def __init__(self,
prefix=None,
optimizer=None,
layers=None,
loss=None,
metrics=None,
epochs=10,
batch_size=32,
callbacks=None,
input_shape=(40, 400, 1)
):
# Default parameter settings -------------------------------
if optimizer is None:
optimizer = {"SGD": {'lr': 0.01,
'decay': 1e-6,
'momentum': 0.9,
'nesterov': True}}
if loss is None:
loss = "mean_squared_error"
if metrics is None:
metrics = ['acc', spIndex]
get_custom_objects().update({"spIndex": spIndex})
# Place input shape on first layer parameter list
# layers[0][1]['input_shape'] = input_shape
# Setting parameters --------------------------------------
self.__dict__ = OrderedDict()
self.__dict__['input_shape'] = input_shape
self.__dict__['prefix'] = prefix
self.__dict__['optimizer'] = Optimizer(optimizer)
self.__dict__['layers'] = Layers()
if layers is not None:
for layer in layers:
# name = layer['type']
# args = layer
# del args['type']
self.layers.add(layer)
self.__dict__['callbacks'] = callbacks
if not callbacks is None:
for args in callbacks:
self.callbacks.add(args)
self.__dict__['loss'] = loss[0] if isinstance(loss, list) else loss
self.__dict__['metrics'] = metrics
self.__dict__['epochs'] = epochs
self.__dict__['batch_size'] = batch_size
def load(self, fullpath=None):
""" Load model """
fullpath = fullpath if fullpath else self.filename
logger.debug("Loading model: '%s'", fullpath)
try:
network = load_model(self.filename, custom_objects=get_custom_objects())
except ValueError as err:
if str(err).lower().startswith("cannot create group in read only mode"):
self.convert_legacy_weights()
return True
logger.warning("Failed loading existing training data. Generating new models")
logger.debug("Exception: %s", str(err))
return False
except OSError as err: # pylint: disable=broad-except
logger.warning("Failed loading existing training data. Generating new models")
logger.debug("Exception: %s", str(err))
return False
self.config = network.get_config()
self.network = network # Update network with saved model
self.network.name = self.type
return True
# Arguments
x: Tensor or variable.
kernel: kernel tensor.
strides: strides tuple.
padding: string, `"same"` or `"valid"`.
data_format: string, `"channels_last"` or `"channels_first"`.
Whether to use Theano or TensorFlow data format
for inputs/kernels/ouputs.
dilation_rate: tuple of 2 integers.
# Returns
A tensor, result of 2D convolution.
# Raises
ValueError: if `data_format` is neither `channels_last` or `channels_first`.
"""
# Transform the filters
transformed_filter = transform_filter_2d_nhwc(w=kernel,
flat_indices=gconv_indices,
shape_info=gconv_shape_info)
return K.conv2d(x=x,
kernel=transformed_filter,
strides=strides,
padding=padding,
data_format=data_format,
dilation_rate=dilation_rate)
get_custom_objects().update({'GConv2D': GConv2D})
def run():
args = initialize_parameters()
get_custom_objects()['PermanentDropout'] = PermanentDropout
model = keras.models.load_model(args.model_file, compile=False)
model.load_weights(args.weights_file)
# model.summary()
df_expr, df_desc = prepare_data(sample_set=args.sample_set,
drug_set=args.drug_set,
use_landmark_genes=args.use_landmark_genes)
print('total available samples: ', df_expr[['Sample']].shape[0])
print('total available drugs: ', df_desc[['Drug']].shape[0])
if args.ns > 0 and args.si > 0:
df_sample_ids = df_expr[['Sample']].iloc[args.si:args.si + args.ns]
elif args.si > 0:
df_sample_ids = df_expr[['Sample']].iloc[args.si:]
elif args.ns > 0:
df_sample_ids = df_expr[['Sample']].head(args.ns)
else:
df_sample_ids = df_expr[['Sample']].copy()
if args.nd > 0:
df_drug_ids = df_desc[['Drug']].head(args.nd)
else:
df_drug_ids = df_desc[['Drug']].copy()
df_sum = cross_join3(df_sample_ids,
df_drug_ids,
df_drug_ids,
suffixes=('1', '2'))
n_samples = df_sample_ids.shape[0]
n_drugs = df_drug_ids.shape[0]
n_rows = n_samples * n_drugs * n_drugs
print(
'Predicting drug response for {} combinations: {} samples x {} drugs x {} drugs'
.format(n_rows, n_samples, n_drugs, n_drugs))
n = args.n_pred
df_sum['N'] = n
df_seq = pd.DataFrame({'Seq': range(1, n + 1)})
df_all = cross_join(df_sum, df_seq)
total = df_sum.shape[0]
for i in tqdm(range(0, total, args.step)):
j = min(i + args.step, total)
x_all_list = []
df_x_all = pd.merge(df_all[['Sample']].iloc[i:j],
df_expr,
on='Sample',
how='left')
x_all_list.append(df_x_all.drop(['Sample'], axis=1).values)
drugs = ['Drug1', 'Drug2']
for drug in drugs:
df_x_all = pd.merge(df_all[[drug]].iloc[i:j],
df_desc,
left_on=drug,
right_on='Drug',
how='left')
x_all_list.append(df_x_all.drop([drug, 'Drug'], axis=1).values)
preds = []
for k in range(n):
y_pred = model.predict(x_all_list,
batch_size=args.batch_size,
verbose=0).flatten()
preds.append(y_pred)
df_all.loc[i * n + k:(j - 1) * n + k:n, 'PredGrowth'] = y_pred
df_all.loc[i * n + k:(j - 1) * n + k:n, 'Seq'] = k + 1
if n > 0:
df_sum.loc[i:j - 1, 'PredGrowthMean'] = np.mean(preds, axis=0)
df_sum.loc[i:j - 1, 'PredGrowthStd'] = np.std(preds, axis=0)
df_sum.loc[i:j - 1, 'PredGrowthMin'] = np.min(preds, axis=0)
df_sum.loc[i:j - 1, 'PredGrowthMax'] = np.max(preds, axis=0)
# df = df_all.copy()
# df['PredCustomComboScore'] = df.apply(lambda x: custom_combo_score(x['PredGrowth'],
# lookup(df, x['Sample'], x['Drug1'], value='PredGrowth'),
# lookup(df, x['Sample'], x['Drug2'], value='PredGrowth')), axis=1)
csv_all = 'comb_pred_{}_{}.all.tsv'.format(args.sample_set, args.drug_set)
df_all.to_csv(csv_all, index=False, sep='\t', float_format='%.4f')
if n > 0:
csv = 'comb_pred_{}_{}.tsv'.format(args.sample_set, args.drug_set)
df_sum.to_csv(csv, index=False, sep='\t', float_format='%.4f')
config = {
'alpha_initializer':
initializers.serialize(self.alpha_initializer),
'alpha_regularizer':
regularizers.serialize(self.alpha_regularizer),
'alpha_constraint': constraints.serialize(self.alpha_constraint),
'beta_initializer': initializers.serialize(self.beta_initializer),
'beta_regularizer': regularizers.serialize(self.beta_regularizer),
'beta_constraint': constraints.serialize(self.beta_constraint),
'shared_axes': self.shared_axes
}
base_config = super(PELU, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
get_custom_objects().update({'PELU': PELU})
class SReLU(Layer):
"""S-shaped Rectified Linear Unit.
It follows:
`f(x) = t^r + a^r(x - t^r) for x >= t^r`,
`f(x) = x for t^r > x > t^l`,
`f(x) = t^l + a^l(x - t^l) for x <= t^l`.
# Input shape
Arbitrary. Use the keyword argument `input_shape`
(tuple of integers, does not include the samples axis)
when using this layer as the first layer in a model.
# we won't even calculate the output of those steps, saving
# some real computational time.
if self.zeros_like_input is None:
self.zeros_like_input = K.zeros_like(inputs,
name='zeros_like_input')
# just because K.any(step_is_active) doesn't work in PlaidML
any_step_is_active = K.greater(K.sum(K.cast(step_is_active, 'int32')),
0)
step_weighted_output = K.switch(
any_step_is_active,
K.expand_dims(halting_prob, -1) * inputs, self.zeros_like_input)
if self.weighted_output is None:
self.weighted_output = step_weighted_output
else:
self.weighted_output += step_weighted_output
return [inputs, self.weighted_output]
def compute_output_shape(self, input_shape):
return [input_shape, input_shape]
def finalize(self):
self.add_loss(self.ponder_cost)
get_custom_objects().update({
'LayerNormalization': LayerNormalization,
'TransformerTransition': TransformerTransition,
'TransformerACT': TransformerACT,
'gelu': gelu,
})
from callbacks import EpochLog
from my_json_encoder import MyJsonEncoder
parser = argparse.ArgumentParser()
parser.add_argument('--bn', action='store_true')
parser.add_argument('--gap', action='store_true')
parser.add_argument('--full', action='store_true')
parser.add_argument('--da', action='store_true')
args = parser.parse_args()
bn_enabled = args.bn
gap_enabled = args.gap
full_enabled = args.full
da_enabled = args.da
# Register custom activation function
get_custom_objects().update({'mish': functions.mish})
# Declare constants
CUR_DIR = os.path.dirname(os.path.abspath(__file__))
BASE_FILE_NAME = os.path.splitext(os.path.basename(__file__))[0]
if bn_enabled:
BASE_FILE_NAME += '-BN'
if gap_enabled:
BASE_FILE_NAME += '-GAP'
if full_enabled:
BASE_FILE_NAME += '-full'
if da_enabled:
BASE_FILE_NAME += '-DA'
LOG_FILE_PATH = f'{CUR_DIR}/epoch_logs/{BASE_FILE_NAME}_log.txt'
# MODEL_FILE_PATH = f'{CUR_DIR}/saved_models/{BASE_FILE_NAME}_model.h5'
MODEL_WEIGHTS_PATH = f'{CUR_DIR}/saved_models/{BASE_FILE_NAME}_weights.h5'
from keras.layers import BatchNormalization
from keras.models import Model
from keras import backend as K
from keras.engine.topology import get_source_inputs
from keras.utils import layer_utils, get_custom_objects
from keras.utils.data_utils import get_file
WEIGHTS_PATH = 'https://github.com/fchollet/deep-learning-models/releases/download/v0.5/inception_v3_weights_tf_dim_ordering_tf_kernels.h5'
WEIGHTS_PATH_NO_TOP = 'https://github.com/fchollet/deep-learning-models/releases/download/v0.5/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5'
def swish(x):
return (K.sigmoid(x) * x)
get_custom_objects().update({'swish': swish})
def conv2d_bn(x,
filters,
num_row,
num_col,
padding='same',
strides=(1, 1),
name=None):
"""Utility function to apply conv + BN.
# Arguments
x: input tensor.
filters: filters in `Conv2D`.
num_row: height of the convolution kernel.
'The layer can be called only with one tensor as an argument')
_, seq_len, d_model = K.int_shape(inputs)
# The first thing we need to do is to perform affine transformations
# of the inputs to get the Queries, the Keys and the Values.
qkv = K.dot(K.reshape(inputs, [-1, d_model]), self.qkv_weights)
# splitting the keys, the values and the queries before further
# processing
pre_q, pre_k, pre_v = [
K.reshape(
# K.slice(qkv, (0, i * d_model), (-1, d_model)),
qkv[:, i * d_model:(i + 1) * d_model],
(-1, seq_len, self.num_heads, d_model // self.num_heads))
for i in range(3)
]
attention_out = self.attention(pre_q,
pre_v,
pre_k,
seq_len,
d_model,
training=kwargs.get('training'))
return attention_out
def compute_output_shape(self, input_shape):
return input_shape
get_custom_objects().update({
'MultiHeadSelfAttention': MultiHeadSelfAttention,
'MultiHeadAttention': MultiHeadAttention,
})
from keras.layers import GlobalMaxPooling2D
from keras.layers import BatchNormalization
from keras.models import Model
from keras import backend as K
from keras.engine.topology import get_source_inputs
from keras.utils import layer_utils, get_custom_objects
from keras.utils.data_utils import get_file
WEIGHTS_PATH = 'https://github.com/fchollet/deep-learning-models/releases/download/v0.5/inception_v3_weights_tf_dim_ordering_tf_kernels.h5'
WEIGHTS_PATH_NO_TOP = 'https://github.com/fchollet/deep-learning-models/releases/download/v0.5/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5'
def swish(x):
return (K.sigmoid(x) * x)
get_custom_objects().update({'swish': swish})
def conv2d_bn(x,
filters,
num_row,
num_col,
padding='same',
strides=(1, 1),
name=None):
"""Utility function to apply conv + BN.
# Arguments
x: input tensor.
filters: filters in `Conv2D`.
num_row: height of the convolution kernel.
num_col: width of the convolution kernel.
self.updates.append(K.update(vhat, vhat_t))
else:
denom = (K.sqrt(v_t) + self.epsilon)
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(v, v_t))
# Partial momentum adaption.
new_p = p - (lr_t * (m_t / (denom ** (self.partial * 2))))
# Apply constraints.
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def get_config(self):
config = {'lr': float(K.get_value(self.lr)),
'beta_1': float(K.get_value(self.beta_1)),
'beta_2': float(K.get_value(self.beta_2)),
'decay': float(K.get_value(self.decay)),
'epsilon': self.epsilon,
'amsgrad': self.amsgrad,
'partial': self.partial}
base_config = super(Padam, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
get_custom_objects().update({'Padam': Padam})
def register_permanent_dropout():
get_custom_objects()['PermanentDropout'] = PermanentDropout
import keras
from keras.utils import get_custom_objects
from keras import backend as K
class TauDecay(keras.callbacks.Callback):
def __init__(self, tau, **kargs):
self.tau = tau
self.anneal_rate = 0.00003
self.min_temperature = 0.1
def on_train_end(self, logs={}):
tau_value = K.get_value(self.tau)
print('Tau Value: %s' % tau_value)
def on_epoch_end(self, epoch, logs={}):
tau_value = K.get_value(self.tau)
print('Epoch %s, Current Tau Value: %s' % (epoch + 1, tau_value))
def on_batch_end(self, batch, logs=None):
if batch % 100 == 0:
decay_rate = np.exp(-self.anneal_rate * batch)
tau_value = K.get_value(self.tau) * decay_rate
tau_value = np.max([tau_value, self.min_temperature])
K.set_value(self.tau, tau_value)
get_custom_objects().update({
'TauDecay': TauDecay,
})
def build(self):
get_custom_objects().update({"swish": Swish(swish)})
# Backend 為 Tensorflow 定義 channel axis 為 3
channel_axis = 3
# 定義 Input 的大小
input_shape = Input(shape=(224, 224, 1), name='data')
# Stem layer
net = self.conv2d_bn(input_shape,
32,
3,
3,
strides=(2, 2),
padding='valid',
activate=self.activate)
net = self.conv2d_bn(net,
32,
3,
3,
strides=(1, 1),
padding='valid',
activate=self.activate)
net = self.conv2d_bn(net,
64,
3,
3,
strides=(1, 1),
activate=self.activate)
branch_0 = MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(net)
net = self.conv2d_bn(branch_0,
80,
3,
3,
strides=(2, 2),
padding='valid',
activate=self.activate)
net = self.conv2d_bn(net,
192,
3,
3,
strides=(1, 1),
padding='valid',
activate=self.activate)
x = MaxPooling2D((3, 3), strides=(2, 2), padding="valid")(net)
# inception1
branch_0 = self.conv2d_bn(x,
96,
1,
1,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(x,
64,
1,
1,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(branch_1,
96,
3,
3,
strides=(1, 1),
activate=self.activate)
branch_2 = self.conv2d_bn(x,
64,
1,
1,
strides=(1, 1),
activate=self.activate)
branch_2 = self.conv2d_bn(branch_2,
96,
3,
3,
strides=(1, 1),
activate=self.activate)
branch_2 = self.conv2d_bn(branch_2,
96,
3,
3,
strides=(1, 1),
activate=self.activate)
x = [branch_0, branch_1, branch_2]
mix1 = concatenate(x, axis=channel_axis)
x = self.conv2d_bn(mix1,
96,
1,
1,
strides=(1, 1),
padding='valid',
activate=self.activate)
# inception2
branch_0 = self.conv2d_bn(x,
64,
3,
3,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(x,
96,
1,
1,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(branch_1,
128,
3,
3,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(branch_1,
160,
3,
3,
strides=(1, 1),
activate=self.activate)
branch_3 = AveragePooling2D((3, 3),
strides=(1, 1),
padding='same',
name="avg_pool_1")(x)
if (self.use_dense_block):
x1 = [x, branch_0, branch_1, branch_3]
else:
x1 = [branch_0, branch_1, branch_3]
mix2 = concatenate(x1, axis=channel_axis)
# inception3
branch_0 = self.conv2d_bn(mix2,
192,
1,
1,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(mix2,
128,
1,
1,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(branch_1,
160,
1,
7,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(branch_1,
160,
7,
1,
strides=(1, 1),
activate=self.activate)
if (self.use_dense_block):
x = [x, mix2, branch_0, branch_1]
else:
x = [mix2, branch_0, branch_1]
mix3 = concatenate(x, axis=channel_axis, name='mixed3')
# translate layer
if (self.use_dense_block):
x = self.conv2d_bn(mix3,
192,
1,
1,
strides=(1, 1),
padding='valid',
activate=self.activate)
x = AveragePooling2D((2, 2), strides=(2, 2))(x)
x1 = BatchNormalization(scale=True, axis=channel_axis)(x)
else:
x1 = self.conv2d_bn(mix3,
192,
1,
1,
strides=(1, 1),
padding='valid',
activate=self.activate)
# inception4
netb00 = self.conv2d_bn(x1,
192,
1,
1,
strides=(1, 1),
padding='same',
activate=self.activate)
netb10 = self.conv2d_bn(x1,
192,
1,
1,
strides=(1, 1),
padding='same',
activate=self.activate)
netb11 = self.conv2d_bn(netb10,
256,
3,
3,
strides=(1, 1),
padding='same',
activate=self.activate)
netb20 = self.conv2d_bn(x1,
160,
1,
1,
strides=(1, 1),
padding='same',
activate=self.activate)
netb21 = self.conv2d_bn(netb20,
192,
3,
3,
strides=(1, 1),
padding='same',
activate=self.activate)
netb22 = self.conv2d_bn(netb21,
256,
3,
3,
strides=(1, 1),
padding='same',
activate=self.activate)
netb30 = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x1)
netb31 = self.conv2d_bn(netb30,
160,
1,
1,
strides=(1, 1),
padding='same',
activate=self.activate)
if (self.use_dense_block):
x = concatenate([x, netb00, netb11, netb22, netb31],
axis=channel_axis,
name='mixed4')
else:
x = concatenate([netb00, netb11, netb22, netb31],
axis=channel_axis,
name='mixed4')
# inception5 * 2
feature_list = [x]
for _ in range(2):
branch_0 = self.conv2d_bn(x,
256,
1,
1,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(x,
128,
1,
3,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(branch_1,
192,
3,
1,
strides=(1, 1),
activate=self.activate)
branch_1 = self.conv2d_bn(branch_1,
256,
1,
3,
strides=(1, 1),
activate=self.activate)
a = [branch_0, branch_1]
mix5 = concatenate(a, axis=channel_axis)
x1 = self.conv2d_bn(mix5,
256,
1,
1,
strides=(1, 1),
padding='valid',
activate=self.activate)
x = concatenate([x, x1], axis=channel_axis)
feature_list.append(x)
if (self.use_global_average_pool):
x = concatenate(feature_list, axis=channel_axis)
if (self.use_global_average_pool):
# GlobalAveragePooling Layer
x = GlobalAveragePooling2D(name='global_avg_pool')(x)
else:
# Fully Connection Layer
x = Dense(2000)(x)
x = Dense(1000)(x)
x = Dense(7, name='Logits')(x)
x = Activation('softmax', name='prob')(x)
model = Model(inputs=input_shape, outputs=x, name='FNet')
return model
import matplotlib.pyplot as plt
import pandas as pd
from sklearn.preprocessing import StandardScaler
def swish(x):
return K.sigmoid(x) * x
class Swish(Activation):
def __init__(self, activation, **kwargs):
super(Swish, self).__init__(activation, **kwargs)
self.__name__ = 'swish'
get_custom_objects().update({"swish": Swish(swish)})
# 模型類別
class CustomModel(object):
# 卷積層 + Batch Normalization Layer + Activate Function
def conv2d_bn(self,
x,
filters,
num_row,
num_col,
padding='same',
strides=(1, 1),
name=None,
activate='relu'):
if name is not None:
from tensorflow.python.keras.models import load_model
from scipy import stats
import random as rnd
from matplotlib import gridspec
def pure_linear(x):
return x
def f(x):
return np.int(x)
f2 = np.vectorize(f)
get_custom_objects().update({'pure_linear': pure_linear})
data_Train_Input = np.loadtxt('/home/andrea/JET/84786_84798_asimV.txt') #targets of the training set
data_Train_Input = data_Train_Input.reshape((-1, 1))
data_Train_Target = np.loadtxt('/home/andrea/JET/84786_84798_efitV.txt') # inputs of the training set
data_Train_Target = data_Train_Target.reshape((-1, 1))
#% ------------------------------------------------------------------
#% Neural Network Layout
#% ------------------------------------------------------------------
data_Train_Input_norm = stats.zscore(data_Train_Input)
data_Train_Input_mean = np.mean(data_Train_Input)
def __init__(self,
prefix="convnet",
optimizer=None,
layers=None,
loss=None,
metrics=None,
epochs=10,
batch_size=32,
callbacks=[],
scale=False,
input_shape=(28, 28, 1)
):
# Default parameter settings -------------------------------
if optimizer is None:
optimizer = ["SGD", {'lr': 0.01,
'decay': 1e-6,
'momentum': 0.9,
'nesterov': True}]
if layers is None:
layers = [["Conv2D", {"filters": 6,
"kernel_size": (4, 4),
"strides": 1,
"data_format": "channels_last",
"padding": "same"
}
],
["Activation", {"activation": "relu"}],
["MaxPooling2D", {"pool_size": (2, 2),
"padding": "valid",
"strides": None}
],
["Flatten", {}],
["Dense", {"units": 50}],
["Activation", {"activation": "tanh"}],
["Dense", {"units": 4}],
["Activation", {"activation": "softmax"}]]
if loss is None:
loss = "mean_squared_error"
if metrics is None:
metrics = ['acc', spIndex]
get_custom_objects().update({"spIndex": spIndex})
# Place input shape on first layer parameter list
layers[0][1]['input_shape'] = input_shape
# Setting parameters --------------------------------------
self.__dict__ = OrderedDict()
self.__dict__['input_shape'] = input_shape
self.__dict__['prefix'] = prefix
self.list2Optimizer(optimizer)
self.list2layers(layers)
self.__dict__['loss'] = loss[0] if isinstance(loss, list) else loss
self.__dict__['metrics'] = metrics
self.__dict__['epochs'] = epochs
self.__dict__['batch_size'] = batch_size
self.__dict__['scale'] = scale
if isinstance(callbacks, Callbacks) or callbacks is None:
self.__dict__['callbacks'] = Callbacks
elif isinstance(layers, list):
self.__dict__['callbacks'] = Callbacks(callbacks)
else:
raise ValueError('layers must be an instance of Layers or list'
'%s of type %s was passed' % (callbacks, type(callbacks)))
regularizers.serialize(self.bias_regularizer),
'activity_regularizer':
regularizers.serialize(self.activity_regularizer),
'kernel_constraint':
constraints.serialize(self.kernel_constraint),
'bias_constraint':
constraints.serialize(self.bias_constraint),
'use_bias':
self.use_bias
}
base_config = super(CosineConvolution2D, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
CosineConv2D = CosineConvolution2D
get_custom_objects().update({'CosineConvolution2D': CosineConvolution2D})
get_custom_objects().update({'CosineConv2D': CosineConv2D})
class SubPixelUpscaling(Layer):
""" Sub-pixel convolutional upscaling layer.
This layer requires a Convolution2D prior to it,
having output filters computed according to
the formula :
filters = k * (scale_factor * scale_factor)
where k = a user defined number of filters (generally larger than 32)
scale_factor = the upscaling factor (generally 2)
This layer performs the depth to space operation on
base_lr = self._optimizer.lr
for param, multiplier in mul_lr_params.items():
self._optimizer.lr = base_lr * multiplier
updates.extend(self._optimizer.get_updates(loss, [param]))
self._optimizer.lr = base_lr
updates.extend(self._optimizer.get_updates(loss, base_lr_params))
return updates
def get_config(self):
config = {
'optimizer': self._class,
'lr_multipliers': self._lr_multipliers
}
base_config = self._optimizer.get_config()
# noinspection PyTypeChecker
return dict(list(base_config.items()) + list(config.items()))
def __getattr__(self, name):
return getattr(self._optimizer, name)
def __setattr__(self, name, value):
if name.startswith('_'):
super(LearningRateMultiplier, self).__setattr__(name, value)
else:
self._optimizer.__setattr__(name, value)
get_custom_objects().update({'LearningRateMultiplier': LearningRateMultiplier})
shape=(d_model, ),
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True)
return super().build(input_shape)
def call(self, inputs, mask=None, training=None):
input_shape = K.shape(inputs)
d_model = input_shape[-1]
step1 = self.activation(
K.bias_add(K.dot(K.reshape(inputs, (-1, d_model)), self.weights1),
self.biases1,
data_format='channels_last'))
if 0.0 < self.dropout_rate < 1.0:
def dropped_inputs():
return K.dropout(step1, self.dropout_rate, K.shape(step1))
step1 = K.in_train_phase(dropped_inputs, step1, training=training)
step2 = K.bias_add(K.dot(step1, self.weights2),
self.biases2,
data_format='channels_last')
result = K.reshape(step2, (-1, input_shape[-2], input_shape[-1]))
return result
get_custom_objects().update({
'TransformerTransition': TransformerTransition,
})
model.add(SineReLU())
model.add(Dropout(0.3))
model.add(Dense(1024))
model.add(SineReLU())
model.add(Dropout(0.5))
model.add(Dense(10, activation = 'softmax'))
```
"""
def __init__(self, epsilon=0.0025, **kwargs):
super(SineReLU, self).__init__(**kwargs)
self.supports_masking = True
self.epsilon = K.cast_to_floatx(epsilon)
def call(self, Z):
m = self.epsilon * (K.sin(Z) - K.cos(Z))
A = K.maximum(m, Z)
return A
def get_config(self):
config = {'epsilon': float(self.epsilon)}
base_config = super(SineReLU, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def compute_output_shape(self, input_shape):
return input_shape
get_custom_objects().update({'SineReLU': SineReLU})
name='kernel',
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint)
def call(self, inputs):
# l2 normalize parameters
norm_x = K.l2_normalize(inputs)
norm_w = K.l2_normalize(self.kernel)
# compute arc distance
output = K.dot(norm_x, norm_w)
return output
def compute_output_shape(self, input_shape):
return (None, self.units)
def get_config(self):
config = {
'units': self.units,
'kernel_initializer':
initializers.serialize(self.kernel_initializer),
'kernel_regularizer':
regularizers.serialize(self.kernel_regularizer),
'kernel_constraint': constraints.serialize(self.kernel_constraint),
}
base_config = super().get_config()
return dict(list(base_config.items()) + list(config.items()))
get_custom_objects().update({'ArcDense': ArcDense})
def call(self, inputs, **kwargs):
block_depth = kwargs.get('step')
sf_depth = kwargs.get('src_fact_step')
if block_depth is None:
raise ValueError("Please, provide current Transformer's step"
"using 'step' keyword argument.")
if sf_depth is None:
raise ValueError("Please, provide src and facts' step"
"using 'src_fact_step' keyword argument.")
result = inputs + self.word_position_embeddings
if block_depth is not None:
result = result + self.depth_embeddings[block_depth]
if sf_depth is not None:
if sf_depth == 0:
result = result + self.sf_depth_embeddings[sf_depth]
elif sf_depth == 1:
result = K.permute_dimensions(result, [0, 2, 1, 3])
result = result + self.sf_depth_embeddings[1:, :]
result = K.permute_dimensions(result, [0, 2, 1, 3])
return result
get_custom_objects().update({
'TransformerCoordinateEmbedding': TransformerCoordinateEmbedding,
'SrcFactTransformerCoordinateEmbedding':
SrcFactTransformerCoordinateEmbedding,
'AddCoordinateEncoding': AddCoordinateEncoding,
'AddPositionalEncoding': AddPositionalEncoding,
})
self.model.save(file)
def load(self, file='D:/TensorFlow/stock_data/lstm_28.h5'):
self.model = load_model(
file, custom_objects={'risk_estimation': risk_estimation})
def predict(self, test):
predict = []
for sample_index in range(test.shape[0]):
test_data = test[sample_index].reshape(1, time_step, input_size)
prev = self.model.predict(test_data)
predict.append(prev)
return np.array(predict)
get_custom_objects().update({'ReLU': ReLU})
model = Model(input_shape=(time_step, input_size), loss=risk_estimation)
net = model.lstmModel()
model.load()
# 训练模型
model.train()
# 储存模型
model.save()
# 读入模型
model.load()
# 预测
predict = model.predict(test_x)
predict = predict.reshape(-1, output_size)
print(predict)
batch_token_ids = []
for i in range(0, len(X)):
token_ids = X[i].tolist()
segment_ids = [0 for token in X[i]]
batch_token_ids.append(token_ids)
batch_segment_ids.append(segment_ids)
return [np.array(batch_token_ids), np.array(batch_segment_ids)]
model_path = "model_75_epochs_no_earlystopping.hd5"
print("Trained model path : %s" % model_path)
test_filename = "data/trial_data.txt"
print("Test dataset path : %s" % test_filename)
results_path = "data/res/submission.txt"
print("Results path : %s" % results_path)
# ALBERT predictions
print("\n === ALBERT predictions ===\n")
X_test, _, _ = create_dataset(test_filename)
model = load_model(model_path, custom_objects=get_custom_objects())
model.summary()
predictions = model.predict(data_creator(X_test))
word_id_lsts, post_lsts, _, _, _, _ = read_data(test_filename)
predictions_unpadded = unpad(np.array(predictions), word_id_lsts)
write_results(word_id_lsts, post_lsts, predictions_unpadded, results_path)
print("Results written")
def __init__(self, save_all_models=False):
Callback.__init__(self)
self.save_all_models = save_all_models
get_custom_objects()['PermanentDropout'] = PermanentDropout
features extract layer , for extracting the CLS vector generally
"""
def __init__(self, index=0, **kwargs):
super(Extract, self).__init__(**kwargs)
self.index = index
self.supports_masking = True
def get_config(self):
config = {
'index': self.index,
}
base_config = super(Extract, self).get_config()
base_config.update(config)
return base_config
def compute_output_shape(self, input_shape):
# [N, emb_dim]
return input_shape[:1] + input_shape[2:]
def compute_mask(self, inputs, mask=None):
return None
def call(self, x, mask=None):
# [N, emd_dim]
return x[:, self.index]
get_custom_objects().update(custom_config())
K.transpose(embedding_matrix))
if self.add_biases:
projected = K.bias_add(projected,
self.biases,
data_format='channels_last')
if 0 < self.projection_dropout < 1:
projected = K.in_train_phase(
lambda: K.dropout(projected, self.projection_dropout),
projected,
training=kwargs.get('training'))
attention = projected #K.dot(projected, K.transpose(embedding_matrix))
if self.scaled_attention:
# scaled dot-product attention, described in
# "Attention is all you need" (https://arxiv.org/abs/1706.03762)
sqrt_d = K.constant(math.sqrt(emb_output_dim), dtype=K.floatx())
attention = attention / sqrt_d
result = K.reshape(
self.activation(attention),
(input_shape_tensor[0], input_shape_tensor[1], emb_input_dim))
return result
def compute_output_shape(self, input_shape):
main_input_shape, embedding_matrix_shape = input_shape
emb_input_dim, emb_output_dim = embedding_matrix_shape
return main_input_shape[0], main_input_shape[1], emb_input_dim
get_custom_objects().update({
'ReusableEmbedding': ReusableEmbedding,
'TiedOutputEmbedding': TiedOutputEmbedding,
})
def get_model(self, pad_id):
self.pad_id = pad_id
inp_src = Input(name='src_input', shape=(None, ), dtype='int32')
inp_answer = Input(
name='answer_input',
shape=(None, ),
dtype='int32',
)
encoder_output = self.__get_encoder(inp_src)
mutual_attn_mask = PaddingMaskLayer(name='decoder_mutual_padding_mask',
src_len=self.args.tar_seq_length,
pad_id=self.pad_id)(inp_src)
decoder_output = self.__get_decoder(inp_answer, encoder_output,
mutual_attn_mask)
# build model part
word_predictions = self.output_softmax_layer(
self.output_layer([decoder_output, self.decoder_embedding_matrix]))
model = Model(inputs=[inp_src, inp_answer], outputs=[word_predictions])
return model
get_custom_objects().update({
'PaddingMaskLayer': PaddingMaskLayer,
'SequenceMaskLayer': SequenceMaskLayer,
})
# some real computational time.
if self.zeros_like_input is None:
self.zeros_like_input = K.zeros_like(
inputs, name='zeros_like_input')
# just because K.any(step_is_active) doesn't work in PlaidML
any_step_is_active = K.greater(
K.sum(K.cast(step_is_active, 'int32')), 0)
step_weighted_output = K.switch(
any_step_is_active,
K.expand_dims(halting_prob, -1) * inputs,
self.zeros_like_input)
if self.weighted_output is None:
self.weighted_output = step_weighted_output
else:
self.weighted_output += step_weighted_output
return [inputs, self.weighted_output]
def compute_output_shape(self, input_shape):
return [input_shape, input_shape]
def finalize(self):
self.add_loss(self.ponder_cost)
get_custom_objects().update({
'TransformerEncoderBlock': TransformerEncoderBlock,
'TransformerDecoderBlock': TransformerDecoderBlock,
'TransformerACT': TransformerACT,
'gelu': gelu,
})
Parameters
----------
inputs: tensor
Input tensor, or list/tuple of input tensors
"""
if get_backend() == "amd":
return inputs * K.sigmoid(inputs * self.beta)
# Native TF Implementation has more memory-efficient gradients
return tf.nn.swish(inputs * self.beta)
def get_config(self):
"""Returns the config of the layer.
Adds the :attr:`beta` to config.
Returns
--------
dict
A python dictionary containing the layer configuration
"""
config = super().get_config()
config["beta"] = self.beta
return config
# Update layers into Keras custom objects
for name, obj in inspect.getmembers(sys.modules[__name__]):
if inspect.isclass(obj) and obj.__module__ == __name__:
get_custom_objects().update({name: obj})
enc_state3 = self.memnn_encoder2([inp_q, inp_fact])
emb_ans = self.decoder_embedding(inp_tar)
emb_fact_ans = self.decoder_embedding(inp_fact_tar)
# task 1: seq2seq, input: question; output: answer
output1, state1 = self.decoder(emb_ans, initial_state=enc_state1)
# task 2: memnn, input: question and facts; output: fact
output2, state2 = self.decoder(emb_fact_ans, initial_state=enc_state2)
# task 3: memnn, input: question and facts; output: answer
output3, state3 = self.decoder(emb_ans, initial_state=enc_state3)
# final output
final_output1 = self.decoder_dense1(output1)
final_output2 = self.decoder_dense2(output2)
final_output3 = self.decoder_dense3(output3)
# define model
model = Model(inputs=[inp_q, inp_tar, inp_fact_tar, inp_fact],
outputs=[final_output1, final_output2, final_output3])
return model
get_custom_objects().update({
'PosEncodeEmbedding': PosEncodeEmbedding,
'MultiTaskModel': MultiTaskModel,
'S2SEncoder': S2SEncoder,
'MemNNEncoder': MemNNEncoder,
})
def custom_function_keras():
get_custom_objects().update({'swish': Activation(swish, name='swish')})
def __init__(self,
latent_dim,
output_dim,
batch_size=1000,
noise_dim=50,
n_epochs=1000,
hidden_layer_depth=3,
hidden_size=200,
activation='swish',
verbose=True,
n_discriminator=5):
self.latent_dim = latent_dim
self.output_dim = output_dim
self.hidden_layer_depth = hidden_layer_depth
self.hidden_size = hidden_size
self.activation = activation
self.batch_size = batch_size
self.n_epochs = n_epochs
self.noise_dim = noise_dim
self.verbose = verbose
self.n_discriminator = n_discriminator
optimizer = RMSprop(lr=0.00005)
# Register swish
get_custom_objects().update({'swish': swish})
# Build generator and discriminator
self.generator = self._build_generator()
self.discriminator = self._build_discriminator()
#
# Construct computational graph for discriminator
#
self.generator.trainable = False
# Real molecule
real_mol = Input(shape=(self.output_dim, ))
# Discriminator input
z_disc = Input(shape=(self.latent_dim + self.noise_dim, ))
fake_mol = self.generator(z_disc)
# conditioning input
z_cond = Input(shape=(self.latent_dim, ))
# Condition both fake and valid.
fake_mol_cond = Concatenate(axis=1)([fake_mol, z_cond])
real_mol_cond = Concatenate(axis=1)([real_mol, z_cond])
# Discriminator does its job
fake = self.discriminator(fake_mol_cond)
valid = self.discriminator(real_mol_cond)
# Interpolated between real and fake molecule.
interp_mol = RandomWeightedAverage(batch_size=self.batch_size)(
[real_mol_cond, fake_mol_cond])
validity_interp = self.discriminator(interp_mol)
# Use Python partial to provide loss function with additional
# 'averaged_samples' argument
partial_gp_loss = partial(self.ـgradient_penalty_loss,
averaged_samples=interp_mol)
partial_gp_loss.__name__ = 'gradient_penalty' # Keras requires function names
self.discriminator_model = Model(
inputs=[real_mol, z_disc, z_cond],
outputs=[valid, fake, validity_interp])
self.discriminator_model.compile(loss=[
self._wasserstein_loss, self._wasserstein_loss, partial_gp_loss
],
optimizer=optimizer,
loss_weights=[1, 1, 10])
#
# Construct computational graph for generator
#
self.discriminator.trainable = False
self.generator.trainable = True
z_gen = Input(shape=(self.latent_dim + self.noise_dim, ))
mol = self.generator(z_gen)
z_cond = Input(shape=(self.latent_dim, ))
mol_cond = Concatenate(axis=1)([mol, z_cond])
valid = self.discriminator(mol_cond)
self.generator_model = Model([z_gen, z_cond], valid)
self.generator_model.compile(loss=self._wasserstein_loss,
optimizer=optimizer)
self.is_fitted = False
