg_activ,
dense_kernel_regularizer=use_reg_weight,
dense_bias_regularizer=use_reg_bias,
dropout_use=use_dropout,
dropout_dropout=dropout,
name='mlp')
self.virt_layer = ks.layers.Dense(out_dim * in_model_dim,
name='virt',
use_bias=False,
activation='linear')
self.resh_layer = tf.keras.layers.Reshape((out_dim, in_model_dim))
self.prop_grad_layer = PropagateNACGradient2(axis=(2, 1))
# Build all layers
self.precomputed_features = False
self.build((None, indim, 3))
self.precomputed_features = precomputed_features
def __init__(self,
states=1,
atoms=2,
invd_index=None,
angle_index=None,
dihed_index=None,
nn_size=100,
depth=3,
activ='selu',
use_reg_activ=None,
use_reg_weight=None,
use_reg_bias=None,
use_dropout=False,
dropout=0.01,
normalization_mode=1,
energy_only=True,
precomputed_features=False,
model_module="mlp_e",
**kwargs):
super(EnergyModel, self).__init__(**kwargs)
self.invd_index = invd_index
self.angle_index = angle_index
self.dihed_index = dihed_index
self.nn_size = nn_size
self.depth = depth
self.activ = activ
self.use_reg_activ = use_reg_activ
self.use_reg_weight = use_reg_weight
self.use_reg_bias = use_reg_bias
self.use_dropout = use_dropout
self.dropout = dropout
self.normalization_mode = normalization_mode
self.out_dim = int(states)
self.in_atoms = int(atoms)
self.energy_only = energy_only
self.model_module = model_module
out_dim = int(states)
indim = int(atoms)
# Allow for all distances, backward compatible
if isinstance(invd_index, bool):
if invd_index:
invd_index = [[i, j] for i in range(0, int(atoms))
for j in range(0, i)]
use_invd_index = len(invd_index) > 0 if isinstance(
invd_index, list) or isinstance(invd_index, np.ndarray) else False
use_angle_index = len(angle_index) > 0 if isinstance(
angle_index, list) or isinstance(angle_index,
np.ndarray) else False
use_dihed_index = len(dihed_index) > 0 if isinstance(
dihed_index, list) or isinstance(dihed_index,
np.ndarray) else False
invd_index = np.array(invd_index,
dtype=np.int64) if use_invd_index else None
angle_index = np.array(angle_index,
dtype=np.int64) if use_angle_index else None
dihed_index = np.array(dihed_index,
dtype=np.int64) if use_dihed_index else None
invd_shape = invd_index.shape if use_invd_index else None
angle_shape = angle_index.shape if use_angle_index else None
dihed_shape = dihed_index.shape if use_dihed_index else None
self.feat_layer = FeatureGeometric(invd_shape=invd_shape,
angle_shape=angle_shape,
dihed_shape=dihed_shape,
name="feat_geo")
self.feat_layer.set_mol_index(invd_index, angle_index, dihed_index)
if normalization_mode == 1:
self.std_layer = tf.keras.layers.BatchNormalization(
name='feat_std')
elif normalization_mode == 2:
self.std_layer = tf.keras.layers.LayerNormalization(
name='feat_std')
else:
self.std_layer = DummyLayer()
self.mlp_layer = MLP(nn_size,
dense_depth=depth,
dense_bias=True,
dense_bias_last=True,
dense_activ=activ,
dense_activ_last=activ,
dense_activity_regularizer=use_reg_activ,
dense_kernel_regularizer=use_reg_weight,
dense_bias_regularizer=use_reg_bias,
dropout_use=use_dropout,
dropout_dropout=dropout,
name='mlp')
self.energy_layer = ks.layers.Dense(out_dim,
name='energy',
use_bias=True,
activation='linear')
# Build all layers
self.precomputed_features = False
self.build((None, indim, 3))
self.precomputed_features = precomputed_features
class NACModel2(ks.Model):
"""Subclassed tf.keras.model for NACs which outputs NACs from coordinates.
The model is supposed to be saved and exported.
"""
def __init__(self,
atoms,
states,
invd_index=None,
angle_index=None,
dihed_index=None,
nn_size=100,
depth=3,
activ="selu",
use_reg_activ=None,
use_reg_weight=None,
use_reg_bias=None,
use_dropout=False,
dropout=0.01,
normalization_mode=1,
precomputed_features=False,
model_module="mlp_nac2",
**kwargs):
"""Initialize a NACModel with hyperparameters.
Args:
hyper (dict): Hyperparamters.
**kwargs (dict): Additional keras.model parameters.
Returns:
tf.keras.model.
"""
super(NACModel2, self).__init__(**kwargs)
self.model_module = model_module
self.in_invd_index = invd_index
self.in_angle_index = angle_index
self.in_dihed_index = dihed_index
self.nn_size = nn_size
self.depth = depth
self.activ = activ
self.use_reg_activ = use_reg_activ
self.use_reg_weight = use_reg_weight
self.use_reg_bias = use_reg_bias
self.use_dropout = use_dropout
self.dropout = dropout
self.normalization_mode = normalization_mode
self.y_atoms = int(atoms)
self.in_states = int(states)
out_dim = int(states * (states - 1) / 2)
indim = int(atoms)
# Allow for all distances, backward compatible
if isinstance(invd_index, bool):
if invd_index:
invd_index = [[i, j] for i in range(0, int(atoms))
for j in range(0, i)]
use_invd_index = len(invd_index) > 0 if isinstance(
invd_index, list) or isinstance(invd_index, np.ndarray) else False
use_angle_index = len(angle_index) > 0 if isinstance(
angle_index, list) or isinstance(angle_index,
np.ndarray) else False
use_dihed_index = len(dihed_index) > 0 if isinstance(
dihed_index, list) or isinstance(dihed_index,
np.ndarray) else False
invd_index = np.array(invd_index,
dtype=np.int64) if use_invd_index else None
angle_index = np.array(angle_index,
dtype=np.int64) if use_angle_index else None
dihed_index = np.array(dihed_index,
dtype=np.int64) if use_dihed_index else None
invd_shape = invd_index.shape if use_invd_index else None
angle_shape = angle_index.shape if use_angle_index else None
dihed_shape = dihed_index.shape if use_dihed_index else None
in_model_dim = 0
if use_invd_index:
in_model_dim += len(invd_index)
if use_angle_index:
in_model_dim += len(angle_index)
if use_dihed_index:
in_model_dim += len(dihed_index)
self.feat_layer = FeatureGeometric(invd_shape=invd_shape,
angle_shape=angle_shape,
dihed_shape=dihed_shape,
name="feat_geo")
self.feat_layer.set_mol_index(invd_index, angle_index, dihed_index)
if normalization_mode == 1:
self.std_layer = tf.keras.layers.BatchNormalization(
name='feat_std')
elif normalization_mode == 2:
self.std_layer = tf.keras.layers.LayerNormalization(
name='feat_std')
else:
self.std_layer = DummyLayer()
self.mlp_layer = MLP(nn_size,
dense_depth=depth,
dense_bias=True,
dense_bias_last=False,
dense_activ=activ,
dense_activ_last=activ,
dense_activity_regularizer=use_reg_activ,
dense_kernel_regularizer=use_reg_weight,
dense_bias_regularizer=use_reg_bias,
dropout_use=use_dropout,
dropout_dropout=dropout,
name='mlp')
self.virt_layer = ks.layers.Dense(out_dim * in_model_dim,
name='virt',
use_bias=False,
activation='linear')
self.resh_layer = tf.keras.layers.Reshape((out_dim, in_model_dim))
self.prop_grad_layer = PropagateNACGradient2(axis=(2, 1))
# Build all layers
self.precomputed_features = False
self.build((None, indim, 3))
self.precomputed_features = precomputed_features
def call(self, data, training=False, **kwargs):
"""Call the model output, forward pass.
Args:
data (tf.tensor): Coordinates.
training (bool, optional): Training Mode. Defaults to False.
Returns:
y_pred (tf.tensor): predicted NACs.
"""
x = data
# Compute predictions
if not self.precomputed_features:
with tf.GradientTape() as tape2:
tape2.watch(x)
feat_flat = self.feat_layer(x)
temp_grad = tape2.batch_jacobian(feat_flat, x)
feat_flat_std = self.std_layer(feat_flat)
temp_hidden = self.mlp_layer(feat_flat_std, training=training)
temp_v = self.virt_layer(temp_hidden)
temp_va = self.resh_layer(temp_v)
# y_pred = ks.backend.batch_dot(temp_va,temp_grad ,axes=(2,1))
y_pred = self.prop_grad_layer([temp_va, temp_grad])
else:
x1 = x[0]
x2 = x[1]
feat_flat_std = self.std_layer(x1)
temp_hidden = self.mlp_layer(feat_flat_std, training=training)
temp_v = self.virt_layer(temp_hidden)
temp_va = self.resh_layer(temp_v)
# y_pred = ks.backend.batch_dot(temp_va, x2, axes=(2, 1))
y_pred = self.prop_grad_layer([temp_va, x2])
return y_pred
@tf.function
def predict_chunk_feature(self, tf_x, training=False):
with tf.GradientTape() as tape2:
tape2.watch(tf_x)
feat_pred = self.feat_layer(tf_x,
training=training) # Forward pass
grad = tape2.batch_jacobian(feat_pred, tf_x)
return feat_pred, grad
def precompute_feature_in_chunks(self, x, batch_size, training=False):
np_x = []
np_grad = []
for j in range(int(np.ceil(len(x) / batch_size))):
a = int(batch_size * j)
b = int(batch_size * j + batch_size)
tf_x = tf.convert_to_tensor(x[a:b], dtype=tf.float32)
feat_pred, grad = self.predict_chunk_feature(tf_x,
training=training)
np_x.append(np.array(feat_pred.numpy()))
np_grad.append(np.array(grad.numpy()))
np_x = np.concatenate(np_x, axis=0)
np_grad = np.concatenate(np_grad, axis=0)
return np_x, np_grad
def fit(self, **kwargs):
return super(NACModel2, self).fit(**kwargs)
def get_config(self):
# conf = super(NACModel2, self).get_config()
conf = {}
conf.update({
'atoms': self.y_atoms,
'states': self.in_states,
'invd_index': self.in_invd_index,
'angle_index': self.in_angle_index,
'dihed_index': self.in_dihed_index,
'nn_size': self.nn_size,
'depth': self.depth,
'activ': self.activ,
'use_reg_activ': self.use_reg_activ,
'use_reg_weight': self.use_reg_weight,
'use_reg_bias': self.use_reg_bias,
'use_dropout': self.use_dropout,
'dropout': self.dropout,
'normalization_mode': self.normalization_mode,
'precomputed_features': self.precomputed_features,
"model_module": self.model_module
})
return conf
def save(self, filepath, **kwargs):
# copy to new model
self_conf = self.get_config()
self_conf['precomputed_features'] = False
copy_model = NACModel2(**self_conf)
copy_model.set_weights(self.get_weights())
# Make graph and test with training data
copy_model.predict(np.ones((1, self.y_atoms, 3)))
tf.keras.models.save_model(copy_model, filepath, **kwargs)
class EnergyModel(ks.Model):
"""Subclassed tf.keras.model for energy/gradient which outputs both energy and gradient from coordinates.
It can also
"""
def __init__(self,
states=1,
atoms=2,
invd_index=None,
angle_index=None,
dihed_index=None,
nn_size=100,
depth=3,
activ='selu',
use_reg_activ=None,
use_reg_weight=None,
use_reg_bias=None,
use_dropout=False,
dropout=0.01,
normalization_mode=1,
energy_only=True,
precomputed_features=False,
model_module="mlp_e",
**kwargs):
super(EnergyModel, self).__init__(**kwargs)
self.invd_index = invd_index
self.angle_index = angle_index
self.dihed_index = dihed_index
self.nn_size = nn_size
self.depth = depth
self.activ = activ
self.use_reg_activ = use_reg_activ
self.use_reg_weight = use_reg_weight
self.use_reg_bias = use_reg_bias
self.use_dropout = use_dropout
self.dropout = dropout
self.normalization_mode = normalization_mode
self.out_dim = int(states)
self.in_atoms = int(atoms)
self.energy_only = energy_only
self.model_module = model_module
out_dim = int(states)
indim = int(atoms)
# Allow for all distances, backward compatible
if isinstance(invd_index, bool):
if invd_index:
invd_index = [[i, j] for i in range(0, int(atoms))
for j in range(0, i)]
use_invd_index = len(invd_index) > 0 if isinstance(
invd_index, list) or isinstance(invd_index, np.ndarray) else False
use_angle_index = len(angle_index) > 0 if isinstance(
angle_index, list) or isinstance(angle_index,
np.ndarray) else False
use_dihed_index = len(dihed_index) > 0 if isinstance(
dihed_index, list) or isinstance(dihed_index,
np.ndarray) else False
invd_index = np.array(invd_index,
dtype=np.int64) if use_invd_index else None
angle_index = np.array(angle_index,
dtype=np.int64) if use_angle_index else None
dihed_index = np.array(dihed_index,
dtype=np.int64) if use_dihed_index else None
invd_shape = invd_index.shape if use_invd_index else None
angle_shape = angle_index.shape if use_angle_index else None
dihed_shape = dihed_index.shape if use_dihed_index else None
self.feat_layer = FeatureGeometric(invd_shape=invd_shape,
angle_shape=angle_shape,
dihed_shape=dihed_shape,
name="feat_geo")
self.feat_layer.set_mol_index(invd_index, angle_index, dihed_index)
if normalization_mode == 1:
self.std_layer = tf.keras.layers.BatchNormalization(
name='feat_std')
elif normalization_mode == 2:
self.std_layer = tf.keras.layers.LayerNormalization(
name='feat_std')
else:
self.std_layer = DummyLayer()
self.mlp_layer = MLP(nn_size,
dense_depth=depth,
dense_bias=True,
dense_bias_last=True,
dense_activ=activ,
dense_activ_last=activ,
dense_activity_regularizer=use_reg_activ,
dense_kernel_regularizer=use_reg_weight,
dense_bias_regularizer=use_reg_bias,
dropout_use=use_dropout,
dropout_dropout=dropout,
name='mlp')
self.energy_layer = ks.layers.Dense(out_dim,
name='energy',
use_bias=True,
activation='linear')
# Build all layers
self.precomputed_features = False
self.build((None, indim, 3))
self.precomputed_features = precomputed_features
def call(self, data, training=False, **kwargs):
"""Call the model output, forward pass.
Args:
data (tf.tensor): Coordinates.
training (bool, optional): Training Mode. Defaults to False.
Returns:
y_pred (list): List of tf.tensor for predicted [energy,gradient]
"""
# Unpack the data
x = data
y_pred = None
# Compute predictions
if self.energy_only and not self.precomputed_features:
feat_flat = self.feat_layer(x)
feat_flat_std = self.std_layer(feat_flat, training=training)
temp_hidden = self.mlp_layer(feat_flat_std, training=training)
temp_e = self.energy_layer(temp_hidden)
y_pred = temp_e
elif not self.energy_only and not self.precomputed_features:
with tf.GradientTape() as tape2:
tape2.watch(x)
feat_flat = self.feat_layer(x)
feat_flat_std = self.std_layer(feat_flat, training=training)
temp_hidden = self.mlp_layer(feat_flat_std, training=training)
temp_e = self.energy_layer(temp_hidden)
temp_g = tape2.batch_jacobian(temp_e, x)
y_pred = [temp_e, temp_g]
elif self.precomputed_features:
x1 = x[0]
feat_flat_std = self.std_layer(x1, training=training)
temp_hidden = self.mlp_layer(feat_flat_std, training=training)
temp_e = self.energy_layer(temp_hidden)
y_pred = temp_e
return y_pred
@tf.function
def predict_chunk_feature(self, tf_x, training=False):
with tf.GradientTape() as tape2:
tape2.watch(tf_x)
feat_pred = self.feat_layer(tf_x,
training=training) # Forward pass
grad = tape2.batch_jacobian(feat_pred, tf_x)
return feat_pred, grad
def precompute_feature_in_chunks(self, x, batch_size, training=False):
np_x = []
np_grad = []
for j in range(int(np.ceil(len(x) / batch_size))):
a = int(batch_size * j)
b = int(batch_size * j + batch_size)
tf_x = tf.convert_to_tensor(x[a:b], dtype=tf.float32)
feat_pred, grad = self.predict_chunk_feature(tf_x,
training=training)
np_x.append(np.array(feat_pred.numpy()))
np_grad.append(np.array(grad.numpy()))
np_x = np.concatenate(np_x, axis=0)
np_grad = np.concatenate(np_grad, axis=0)
return np_x, np_grad
def fit(self, **kwargs):
return super(EnergyModel, self).fit(**kwargs)
def get_config(self):
conf = {}
conf.update({
'states': self.out_dim,
'atoms': self.in_atoms,
'invd_index': self.invd_index,
'angle_index': self.angle_index,
'dihed_index': self.dihed_index,
'nn_size': self.nn_size,
'depth': self.depth,
'activ': self.activ,
'use_reg_activ': self.use_reg_activ,
'use_reg_weight': self.use_reg_weight,
'use_reg_bias': self.use_reg_bias,
'use_dropout': self.use_dropout,
'dropout': self.dropout,
'normalization_mode': self.normalization_mode,
"energy_only": self.energy_only,
"precomputed_features": self.precomputed_features,
"model_module": self.model_module
})
return conf
def save(self, filepath, **kwargs):
# copy to new model
self_conf = self.get_config()
self_conf['precomputed_features'] = False
copy_model = EnergyModel(**self_conf)
copy_model.set_weights(self.get_weights())
# Make graph and test with training data
copy_model.predict(np.ones((1, self.in_atoms, 3)))
tf.keras.models.save_model(copy_model, filepath, **kwargs)