Time per epoch on CPU (Core i7): ~150s.
'''
max_features = 3
maxlen = 19200 # cut texts after this number of words (among top max_features most common words)
batch_size = 10
nb_epoch = 1
row_count = 120
column_count = 160
number_of_training_data = 455
DEBUG = False
N_HIDDEN = 20
X_train, y_train, X_val, y_val, X_test, y_test = dt.load_dataset(
number_of_training_data, row_count, column_count, max_features, DEBUG)
dt.saveImage(y_train[0], "gt.png", row_count, column_count, 3, True)
print(len(X_train), 'train sequences')
print(len(X_test), 'test sequences')
print("Pad sequences (samples x time)")
#X_train = sequence.pad_sequences(X_train, maxlen=maxlen)
#X_test = sequence.pad_sequences(X_test, maxlen=maxlen)
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)
y_train = np.array(y_train)
y_test = np.array(y_test)
print('Build model...')
model = Graph()
model.add_input(name='input', input_shape=(None, max_features), dtype='float')
def main(num_epochs=NUM_EPOCHS):
print("Building network ...")
l_in = lasagne.layers.InputLayer((None, MAX_LENGTH, 3))
batchsize, seqlen, _ = l_in.input_var.shape
l_forward = lasagne.layers.LSTMLayer(
l_in, N_HIDDEN, grad_clipping=GRAD_CLIP,
nonlinearity=lasagne.nonlinearities.tanh)
l_backward = lasagne.layers.LSTMLayer(
l_in, N_HIDDEN, grad_clipping=GRAD_CLIP,
nonlinearity=lasagne.nonlinearities.tanh, backwards=True)
l_recurrent = lasagne.layers.ElemwiseMergeLayer([l_forward, l_backward], T.mul)
softmax = lasagne.nonlinearities.softmax
l_reshape = lasagne.layers.ReshapeLayer(l_recurrent,(-1, N_HIDDEN))
l_drop_out = lasagne.layers.DropoutLayer(l_reshape, p=0.95)
l_dense = lasagne.layers.DenseLayer(l_drop_out, num_units=1, nonlinearity=lasagne.nonlinearities.tanh)
l_drop_out_2 = lasagne.layers.DropoutLayer(l_dense, p=0.95)
#l_drop_out_2 = lasagne.layers.DropoutLayer(l_reshape, p=0.5)
l_softmax = lasagne.layers.DenseLayer(l_drop_out_2, num_units=2, nonlinearity = softmax)
l_out = lasagne.layers.ReshapeLayer(l_softmax, (batchsize, seqlen, 2))
# Now, we'll concatenate the outputs to combine them.
#l_sum = lasagne.layers.ConcatLayer([l_forward, l_backward], 2)
#l_shp = lasagne.layers.ReshapeLayer(l_sum, (-1, N_HIDDEN))
# Our output layer is a simple dense connection, with 1 output unit
#l_final = lasagne.layers.DenseLayer(l_shp, num_units=1, nonlinearity=lasagne.nonlinearities.tanh)
#l_out = lasagne.layers.ReshapeLayer(l_final, (batchsize, seqlen, 1))
target_values = T.tensor3('target_output')
# lasagne.layers.get_output produces a variable for the output of the net
network_output = lasagne.layers.get_output(l_out)
# The value we care about is the final value produced for each sequence
#predicted_values = T.argmax(network_output, axis = 2, keepdims = True)
predicted_values = network_output
# Our cost will be mean-squared error
cost = T.mean((T.argmax(predicted_values, axis = 2, keepdims = True) - target_values)**2)
#cost = lasagne.objectives.squared_error(T.argmax(predicted_values, axis = 2)+1, target_values).mean()
#cost = cost.mean()
acc = T.mean(T.eq(T.argmax(predicted_values, axis = 2, keepdims = True), target_values),
dtype=theano.config.floatX)
# Retrieve all parameters from the network
all_params = lasagne.layers.get_all_params(l_out)
# Compute SGD updates for training
print("Computing updates ...")
updates = lasagne.updates.nesterov_momentum(cost, all_params, learning_rate=LEARNING_RATE)
# Theano functions for training and computing cost
print("Compiling functions ...")
train = theano.function([l_in.input_var, target_values],
cost, updates=updates)
compute_cost = theano.function(
[l_in.input_var, target_values], cost)
compute_acc = theano.function(
[l_in.input_var, target_values], acc)
get_out = T.argmax(predicted_values, axis = 2, keepdims = True)
get_prediction = theano.function([l_in.input_var], get_out)
get_prediction_2 = theano.function([l_in.input_var], predicted_values)
# We'll use this "validation set" to periodically check progress
X_train, y_train, X_val, y_val, X_test, y_test = dt.load_dataset(BATCH_SIZE, row_count, column_count, plane_count, DEBUG)
print("Training ...")
#print(get_prediction(X_train[0:1]))
try:
index = 0 #*len(dt.labels_rev)
dt.saveImage(y_test[0], "results/y_GT.png",row_count, column_count, plane_count)
for epoch in range(num_epochs):
X = X_train[EPOCH_SIZE*epoch:EPOCH_SIZE*(epoch+1)]
y = y_train[EPOCH_SIZE*epoch:EPOCH_SIZE*(epoch+1)]
train(X, y)
cost_val = compute_cost(X_val, y_val)
cost_test = compute_acc(X_test, y_test)*100
#print(y_test[0])
#print(get_prediction(X_test)[0])
#print(get_prediction_2(X_test)[0])
print("Epoch {} validation cost = {} test acc = {} %".format(epoch, cost_val, cost_test))
dt.saveImage(get_prediction(X_test)[0], "results/y_output_{}.png".format(epoch), row_count, column_count, plane_count, True)
dt.saveImage(get_prediction(X_test)[0], "results/y_output.png", row_count, column_count, plane_count, True)
except KeyboardInterrupt:
pass
Output after 4 epochs on CPU: ~0.8146
Time per epoch on CPU (Core i7): ~150s.
'''
max_features = 3
maxlen = 19200 # cut texts after this number of words (among top max_features most common words)
batch_size = 10
nb_epoch = 1
row_count = 120
column_count = 160
number_of_training_data = 455
DEBUG = False
N_HIDDEN = 20
X_train, y_train, X_val, y_val, X_test, y_test = dt.load_dataset(number_of_training_data, row_count, column_count, max_features, DEBUG)
dt.saveImage(y_train[0], "gt.png", row_count, column_count, 3, True)
print(len(X_train), 'train sequences')
print(len(X_test), 'test sequences')
print("Pad sequences (samples x time)")
#X_train = sequence.pad_sequences(X_train, maxlen=maxlen)
#X_test = sequence.pad_sequences(X_test, maxlen=maxlen)
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)
y_train = np.array(y_train)
y_test = np.array(y_test)
print('Build model...')