def create_encoder_mutinfo(params):
""" Encoder without batch correction."""
input_shape = (params['datadims'],)
latent_dim = params['latentdims']
encoder_inputs = keras.Input(shape=input_shape,
name='input_data')
x = encoder_inputs
#xinit = layers.Dropout(params['inputdropout'])(x)
xinit = x
nhidden_e = params['nhidden_e']
xinit = layers.Dense(nhidden_e, activation="relu")(xinit)
for _ in range(params['nlayers_e']):
x = layers.Dense(nhidden_e, activation="relu")(xinit)
#x = layers.Dropout(params['hidden_e_dropout'])(x)
x = layers.Dense(nhidden_e)(x)
x = layers.Add()([x, xinit])
xinit = layers.Activation(activation='relu')(x)
x = xinit
z_mean = layers.Dense(latent_dim, name="z_mean")(x)
z_log_var = layers.Dense(latent_dim, name="z_log_var")(x)
z_log_var = ClipLayer(-10., 10.)(z_log_var)
z_mean = MutInfoLayer()(z_mean)
z_mean, z_log_var = KLlossLayer()([z_mean, z_log_var])
z = Sampling(params['nsamples'], name='random_latent')([z_mean, z_log_var])
encoder = keras.Model(encoder_inputs, z, name="encoder")
return encoder
def create_batch_encoder_gan_lastlayer(params):
""" With batch-adversarial learning on last latent dims."""
input_shape = (params['datadims'],)
latent_dim = params['latentdims']
encoder_inputs = keras.Input(shape=input_shape,
name='input_data')
x = encoder_inputs
xinit = layers.Dropout(params['inputdropout'])(x)
batches = []
nhidden_e = params['nhidden_e']
xinit = layers.Dense(nhidden_e)(xinit)
#batches.append(create_batch_net(xinit, params, '00'))
xinit = layers.Activation(activation='relu')(xinit)
for i in range(params['nlayers_e']):
x = layers.Dense(nhidden_e, activation="relu")(xinit)
x = layers.Dropout(params['hidden_e_dropout'])(x)
x = layers.Dense(nhidden_e)(x)
x = layers.Add()([x, xinit])
#batches.append(create_batch_net(x, params, f'1{i}'))
xinit = layers.Activation(activation='relu')(x)
x = xinit
z_mean = layers.Dense(latent_dim, name="z_mean")(x)
z_log_var = layers.Dense(latent_dim, name="z_log_var")(x)
z_log_var = ClipLayer(-10., 10.)(z_log_var)
z_mean, z_log_var = KLlossLayer()([z_mean, z_log_var])
z = Sampling(params['nsamples'], name='random_latent')([z_mean, z_log_var])
batches.append(create_batch_net(z, params, f'20'))
pred_batches = combine_batch_net(batches)
batch_inputs = [keras.Input(shape=(ncat,), name='batch_input') for bname, ncat in zip(params['batchnames'], params['nbatchcats'])]
true_batch_layer = [ExpandDims()(l) for l in batch_inputs]
batch_loss = BatchLoss(name='batch_loss')([pred_batches, true_batch_layer])
encoder = keras.Model([encoder_inputs, batch_inputs], [z, batch_loss], name="encoder")
return encoder
def create_batchlatent_decoder(params):
warnings.warn("create_batchlatent_decoder is experimental and may be removed.",
type=DeprecationWarning)
nsamples = params['nsamples']
input_shape = (params['datadims'],)
latent_dim = params['latentdims']
batch_dim = params['batchnames']
latent_inputs = keras.Input(shape=(nsamples, latent_dim,), name='latent_input')
batch_latent = keras.Input(shape=(params['nlasthiddenbatcher'],), name='batch_latent')
batch_layer = layers.RepeatVector(nsamples)(batch_latent)
#print(batch_layer)
x = layers.Concatenate()([latent_inputs, batch_layer])
#print(x)
for nhidden in range(params['nlayers_d']):
x = layers.Dense(params['nhiddendecoder'], activation="relu")(x)
x = layers.Dropout(params['hidden_d_dropout'])(x)
target_inputs = keras.Input(shape=input_shape, name='targets')
targets = layers.Reshape((1, params['datadims']))(target_inputs)
# multinomial part
logits = layers.Dense(params['datadims'],
activation='linear', name='logits',
use_bias=False)(x)
logits = AddBiasLayer(name='extra_bias')(logits)
# dispersion parameter
r = ScalarBiasLayer()(x)
r = layers.Activation(activation=tf.math.softplus)(r)
r = ClipLayer(1e-10, 1e5)(r)
prob_loss = NegativeMultinomialEndpoint()([logits, r, targets])
decoder = keras.Model([latent_inputs, target_inputs, batch_latent],
prob_loss, name="decoder")
decoder.summary()
return decoder
def create_batch_decoder(params):
nsamples = params['nsamples']
input_shape = (params['datadims'],)
latent_dim = params['latentdims']
batch_dim = params['batchnames']
latent_inputs = keras.Input(shape=(nsamples, latent_dim,), name='latent_input')
batch_inputs = [keras.Input(shape=(ncat,), name='batch_input') for bname, ncat in zip(params['batchnames'], params['nbatchcats'])]
batch_layer = batch_inputs
if len(batch_layer)>1:
batch_layer = layers.Concatenate()(batch_layer)
else:
batch_layer = batch_layer[0]
batch_layer = layers.RepeatVector(nsamples)(batch_layer)
x = layers.Concatenate()([latent_inputs, batch_layer])
for nhidden in range(params['nlayers_d']):
x = layers.Dense(params['nhiddendecoder'], activation="relu")(x)
x = layers.Dropout(params['hidden_d_dropout'])(x)
target_inputs = keras.Input(shape=input_shape, name='targets')
targets = layers.Reshape((1, params['datadims']))(target_inputs)
# multinomial part
logits = layers.Dense(params['datadims'],
activation='linear', name='logits',
use_bias=False)(x)
logits = AddBiasLayer(name='extra_bias')(logits)
# dispersion parameter
r = ScalarBiasLayer()(x)
r = layers.Activation(activation=tf.math.softplus)(r)
r = ClipLayer(1e-10, 1e5)(r)
prob_loss = NegativeMultinomialEndpoint()([logits, r, targets])
decoder = keras.Model([latent_inputs, target_inputs, batch_inputs],
prob_loss, name="decoder")
return decoder
def create_batch_encoder_alllayers(params):
""" Condition on batches in all hidden layers."""
input_shape = (params['datadims'],)
latent_dim = params['latentdims']
encoder_inputs = keras.Input(shape=input_shape,
name='input_data')
x = encoder_inputs
xinit = layers.Dropout(params['inputdropout'])(x)
batch_inputs = [keras.Input(shape=(ncat,), name='batch_input') for bname, ncat in zip(params['batchnames'], params['nbatchcats'])]
batch_layer = batch_inputs
if len(batch_layer)>1:
batch_layer = layers.Concatenate()(batch_layer)
else:
batch_layer = batch_layer[0]
xinit = layers.Concatenate()([xinit, batch_layer])
nhidden_e = params['nhidden_e']
xinit = layers.Dense(nhidden_e, activation="relu")(xinit)
for _ in range(params['nlayers_e']):
x = layers.Dense(nhidden_e, activation="relu")(xinit)
x = layers.Dropout(params['hidden_e_dropout'])(x)
x = layers.Dense(nhidden_e)(x)
xbatch = layers.Dense(nhidden_e)(batch_layer)
x = layers.Add()([x, xinit, xbatch])
xinit = layers.Activation(activation='relu')(x)
x = xinit
z_mean = layers.Dense(latent_dim, name="z_mean")(x)
z_log_var = layers.Dense(latent_dim, name="z_log_var")(x)
z_log_var = ClipLayer(-10., 10.)(z_log_var)
z_mean, z_log_var = KLlossLayer()([z_mean, z_log_var])
z = Sampling(params['nsamples'], name='random_latent')([z_mean, z_log_var])
encoder = keras.Model([encoder_inputs, batch_inputs], z, name="encoder")
return encoder