def jointModel():
shared_input_layer = TemporalInputFeatures(inputJointFeatures)
shared_hidden_layer = LSTM('tanh','sigmoid','orthogonal',4,128)
shared_layers = [shared_input_layer,shared_hidden_layer]
human_layers = [ConcatenateFeatures(inputHumanFeatures),LSTM('tanh','sigmoid','orthogonal',4,256),softmax(num_sub_activities)]
object_layers = [ConcatenateFeatures(inputObjectFeatures),LSTM('tanh','sigmoid','orthogonal',4,256),softmax(num_affordances)]
[X,Y,num_classes,class_ids_reverse] = createTrain('shakespeare_input.txt',num_samples,len_samples)
inputD = num_classes
outputD = num_classes
permutation = permute(num_samples)
X = X[:,permutation]
Y = Y[:,permutation]
X_tr = X[:,:num_train]
Y_tr = Y[:,:num_train]
X_valid = X[:,num_train:]
Y_valid = Y[:,num_train:]
# Creating network layers
layers = [OneHot(num_classes),LSTM(),LSTM(),LSTM(),softmax(num_classes)]
trY = T.lmatrix()
# Initializing network
rnn = RNN(layers,softmax_loss,trY,1e-3)
# Fitting model
rnn.fitModel(X_tr,Y_tr,1,'checkpoints/',epochs,batch_size,learning_rate_decay,decay_after)
# Printing a generated sentence
out = rnn.predict_language_model(X_valid[:,:1],1000,OutputSampleFromDiscrete)
# Print the sentence here
text_produced = text_prediction(class_ids_reverse,out)
print 'Number of classes ',outputD
print 'Feature dimension ',inputD
epochs = 200
batch_size = num_train
learning_rate_decay = 0.97
decay_after = 5
use_pretrained = False
train_more = False
global rnn
if not use_pretrained:
# Creating network layers
layers = [TemporalInputFeatures(inputD),LSTM('tanh','sigmoid','orthogonal',4,512),softmax(num_classes)]
trY = T.lmatrix()
# Initializing network
rnn = RNN(layers,softmax_loss,trY,1e-3)
if not os.path.exists('{1}/{0}/'.format(index,path_to_checkpoints)):
os.mkdir('{1}/{0}/'.format(index,path_to_checkpoints))
# Fitting model
rnn.fitModel(X_tr,Y_tr,1,'{1}/{0}/'.format(index,path_to_checkpoints),epochs,batch_size,learning_rate_decay,decay_after)
else:
checkpoint = sys.argv[3]
# Prediction
rnn = load('{2}/{0}/checkpoint.{1}'.format(index,checkpoint,path_to_checkpoints))
