def main_okapi():
from OkapiV2.Core import Model, Branch
from OkapiV2.Layers.Basic import FullyConnected, Dropout
from OkapiV2.Layers.Activations import ActivationLayer, PReLULayer
from OkapiV2.Layers.Convolutional import Convolutional, MaxPooling
from OkapiV2 import Activations, Datasets
from OkapiV2 import Optimizers
X_train, y_train, X_val, y_val, X_test, y_test = Datasets.load_mnist()
tree = Branch()
tree.add_layer(Convolutional(num_filters, filter_size, filter_size, pad=pad))
# tree.add_layer(BatchNorm())
tree.add_layer(PReLULayer())
tree.add_layer(MaxPooling(pool_size, pool_size))
tree.add_layer(Convolutional(num_filters, filter_size, filter_size, pad=pad))
# tree.add_layer(BatchNorm())
tree.add_layer(PReLULayer())
tree.add_layer(MaxPooling(pool_size, pool_size))
tree.add_layer(Convolutional(num_filters, filter_size, filter_size, pad=pad))
# tree.add_layer(BatchNorm())
tree.add_layer(PReLULayer())
tree.add_layer(MaxPooling(pool_size, pool_size))
'''tree.add_layer(Convolutional(num_filters, filter_size, filter_size, pad=pad))
# tree.add_layer(BatchNorm())
tree.add_layer(PReLULayer())
tree.add_layer(MaxPooling(pool_size, pool_size))'''
'''tree.add_layer(Dropout(0.25))
tree.add_layer(FullyConnected((h_size, 1, 1, 1)))
tree.add_layer(PReLULayer())
# tree.add_layer(BatchNorm())'''
tree.add_layer(Dropout(0.5))
tree.add_layer(FullyConnected())
tree.add_layer(ActivationLayer(Activations.alt_softmax))
tree.add_input(X_train)
model = Model()
model.set_tree(tree)
model.set_optimizer(Optimizers.RMSprop(learning_rate=learning_rate,
momentum=momentum))
model.train([X_train], y_train, num_epochs=num_epochs,
batch_size=batch_size)
# ok.save_model(model, 'okapi_mnist.pk')
okapi_accuracy = model.get_accuracy([X_test], y_test)
print("Test Accuracy: {}%"
.format(round(okapi_accuracy, 2)))
return okapi_accuracy
def main_okapi():
import OkapiV2.Core as ok
from OkapiV2.Core import Model
from OkapiV2.Layers.Basic import FullyConnected, Dropout, BatchNorm
from OkapiV2.Layers.Activations import ActivationLayer, PReLULayer
from OkapiV2.Layers.Recurrent import LSTM
from OkapiV2 import Activations, Optimizers, Losses
path = 'data/lear.txt'
text = open(path).read().lower() # [0:corpus_length]
print('Corpus length:', len(text))
chars = set(text)
print('Total Characters:', len(chars))
char_to_index = dict((c, i) for i, c in enumerate(chars))
index_to_char = dict((i, c) for i, c in enumerate(chars))
# cut the text in semi-redundant sequences of maxlen characters
sentences = []
next_chars = []
for i in range(0, len(text) - maxlen, step):
sentences.append(text[i: i + maxlen])
next_chars.append(text[i + maxlen])
print('Total Sequences:', len(sentences))
print('Vectorization...')
X = np.zeros((len(sentences), maxlen, len(chars)), dtype=np.bool)
y = np.zeros((len(sentences), len(chars)), dtype=np.bool)
for i, sentence in enumerate(sentences):
for t, char in enumerate(sentence):
X[i, t, char_to_index[char]] = 1
y[i, char_to_index[next_chars[i]]] = 1
def sample(a, temperature=1.0):
# helper function to sample an index from a probability array
a = np.log(a) / temperature
a = np.exp(a) / np.sum(np.exp(a)) - 1e-7
return np.argmax(np.random.multinomial(1, a, 1))
model = Model()
model.add(LSTM((h_layer_size, 1, 1, 1)))
model.add(PReLULayer())
model.add(Dropout(0.2))
model.add(BatchNorm())
model.add(LSTM((h_layer_size, 1, 1, 1)))
model.add(PReLULayer())
model.add(Dropout(0.2))
model.add(BatchNorm())
model.add(FullyConnected())
model.add(ActivationLayer(Activations.alt_softmax))
model.set_loss(Losses.Crossentropy())
model.set_optimizer(Optimizers.RMSprop(learning_rate=learning_rate))
for iteration in range(0, num_iterations):
print()
print('-' * 50)
print('Iteration', iteration + 1)
model.train(X, y, batch_size=batch_size, num_epochs=1,
params_filename='okapi_shakespeare_params.pk')
start_index = random.randint(0, len(text) - maxlen - 1)
for diversity in diversities:
print()
print('----- diversity:', diversity)
generated = ''
sentence = text[start_index: start_index + maxlen]
generated += sentence
print('----- Generating with seed: "' + sentence + '"')
sys.stdout.write(generated)
for iteration in range(num_chars):
x = np.zeros((1, maxlen, len(chars)))
for t, char in enumerate(sentence):
x[0, t, char_to_index[char]] = 1.
preds = model.predict(x)
preds = preds[0]
next_index = sample(preds, diversity)
next_char = index_to_char[next_index]
generated += next_char
sentence = sentence[1:] + next_char
sys.stdout.write(next_char)
sys.stdout.flush()
print()
ok.save_model(model, 'okapi_shakespeare_model.pk')
from OkapiV2.Core import Model, Branch
from OkapiV2.Layers.Basic import FullyConnected
from OkapiV2.Layers.Activations import ActivationLayer, PReLULayer
from OkapiV2 import Activations, Datasets, Optimizers
x_train, y_train, x_val, y_val, x_test, y_test = Datasets.load_mnist()
tree = Branch()
tree.add_layer(FullyConnected((512, 1, 1, 1)))
tree.add_layer(PReLULayer())
tree.add_layer(FullyConnected((512, 1, 1, 1)))
tree.add_layer(PReLULayer())
tree.add_layer(FullyConnected((512, 1, 1, 1)))
tree.add_layer(PReLULayer())
tree.add_layer(FullyConnected())
tree.add_layer(ActivationLayer(Activations.softmax))
tree.add_input(x_train)
model = Model()
model.set_tree(tree)
model.set_optimizer(Optimizers.RMSprop(learning_rate=0.00005))
index = 60000
model.train([x_train[:index, :, :, :]], y_train[:index, :], 24)
accuracy = model.get_accuracy([x_train[:index, :, :, :]], y_train[:index])
print("Accuracy: {}%".format(accuracy))
test_accuracy = model.get_accuracy([x_test], y_test)
print("Test accuracy: {}%".format(test_accuracy))
from OkapiV2.Core import Model, Branch
from OkapiV2.Layers.Basic import FullyConnected
from OkapiV2.Layers.Activations import ActivationLayer, PReLULayer
from OkapiV2 import Activations, Datasets, Optimizers
X_train, y_train, X_val, y_val, X_test, y_test = Datasets.load_mnist()
tree = Branch()
tree.add_layer(FullyConnected((512, 1, 1, 1)))
tree.add_layer(PReLULayer())
tree.add_layer(FullyConnected((512, 1, 1, 1)))
tree.add_layer(PReLULayer())
tree.add_layer(FullyConnected((512, 1, 1, 1)))
tree.add_layer(PReLULayer())
tree.add_layer(FullyConnected())
tree.add_layer(ActivationLayer(Activations.softmax))
tree.add_input(X_train)
model = Model()
model.set_tree(tree)
model.set_optimizer(Optimizers.RMSprop(learning_rate=0.00005))
index = 60000
model.train([X_train], y_train, 24)
accuracy = model.get_accuracy([X_train], y_train)
print('Accuracy: {}%'.format(accuracy))
test_accuracy = model.get_accuracy([X_test], y_test)
print('Test accuracy: {}%'.format(test_accuracy))
model = Model()
model.set_tree(tree)
def sample(a, temperature=1.0):
# helper function to sample an index from a probability array
a = np.log(a) / temperature
a = np.exp(a) / np.sum(np.exp(a))
return np.argmax(np.random.multinomial(1, a, 1))
# train the model, output generated text after each iteration
for iteration in range(1, 60):
print()
print('-' * 50)
print('Iteration', iteration)
model.train([X], y, num_epochs=1)
start_index = random.randint(0, len(text) - maxlen - 1)
for diversity in [0.2, 0.5, 1.0, 1.2]:
print()
print('----- diversity:', diversity)
generated = ''
sentence = text[start_index: start_index + maxlen]
generated += sentence
print('----- Generating with seed: "' + sentence + '"')
sys.stdout.write(generated)
for iteration in range(400):
x = np.zeros((1, maxlen, len(chars)))
tree.add_input(X_reward[1])
model = Model()
model.set_tree(tree)
model.set_loss(Losses.MeanSquared())
learning_rate = 0.00002
model.set_optimizer(Optimizers.RMSprop(learning_rate=learning_rate))
# model.compile(X_reward, y_reward)
# model.train(X_reward, y_reward, 24)
reinforce_index = X_obs.shape[0]
X_batches, y_batches, num_batches = ok.make_batches([X_train], y_train, batch_size=10000)
for i in range(8):
print('\n---Iteration {}---'.format(i + 1))
for X_batch, y_batch in zip(X_batches, y_batches):
model.train(X_reward, y_reward, 24)
accuracy, preds = model.get_dream_accuracy([X_batch[0], None], y_batch)
preds = preds[0]
preds += Initializers.normal(preds.shape, 0.01)
print('Accuracy: {}%'.format(accuracy))
preds_reward = reward(preds.reshape(preds.shape[0], preds.shape[1]).astype('float32'),
y_batch.reshape(preds.shape[0], preds.shape[1]).astype('float32'))
print('Avg Reward: {}'.format(np.mean(preds_reward)))
X_reward[0] = np.append(X_reward[0], X_batch[0], axis=0)
X_reward[1] = np.append(X_reward[1], preds, axis=0)
y_reward = np.append(y_reward, preds_reward.reshape(preds_reward.shape[0], 1), axis=0)
params = model.get_params_as_vec()
file = open('reward_model_params_vec.pk', 'wb')
pickle.dump(params, file, protocol=pickle.HIGHEST_PROTOCOL)
file.close()
