def get_loss_weight_metrics_for_plot(self):
metrics, names = [], []
contrast_increment = 0 # increment to help loss weight visualization
for l, loss in self.losses.items():
metrics.append(np.array(loss.weight_history) + contrast_increment)
names.append(l)
contrast_increment += 0.005
return [metrics], [1], [names], ['lines']
def get_metrics_for_plot(self):
metrics, iters, names, types = [], [], [], []
for m, metric in self.metrics.items():
data = metric.get_data_for_plot()
metrics.append(data)
names.append(m)
iters.append(self.notify_every)
types.append(metric.plot_type)
return metrics, iters, names, types
def get_loss_metrics_for_plot(self, plot_organization):
metrics, iters, names = [], [], []
for l in plot_organization:
if isinstance(l, (tuple, list)):
submetrics, subiters, subnames, _ = self.get_loss_metrics_for_plot(
l)
metrics.append(submetrics)
iters.append(subiters[0])
names.append(subnames)
else:
metrics.append(self.losses[l].history)
iters.append(1)
names.append(l)
return metrics, iters, names, ['lines'] * len(metrics)
def configureModel(input, outputLen = len(RD['labels'])):
print(' Initializing and compiling...')
alphaSize = input.shape[2]
model = Sequential()
'''
if RD['use_embedding']:
# second value in nomiSize tuple is shift while using embedding
model.add(Embedding(1 << nomiSize[1], RP['embedding_outputs']))
model.add(TimeDistributed(Dense(int(RP['td_layer_multiplier'] * (alphaSize +
nomiSize[0])), activation = 'tanh', trainable = RP['trainable_inner'])))
else:
'''
model.add(TimeDistributed(Dense(300*RG['ratios'][0], activation = 'tanh', trainable = RP['trainable_inner']), input_shape = (None, alphaSize )))
model.add(Dropout(0.30))
model.add(GRU(300*RG['ratios'][1], trainable = RP['trainable_inner'], return_sequences = True))
model.add(Activation('tanh', trainable = RP['trainable_inner']))
model.add(Dropout(0.30))
model.add(GRU(300*RG['ratios'][2], trainable = RP['trainable_inner']))
model.add(Activation('tanh', trainable = RP['trainable_inner']))
model.add(Dropout(0.30))
model.add(Dense(outputLen))
# molweight
# model = utility.loadModel('b3d9609da78bfbf0ad1a62ee6740df3b52f104b4', 'mol_')
# all compounds
# model = utility.loadModel('eab15a05a70b35d119c02fcc36b1cfaf27a0f36a', 'mol_')
# maccs
# model = utility.loadModel('67b51a1543b5d32b05671e4a08d193eed702ca54', 'mol_')
# model.pop()
# model.pop()
# for i in xrange(len(model.layers)):
# model.layers[0].trainable = False
'''
model.add(Dropout(0.50))
model.add(Dense(500))
model.add(Activation('relu'))
model.add(Dropout(0.50))
model.add(Dense(500))
model.add(Activation('relu'))
model.add(Dropout(0.30))
'''
# model.add(Dense(outputLen))
if RP['classify']:
model.add(Activation(RP['classify_activation'], trainable = RP['trainable_inner']))
metrics = []
if RP['classify']:
metrics.append('accuracy')
model.compile(loss = RP['objective'], optimizer = OPTIMIZER, metrics = metrics)
print(' ...done')
return model
def configureEdgeModel(inputSmiles, inputFasta):
print(' Initializing edge model and compiling...')
smilesGRUInputShape = (None, inputSmiles.shape[2])
# smilesGRUSize = int(RP['gru_layer_multiplier'] * smilesGRUInputShape[1])
fastaGRUInputShape = (None, inputFasta.shape[2])
# fastaGRUSize = int(RP['fasta_gru_layer_multiplier'] * fastaGRUInputShape[1])
mergedOutputLen = len(RD['labels'])
smilesModel = Sequential()
smilesModel.add(TimeDistributed(Dense(300, activation = 'tanh', trainable = RP['trainable_inner']), input_shape = smilesGRUInputShape))
smilesModel.add(Dropout(0.30))
smilesModel.add(GRU(300, trainable = RP['trainable_inner'], return_sequences = True))
smilesModel.add(Activation('tanh', trainable = RP['trainable_inner']))
smilesModel.add(Dropout(0.30))
smilesModel.add(GRU(300, trainable = RP['trainable_inner']))
smilesModel.add(Activation('tanh', trainable = RP['trainable_inner']))
# utility.setModelConsumeLess(smilesModel, 'mem')
'''
smilesModel = utility.loadModel('24e62794bb6d5b5c562e41a3a2cccc3525fa625f', 'smiles_')
smilesModel.pop() # output
smilesModel.pop() # dropout
'''
# utility.setModelConsumeLess(smilesModel, 'gpu')
fastaModel = Sequential()
fastaModel.add(TimeDistributed(Dense(300, activation = 'tanh', trainable = RP['trainable_inner']), input_shape = fastaGRUInputShape))
fastaModel.add(Dropout(0.30))
fastaModel.add(GRU(300, trainable = RP['trainable_inner'], return_sequences = True))
fastaModel.add(Activation('tanh', trainable = RP['trainable_inner']))
fastaModel.add(Dropout(0.30))
fastaModel.add(GRU(300, trainable = RP['trainable_inner']))
fastaModel.add(Activation('tanh', trainable = RP['trainable_inner']))
# utility.setModelConsumeLess(fastaModel, 'mem')
'''
fastaModel = utility.loadModel('e6beb8b7e146b9ab46a71db8f3001bf62d96ff08', 'fasta_')
fastaModel.pop() # activation
fastaModel.pop() # output
fastaModel.pop() # dropout
'''
# utility.setModelConsumeLess(fastaModel, 'gpu')
merged = Merge([smilesModel, fastaModel], mode='concat')
mergedModel = Sequential()
mergedModel.add(merged)
mergedModel.add(Dense(300))
mergedModel.add(Activation('relu'))
mergedModel.add(Dropout(0.3))
mergedModel.add(Dense(300))
mergedModel.add(Activation('relu'))
mergedModel.add(Dropout(0.3))
mergedModel.add(Dense(mergedOutputLen))
if RP['classify']:
mergedModel.add(Activation(RP['classify_activation'], trainable = RP['trainable_inner']))
metrics = []
if RP['classify']:
metrics.append('accuracy')
mergedModel.compile(loss = RP['objective'], optimizer = OPTIMIZER, metrics = metrics)
print(' ...done')
return mergedModel
