class Classifier:
def __init__(self):
self.mlp = None
def predict_input_img(self, img):
plt.imshow(img, cmap="hot", interpolation="nearest")
plt.show()
preprocessing_normalized = preprocessing(
dataset=tf.keras.datasets.mnist.load_data(),
symetric_dataset=True,
dimension_reduction=False,
data_whitening=False,
normalize=True,
)
X_train, y_train, X_test, y_test = preprocessing_normalized.preprocess_mnist(
[img]
)
if self.mlp is None:
print("Training the model.\nThis will only happen once.")
self.mlp = MLP(ReLU, X_train.shape[1], [128, 128])
gen = self.mlp.fit(X_train, y_train, 25, 50)
ls = list(gen)
pred = self.mlp.predict(X_test)
print(f"Our MultiLayer Perceptron thinks the digit is: {pred[0]}")
ascii_banner = pyfiglet.figlet_format(str(pred[0]))
print(ascii_banner)
def load_checkpoint(filename, Model, MiddleModel: nn.Module = None):
with bz2.BZ2File(filename, "rb") as f:
obj = pickle.load(f)
pose_autoencoder = MLP.load_checkpoint(obj["pose_autoencoder_path"])
cost_encoder = MLP.load_checkpoint(obj["cost_encoder_path"])
generationModel = Model.load_checkpoint(
obj["motionGenerationModelPath"])
model = MotionGenerationModelRNN(config=obj["config"],
feature_dims=obj["feature_dims"],
input_slicers=obj["in_slices"],
output_slicers=obj["out_slices"],
name=obj["name"])
if MiddleModel is None:
MiddleModel = nn.Linear(
in_features=pose_autoencoder.dimensions[-1],
out_features=pose_autoencoder.dimensions[-11])
MiddleModel.load_state_dict(obj["middle_layer_dict"])
model.in_slices = obj["in_slices"]
model.out_slices = obj["out_slices"]
model.pose_autoencoder = pose_autoencoder
model.cost_encoder = cost_encoder
model.generationModel = generationModel
return model
def Main():
x_train = pd.read_csv('Dataset/xtrain_3spirals.txt', sep=' ', header=None)
x_test = pd.read_csv('Dataset/xtest_3spirals.txt', sep=' ', header=None)
d_train = pd.read_csv('Dataset/dtrain_3spirals.txt', sep=',', header=None)
d_test = pd.read_csv('Dataset/dtest_3spirals.txt', sep=',', header=None)
## Aplication of MLP algorithm
mlp = MLP(15000, 0.15, 0.000001, [4, 3], 0.5)
mlp.train(x_train.to_numpy(), d_train.to_numpy())
new_classes = mlp.application(x_test.to_numpy())
comparative = np.concatenate((d_test.to_numpy(), new_classes), 1)
print("Matrix of comparative between classes")
print(comparative)
print("------------------------------")
hit_table = np.zeros((len(new_classes), 1))
for row in range(len(new_classes)):
if all(d_test.to_numpy()[row] == new_classes[row]):
hit_table[row] = 1
tax_hit = sum(hit_table) / len(new_classes)
print("------------------------------")
print("Matrix of hits")
print(hit_table)
print("------------------------------")
print("Tax of hits: " + str(tax_hit))
def __init__(self):
self.type = None
self.nn = MLP()
self.training_method = None
self.activation_function = None
self.dropout_rate = 0.0
self.training = True
self.learning_rate = 0.1
self.fitness_threshold = 0.75
self.epoch_threshold = -1
self.batch_size = 100
self.shuffle_rate = 2500
self.display_step = 1000
self.epoch = 0
self.layers = []
self.data_set = None
self.bed = BinaryEncoderDecoder()
self.utils = Utilities()
self.debug_mode = False
# Plotting variables
self.losses = []
self.fitnesses = []
self.iterations = []
self.save_location = './nn/log/'
def __init__(self, x, model_file):
"""
Sampling works as follows.
You feed it a model and a dataset, and model_files.
The model_files allow you to load in models from different checkpoints.
A feed through function is compiled that samples from the test dataset.
It calculates the error and the output for each element in test dataset.
It generates two distributions -- output for signal, and output for background.
args:
model: MLP object
dataset Dataset object
model_files: list of files corresponding to saved models
"""
self.model_file = model_file
self.param = self.detect_params(self.model_file)
self.dataset = Dataset(self.param['dataset'])
self.dataset.set_indexing(self.param['indexing'])
self.shared_train_x = self.dataset.train_x
self.shared_train_y = self.dataset.train_y
self.shared_test_x = self.dataset.test_x
self.shared_test_y = self.dataset.test_y
try:
self.train_labels = self.dataset.train_labels
self.test_labels = self.dataset.test_labels
except AttributeError:
print(
"You're used a dataset without labels. You won't be able to call gen_labeled_outputs"
)
mlp = MLP(x, [self.param['h0'], self.param['h1'], 2],
np.random.RandomState(1234),
transfer_func=T.nnet.relu)
mlp.load_params(self.model_file, mode='hdf5')
self.model = mlp
self.predicted = dict()
def __init__(self, word_embeddings_dim, tag_embeddings_dim, vocabulary_size, tag_uniqueCount, label_uniqueCount, pretrainedWordEmbeddings=None, pretrainedTagEmbeddings=None):
super().__init__()
self.word_embeddings = nn.Embedding(vocabulary_size, word_embeddings_dim)
if pretrainedWordEmbeddings.any():
assert pretrainedWordEmbeddings.shape == (vocabulary_size, word_embeddings_dim)
self.word_embeddings.weight.data.copy_(torch.from_numpy(pretrainedWordEmbeddings))
self.tag_embeddings = nn.Embedding(tag_uniqueCount, tag_embeddings_dim)
if pretrainedTagEmbeddings.any():
assert pretrainedTagEmbeddings.shape == (tag_uniqueCount, tag_embeddings_dim)
self.tag_embeddings.weight.data.copy_(torch.from_numpy(pretrainedTagEmbeddings))
# Save computation time by not training already trained word vectors
# disableTrainingForEmbeddings(self.word_embeddings, self.tag_embeddings)
# Now we need to train the embeddings for <root> and <unk>
self.inputSize = word_embeddings_dim + tag_embeddings_dim # The number of expected features in the input x
self.hiddenSize = self.inputSize #* 2 # 512? is this the same as outputSize?
self.nLayers = 2
self.biLstm = nn.LSTM(self.inputSize, self.hiddenSize, self.nLayers, bidirectional=True)
self.nDirections = 2
self.batch = 1 # this is per recommendation
# Input size of the MLP for arcs scores is the size of the output from previous step concatenated with another of the same size
biLstmOutputSize = self.hiddenSize * self.nDirections
mlpForScoresInputSize = biLstmOutputSize * 2
self.mlpArcsScores = MLP(mlpForScoresInputSize, hidden_size=mlpForScoresInputSize, output_size=1)
# MLP for labels
self.label_uniqueCount = label_uniqueCount
self.mlpLabels = MLP(mlpForScoresInputSize, hidden_size=mlpForScoresInputSize, output_size=self.label_uniqueCount)
def add_models(self,
input_dims: list = None,
pose_labels: list = None,
freeze=False):
n = len(self.models) + 1
if pose_labels is not None:
self.models += [
MLP(config=self.config,
dimensions=[input_dims[i]],
pose_labels=pose_labels[i],
name="M" + str(i + n),
single_module=0) for i in range(len(input_dims))
]
else:
self.models += [
MLP(config=self.config,
dimensions=[input_dims[i]],
name="M" + str(i + n),
single_module=0) for i in range(len(input_dims))
]
if freeze:
for model in self.active_models:
model.freeze(True)
self.active_models = self.models[n - 1:]
self.input_dims = input_dims
else:
self.active_models = self.models
self.input_dims += input_dims
self.input_slice = [0] + list(accumulate(add, self.input_dims))
def load_pretrain_weights(self):
"""Loading weights from trained MLP model & GMF model"""
config = self.config
config['latent_dim'] = config['latent_dim_mlp']
mlp_model = MLP(config)
if config['use_cuda'] is True:
mlp_model.cuda()
resume_checkpoint(mlp_model,
model_dir=config['pretrain_mlp'],
device_id=config['device_id'])
self.embedding_user_mlp.weight.data = mlp_model.embedding_user.weight.data
self.embedding_item_mlp.weight.data = mlp_model.embedding_item.weight.data
for idx in range(len(self.fc_layers)):
self.fc_layers[idx].weight.data = mlp_model.fc_layers[
idx].weight.data
config['latent_dim'] = config['latent_dim_mf']
gmf_model = GMF(config)
if config['use_cuda'] is True:
gmf_model.cuda()
resume_checkpoint(gmf_model,
model_dir=config['pretrain_mf'],
device_id=config['device_id'])
self.embedding_user_mf.weight.data = gmf_model.embedding_user.weight.data
self.embedding_item_mf.weight.data = gmf_model.embedding_item.weight.data
self.affine_output.weight.data = 0.5 * torch.cat([
mlp_model.affine_output.weight.data,
gmf_model.affine_output.weight.data
],
dim=-1)
self.affine_output.bias.data = 0.5 * (
mlp_model.affine_output.bias.data +
gmf_model.affine_output.bias.data)
def __init__(self, env, alpha, gamma, episode_num, target_reward,
step_count, minbatch, memory_size, flag):
self.env = env
self.alpha = alpha
self.gamma = gamma
self.episode_num = episode_num
self.target_reward = target_reward
self.step_count = step_count
# self.test_step=test_step
self.minbatch = minbatch
self.memory_size = memory_size
self.flag = flag
self.Q = MLP()
self.state_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.spaces[
0].n * env.action_space.spaces[1].n
# self.action_dim = env.action_space.n
self.Q.creat2(self.state_dim, env.action_space.spaces[0].n,
env.action_space.spaces[1].n)
self.memory_num = 0
self.memory = np.zeros((memory_size, self.state_dim * 2 + 4))
self.optimizer = torch.optim.Adam(self.Q.parameters(), lr=alpha)
self.loss_func = nn.MSELoss()
def __init__(self,x, model_file):
"""
Sampling works as follows.
You feed it a model and a dataset, and model_files.
The model_files allow you to load in models from different checkpoints.
A feed through function is compiled that samples from the test dataset.
It calculates the error and the output for each element in test dataset.
It generates two distributions -- output for signal, and output for background.
args:
model: MLP object
dataset Dataset object
model_files: list of files corresponding to saved models
"""
self.model_file = model_file
self.param = self.detect_params(self.model_file)
self.dataset = Dataset(self.param['dataset'])
self.dataset.set_indexing(self.param['indexing'])
self.shared_train_x = self.dataset.train_x
self.shared_train_y = self.dataset.train_y
self.shared_test_x = self.dataset.test_x
self.shared_test_y = self.dataset.test_y
try:
self.train_labels = self.dataset.train_labels
self.test_labels = self.dataset.test_labels
except AttributeError:
print("You're used a dataset without labels. You won't be able to call gen_labeled_outputs")
mlp = MLP(x,[self.param['h0'],self.param['h1'],2],np.random.RandomState(1234), transfer_func=T.nnet.relu)
mlp.load_params(self.model_file,mode='hdf5')
self.model = mlp
self.predicted = dict()
def main(argv):
argv = FLAGS(argv)
inputs, outputs = load_CIFAR_train(FLAGS.datapath)
X_test, y_test = load_CIFAR_test(FLAGS.datapath)
nn = MLP(3072, FLAGS.hidden_dim, 10, FLAGS.activation, FLAGS.loss_type, FLAGS.layer_num)
nn.fit(inputs, outputs, FLAGS.epoch, FLAGS.batch, [FLAGS.lr_W, FLAGS.lr_b],
X_test, y_test)
print nn.test(X_test, y_test)
def test():
model_to_be_restored = MLP()
checkpoint = tf.train.Checkpoint(myAwesomeModel=model_to_be_restored)
checkpoint.restore(tf.train.latest_checkpoint('./check_point'))
y_pred = np.argmax(model_to_be_restored.predict(data_loader.test_data),
axis=-1)
print("test accuracy: %f" %
(sum(y_pred == data_loader.test_label) / data_loader.num_test_data))
def __init__(self,
config: dict = None,
Model=None,
pose_autoencoder=None,
feature_dims=None,
input_slicers: list = None,
output_slicers: list = None,
train_set=None,
val_set=None,
test_set=None,
name="MotionGeneration"):
super().__init__()
self.feature_dims = feature_dims
self.config = config
self.loss_fn = config[
"loss_fn"] if "loss_fn" in config else nn.functional.mse_loss
self.opt = config[
"optimizer"] if "optimizer" in config else torch.optim.Adam
self.scheduler = config["scheduler"] if "scheduler" in config else None
self.scheduler_param = config[
"scheduler_param"] if "scheduler_param" in config else None
self.batch_size = config["batch_size"]
self.learning_rate = config["lr"]
self.best_val_loss = np.inf
self.phase_smooth_factor = 0.9
self.pose_autoencoder = pose_autoencoder if pose_autoencoder is not None else \
MLP(config=config, dimensions=[feature_dims["pose_dim"]], name="PoseAE")
self.use_label = pose_autoencoder is not None and pose_autoencoder.use_label
cost_hidden_dim = config["cost_hidden_dim"]
self.cost_encoder = MLP(config=config,
dimensions=[
feature_dims["cost_dim"], cost_hidden_dim,
cost_hidden_dim, cost_hidden_dim
],
name="CostEncoder",
single_module=-1)
self.generationModel = Model(config=config,
dimensions=[
feature_dims["g_input_dim"],
feature_dims["g_output_dim"]
],
phase_input_dim=feature_dims["phase_dim"])
self.input_dims = input_slicers
self.output_dims = output_slicers
self.in_slices = [0] + list(accumulate(add, input_slicers))
self.out_slices = [0] + list(accumulate(add, output_slicers))
self.train_set = train_set
self.val_set = val_set
self.test_set = test_set
self.name = name
def experiment_learning_curves_error():
train_test = [0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1]
use_validation_set = True
num_hidden_nodes_layer_1 = 20
num_iterations = 1000
learning_rate = 0.001
verbose = False
cases = [1, 2, 3, 4]
train_MSE = []
val_MSE = []
for case in cases:
[inputs, inputs_labels, input_validation,
input_validation_labels] = Utils.create_non_linearly_separable_data_2(
use_validation_set=use_validation_set, case=case)
print(case)
current_train = []
current_validation = []
for check in train_test:
X_train, X_test, y_train, y_test = train_test_split(
inputs.T, inputs_labels, test_size=check, random_state=42)
mlp_batch = MLP(inputs=X_train.T,
inputs_labels=y_train,
input_validation=input_validation,
input_validation_labels=input_validation_labels,
num_nodes_hidden_layer=num_hidden_nodes_layer_1,
num_iterations=num_iterations,
learning_rate=learning_rate,
batch_train=True,
verbose=verbose)
[_, _, mse_batch] = mlp_batch.fit()
current_train.append(mlp_batch.mse[-1])
current_validation.append(mlp_batch.validation_mse[-1])
train_MSE.append(current_train)
val_MSE.append(current_validation)
legend_names = [
'train mse error case 1', 'train mse error case 2',
'train mse error case 3', 'train mse error case 4',
'validation mse error case 1', 'validation mse error case 2',
'validation mse error case 3', 'validation mse error case 4'
]
Utils.plot_learning_curves(
train_MSE,
legend_names=legend_names,
train_size=train_test,
title='Learning curve with lr = {0}, iterations = {1} '.format(
learning_rate, num_iterations),
loss=val_MSE)
def runTest(self):
inputs = np.array([[1, 0], [0, 1], [1, 1], [0, 0]], dtype = util.FLOAT).T
outputs = np.array([[0, 1], [0, 1], [1, 0], [1, 0]], dtype = util.FLOAT).T
mlp = MLP([2, 5, 2], Lambda = 0.0001, alpha = .1, activation = "sigmoid", costType = "mse")
mlp.fit(inputs, outputs, 20000, 4)
mlp_result, mlp.prediction = mlp.predict(inputs, outputs)
loss = np.mean( (mlp_result - outputs)**2)
print "Prediction: ", mlp_result
print "Loss: ", loss
np.testing.assert_almost_equal(loss, 0, 2)
def experiment_train_validation_nodes():
use_validation_set = True
num_iterations = 1000
learning_rate = 0.002
verbose = False
nodes = [1, 5, 10, 20, 25]
cases = [1, 2, 3, 4]
train_MSE = []
val_MSE = []
for case in cases:
print(case)
[inputs, inputs_labels, input_validation,
input_validation_labels] = Utils.create_non_linearly_separable_data_2(
use_validation_set=use_validation_set, case=case)
current_mse = []
current_val_mse = []
for node in nodes:
mlp_batch = MLP(inputs=inputs,
inputs_labels=inputs_labels,
input_validation=input_validation,
input_validation_labels=input_validation_labels,
num_nodes_hidden_layer=node,
num_iterations=num_iterations,
learning_rate=learning_rate,
batch_train=True,
verbose=verbose)
[_, _, mse_batch] = mlp_batch.fit()
current_mse.append(mlp_batch.mse[-1])
current_val_mse.append(mlp_batch.validation_mse[-1])
train_MSE.append(current_mse)
val_MSE.append(current_val_mse)
legend_names = [
'train mse error case 1', 'train mse error case 2',
'train mse error case 3', 'train mse error case 4',
'validation mse error case 1', 'validation mse error case 2',
'validation mse error case 3', 'validation mse error case 4'
]
Utils.plot_error_hidden_nodes(
train_MSE,
legend_names=legend_names,
hidden_nodes=nodes,
title='MLP with learning rate {0}, iterations {1} '.format(
learning_rate, num_iterations),
loss=val_MSE)
def run_test():
print 'starting ....'
Y = read_image_data('data/observations.arff')
print 'image reading is finished...'
print len(Y)
#Allocate an n-by-k matrix, X, to hold intrinsic vectors
X , K = list(), 2
for _ in xrange(len(Y)):
X.append([0.0*x for x in xrange(K)])
w,h= 64,48
us_mlp = UnsupervisedMLP([4, 12, 12, 3])
us_mlp.set_image_dim(w, h)
us_mlp.train(Y, X, K)
print 'USMLP training is finished'
plot_intrinsic(X, w, h)
print 'intrinsic plotting is done'
#2nd MLP
normalize(X)
print 'reading action file'
A = read_action_data('data/actions.arff')
print len(A), len(X)
for i in xrange(len(A)-1):
row = A[i]
row.extend(X[i])
row.extend(X[i+1])
del A[-1]
print 'training mlp with action '
mlp = MLP([6,6,2])
mlp.train1(A, K)
print 'training done'
#operate the crane now!
#'a':[1.0,0.0,0.0,0.0],'c': [0.0,0.0,1.0,0.0]
s = [1.0,0.0,0.0,0.0]
s.extend(X[0])
for i in range(5):
predict=mlp.predict(s)
s[4] = predict[0]
s[5] = predict[1]
# image_generation(us_mlp, s, w, h, 'frame'+str(i))
#up
s[0]=0 ; s[2]=1.0
for i in range(5,10):
predict=mlp.predict(s)
s[4] = predict[0]
s[5] = predict[1]
image_generation(us_mlp, s, w, h, 'frame'+str(i+1))
def __init__(self, episode_size=150):
self.model = MLP((SCREEN_HEIGHT_g, SCREEN_WIDTH_g), 300)
#self.load("models/model_1185.npz")
self.activations = []
self.frames = []
self.states_alive = []
self.episode_size = episode_size
self.episode_decisions = np.zeros((8))
self.episodes_wins = 0
self.episodes_nb = 0
self.iter = 0
def test_MLP_Layer_size(X, Y, Z, num_hidden_nodes_layer_1):
targets = np.reshape(Z, (1, (len(X) * len(Y)))).T
n_X = len(X)
n_Y = len(Y)
num_data = n_X * n_Y
xx = np.reshape(X, (1, (num_data)))
yy = np.reshape(Y, (1, (num_data)))
patterns = np.vstack((xx, yy)).T
num_iterations = 500
learning_rate = 0.01
verbose = False
X_train, X_test, y_train, y_test = train_test_split(patterns,
targets,
test_size=0.2,
random_state=42)
X_train, X_test = X_train.T, X_test.T
MSEs = []
Models = []
for layers in num_hidden_nodes_layer_1:
mlp_batch = MLP(inputs=X_train,
inputs_labels=y_train,
num_nodes_hidden_layer=layers,
num_iterations=num_iterations,
learning_rate=learning_rate,
batch_train=True,
verbose=verbose,
binary=False,
num_output_layers=1)
mlp_batch.fit()
o_out = mlp_batch.predict(patterns.T)
# print(o_out.shape)
Z = np.reshape(o_out, (n_X, n_Y))
[_, mse] = Utils.compute_error(targets, o_out, False)
MSEs.append(mse)
Models.append(mlp_batch)
title = 'Number of hidden layer: {0} and MSE: {1}'.format(
layers, round(mse, 4))
Utils.plot_3d_data(X, Y, Z, title)
return MSEs, Models
def check_MlP_test_sizes(X, Y, Z):
targets = np.reshape(Z, (1, (len(X) * len(Y)))).T
n_X = len(X)
n_Y = len(Y)
num_data = n_X * n_Y
xx = np.reshape(X, (1, (num_data)))
yy = np.reshape(Y, (1, (num_data)))
patterns = np.vstack((xx, yy)).T
num_hidden_nodes_layer_1 = 20
num_iterations = 5000
learning_rate = 0.001
verbose = False
train_test = [0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8]
MSEs = []
Models = []
for check in train_test:
X_train, X_test, y_train, y_test = train_test_split(patterns,
targets,
test_size=check,
random_state=42)
X_train, X_test = X_train.T, X_test.T
print(X_train.shape)
mlp_batch = MLP(inputs=X_train,
inputs_labels=y_train,
num_nodes_hidden_layer=num_hidden_nodes_layer_1,
num_iterations=num_iterations,
learning_rate=learning_rate,
batch_train=True,
verbose=verbose,
binary=False,
num_output_layers=1)
mlp_batch.fit()
o_out = mlp_batch.predict(patterns.T)
Z = np.reshape(o_out, (n_X, n_Y))
[_, mse] = Utils.compute_error(targets, o_out, False)
MSEs.append(mse)
Models.append(mlp_batch)
title = 'MSE: {0}'.format(round(mse, 4))
# Utils.plot_3d_data(X, Y, Z, title)
return MSEs, Models
def treinar(self, base, qtd_hidden):
n_x = len(base[1][0])
n_y = len(base[1][1])
#print([n_x, n_y])
rna = MLP(qtd_input=n_x, qtd_hidden=qtd_hidden, qtd_output=n_y)
for e in range(0, 300000):
erroEpoca = 0
erro_classificacao = 0
for a in range(0, len(base)):
amostra = base[a]
x = amostra[0]
y = amostra[1]
# Saída da Rede Neural
out = rna.treinar(x, y)
# Calculo do erro simples
err = []
for er in zip(out, y):
err.append(abs(er[0] - er[1]))
# Erro de aproximação
for er in err:
erroEpoca += er
# Erro de Classificação 01
vet_cla = []
for o in out:
if o >= 0.5:
vet_cla.append(1)
else:
vet_cla.append(0)
aux = 0
for v in zip(vet_cla, y):
aux += abs(v[1] - v[0])
ecl = 0
if aux > 0:
ecl = 1
erro_classificacao += ecl
print('Época: {}'.format(e + 1), end=' --- ')
print('Erro: {}'.format(erroEpoca), end=' --- ')
print('Classificação: {}'.format(erro_classificacao))
def __init__(self,
num_input,
num_units,
num_classes,
batch_size=100,
num_epochs=10,
display=False,
blacklist=[],
whitelist=[]):
"""Creates classifier for finding the version"""
# Network parameters
self.l_rate = 0.001
self.dropout_prob = 1
self.training_epochs = num_epochs
self.display_step = 10
self.batch_size = batch_size
self.display = display
self.blacklist = blacklist
self.whitelist = whitelist
assert not (self.blacklist and self.whitelist), (
'Both whitelist and blacklist are defined')
# Placeholders
self.X = tf.placeholder('float', [None, num_input], name='X')
self.Y = tf.placeholder('int64', [None], name='Y')
self.keep_prob = tf.placeholder('float')
# Create Network
self.mlp = MLP([num_input, num_units, num_classes],
[tf.nn.relu, tf.identity])
logits = self.mlp.create_network(self.X, self.keep_prob)
self.cost = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits, self.Y)
self.optimizer = tf.train.AdamOptimizer(learning_rate=self.l_rate)
self.optimizer = self.optimizer.minimize(self.cost)
# for evaluation
predictions = tf.equal(tf.argmax(logits, 1), self.Y)
self.accuracy = tf.reduce_mean(tf.cast(predictions, 'float'))
self.init_op = tf.initialize_all_variables()
self.saver = tf.train.Saver()
# for gpu
self.config = tf.ConfigProto(log_device_placement=False)
def test(args, device):
full_data = get_data_loader(args)
if args.model_type == "CNN":
from CNN import CNN
model = CNN(args).to(device)
elif args.model_type == "MLP":
from MLP import MLP
model = MLP(args).to(device)
elif args.model_type == "LSTM":
from LSTM import LSTM
model = LSTM(args).to(device)
optimiser = optim.Adam(
model.parameters(), lr=args.learning_rate)
state = torch.load(args.model_path, map_location=device)
model.load_state_dict(state['model'])
optimiser.load_state_dict(state['optimiser'])
total_difference = 0
n = 0
for batch_num, data in enumerate(full_data):
x, y = data[0].float().to(device), data[1].float().to(device)
num_of_predictions = x.shape[0]
pred = model(x)
pred = pred.reshape(y.shape)
total_difference += sum((abs(pred - y)/y) * 100)
n += num_of_predictions
return total_difference/n
def __init__(self,
config: dict = None,
input_dims: list = None,
pose_labels=None,
train_set=None,
val_set=None,
test_set=None,
name: str = "model",
save_period=5,
workers=6):
super(RBF, self).__init__()
M = len(input_dims)
self.name = name
self.input_dims = input_dims
self.input_slice = [0] + list(accumulate(add, input_dims))
self.act = nn.ELU
self.save_period = save_period
self.workers = workers
self.pose_labels = pose_labels if pose_labels is not None else [
None for _ in range(M)
]
self.config = config
self.basis_func = basis_func_dict()[config["basis_func"]]
self.hidden_dim = config["hidden_dim"]
self.keep_prob = config["keep_prob"]
self.k = config["k"]
self.learning_rate = config["lr"]
self.batch_size = config["batch_size"]
self.loss_fn = config[
"loss_fn"] if "loss_fn" in config else nn.functional.mse_loss
self.opt = config[
"optimizer"] if "optimizer" in config else torch.optim.Adam
self.scheduler = config["scheduler"] if "scheduler" in config else None
self.scheduler_param = config[
"scheduler_param"] if "scheduler_param" in config else None
self.models = [
MLP(config=config,
dimensions=[input_dims[i]],
pose_labels=self.pose_labels[i],
name="M" + str(i),
single_module=0) for i in range(M)
]
self.active_models = self.models
self.cluster_model = RBF_Layer(in_features=self.k,
out_features=self.k,
basis_func=self.basis_func)
self.train_set = train_set
self.val_set = val_set
self.test_set = test_set
self.best_val_loss = np.inf
def test_units_accuracy(units, steps, epochs):
accuracies = []
units = range(1, units, steps)
for unit in units:
mlp = MLP(ReLU, X_train.shape[1], [unit])
accuracies.append(test_mlp_model(mlp, epochs, 30, plot=False))
plot_unit_accuracy(accuracies, units)
def mlp_test(test_set, Model, n_input=2030, n_output=150, n_hidden=50):
datasets = load_data(test_set, test_set, test_set)
test_set_x, test_set_y = datasets[0]
index = T.lscalar() # index to a [mini]batch
x = T.vector('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
rng = numpy.random.RandomState(1234)
# construct the MLP class
classifier = MLP(rng=rng,
input=x,
n_in=n_input,
n_hidden=n_hidden,
n_out=n_output,
Model=Model)
#classifier.hiddenLayer.__setstate__((Model['hidden_W'], Model['hidden_b']))
#classifier.logRegressionLayer.__setstate__((Model['logRegression_W'], Model['logRegression_b']))
test_model = theano.function(inputs=[index],
outputs=classifier.predictAll,
givens={
x: test_set_x[index],
})
out = test_model(0).tolist()
return out
def experiment_train_validation_error():
use_validation_set = True
num_hidden_nodes_layer_1 = 20
num_iterations = 1000
learning_rate = 0.002
verbose = False
cases = [1, 2, 3, 4]
cases = [1, 2, 3, 4]
train_MSE = []
val_MSE = []
mse = []
for case in cases:
[inputs, inputs_labels, input_validation,
input_validation_labels] = Utils.create_non_linearly_separable_data_2(
use_validation_set=use_validation_set, case=case)
# Utils.plot_initial_data(inputs.T, inputs_labels)
mlp_batch = MLP(inputs=inputs,
inputs_labels=inputs_labels,
input_validation=input_validation,
input_validation_labels=input_validation_labels,
num_nodes_hidden_layer=num_hidden_nodes_layer_1,
num_iterations=num_iterations,
learning_rate=learning_rate,
batch_train=True,
verbose=verbose)
[_, _, mse_batch] = mlp_batch.fit()
mse.append(mlp_batch.mse)
mse.append(mlp_batch.validation_mse)
legend_names = [
'train mse error case 1', 'validation mse error case 1',
'train mse error case 2', 'validation mse error case 2',
'train mse error case 3', 'validation mse error case 3',
'train mse error case 4', 'validation mse error case 4'
]
Utils.plot_error_with_epochs(
mse,
legend_names=legend_names,
num_epochs=num_iterations,
title='MLP with lr = {0}, iterations = {1} , hidden nodes = {2} '.
format(learning_rate, num_iterations, num_hidden_nodes_layer_1))
def main(client_id):
while True:
try:
response = requests.get(url + 'deepLearning')
json_data = json.loads(response.text)
# response_message = response.content().decode('utf-8')
image_file_index = int(json_data['image_file_index'])
epoch_number = int(json_data['epoch_number'])
print('image_index_file: ' + str(image_file_index))
print('epoch_number: ' + str(epoch_number))
mode = str(json_data['mode'])
print('mode: ' + mode)
if mode == 'stop':
return
if mode == 'wait':
time.sleep(1.5)
continue
client = MongoClient(
'mongodb://*****:*****@ds111529.mlab.com:11529/primre')
_db = client.primre
print('start download network')
try:
network = _db.Network.find_one({'id': 1})
l1_w_list = network['l1_list']
l2_w_list = network['l2_list']
except:
_db.GlobalParameters.update_one(
{'id': 1}, {'$inc': {
'image_file_index': -1
}})
continue
print('finish download network')
lin_neural_network_l1 = L.Linear(784, 300)
lin_neural_network_l2 = L.Linear(300, 10)
for i in range(300):
for j in range(784):
lin_neural_network_l1.W.data[i][j] = l1_w_list[i][j]
for i in range(10):
for j in range(300):
lin_neural_network_l2.W.data[i][j] = l2_w_list[i][j]
mlp = MLP(lin_neural_network_l1, lin_neural_network_l2)
file_images_name = '~/images_train/image_' + str(image_file_index)
file_labels_name = '~/labels_train/label_' + str(image_file_index)
if mode == 'test':
file_images_name = '~/images_test/images_' + str(
image_file_index)
file_labels_name = '~/labels_test/label_' + str(
image_file_index)
if mode == "train":
train(_db, client_id, mlp, file_images_name, file_labels_name,
l1_w_list, l2_w_list)
else:
validate_test(_db, mode, mlp, epoch_number, file_images_name,
file_labels_name)
except:
continue
def runTest(self):
MNIST_DIR = "../Data/mnist.pkl.gz"
f = gzip.open(MNIST_DIR, "rb")
train_set, valid_set, test_set = cPickle.load(f)
f.close()
X_train, y_train = translate(train_set)
X_test, y_test = translate(test_set)
mlp = MLP([784, 800, 300, 10], 0, 0.2, "sigmoid", "mse", load="True")
mlp.fit(X_train, y_train, 100, 500000)
mlp_result, mlp_prediction = mlp.predict(X_test, y_test)
mlp_result, mlp_prediction = mlp.predict(X_train, y_train)
loss = np.mean( (mlp_result - y_train)**2)
print "Loss: ", loss
error = sum([mlp_prediction[i] != train_set[1][i] for i in xrange(len(mlp_prediction))])
error /= float(len(mlp_prediction))
print "Error: ", error
self.assertTrue(error < .1)
def test_depth_accuracy(depth, epochs):
accuracies = []
depths = range(depth)
for depth in depths:
print(f"MLP depth: {depth}")
mlp = MLP(ReLU, X_train.shape[1], [128] * depth)
accuracies.append(test_mlp_model(mlp, epochs, 30, plot=False))
plot_depth_accuracy(accuracies, depths)
def get_mlp(self):
try:
hidden_nodes = int(self.hidden_nodes_input.get())
learning_rate = float(self.learning_rate_input.get())
return MLP(input_nodes=784, hidden_nodes=hidden_nodes,
output_nodes=10, learning_rate=learning_rate)
except ValueError:
return None
def tune(n):
model = MLP((4, ), training_epochs=5000, beta=betas[n], debug=False)
m = Model(model, transfs, gen_x, gen_y, RMSE)
window = [1, 4, 12]
ret = m.expanding_window(X_train, y_train, TRAIN_OFFSET, window, 'dynamic')
print(betas[n])
return betas[n], ret[1][3].iloc[-1, 0], ret[1][0], ret[4][0], ret[12][0]
def main():
"""
Run the test case XOR
"""
print("...Reading dataset")
dataset = load_dataset("datasets/xor.dat")
print("...done!")
print("...Spliting the dataset")
training_samples, testing_samples, labels_train, labels_test = split_dataset(
dataset)
print("...done!")
print("...Creating the classifier")
clf = MLP(input_layer=2, hidden=2, output=1)
print("...done!")
print("...Fiting the clf")
clf.fit(training_samples, labels_train, verbose_error=True)
print("...done!")
print("...Made a prediction")
pred = clf.predict(testing_samples)
print("...done!")
print('Convergence: with MSE:{}'.format(clf.error))
print(clf)
print(
pd.DataFrame.from_items([('Expected', labels_test),
('Obtained', pred)]))
clf.plot_errors()
def __init__(self, img=None, model_path='my_model.npz'):
self.origin_img = img
self.preprocess_img = None
self.detected_number = None
self.number_model_path = model_path
self.net = L.Classifier(MLP(1000, 10))
self.data_directory = "data"
self.img_directory = "imgs" # 画像を保管するディレクトリ
self.setup()
def generateModel(learningRate):
layer0 = Layer(4, 0, 4, ActivationFunction.linear)
layer1 = Layer(3, 1, 4, ActivationFunction.sigmoid)
layer2 = Layer(3, 2, 3, ActivationFunction.sigmoid)
layers = []
layers.append(layer0)
layers.append(layer1)
layers.append(layer2)
return MLP(layers, learningRate)
def __init__ (self, size_x, size_y, beta, hidden, learning_rate, reward):
Bot.__init__(self)
self.bot_name = "Bot_RL_MLP"
self.mlp = MLP (size_x * size_y, hidden, size_x * size_y, learning_rate)
self.reward = reward[:]
#hoher Wert für beta (50?): exploitation
#niedriger Wert für beta : exploration
self.beta = beta
def RecognizerFactory(recognizer_data):
"""Initialize the recognizer. Recognizer data may describe
either an SLP or an MLP, so determine which and represent it
appropriately."""
if LinearRecognizer.validate_input(recognizer_data):
return LinearRecognizer(recognizer_data)
elif SLP.validate_input(recognizer_data): #Should never eval to true
return SLP(recognizer_data)
elif MLP.validate_input(recognizer_data):
return MLP(recognizer_data)
else:
raise Exception("Couldn't parse recognizer data")
def playMLP(self):
# allow to test all combinations of settings
i = 1 ## number hidden layers
step_epochs = 5 ## number of epochs
################################################################################################################
######## To calculate the number of hidden nodes we use a general rule of: (Number of inputs + outputs) x 2/3###
################################################################################################################
k = 3 ## number of hidden neurons
l = 0.0001 ## eta learning rate
s = 0.1 ## step
for k in range(4, 100):
for j in range(1, 20):
self.obj_mlp = MLP(self.int_num_features, self.int_num_classes, i, j * step_epochs, k, l)
self.obj_mlp.train(self.training)
#self.obj_mlp.plotMSE()
self.obj_mlp.test(self.testing)
k += 5
def __init__ (self, size_x = 3, size_y = 3, beta = 1, hidden = 20, learning_rate = 0.1, reward = [0, 1.0, -1.0], initial_field = [0], player_ID = 1):
Bot.__init__(self)
self.initial_field = initial_field
self.player_ID = player_ID
self.bot_name = "Bot_RL_MLP"
self.version = 1
self.counter = 0
self.optimization = []
self.reward = reward[:]
self.first_action = True
self.beta = beta
#hoher Wert für beta (50?): exploitation
#niedriger Wert für beta : exploration
self.mlp = MLP (size_x * size_y, hidden, size_x * size_y, learning_rate)
self.new_game()
def unit_test2():
mlp = MLP([1,2,1]);
mlp.layers[1].set_params([[-.3],[4]],[.15,-2])
mlp.layers[2].set_params([[0,2]],[-1])
num_labels=1
data= [1.0, 1.0]
predicted = mlp.predict(data[:-num_labels])
print 'predicted: ', predicted
mlp.output_error_calculation(data[-num_labels:])
mlp.backpropagation()
print 'errors:'
for layer in mlp.layers:
print layer.e
mlp.gradient_descend(.5)
print 'weights:'
for layer in mlp.layers:
print zip( layer.b, layer.weight)
def unit_test():
mlp = MLP([2,2,2,2]);
mlp.layers[1].set_params([[.2,-.1],[.3,-.3]],[.1,-.2])
mlp.layers[2].set_params([[-.2,-.3],[-.1,.3]],[.1,.2])
mlp.layers[3].set_params([[-.1,.3],[-.2,-.3]],[.2,.1])
num_labels=2
data= [.3, .7, .1, 1.0]
predicted = mlp.predict(data[:-num_labels])
print 'predicted: ', predicted
mlp.output_error_calculation(data[-num_labels:])
mlp.backpropagation()
print 'errors:'
for layer in mlp.layers:
print layer.e
mlp.gradient_descend(.1)
print 'weights:'
for layer in mlp.layers:
print zip( layer.b, layer.weight)
def main():
inputs = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
targets = np.array([0, 1, 1, 0])
# initialize
mlp = MLP(n_input_units=2, n_hidden_units=3, n_output_units=1)
mlp.print_configuration()
# training
mlp.fit(inputs, targets)
print('--- training ---')
print('first layer weight: ')
print(mlp.v)
print('second layer weight: ')
print(mlp.w)
# predict
print('--- predict ---')
for i in [[0, 0], [0, 1], [1, 0], [1, 1]]:
print(i, mlp.predict(i))
def main():
digits = load_digits()
X = digits.data
y = digits.target
X -= X.min()
X /= X.max()
mlp = MLP(64, 100, 10)
mlp.print_configuration()
X_train, X_test, y_train, y_test = train_test_split(X, y)
labels_train = LabelBinarizer().fit_transform(y_train)
labels_test = LabelBinarizer().fit_transform(y_test)
mlp.fit(X_train, labels_train)
predictions = []
for i in range(X_test.shape[0]):
o = mlp.predict(X_test[i])
predictions.append(np.argmax(o))
print confusion_matrix(y_test, predictions)
print classification_report(y_test, predictions)
def train(self, x_input, y_input):
"""
Demonstrate stochastic gradient descent optimization for a multilayer
perceptron
:type learning_rate: float
:param learning_rate: learning rate used (factor for the stochastic gradient
:type L1_reg: float
:param L1_reg: L1-norm's weight when added to the cost (see
regularization)
:type L2_reg: float
:param L2_reg: L2-norm's weight when added to the cost (see
regularization)
:type n_epochs: int
:param n_epochs: maximal number of epochs to run the optimizer
:type dataset: string
:param dataset: /datasets/ACEInhibitors_processed.csv
"""
index = T.lscalar('index') # index to a [mini]batch
train_size = x_input.shape[0] * self.train_percent
max_size = x_input.shape[0] - (x_input.shape[0] % 10)
train_set_x = x_input[:train_size, :]
train_set_y = y_input[:train_size]
valid_set_x = x_input[(train_size + 1 ):max_size, :]
valid_set_y = y_input[(train_size + 1):max_size]
#compute number of minibatches for training, validation and testing
n_train_batches = int(train_set_x.shape[0] / self.batch_size)
n_valid_batches = int(valid_set_x.shape[0] / self.batch_size)
# n_test_batches = int(test_set_x.shape[0] / batch_size)
number_in = train_set_x.shape[1]
valid_set_x = theano.shared(valid_set_x, 'valid_set_x')
valid_set_y = theano.shared(valid_set_y, 'valid_set_y')
train_set_x = theano.shared(train_set_x, 'train_set_x')
train_set_y = theano.shared(train_set_y, 'train_set_y')
# start-snippet-4
# the cost we minimize during training is the negative log likelihood of
# the model plus the regularization terms (L1 and L2); cost is expressed
# here symbolically
self.mlp = MLP(
rng= numpy.random.RandomState(),
input=self.x,
n_in = number_in,
n_out=self.output_size,
a_function = self.activation_function,
n_hidden_sizes=self.hidden_layer_sizes,
dropout=self.dropout,
dropout_rate=self.dropout_rate
)
cost = (
self.mlp.negative_log_likelihood(self.y)
+ self.L1_reg * self.mlp.L1
+ self.L2_reg * self.mlp.L2_sqr
)
# end-snippet-4
# compiling a Theano function that computes the mistakes that are made
# by the model on a minibatch
validate_model = theano.function(
inputs=[index],
outputs=self.mlp.errors(self.y),
givens={
self.x: valid_set_x[index * self.batch_size:(index + 1) * self.batch_size],
self.y: valid_set_y[index * self.batch_size:(index + 1) * self.batch_size]
}
)
training_errors = theano.function(
inputs=[index],
outputs=self.mlp.errors(self.y),
givens={
self.x: train_set_x[index * self.batch_size:(index + 1) * self.batch_size],
self.y: train_set_y[index * self.batch_size:(index + 1) * self.batch_size]
}
)
# start-snippet-5
# compute the gradient of cost with respect to theta (sotred in params)
# the resulting gradients will be stored in a list gparams
parameter_gradients = [T.grad(cost, param) for param in self.mlp.params]
# specify how to update the parameters of the model as a list of
# (variable, update expression) pairs
updates = []
if self.momentum:
delta_before=[]
for param_i in self.mlp.params:
delta_before_i=theano.shared(value=numpy.zeros(param_i.get_value().shape))
delta_before.append(delta_before_i)
for param, parameter_gradients, delta_before_i in zip(self.mlp.params, parameter_gradients, delta_before):
delta_i = -self.learning_rate * parameter_gradients + self.momentum_term*delta_before_i
updates.append((param, param + delta_i))
updates.append((delta_before_i,delta_i))
else:
for param, parameter_gradients in zip(self.mlp.params, parameter_gradients):
updates.append((param, param - self.learning_rate * parameter_gradients))
# compiling a Theano function `train_model` that returns the cost, but
# in the same time updates the parameter of the model based on the rules
# defined in `updates`
train_model = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
self.x: train_set_x[index * self.batch_size: (index + 1) * self.batch_size],
self.y: train_set_y[index * self.batch_size: (index + 1) * self.batch_size]
}
)
###############
# TRAIN MODEL #
###############
print('... training')
# early-stopping parameters
patience = 1000000 # look as this many examples regardless
patience_increase = 2 # wait this much longer when a new best is
# found
improvement_threshold = 0.995 # a relative improvement of this much is
# considered significant
validation_frequency = min(n_train_batches, patience / 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
best_validation_loss = numpy.inf
best_iter = 0
test_score = 0.
start_time = timeit.default_timer()
epoch = 0
done_looping = False
while (epoch < self.n_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in range(n_train_batches):
minibatch_avg_cost = train_model(minibatch_index)
# iteration number
iter = (epoch - 1) * n_train_batches + minibatch_index
if (iter + 1) % validation_frequency == 0:
# compute zero-one loss on validation set
validation_losses = [validate_model(i) for i
in range(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
print(
'epoch %i, minibatch %i/%i, validation error %f %%, cost %f' %
(
epoch,
minibatch_index + 1,
n_train_batches,
this_validation_loss * 100.,
cost
)
)
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
#improve patience if loss improvement is good enough
if (
this_validation_loss < best_validation_loss *
improvement_threshold
):
patience = max(patience, iter * patience_increase)
best_validation_loss = this_validation_loss
best_iter = iter
print((' epoch %i, minibatch %i/%i, test error of '
'best model %f %%') %
(epoch, minibatch_index + 1, n_train_batches,
test_score * 100.))
if patience <= iter:
done_looping = True
break
end_time = timeit.default_timer()
print(('Optimization complete. Best validation score of %f %% '
'obtained at iteration %i, with test performance %f %%') %
(best_validation_loss * 100., best_iter + 1, test_score * 100.))
print(sys.stderr, ('The code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.)))
def test_mlp(learning_rate=0.01, L1_reg=0.0, L2_reg=0.0001, n_epochs=1000,
dataset='mnist.pkl.gz', batch_size=20, n_hidden=500):
# get the datasets
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# compute number of minibatches for training, validation & testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0] // batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] // batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] // batch_size
######################
# BUILD ACTUAL MODEL #
######################
print('... building the model')
index = T.lscalar() # index to minibatch
x = T.matrix(name='x')
y = T.ivector(name='y')
rng = numpy.random.RandomState(1234)
# MLP class
classifier = MLP(rng, x, 28*28, n_hidden, 10)
# cost
cost = (classifier.negative_log_likelihood(y) + \
L1_reg * classifier.L1 + L2_reg * classifier.L2_sqr)
# test function on minibatch
test_model = theano.function(inputs=[index], outputs=classifier.errors(y), \
givens={
x: test_set_x[index*batch_size : (index+1)*batch_size],
y: test_set_y[index*batch_size : (index+1)*batch_size]
})
# validation function on minibatch
validate_model = theano.function(inputs=[index], outputs=classifier.errors(y), \
givens={
x: valid_set_x[index*batch_size : (index+1)*batch_size],
y: valid_set_y[index*batch_size : (index+1)*batch_size]
})
# gradient params
gparams = [T.grad(cost, param) for param in classifier.params]
# updates for training
updates = [(param, param - learning_rate * gparam) \
for param, gparam in zip(classifier.params, gparams)]
# train model on minibatch
train_model = theano.function(inputs=[index], outputs=cost, updates=updates,
givens={
x: train_set_x[index*batch_size : (index+1)*batch_size],
y: train_set_y[index*batch_size : (index+1)*batch_size]
})
###############
# TRAIN MODEL #
###############
print('... training')
# early-stopping parameters
patience = 10000 # look as this many examples regardless
patience_increase = 2 # wait this much longer when a new best is
# found
improvement_threshold = 0.995 # a relative improvement of this much is
# considered significant
validation_frequency = min(n_train_batches, patience // 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
best_validation_loss = numpy.inf
best_iter = 0
test_score = 0.
start_time = timeit.default_timer()
epoch = 0
done_looping = False
while (epoch < n_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in range(n_train_batches):
minibatch_avg_cost = train_model(minibatch_index)
# iteration number
iter = (epoch - 1) * n_train_batches + minibatch_index
if (iter + 1) % validation_frequency == 0:
# compute zero-one loss on validation set
validation_losses = [validate_model(i) for i
in range(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
print(
'epoch %i, minibatch %i/%i, validation error %f %%' %
(
epoch,
minibatch_index + 1,
n_train_batches,
this_validation_loss * 100.
)
)
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
#improve patience if loss improvement is good enough
if (
this_validation_loss < best_validation_loss *
improvement_threshold
):
patience = max(patience, iter * patience_increase)
best_validation_loss = this_validation_loss
best_iter = iter
# test it on the test set
test_losses = [test_model(i) for i
in range(n_test_batches)]
test_score = numpy.mean(test_losses)
print((' epoch %i, minibatch %i/%i, test error of '
'best model %f %%') %
(epoch, minibatch_index + 1, n_train_batches,
test_score * 100.))
if patience <= iter:
done_looping = True
break
end_time = timeit.default_timer()
print(('Optimization complete. Best validation score of %f %% '
'obtained at iteration %i, with test performance %f %%') %
(best_validation_loss * 100., best_iter + 1, test_score * 100.))
print(('The code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.)), file=sys.stderr)
def trainLeNet(train_x, train_y, validation_x, validation_y, test_x, test_y,
convolution_layer_size = None, rate = 0.1, batch_size = 500, n_epochs = 200):
rng = np.random.RandomState(274563533)
x = T.matrix('x')
y = T.ivector('y')
layer_0_input = x.reshape((batch_size, 1, 28, 28))
layer_0 = LeNetConvPoolLayer(rng, input = layer_0_input,
layer_shape = (convolution_layer_size[0], 1, 5, 5),
input_shape = (batch_size, 1, 28, 28),
pool_size = (2,2))
layer_1 = LeNetConvPoolLayer(rng, input = layer_0.output,
layer_shape = (convolution_layer_size[1], convolution_layer_size[0], 5, 5),
input_shape = (batch_size, convolution_layer_size[0], 12, 12),
pool_size = (2,2))
MLP_input = layer_1.output.flatten(2)
layer_final = MLP(MLP_input, convolution_layer_size[1] * 4 * 4, 500, 10)
cost = layer_final.negativeLogLikelihood(y)
error = layer_final.errors(y)
index = T.lscalar('index')
validation_model = function([index], error,
givens={x: validation_x[index * batch_size : (index + 1) * batch_size],
y: validation_y[index * batch_size : (index + 1) * batch_size]})
test_model = function([index], error,
givens={x: test_x[index * batch_size : (index + 1) * batch_size],
y: test_y[index * batch_size : (index + 1) * batch_size]})
params = layer_final.params + layer_1.params + layer_0.params
#for param in params:
# pickle.dump(param, serial)
param_grad = T.grad(cost, params)
updates = [(p, p - rate * pg) for p, pg in zip(params, param_grad)]
train_model = function([index], cost,
givens={x:train_x[index * batch_size : (index + 1) * batch_size],
y:train_y[index * batch_size : (index + 1) * batch_size]},
updates = updates)
n_train_batches = train_x.get_value().shape[0] // batch_size
n_test_batches = test_x.get_value().shape[0] // batch_size
n_validation_batches = validation_x.get_value().shape[0] // batch_size
epoch = 0
best_validation_cost = np.Inf
patience = 1000000
improvement_thread = 0.995
patience_increase = 2
validation_frequency = min(n_train_batches, patience / 2)
loop_done = False
while epoch <= n_epochs and not loop_done:
epoch += 1
for minibatch_index in range(n_train_batches):
batch_cost = train_model(minibatch_index)
iter = (epoch - 1) * n_train_batches + minibatch_index
if iter % 100 == 0:
print 'training @ iter = ', iter
if(iter + 1) % validation_frequency == 0:
validation_losses = [validation_model(i) for i in range(n_validation_batches)]
this_validation_loss = np.mean(validation_losses)
print 'epoch %i, minibatch %i / %i, validation error %f %%' \
% (epoch, minibatch_index+1, n_train_batches, this_validation_loss * 100)
if this_validation_loss < best_validation_cost:
with open('LeNet_params.pkl', 'w') as serial:
pickle.dump(params, serial)
if this_validation_loss < best_validation_cost * improvement_thread:
patience = max(patience, iter * patience_increase)
best_validation_cost = this_validation_loss
test_losses = [test_model(i) for i in range(n_test_batches)]#lkfanldnfaklfnklasnfklasnklfnalksdfnkl
test_loss = np.mean(test_losses)
print 'test error: %f %%'%(test_loss * 100)
if patience <= iter:
loop_done = True
break
def train_model(filename):
learning_rate = 0.05
patience = 10000
size = 1000
batch = 100
loader = DataLoader(filename, batch)
rng = numpy.random.RandomState()
print '... building the model'
x = T.matrix('x')
y = T.ivector('y')
# construct the MLP class
classifier = MLP(
rng=rng,
input=x,
n_in=12*12*5,
n_hidden=size,
n_out=12
)
cost = (
classifier.negative_log_likelihood(y)
)
gparams = [T.grad(cost, param) for param in classifier.params]
updates = [
(param, param - learning_rate * gparam)
for param, gparam in zip(classifier.params, gparams)
]
print '... training'
for i in xrange(patience):
ip, op = loader.get_data();
test_model = theano.function(
inputs=[],
outputs=classifier.errors(y),
givens={
x: ip,
y: op
}
)
train_model = theano.function(
inputs=[],
outputs=cost,
updates=updates,
givens={
x: ip,
y: op
}
)
before = test_model()
train_model()
after = test_model()
print 100.0 * i / patience, '%', before, after
W1 = classifier.params[0].get_value()
b1 = classifier.params[1].get_value()
W2 = classifier.params[2].get_value()
b2 = classifier.params[3].get_value()
W3 = classifier.params[4].get_value()
b3 = classifier.params[5].get_value()
out = open('W1.txt', 'w')
print >> out, '\n'.join(['\t'.join(['%.6f'%item for item in row]) for row in W1])
out.close()
out = open('b1.txt', 'w')
print >> out, '\n'.join(['%.6f'%item for item in b1])
out.close()
out = open('W2.txt', 'w')
print >> out, '\n'.join(['\t'.join(['%.6f'%item for item in row]) for row in W2])
out.close()
out = open('b2.txt', 'w')
print >> out, '\n'.join(['%.6f'%item for item in b2])
out.close()
out = open('W3.txt', 'w')
print >> out, '\n'.join(['\t'.join(['%.6f'%item for item in row]) for row in W3])
out.close()
out = open('b3.txt', 'w')
print >> out, '\n'.join(['%.6f'%item for item in b3])
out.close()
def evaluate_lenet5(learning_rate=0.33, n_epochs=200, dataset="mnist.pkl.gz", nkerns=[32, 32, 32], batch_size=500):
""" Demonstrates lenet on CIFAR-10 dataset
:type learning_rate: float
:param learning_rate: learning rate used (factor for the stochastic
gradient)
:type n_epochs: int
:param n_epochs: maximal number of epochs to run the optimizer
:type nkerns: list of ints
:param nkerns: number of kernels on each layer
"""
rng = numpy.random.RandomState(23455)
def shared_dataset(data_xy, borrow=True):
""" Function that loads the dataset into shared variables
The reason we store our dataset in shared variables is to allow
Theano to copy it into the GPU memory (when code is run on GPU).
Since copying data into the GPU is slow, copying a minibatch everytime
is needed (the default behaviour if the data is not in a shared
variable) would lead to a large decrease in performance.
"""
data_x, data_y = data_xy
shared_x = theano.shared(numpy.asarray(data_x, dtype=theano.config.floatX), borrow=borrow)
shared_y = theano.shared(numpy.asarray(data_y, dtype=theano.config.floatX), borrow=borrow)
# When storing data on the GPU it has to be stored as floats
# therefore we will store the labels as ``floatX`` as well
# (``shared_y`` does exactly that). But during our computations
# we need them as ints (we use labels as index, and if they are
# floats it doesn't make sense) therefore instead of returning
# ``shared_y`` we will have to cast it to int. This little hack
# lets ous get around this issue
return shared_x, T.cast(shared_y, "int32")
data_batch_1 = unpickle("cifar-10-batches-py/data_batch_1")
data_batch_2 = unpickle("cifar-10-batches-py/data_batch_2")
data_batch_3 = unpickle("cifar-10-batches-py/data_batch_3")
data_batch_4 = unpickle("cifar-10-batches-py/data_batch_4")
data_batch_5 = unpickle("cifar-10-batches-py/data_batch_5")
test = unpickle("cifar-10-batches-py/test_batch")
train_set_1 = data_batch_1["data"]
train_set_2 = data_batch_2["data"]
train_set_3 = data_batch_3["data"]
train_set_4 = data_batch_4["data"]
train_set_5 = data_batch_5["data"]
X_train = numpy.concatenate((train_set_1, train_set_2, train_set_3, train_set_4, train_set_5), axis=0)
y_train = numpy.concatenate(
(
data_batch_1["labels"],
data_batch_2["labels"],
data_batch_3["labels"],
data_batch_4["labels"],
data_batch_5["labels"],
)
)
test_set = test["data"]
Xte_rows = test_set.reshape(train_set_1.shape[0], 32 * 32 * 3)
Yte = numpy.asarray(test["labels"])
Xval_rows = X_train[:7500, :] # take first 1000 for validation
Yval = y_train[:7500]
Xtr_rows = X_train[7500:50000, :] # keep last 49,000 for train
Ytr = y_train[7500:50000]
mean_train = Xtr_rows.mean(axis=0)
stdv_train = Xte_rows.std(axis=0)
Xtr_rows = (Xtr_rows - mean_train) / stdv_train
Xval_rows = (Xval_rows - mean_train) / stdv_train
Xte_rows = (Xte_rows - mean_train) / stdv_train
learning_rate = theano.shared(learning_rate)
"""whitening"""
"""
Xtr_rows -= numpy.mean(Xtr_rows, axis=0) # zero-center the data (important)
cov = numpy.dot(Xtr_rows.T, Xtr_rows) / Xtr_rows.shape[0]
U,S,V = numpy.linalg.svd(cov)
Xrot = numpy.dot(Xtr_rows, U)# decorrelate the data
Xrot_reduced = numpy.dot(Xtr_rows, U[:,:100])
# whiten the data:
# divide by the eigenvalues (which are square roots of the singular values)
Xwhite = Xrot / numpy.sqrt(S + 1e-5)"""
"""whitening"""
# Xtr_rows = whiten(Xtr_rows)
# zero-center the data (important)
"""cov = numpy.dot(Xtr_rows.T, Xtr_rows) / Xtr_rows.shape[0]
U,S,V = numpy.linalg.svd(cov)
Xrot = numpy.dot(Xtr_rows, U)
Xtr_rows = Xrot / numpy.sqrt(S + 1e-5)
Xval_rot = numpy.dot(Xval_rows,U)
Xval_rows = Xval_rot / numpy.sqrt(S + 1e-5)
Xte_rot = numpy.dot(Xte_rows,U)
Xte_rows = Xte_rot / numpy.sqrt(S + 1e-5)
"""
train_set = (Xtr_rows, Ytr)
valid_set = (Xval_rows, Yval)
test_set = (Xte_rows, Yte)
test_set_x, test_set_y = shared_dataset(test_set)
valid_set_x, valid_set_y = shared_dataset(valid_set)
train_set_x, train_set_y = shared_dataset(train_set)
datasets = [(train_set_x, train_set_y), (valid_set_x, valid_set_y), (test_set_x, test_set_y)]
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0]
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0]
n_test_batches = test_set_x.get_value(borrow=True).shape[0]
n_train_batches //= batch_size
n_valid_batches //= batch_size
n_test_batches //= batch_size
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
# start-snippet-1
x = T.matrix("x") # the data is presented as rasterized images
y = T.ivector("y") # the labels are presented as 1D vector of [int] labels
######################
# BUILD ACTUAL MODEL #
######################
print("... building the model")
# Reshape matrix of rasterized images of shape (batch_size, 28 * 28)
# to a 4D tensor, compatible with our LeNetConvPoolLayer
# (28, 28) is the size of MNIST images.
layer0_input = x.reshape((batch_size, 3, 32, 32))
# Construct the first convolutional pooling layer:
# filtering reduces the image size to (32-5+1 , 32-5+1) = (28, 28)
# maxpooling reduces this further to (28/2, 28/2) = (14, 14)
# 4D output tensor is thus of shape (batch_size, nkerns[0], 14, 14)
layer0 = LeNetConvPoolLayer(
rng, input=layer0_input, image_shape=(batch_size, 3, 32, 32), filter_shape=(nkerns[0], 3, 5, 5), poolsize=(2, 2)
)
# Construct the second convolutional pooling layer
# filtering reduces the image size to (14-5+1, 14-5+1) = (10, 10)
# maxpooling reduces this further to (10/2, 10/2) = (5, 5)
# 4D output tensor is thus of shape (batch_size, nkerns[1], 5, 5)
layer1 = LeNetConvPoolLayer(
rng,
input=layer0.output,
image_shape=(batch_size, nkerns[0], 14, 14),
filter_shape=(nkerns[1], nkerns[0], 5, 5),
poolsize=(2, 2),
)
# Construct the third convolutional pooling layer
# filtering reduces the image size to (5-2+1, 5-2+1) = (4, 4)
# maxpooling reduces this further to (4/2, 4/2) = (2, 2)
# 4D output tensor is thus of shape (batch_size, nkerns[2], 2, 2)
layer2conv = LeNetConvPoolLayer(
rng,
input=layer1.output,
image_shape=(batch_size, nkerns[1], 5, 5),
filter_shape=(nkerns[2], nkerns[1], 2, 2),
poolsize=(2, 2),
)
# the HiddenLayer being fully-connected, it operates on 2D matrices of
# shape (batch_size, num_pixels) (i.e matrix of rasterized images).
# This will generate a matrix of shape (batch_size, nkerns[1] * 4 * 4),
# or (500, 50 * 4 * 4) = (500, 800) with the default values.
layer3_input = layer2conv.output.flatten(2)
print(layer3_input.shape)
# construct a fully-connected sigmoidal layer
layer3 = HiddenLayer(rng, input=layer3_input, n_in=nkerns[2] * 2 * 2, n_out=64, activation=relu)
layer3_1 = MLP(rng, input=layer3.output, n_in=64, n_hidden=200, n_out=10)
# classify the values of the fully-connected sigmoidal layer
# layer4 = LogisticRegression(input=layer3_1.output, n_in=10, n_out=10)
# the cost we minimize during training is the NLL of the model
L2_reg = 0.005
L2_sqr_model = (
(layer0.W ** 2).sum()
+ (layer1.W ** 2).sum()
+ (layer2conv.W ** 2).sum()
+ (layer3.W ** 2).sum()
+ (layer0.W ** 2).sum()
+ (layer3_1.L2_sqr ** 2).sum()
)
cost = layer3_1.negative_log_likelihood(y) + L2_reg * L2_sqr_model
# create a function to compute the mistakes that are made by the model
test_model = theano.function(
[index],
layer3_1.errors(y),
givens={
x: test_set_x[index * batch_size : (index + 1) * batch_size],
y: test_set_y[index * batch_size : (index + 1) * batch_size],
},
)
validate_model = theano.function(
[index],
layer3_1.errors(y),
givens={
x: valid_set_x[index * batch_size : (index + 1) * batch_size],
y: valid_set_y[index * batch_size : (index + 1) * batch_size],
},
)
# create a list of all model parameters to be fit by gradient descent
params = layer3_1.params + layer3.params + layer2conv.params + layer1.params + layer0.params
# create a list of gradients for all model parameters
grads = T.grad(cost, params)
# train_model is a function that updates the model parameters by
# SGD Since this model has many parameters, it would be tedious to
# manually create an update rule for each model parameter. We thus
# create the updates list by automatically looping over all
# (params[i], grads[i]) pairs.
updates = [(param_i, param_i - learning_rate * grad_i) for param_i, grad_i in zip(params, grads)]
train_model = theano.function(
[index],
cost,
updates=updates,
givens={
x: train_set_x[index * batch_size : (index + 1) * batch_size],
y: train_set_y[index * batch_size : (index + 1) * batch_size],
},
)
# end-snippet-1
###############
# TRAIN MODEL #
###############
print("... training")
# early-stopping parameters
patience = 10000 # look as this many examples regardless
patience_increase = 2 # wait this much longer when a new best is
# found
improvement_threshold = 0.995 # a relative improvement of this much is
# considered significant
validation_frequency = min(n_train_batches, patience // 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
best_validation_loss = numpy.inf
best_iter = 0
test_score = 0.0
start_time = timeit.default_timer()
epoch = 0
done_looping = False
epoch_loss_list = []
epoch_val_list = []
while (epoch < n_epochs) and (not done_looping):
epoch += 1
if epoch == 20:
learning_rate.set_value(0.1)
if epoch >= 21 and learning_rate.get_value() >= 0.1 * (0.9 ** 6):
learning_rate.set_value(learning_rate.get_value() * 0.9)
if epoch > 3:
epoch_loss_np = numpy.reshape(epoch_loss_list, newshape=(len(epoch_loss_list), 3))
epoch_val_np = numpy.reshape(epoch_val_list, newshape=(len(epoch_val_list), 3))
numpy.savetxt(fname="epoc_cost.csv", X=epoch_loss_np, fmt="%1.3f")
numpy.savetxt(fname="epoc_val_error.csv", X=epoch_val_np, fmt="%1.3f")
for minibatch_index in range(n_train_batches):
iter = (epoch - 1) * n_train_batches + minibatch_index
if iter % 100 == 0:
print("training @ iter = ", iter)
cost_ij = train_model(minibatch_index)
epoch_loss_entry = [iter, epoch, float(cost_ij)]
epoch_loss_list.append(epoch_loss_entry)
if (iter + 1) % validation_frequency == 0:
# compute zero-one loss on validation set
validation_losses = [validate_model(i) for i in range(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
print(
"epoch %i, minibatch %i/%i, validation error %f %%"
% (epoch, minibatch_index + 1, n_train_batches, this_validation_loss * 100.0)
)
epoch_val_entry = [iter, epoch, this_validation_loss]
epoch_val_list.append(epoch_val_entry)
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
# improve patience if loss improvement is good enough
if this_validation_loss < best_validation_loss * improvement_threshold:
patience = max(patience, iter * patience_increase)
# save best validation score and iteration number
best_validation_loss = this_validation_loss
best_iter = iter
# test it on the test set
test_losses = [test_model(i) for i in range(n_test_batches)]
test_score = numpy.mean(test_losses)
print(
(" epoch %i, minibatch %i/%i, test error of " "best model %f %%")
% (epoch, minibatch_index + 1, n_train_batches, test_score * 100.0)
)
if patience <= iter:
done_looping = True
break
end_time = timeit.default_timer()
print("Optimization complete.")
print(
"Best validation score of %f %% obtained at iteration %i, "
"with test performance %f %%" % (best_validation_loss * 100.0, best_iter + 1, test_score * 100.0)
)
print(
("The code for file " + os.path.split(__file__)[1] + " ran for %.2fm" % ((end_time - start_time) / 60.0)),
file=sys.stderr,
)
epoch_loss_np = numpy.reshape(epoch_loss_list, newshape=(len(epoch_loss_list), 3))
epoch_val_np = numpy.reshape(epoch_val_list, newshape=(len(epoch_val_list), 3))
epoch_loss = pandas.DataFrame(
{"iter": epoch_loss_np[:, 0], "epoch": epoch_loss_np[:, 1], "cost": epoch_loss_np[:, 2]}
)
epoch_vall = pandas.DataFrame(
{"iter": epoch_val_np[:, 0], "epoch": epoch_val_np[:, 1], "val_error": epoch_val_np[:, 2]}
)
epoc_avg_loss = pandas.DataFrame(epoch_loss.groupby(["epoch"]).mean()["cost"])
epoc_avg_val = pandas.DataFrame(epoch_vall.groupby(["epoch"]).mean()["val_error"])
epoc_avg_loss = pandas.DataFrame({"epoch": epoc_avg_loss.index.values, "cost": epoc_avg_loss["cost"]})
epoc_avg_loss_val = pandas.DataFrame({"epoch": epoc_avg_val.index.values, "val_error": epoc_avg_val["val_error"]})
epoc_avg_loss.plot(kind="line", x="epoch", y="cost")
plt.show()
epoc_avg_loss_val.plot(kind="line", x="epoch", y="val_error")
plt.show()
class Bot_RL_MLP (Bot):
def __init__ (self, size_x = 3, size_y = 3, beta = 1, hidden = 20, learning_rate = 0.1, reward = [0, 1.0, -1.0], initial_field = [0], player_ID = 1):
Bot.__init__(self)
self.initial_field = initial_field
self.player_ID = player_ID
self.bot_name = "Bot_RL_MLP"
self.version = 1
self.counter = 0
self.optimization = []
self.reward = reward[:]
self.first_action = True
self.beta = beta
#hoher Wert für beta (50?): exploitation
#niedriger Wert für beta : exploration
self.mlp = MLP (size_x * size_y, hidden, size_x * size_y, learning_rate)
self.new_game()
"""
Initializes a new game
"""
def new_game(self):
self.first_action = True
self.counter += 1
self.mlp.new_game()
"""
Loads
"""
def load_data(self, filename):
fo = open(filename , "r")
data = json.loads(fo.read())
fo.close()
if (data["bot"] == self.bot_name):
if (data["version"] <= self.version):
self.player_ID = data["player_ID"]
self.initial_field = data["initial_field"]
self.counter = data["counter"]
self.optimization = data["optimization"]
self.reward = data["reward"]
self.first_action = data["first_action"]
self.beta = data["beta"]
self.mlp.set_data(data["MLP"])
else:
raise ValueError('dataset is not usable by Bot : different Bot identifier')
else:
raise ValueError('dataset is not usable by this Bot version : dataset version is higher than Bot version')
return data
"""
Saves
"""
def save_data(self, filename):
data = {"bot" : self.bot_name,
"version" : self.version,
"player_ID" : self.player_ID,
"initial_field" : self.initial_field,
"counter" : self.counter,
"optimization" : self.optimization,
"reward" : self.reward,
"first_action" : self.first_action,
"beta" : self.beta,
"MLP" : self.mlp.get_data()}
fo = open(filename , "w")
fo.write(json.dumps(data))
fo.close()
"""
Returns an action depending on the given world
"""
def get_action(self, world_old):
self.info_tic = world_old.get_sensor_info()
self.h_tic = self.mlp.get_action(self.info_tic)
#for i in range(len(self.h_tic)):
# if (self.info_tic[i] > 0):
# self.h_tic[i] = -100000
#Workaround: Wenn nur noch 1 Zug möglich ist, automatisch setzen
moves = world_old.get_moves()
if (len(moves) == 1):
self.act_tic = moves[0]
else:
#Auswahl wiederholen bis ein gültiger Zug ausgewählt wurde
validation = False
while (validation == False):
new_h_tic = []
for i in range(len(self.h_tic)):
if (i in moves):
new_h_tic.append(self.h_tic[i])
self.act_tic = moves[self.rand_winner (new_h_tic, self.beta)] # choose action
#print self.info, self.act
#print "----------\n",self.h_tic, "\n",moves, "\n",new_h_tic, "\n",self.act_tic
x = self.act_tic % world_old.size_x
y = self.act_tic / world_old.size_y
validation = world_old.check_action(x, y)
#Umrechnen 1D -> 2D
x = self.act_tic % world_old.size_x
y = self.act_tic / world_old.size_y
#print "--------------------------"
#print self.h, "->", self.act, "->", x, ",", y
#print "--------------------------"
return (x, y)
"""
Adapts the MLP considering the results (world_new) of its last action
"""
def evaluate_action(self, world_new):
if (self.first_action == False):
r = self.get_reward(world_new.get_winner()) # read reward
#Berechnen der Q-Werte vor und nach der Aktion
q0 = self.h[self.act]
q1 = self.mlp.get_action(world_new.get_sensor_info())[self.act_tic]
#Berechnen der Belohnung auf dem neuen Feld
r = self.get_reward(world_new.get_winner()) # read reward
if (r == self.get_reward(1)): # This is cleaner than defining
target = r # target as r + 0.9 * q1,
else: # because weights now converge.
target = 0.9 * q1 # gamma = 0.9
delta = target - q0 # prediction error
#Wichtig : nur das delta an der Position der Aktion wird als Fehler betrachtet, für alle anderen
#Positionen ist der Fehler 0
error = np.zeros (self.mlp.input_size)
error[self.act] = delta
#Wichtig : Das Lernen erfolgt mittels des Fehlers und der Welt VOR der Aktion
self.mlp.evaluate_action_RL(self.info, error)
#print q0, q1, delta
self.info = self.info_tic
self.h = self.h_tic
self.act = self.act_tic
self.first_action = False
"""
Selects an action
"""
def rand_winner (self, S_from, beta):
#for i in range (len(S_from)):
# if S_from[i] > 200:
# print S_from
# time.sleep(0.2)
#print "--------------------\n",S_from
#time.sleep(0.2)
sum = 0.0
p_i = 0.0
rnd = np.random.random()
d_r = len (S_from)
sel = 0
try:
for i in range (d_r):
sum += np.exp (beta * min(S_from[i],200))
#if field is empty, set reward to 1 for all fields
#to get a probablity higher than 0
if (sum == 0):
sum = d_r
S_from = [1]*d_r
for i in range (d_r):
p_i += np.exp (beta * min(S_from[i],200)) / sum
if p_i > rnd:
sel = i
rnd = 1.1 # out of reach, so the next will not be turned ON
except Exception:
print beta, S_from[i], S_from, sum
return sel
"""
Calculates the reward for the actual board setup
"""
def get_reward (self, winner):
if ((winner >= 0) and (winner <= 2)):
return self.reward[int(winner)]
else:
return 0.0
return [(a-m)*b-2 for a,b,m in zip(x,norm,minv)]
####### all files used
ftr = open(('./'+ feature +'/train.ark') , 'r') #training set
fte = open(('./'+ feature +'/test.ark') , 'r') # testing set
if useState:
flab = open('./state_label/train.lab' , 'r') # label
else:
flab = open('./label/train.lab' , 'r') # label
fmap = open('./phones/state_48_39.map' , 'r') # label mapping 48-39
######## model initialization
model = MLP(
n_in=raw_feat_num*(pre+post+1) ,
n_out=1943 if useState else 48 ,
hidStruct = hiddenStruct ,
hidAct = hiddenActFunc ,
pDrop = dropout if doDrop else 0.0
)
###### prediction function (theano type)
x = T.matrix('x')
pred = theano.function([x] , model.predict(x))
######## model trainer initialization
trainer = MLPtrainer.MLPtrainer(
net = model ,
learning_rate = alpha ,
momentum = momentum ,
L1 = L1 ,
L2 = L2 )
def test_pickle_mlp(learning_rate=0.01, L1_reg=0.00, L2_reg=0.0001,
n_epochs=10, dataset='../data/mnist.pkl.gz', batch_size=20,
pickle_file='/scratch/z/zhaolei/lzamparo/gpu_tests/mlp_results/MLP_pickle.pkl',n_hidden=500):
""" Interrupt the training of an MLP, pickle the MLP object, unpickle, and continue """
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# compute number of minibatches for each set
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
### Build the model ###
print '... building the model'
# allocate symbolic variables for the data
index = T.lscalar()
x = T.matrix('x')
y = T.ivector('y')
rng = numpy.random.RandomState(1234)
# construct the MLP class
classifier = MLP(rng = rng, input = x, n_in=28*28, n_hidden=n_hidden, n_out=10)
# cost to be minimized
cost = classifier.negative_log_likelihood(y) \
+ L1_reg * classifier.L1 \
+ L2_reg * classifier.L2_sqr
# theano function that computes the mistakes made by the model on a minibatch
test_model = theano.function(inputs=[index],
outputs = classifier.errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]})
# theano function to validate the model
validate_model = theano.function(inputs=[index],
outputs = classifier.errors(y),
givens = {
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]})
# compute the gradient of the cost function w.r.t theta
gparams = []
for param in classifier.params:
gparam = T.grad(cost, param)
gparams.append(gparam)
# build the list of parameter updates. This consists of tuples of paramters and values
updates = []
for param, gparam in zip(classifier.params, gparams):
updates.append((param, param - learning_rate * gparam))
# compile a Theano function to return the cost, update the parameters based on the
# updates list
train_model = theano.function(inputs=[index], outputs=cost,
updates=updates,
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]})
### train the model ###
print '... training'
# early-stopping parameters
patience = 10000 # look at this number of examples regardless
patience_increase = 2 # wait this many more epochs when a new best comes up
improvement_threshold = 0.995 # a relative improvement threshold for significance
validation_frequency = min(n_train_batches, patience / 2)
# train for this many minibatches before checking the model on the validation set
best_params = None
best_validation_loss = numpy.inf
best_iter = 0
test_score = 0.
start_time = time.clock()
epoch = 0
done_looping = False
halfway_point = n_epochs / 2
while (epoch < halfway_point) and (not done_looping):
epoch = epoch + 1
for minibatch_index in xrange(n_train_batches):
minibatch_avg_cost = train_model(minibatch_index)
iter = (epoch - 1) * n_train_batches + minibatch_index
# do we validate?
if (iter + 1) % validation_frequency == 0:
validation_losses = [validate_model(i) for i in xrange(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
print('epoch %i, minibatch %i/%i, validation error %f %%' %
(epoch, minibatch_index + 1, n_train_batches,
this_validation_loss * 100.))
if this_validation_loss < best_validation_loss:
# increase patience if loss improvement is good enough
if this_validation_loss < best_validation_loss * \
improvement_threshold:
patience = max(patience, iter * patience_increase)
best_validation_loss = this_validation_loss
best_iter = iter
# test on the test set
test_losses = [test_model(i) for i in xrange(n_test_batches)]
test_scores = numpy.mean(test_losses)
print((' epoch %i, minibatch %i/%i, test error of '
'best model %f %%') %
(epoch, minibatch_index + 1, n_train_batches,
test_score * 100.))
if patience <= iter:
done_looping = True
break
end_time = time.clock()
print(('Halfway point reached. Best validation score of %f %% '
'obtained at iteration %i, with test performance %f %%') %
(best_validation_loss * 100., best_iter + 1, test_score * 100.))
print >> sys.stderr, ('The code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
print "Pickling model..."
f = file(pickle_file, 'wb')
cPickle.dump(classifier, f, protocol=cPickle.HIGHEST_PROTOCOL)
f.close()
print "Unpickling the model..."
f = file(pickle_file, 'rb')
unpickled_classifier = cPickle.load(f)
unpickled_classifier.reconstruct_state(x, T.tanh)
f.close()
### Re-establish the cost, grad, parameter updates ###
# cost to be minimized
cost = unpickled_classifier.negative_log_likelihood(y) \
+ L1_reg * unpickled_classifier.L1 \
+ L2_reg * unpickled_classifier.L2_sqr
# theano function that computes the mistakes made by the model on a minibatch
test_model = theano.function(inputs=[index],
outputs = unpickled_classifier.errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]})
# theano function to validate the model
validate_model = theano.function(inputs=[index],
outputs = unpickled_classifier.errors(y),
givens = {
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]})
# compute the gradient of the cost function w.r.t theta
gparams = []
for param in unpickled_classifier.params:
gparam = T.grad(cost, param)
gparams.append(gparam)
# build the list of parameter updates. This consists of tuples of paramters and values
updates = []
for param, gparam in zip(unpickled_classifier.params, gparams):
updates.append((param, param - learning_rate * gparam))
print(("Continue training for %i epochs ") % (n_epochs - epoch))
start_time = time.clock()
while (epoch < n_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in xrange(n_train_batches):
minibatch_avg_cost = train_model(minibatch_index)
iter = (epoch - 1) * n_train_batches + minibatch_index
# do we validate?
if (iter + 1) % validation_frequency == 0:
validation_losses = [validate_model(i) for i in xrange(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
print('epoch %i, minibatch %i/%i, validation error %f %%' %
(epoch, minibatch_index + 1, n_train_batches,
this_validation_loss * 100.))
if this_validation_loss < best_validation_loss:
# increase patience if loss improvement is good enough
if this_validation_loss < best_validation_loss * \
improvement_threshold:
patience = max(patience, iter * patience_increase)
best_validation_loss = this_validation_loss
best_iter = iter
# test on the test set
test_losses = [test_model(i) for i in xrange(n_test_batches)]
test_scores = numpy.mean(test_losses)
print((' epoch %i, minibatch %i/%i, test error of '
'best model %f %%') %
(epoch, minibatch_index + 1, n_train_batches,
test_score * 100.))
if patience <= iter:
done_looping = True
break
end_time = time.clock()
print(('End point reached. Best validation score of %f %% '
'obtained at iteration %i, with test performance %f %%') %
(best_validation_loss * 100., best_iter + 1, test_score * 100.))
print >> sys.stderr, ('The code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
class Controller:
# dataset
data = []
training = []
testing = []
# characteristics
int_set_size = 0
int_num_features = 0
int_training_size = 30
int_testing_size = 20
int_num_per_class = 50
int_num_classes = 50
# composition
obj_mlp = MLP
# initializiation
def __init__(self):
iris = load_iris()
self.prepareDataSet(iris)
# normalize or scale data
# divide data to training and testing
def prepareDataSet(self, iris):
self.int_num_classes = numpy.unique(iris.target).shape[0]
self.int_set_size = iris.data.shape[0]
self.int_num_features = iris.data.shape[1]
## normalize data
self.data = preprocessing.normalize(iris.data)
self.data = preprocessing.minmax_scale(self.data, (-1, 1))
# load data in arrays
for i in range(0, len(self.data)):
Y = iris.target[i]
X = numpy.append(self.data[i], 1) ## bias input = 1
check_i = i % self.int_num_per_class
if check_i < self.int_training_size:
self.training.append([X, Y])
else:
self.testing.append([X, Y])
numpy.random.shuffle(self.training)
numpy.random.shuffle(self.testing)
# play the MLP
def playMLP(self):
# allow to test all combinations of settings
i = 1 ## number hidden layers
step_epochs = 5 ## number of epochs
################################################################################################################
######## To calculate the number of hidden nodes we use a general rule of: (Number of inputs + outputs) x 2/3###
################################################################################################################
k = 3 ## number of hidden neurons
l = 0.0001 ## eta learning rate
s = 0.1 ## step
for k in range(4, 100):
for j in range(1, 20):
self.obj_mlp = MLP(self.int_num_features, self.int_num_classes, i, j * step_epochs, k, l)
self.obj_mlp.train(self.training)
#self.obj_mlp.plotMSE()
self.obj_mlp.test(self.testing)
k += 5
class Bot_RL_MLP (Bot):
def __init__ (self, size_x, size_y, beta, hidden, learning_rate, reward):
Bot.__init__(self)
self.bot_name = "Bot_RL_MLP"
self.mlp = MLP (size_x * size_y, hidden, size_x * size_y, learning_rate)
self.reward = reward[:]
#hoher Wert für beta (50?): exploitation
#niedriger Wert für beta : exploration
self.beta = beta
"""
Returns an action depending on the given world
"""
def get_action(self, world):
self.info = world.get_sensor_info()
self.h = self.mlp.get_action(self.info)
for i in range(len(self.h)):
if (self.info[i] > 0):
self.h[i] = -10
#Workaround: Wenn nur noch 1 Zug möglich ist, automatisch setzen
moves = world.get_moves()
if (len(moves) == 1):
self.act = moves[0]
else:
#Auswahl wiederholen bis ein gültiger Zug ausgewählt wurde
validation = False
while (validation == False):
self.act = self.rand_winner (self.h, self.beta) # choose action
#print self.info, self.act
x = self.act % world.size_x
y = self.act / world.size_y
validation = world.check_action(x, y)
#Umrechnen 1D -> 2D
x = self.act % world.size_x
y = self.act / world.size_y
#print "--------------------------"
#print self.h, "->", self.act, "->", x, ",", y
#print "--------------------------"
return (x, y)
"""
Adapts the MLP considering the results (world_new) of its last action
"""
def evaluate_action(self, world_new):
#Erstellen des Aktions-Vektors
act_vec = np.zeros (self.mlp.input_size)
act_vec[self.act] = 1.0
#Berechnen der Q-Werte vor und nach der Aktion
q0 = self.h[self.act]
q1 = self.mlp.get_action(world_new.get_sensor_info())[self.act]
#Berechnen der Belohnung auf dem neuen Feld
r = self.get_reward(world_new.get_winner()) # read reward
if (r == self.get_reward(1)): # This is cleaner than defining
target = r # target as r + 0.9 * q1,
else: # because weights now converge.
target = 0.9 * q1 # gamma = 0.9
delta = target - q0 # prediction error
#Wichtig : nur das delta an der Position der Aktion wird als Fehler betrachtet, für alle anderen
#Positionen ist der Fehler 0
error = np.zeros (self.mlp.input_size)
error[self.act] = delta
#Wichtig : Das Lernen erfolgt mittels des Fehlers und der Welt VOR der Aktion
self.mlp.evaluate_action(self.info, error)
"""
Selects an action
"""
def rand_winner (self, S_from, beta):
#for i in range (len(S_from)):
# if S_from[i] > 200:
# print S_from
# time.sleep(0.2)
#print "--------------------\n",S_from
#time.sleep(0.2)
sum = 0.0
p_i = 0.0
rnd = np.random.random()
d_r = len (S_from)
sel = 0
try:
for i in range (d_r):
sum += np.exp (beta * min(S_from[i],200))
#if field is empty, set reward to 1 for all fields
#to get a probablity higher than 0
if (sum == 0):
sum = d_r
S_from = [1]*d_r
for i in range (d_r):
p_i += np.exp (beta * min(S_from[i],200)) / sum
if p_i > rnd:
sel = i
rnd = 1.1 # out of reach, so the next will not be turned ON
except Exception:
print beta, S_from[i], S_from, sum
return sel
"""
Calculates the reward for the actual board setup
"""
def get_reward (self, winner):
if ((winner >= 0) and (winner <= 2)):
return self.reward[int(winner)]
else:
return 0.0
"""
Loads
"""
#def load_data(self, filename):
# fo = open(filename , "r")
# #self.w_mot = json_tricks.load(fo.read())["w_mot"]
# data = json_tricks.load(fo.read())
# fo.close()
#
# return data
"""
Saves
"""
#def save_data(self, filename):
# data = {"bot" : "Bot_RL_MLP",
# "version" : 1,
# "mlp" : self.mlp}
# fo = open(filename , "w")
# fo.write(json_tricks.dumps(data))
# fo.close()
return [(a-m)/b for a,b,m in zip(x,std,mean)]
####### all files used
ftr = open(('./'+ feature +'/train.ark') , 'r') #training set
fte = open(('./'+ feature +'/test.ark') , 'r') # testing set
if useState:
flab = open('./state_label/train.lab' , 'r') # label
else:
flab = open('./label/train.lab' , 'r') # label
fmap = open('./phones/state_48_39.map' , 'r') # label mapping 48-39
######## model initialization
model = MLP(
n_in=raw_feat_num*(pre+post+1) ,
n_out=1943 if useState else 48 ,
hidStruct = hiddenStruct ,
hidAct = hiddenActFunc ,
pDrop = dropout if doDrop else 0.0
)
###### prediction function (theano type)
x = T.matrix('x')
pred = theano.function([x] , model.predict(x))
pred_max = theano.function([x] , model.predict_max(x))
######## model trainer initialization
trainer = MLPtrainer.MLPtrainer(
net = model ,
learning_rate = alpha ,
momentum = momentum ,
L1 = L1 ,
momentum = 0.98
L1 = 0.0
L2 = 0.0
dropout = 0.3
outfilename = 'result_wav.csv'
feat_num = 400
####### all files used
wavDir = './wav/'
flab = open('./label/train.lab' , 'r') # label
fmap = open('./phones/48_39.map' , 'r') # label mapping 48-39
######## model initialization
model = MLP(
n_in=feat_num ,
n_out=48 ,
hidStruct = hiddenStruct
)
###### prediction function (theano type)
x = T.vector('x')
pred = theano.function([x] , model.predict(x))
######## model trainer initialization
trainer = MLPtrainer.MLPtrainer(
net = model ,
learning_rate = alpha ,
momentum = momentum ,
L1 = L1 ,
L2 = L2 )
def MLP_demo(learning_rate=0.01, L1_reg=0.00, L2_reg=0.0001, n_epochs=1000, dataset='mnist.pkl.gz', batch_size=1, n_hidden=309):
datasets = load_multi()
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
print '... building the model'
index = T.lscalar()
x = T.matrix('x')
y = T.ivector('y')
rng = np.random.RandomState(1234)
classifier = MLP(rng, x, y, n_in=103, n_hidden=n_hidden, n_out=9)
test_model = theano.function(inputs=[index],
outputs=classifier.errors(),
givens={x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]})
validate_model = theano.function(inputs=[index],
outputs=classifier.errors(),
givens={x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]})
cost, updates = classifier.get_cost_updates(learning_rate=learning_rate, L1_reg=L1_reg, L2_reg=L2_reg)
train_model = theano.function(inputs=[index],
outputs=cost,
updates=updates,
givens={x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]})
print '... training'
patience = 10000
patience_increase = 2
improvement_threshold = 0.995
validation_frequency = min(n_train_batches, patience / 2)
best_validation_loss = np.inf
best_iter = 0
test_score = 0.
start_time = time.clock()
epoch = 0
done_looping = False
while (epoch < n_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in xrange(n_train_batches):
train_model(minibatch_index)
iter = (epoch - 1) * n_train_batches + minibatch_index
if (iter + 1) % validation_frequency == 0:
validation_losses = [validate_model(i) for i in xrange(n_valid_batches)]
this_validation_loss = np.mean(validation_losses)
print('epoch %i, minibatch %i/%i, validation error %f %%' %
(epoch, minibatch_index + 1, n_train_batches,
this_validation_loss * 100.))
if this_validation_loss < best_validation_loss:
if this_validation_loss < best_validation_loss * improvement_threshold:
patience = max(patience, iter * patience_increase)
best_validation_loss = this_validation_loss
best_iter = iter
test_losses = [test_model(i) for i in xrange(n_test_batches)]
test_score = np.mean(test_losses)
print((' epoch %i, minibatch %i/%i, test error of '
'best model %f %%') %
(epoch, minibatch_index + 1, n_train_batches,
test_score * 100.))
if patience <= iter:
done_looping = True
break
end_time = time.clock()
print(('Optimization complete. Best validation score of %f %% '
'obtained at iteration %i, with test performance %f %%') %
(best_validation_loss * 100., best_iter + 1, test_score * 100.))
print >> sys.stderr, ('The code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
class NeuralNet():
"""
Attributes:
features: Numpy array matrix that represents features
targets: Numpy array matrix that represents the
"""
def __init__(self, n_hidden_units, batch_size, output_size, metric_list="none", learning_rate=1, l1_term=0,
l2_term=0, n_epochs=100, activation_function='tanh', train_p=.6, dropout=False, dropout_rate=.5,
momentum=False, momentum_term=.9, adaptive_learning_rate=False):
# allocate symbolic variables for the data
self.x = T.matrix('x') #
self.y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
self.dropout = dropout
self.dropout_rate = dropout_rate
if metric_list == "none":
self.metrics = {"F1": 0, "Accuracy": 0, "AUC": 0, "Precision": 0, "Recall": 0}
else:
self.metrics = metric_list
self.learning_rate = learning_rate
self.L1_reg = l1_term
self.L2_reg = l2_term
self.n_epochs = n_epochs
self.batch_size = batch_size
self.train_percent = train_p
#Define new ReLU activation function
def relu(x):
return T.switch(x < 0, 0, x)
if activation_function == 'relu':
self.activation_function = relu
elif activation_function == 'tanh':
self.activation_function = T.tanh
elif activation_function == 'sigmoid':
self.activation_function = T.nnet.sigmoid
self.output_size = output_size
self.hidden_layer_sizes = n_hidden_units
self.n_epochs = n_epochs
self.momentum = momentum
self.momentum_term = momentum_term
def train(self, x_input, y_input):
"""
Demonstrate stochastic gradient descent optimization for a multilayer
perceptron
:type learning_rate: float
:param learning_rate: learning rate used (factor for the stochastic gradient
:type L1_reg: float
:param L1_reg: L1-norm's weight when added to the cost (see
regularization)
:type L2_reg: float
:param L2_reg: L2-norm's weight when added to the cost (see
regularization)
:type n_epochs: int
:param n_epochs: maximal number of epochs to run the optimizer
:type dataset: string
:param dataset: /datasets/ACEInhibitors_processed.csv
"""
index = T.lscalar('index') # index to a [mini]batch
train_size = x_input.shape[0] * self.train_percent
max_size = x_input.shape[0] - (x_input.shape[0] % 10)
train_set_x = x_input[:train_size, :]
train_set_y = y_input[:train_size]
valid_set_x = x_input[(train_size + 1 ):max_size, :]
valid_set_y = y_input[(train_size + 1):max_size]
#compute number of minibatches for training, validation and testing
n_train_batches = int(train_set_x.shape[0] / self.batch_size)
n_valid_batches = int(valid_set_x.shape[0] / self.batch_size)
# n_test_batches = int(test_set_x.shape[0] / batch_size)
number_in = train_set_x.shape[1]
valid_set_x = theano.shared(valid_set_x, 'valid_set_x')
valid_set_y = theano.shared(valid_set_y, 'valid_set_y')
train_set_x = theano.shared(train_set_x, 'train_set_x')
train_set_y = theano.shared(train_set_y, 'train_set_y')
# start-snippet-4
# the cost we minimize during training is the negative log likelihood of
# the model plus the regularization terms (L1 and L2); cost is expressed
# here symbolically
self.mlp = MLP(
rng= numpy.random.RandomState(),
input=self.x,
n_in = number_in,
n_out=self.output_size,
a_function = self.activation_function,
n_hidden_sizes=self.hidden_layer_sizes,
dropout=self.dropout,
dropout_rate=self.dropout_rate
)
cost = (
self.mlp.negative_log_likelihood(self.y)
+ self.L1_reg * self.mlp.L1
+ self.L2_reg * self.mlp.L2_sqr
)
# end-snippet-4
# compiling a Theano function that computes the mistakes that are made
# by the model on a minibatch
validate_model = theano.function(
inputs=[index],
outputs=self.mlp.errors(self.y),
givens={
self.x: valid_set_x[index * self.batch_size:(index + 1) * self.batch_size],
self.y: valid_set_y[index * self.batch_size:(index + 1) * self.batch_size]
}
)
training_errors = theano.function(
inputs=[index],
outputs=self.mlp.errors(self.y),
givens={
self.x: train_set_x[index * self.batch_size:(index + 1) * self.batch_size],
self.y: train_set_y[index * self.batch_size:(index + 1) * self.batch_size]
}
)
# start-snippet-5
# compute the gradient of cost with respect to theta (sotred in params)
# the resulting gradients will be stored in a list gparams
parameter_gradients = [T.grad(cost, param) for param in self.mlp.params]
# specify how to update the parameters of the model as a list of
# (variable, update expression) pairs
updates = []
if self.momentum:
delta_before=[]
for param_i in self.mlp.params:
delta_before_i=theano.shared(value=numpy.zeros(param_i.get_value().shape))
delta_before.append(delta_before_i)
for param, parameter_gradients, delta_before_i in zip(self.mlp.params, parameter_gradients, delta_before):
delta_i = -self.learning_rate * parameter_gradients + self.momentum_term*delta_before_i
updates.append((param, param + delta_i))
updates.append((delta_before_i,delta_i))
else:
for param, parameter_gradients in zip(self.mlp.params, parameter_gradients):
updates.append((param, param - self.learning_rate * parameter_gradients))
# compiling a Theano function `train_model` that returns the cost, but
# in the same time updates the parameter of the model based on the rules
# defined in `updates`
train_model = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
self.x: train_set_x[index * self.batch_size: (index + 1) * self.batch_size],
self.y: train_set_y[index * self.batch_size: (index + 1) * self.batch_size]
}
)
###############
# TRAIN MODEL #
###############
print('... training')
# early-stopping parameters
patience = 1000000 # look as this many examples regardless
patience_increase = 2 # wait this much longer when a new best is
# found
improvement_threshold = 0.995 # a relative improvement of this much is
# considered significant
validation_frequency = min(n_train_batches, patience / 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
best_validation_loss = numpy.inf
best_iter = 0
test_score = 0.
start_time = timeit.default_timer()
epoch = 0
done_looping = False
while (epoch < self.n_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in range(n_train_batches):
minibatch_avg_cost = train_model(minibatch_index)
# iteration number
iter = (epoch - 1) * n_train_batches + minibatch_index
if (iter + 1) % validation_frequency == 0:
# compute zero-one loss on validation set
validation_losses = [validate_model(i) for i
in range(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
print(
'epoch %i, minibatch %i/%i, validation error %f %%, cost %f' %
(
epoch,
minibatch_index + 1,
n_train_batches,
this_validation_loss * 100.,
cost
)
)
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
#improve patience if loss improvement is good enough
if (
this_validation_loss < best_validation_loss *
improvement_threshold
):
patience = max(patience, iter * patience_increase)
best_validation_loss = this_validation_loss
best_iter = iter
print((' epoch %i, minibatch %i/%i, test error of '
'best model %f %%') %
(epoch, minibatch_index + 1, n_train_batches,
test_score * 100.))
if patience <= iter:
done_looping = True
break
end_time = timeit.default_timer()
print(('Optimization complete. Best validation score of %f %% '
'obtained at iteration %i, with test performance %f %%') %
(best_validation_loss * 100., best_iter + 1, test_score * 100.))
print(sys.stderr, ('The code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.)))
def setup_labels(self, y):
assert "There is no need to relabel if n_classes < 2 ", y < 2
negative_example_label = 2
#Transform matrices and relabel them for the neural network
for i, yi in enumerate(y):
if i > 0:
negative_example_label = negative_example_label+2
positive_example_label = negative_example_label+1
relabeled_y = yi
relabeled_y[relabeled_y == 0] = negative_example_label
relabeled_y[relabeled_y == 1] = positive_example_label
if i == 0:
neural_net_y = relabeled_y
else:
neural_net_y = numpy.hstack((neural_net_y, relabeled_y))
neural_net_y = numpy.intc(neural_net_y)
return neural_net_y
def test(self, x, y):
prediction = self.predict(x)
f1 = f1_score(y, prediction)
precision = precision_score(y, prediction)
recall = recall_score(y, prediction)
auc = roc_auc_score(y, prediction)
accuracy = accuracy_score(y, prediction)
self.metrics["F1"] = f1
self.metrics["Precision"] = precision
self.metrics["Recall"] = recall
self.metrics["AUC"] = auc
self.metrics["Accuracy"] = accuracy
def predict(self, x):
#Create a theano shared variable for the input x: the data to be predicted
test_set_x = theano.shared(x, 'test_set_x')
input = test_set_x
#Iterate over all the hidden layers in the MLP
for i_hidden_layer, hidden_layer in enumerate(self.mlp.hidden_layers):
hl_W = hidden_layer.W
hl_b = hidden_layer.b
if self.dropout:
hl_W *= self.dropout_rate
weight_matrix = self.activation_function(T.dot(input, hl_W) + hl_b)
#Multiply the weights by the expected value of the dropout which is just the
#dropoutrate so in most cases half the weights but only at test time
input = weight_matrix
#Get the weights and bias from the softmax output layer
W = self.mlp.logRegressionLayer.W
b = self.mlp.logRegressionLayer.b
#compile the thenao function for calculating the outputs from the softmax layer
get_y_prediction = theano.function(
inputs=[],
outputs=T.argmax(T.nnet.softmax(T.dot(weight_matrix, W) + b), axis=1),
on_unused_input='ignore',
)
return get_y_prediction()
def transfer_learned_weights(self, x):
a_function = self.activation_function
final_hidden_layer = self.mlp.hidden_layers[-1]
w = final_hidden_layer.W
b = final_hidden_layer.b
if self.dropout:
transformation_function = theano.function(
inputs=[],
outputs=a_function(T.dot(x, (w * self.dropout_rate)) + b),
on_unused_input='ignore',
)
else:
transformation_function = theano.function(
inputs=[],
outputs=a_function(T.dot(x, w) + b),
on_unused_input='ignore',
)
return transformation_function()
def __str__(self):
return "MLP:\nF1 Score: {}\nPrecision: {}\n" \
"Recall: {}\nAccuracy: {}\nROC: {}\n".format(self.metrics['F1'],
self.metrics['Precision'],
self.metrics['Recall'],
self.metrics['Accuracy'],
self.metrics['AUC'])
valid_labels,
clf.predict(valid_dataset)
)
logging.info('Epoch [%d] Cost %f, Validation Accurary %.2f%%'%(iepoch, cost, valid_acc * 100))
if epoch_costs and iepoch > 0 and iepoch % model_commit_freq == 0:
commit_model(clf)
end_time = time.time()
logging.info('Model training completed in %.0fs'%(end_time - start_time))
plt.plot(np.arange(0, num_epochs), avg_costs)
plt.show()
commit_model(clf)
return
clf = load_model()
if clf is None:
clf = MLP(beta=beta, n_in=image_size * image_size,
n_hidden=1024,
num_hidden_layers=num_hidden_layers,
n_out=num_labels,
activation=theano.tensor.nnet.sigmoid)
train_model(clf)
elif force_training:
logging.info('force_training is set. Model will be retrained')
train_model(clf)
acc = accuracy_score(test_labels, clf.predict(test_dataset)) * 100
logging.info('Model accuracy %.2f%%'%acc)
from MLP import MLP
from datasetmanager import download_extract_randomize_save, get_and_reformat_all_datasets
# ===== MAIN =====
# DOWNLOAD DATASETS
# download_extract_randomize_save()
# DEFINE MLP
""" With those parameters, I get 91.7% of accuracy """
image_size = 28
num_labels = 10
network_shape = [image_size * image_size,600,300,150,num_labels]
initial_learning_rate = 0.0001
decay_steps = 0
decay_rate = 0
regularization_parameter = 0.0
dropout_keep_prob = 0.5
mlp = MLP(network_shape, initial_learning_rate, decay_steps, decay_rate, regularization_parameter, dropout_keep_prob)
# GET AND REFORMAT ALL DATASETS
train_dataset, train_labels, valid_dataset, valid_labels, test_dataset, test_labels = get_and_reformat_all_datasets()
print('Training set', train_dataset.shape, train_labels.shape)
print('Validation set', valid_dataset.shape, valid_labels.shape)
print('Test set', test_dataset.shape, test_labels.shape)
# RUN TRAINING
batch_size = 150
num_epochs = 3
num_steps = len(train_dataset)/batch_size * num_epochs
print('Steps : ', num_steps)
mlp.train(train_dataset, train_labels, valid_dataset, valid_labels, test_dataset, test_labels, batch_size, num_steps)
def flush(self):
self.log.flush()
self.terminal.flush()
pass
sys.stdout = Logger()
train_count = 60000
train_labels = MNIST.label(*range(train_count), setname='train', decode=True)
train_samples = MNIST.image(*range(train_count), setname='train', flat=True, normalize=True)
test_count = 9000
test_labels = MNIST.label(*range(test_count), setname='test', decode=True)
test_samples = MNIST.image(*range(test_count), setname='test', flat=True, normalize=True)
mlp = MLP([784, 30, 10])
epoch_count = 10
train_errors = np.zeros(epoch_count)
test_errors = np.zeros(epoch_count)
begin_time = time.time()
print(time.strftime("Began training at %Y%m%d-%H%M%S"))
print()
for epoch_idx in range(1, epoch_count+1):
mlp.train(train_samples, train_labels, epochs=1, block_size=1, learn_rate=1/epoch_idx)
train_errors[epoch_idx-1] = mlp.validate(train_samples, train_labels) #training error
test_errors[epoch_idx-1] = mlp.validate(test_samples, test_labels) #test error
print("Epoch %d done at " % (epoch_idx) + time.strftime("%Y%m%d-%H%M%S"))
print("Training Accuracy: %.2f" % (train_errors[epoch_idx-1]))