def train(sp=-1):
model = KerasModel()
train = DataGenerator('train')
test = DataGenerator('test')
if sp != -1:
chkp = '.' + os.sep + 'models' + os.sep + 'save_' + str(sp) + '.h5'
model.load_weights(chkp)
print('start point: %d'%sp)
print("开始训练：")
model.fit_generator(
generator=train,
#steps_per_epoch：在声明一个epoch完成并开始下一个epoch之前从generator产生的总步数，它通常应该等于你的数据集的样本数量除以批量大小，对于Sequence，它是可选的，如果未指定，将使用len（generator）作为步数
#按理说是按照上面的标准进行的，但是因为我们数据集较少，所以我们使用更多的数量跑
#在测试的时候我运用的是14是完全按照理论做的
#如果现在的次数不能够正常运行请重新改回14并增加时代
steps_per_epoch=3000，
#epochs 确定世代的次数
epochs=epochs_time,
verbose=1,
#validation_steps是步长
validation_steps=100
#想要了解更多fit_generator的参数：https://blog.csdn.net/qq_32951799/article/details/82918098
)
#每个世代的模型都做保存
model.save_weights('.' + os.sep + 'models' + os.sep + 'save_' +str(epochs_time)+ '.h5' )
def generate_dataset(dataset_path, cols_size):
for size in cols_size:
block_cnt = 20
block_len = 30
dgen = DataGenerator(block_cnt, block_len, 9999, 0)
random.seed(1)
f = open(dataset_path + "/" + str(size) + ".txt", 'w')
for i in range(1, block_cnt + 1):
block_to_use = i
for col in dgen.generate_columns(size // 20, block_to_use):
f.write(str(col) + '\n')
def run(self):
clean_data = CleanData()
clean_data.run()
calc_corr = CalcCorrMatrix()
calc_corr.run()
data_generator = DataGenerator()
train_feature, train_label, test_feature, test_label = data_generator.run()
return train_feature, train_label, test_feature, test_label
def __init__(self, env, cur_length, full_dataset_length, max_length,
num_samples, epsilon):
self.env = env
self.cur_length = cur_length
self.max_length = max_length
self.num_samples = num_samples
self.epsilon = epsilon
self.full_dataset_length = full_dataset_length
self.generator = DataGenerator(env, epsilon)
self.states_full, self.actions, _, _ = self.generator.calc_data_full(
self.full_dataset_length)
self.reinitialize()
def __init__(self, env, cur_length, full_dataset_length, max_length,
num_samples, epsilon):
self.env = env
self.cur_length = cur_length
self.max_length = max_length
self.num_samples = num_samples
self.epsilon = epsilon
self.full_dataset_length = full_dataset_length
self.generator = DataGenerator(env, epsilon)
self.states_full, self.actions, self.next_states_full,\
self.masks_full = self.generator.calc_data_full(self.full_dataset_length)
n = self.cur_length - self.full_dataset_length
if n > 0:
self.states_rand, self.next_states_rand, self.masks_rand\
= self.generator.calc_data_rand(self.states_full[-1].view(-1, 2, 2, 2), self.actions, n)
def main(modelName):
INPUT_SHAPE = (6000, 12)
NUM_CLASSES = 5
BATCH_SIZE = 16
EPOCHS = 80
labels = load_obj('sampleLabels')
params = {
'dim': INPUT_SHAPE[0],
'batchSize': BATCH_SIZE,
'numClasses': NUM_CLASSES,
'numChannels': INPUT_SHAPE[1],
'shuffle': True
}
callbacksList = [
keras.callbacks.EarlyStopping(monitor='val_loss',
mode='min',
verbose=1,
patience=25),
keras.callbacks.ModelCheckpoint('E1/' + modelName + '_wts_b.hdf5',
monitor='val_acc',
verbose=1,
save_best_only=True,
mode='max')
]
trainIDs = load_obj('trainIDs2')
valIDs = load_obj('valIDs2')
testIDs = load_obj('testIDs2')
trainingGen = DataGenerator(trainIDs, labels, **params)
validationGen = DataGenerator(valIDs, labels, **params)
testingGen = DataGenerator(testIDs, labels, **params)
keras.backend.clear_session()
model = Model.factory(modelName, INPUT_SHAPE, NUM_CLASSES)
model.fit(BATCH_SIZE, EPOCHS, trainingGen, validationGen, callbacksList)
model.load_weights('E1/' + modelName + '_wts_b.hdf5')
model.test(testingGen)
def train(sp=-1):
model = KerasModel()
train = DataGenerator('train')
test = DataGenerator('test')
if sp != -1:
chkp = '.' + os.sep + 'models' + os.sep + 'save_' + str(sp) + '.h5'
model.load_weights(chkp)
print('start point: %d' % sp)
for i in range(sp + 1, 100):
model.fit_generator(generator=train,
samples_per_epoch=3000,
nb_epoch=1,
validation_data=test,
nb_val_samples=100)
model.save_weights('.' + os.sep + 'models' + os.sep + 'save_' +
str(i) + '.h5')
class RandomStateDataset(data.Dataset):
def __init__(self, env, cur_length, full_dataset_length, max_length,
num_samples, epsilon):
self.env = env
self.cur_length = cur_length
self.max_length = max_length
self.num_samples = num_samples
self.epsilon = epsilon
self.full_dataset_length = full_dataset_length
self.generator = DataGenerator(env, epsilon)
self.states_full, self.actions, self.next_states_full,\
self.masks_full = self.generator.calc_data_full(self.full_dataset_length)
n = self.cur_length - self.full_dataset_length
if n > 0:
self.states_rand, self.next_states_rand, self.masks_rand\
= self.generator.calc_data_rand(self.states_full[-1].view(-1, 2, 2, 2), self.actions, n)
def reinitialize(self):
self.states_full, self.actions, self.next_states_full,\
self.masks_full = self.generator.calc_data_full(self.full_dataset_length)
n = self.cur_length - self.full_dataset_length
if n > 0:
self.states_rand, self.next_states_rand, self.masks_rand\
= self.generator.calc_data_rand(self.states_full[-1].view(-1, 2, 2, 2), self.actions, n)
def __len__(self):
return self.num_samples
def __getitem__(self, _):
length = torch.randint(0, self.cur_length, ())
if length < self.full_dataset_length:
idx = torch.randint(0, len(self.states_full[length]), ())
state = self.states_full[length][idx]
next_states = self.next_states_full[length][idx]
mask = self.masks_full[length][idx]
else:
length = length - self.full_dataset_length
idx = torch.randint(0, len(self.states_rand[length]), ())
state = self.states_rand[length][idx]
next_states = self.next_states_rand[length][idx]
mask = self.masks_rand[length][idx]
return {'state': state, 'next_states': next_states, 'mask': mask}
def run(self):
if self.preprocess == 1:
print "cleaning data...\n"
clean_data = CleanData(self.file_path)
clean_data.run()
print "cleaning data successfully\n"
print "calculating correlation matrix...\n"
calc_corr = CalcCorrMatrix()
corr_matrix = calc_corr.run()
print "calculatE successfully\n"
print "generating dataset...\n"
data_generator = DataGenerator('../data/', self.code, self.ratio,
corr_matrix, self.train_set_choice)
train_feature, train_label, test_feature, test_label = data_generator.run(
)
print "generate successfully\n"
return train_feature, train_label, test_feature, test_label, corr_matrix
def compute(self, text_tokenizer, inv_map, k_means, embed, num_words,
comb):
input_1 = Input(shape=(num_words, ))
#process_1 = Dense(1024)(input_1)
#process_2 = Activation('tanh')(process_1)
process_3 = Dense(1)(input_1)
output_1 = Activation('tanh')(process_3)
model_1 = Model(input_1, output_1)
print(model_1.summary())
input_2 = Input(shape=(4, num_words))
process_4 = TimeDistributed(model_1)(input_2)
process_5 = Flatten()(process_4)
process_6 = Dense(1)(process_5)
output_2 = Activation('sigmoid')(process_6)
model_2 = Model(input_2, output_2)
print(model_2.summary())
train_generator = DataGenerator(self.train, self.train_labels,
text_tokenizer, inv_map, k_means,
embed, num_words)
valid_generator = DataGenerator(self.valid, self.valid_labels,
text_tokenizer, inv_map, k_means,
embed, num_words)
model_2.compile(optimizer='Adam',
loss='binary_crossentropy',
metrics=['accuracy'])
model_2.fit_generator(generator=train_generator,
validation_data=valid_generator,
epochs=2,
verbose=1)
# a = model_2.predict_generator(data_yield_1(valid_sequences,valid_results,valid_clusters,992),steps=62)
#print(a[:10])
#a = np.argmax(a,axis=1)
#b = np.argmax(valid_results,axis=1)
#print(a[:100],valid_results[:10],b[:10])
#print(classification_report(b,a))
return model_2
class TestDataset(data.Dataset):
def __init__(self, env, cur_length, full_dataset_length, max_length,
num_samples, epsilon):
self.env = env
self.cur_length = cur_length
self.max_length = max_length
self.num_samples = num_samples
self.epsilon = epsilon
self.full_dataset_length = full_dataset_length
self.generator = DataGenerator(env, epsilon)
self.states_full, self.actions, _, _ = self.generator.calc_data_full(
self.full_dataset_length)
self.reinitialize()
def reinitialize(self):
n = self.cur_length - self.full_dataset_length
if n > 0:
self.states_rand, _, _ = self.generator.calc_data_rand(
self.states_full[-1].view(-1, 2, 2, 2), self.actions, n)
def __len__(self):
return self.num_samples
def __getitem__(self, _):
length = torch.randint(0, self.cur_length, ())
if length < self.full_dataset_length:
idx = torch.randint(0, len(self.states_full[length]), ())
state = self.states_full[length][idx]
return {'state': state, 'length': length + 1}
else:
length = length - self.full_dataset_length
idx = torch.randint(0, len(self.states_rand[length]), ())
state = self.states_rand[length][idx]
return {
'state': state,
'length': length + self.full_dataset_length + 1
}
def main(inputs_path='data/inputs/',
targets_path='data/5-shot-targets/',
save_confusion=True):
"""Convolutional Neural Network Creation and Evaluation. Uses data collected from the Physiobank data set and stored
in the specified input and target folders"""
print("Running...")
model = define_model()
train_files, test_files, validation_files = get_input_file_names(
inputs_path)
# generate training/test/validation data in batches
train_generator = DataGenerator(inputs_path, targets_path, train_files)
validation_generator = DataGenerator(inputs_path, targets_path,
validation_files)
test_generator = DataGenerator(inputs_path, targets_path, test_files)
# run model training and evaluation
es = EarlyStopping(monitor='val_acc',
mode='max',
patience=5,
verbose=1,
restore_best_weights=True)
history = model.fit_generator(train_generator,
validation_data=validation_generator,
epochs=100,
verbose=1,
callbacks=[es])
_, accuracy = model.evaluate_generator(test_generator, verbose=0)
# create test set and targets
x_test, y_test = [], []
test_generator.on_epoch_end()
for i in range(len(test_generator)):
x_batch, y_batch = test_generator[i]
x_test.extend(x_batch)
y_test.extend(y_batch)
x_test = np.array(x_test)
y_test = np.array(y_test)
y_prediction = model.predict(x_test)
save_model(model)
evaluate_model(model, history, accuracy, y_test, y_prediction,
save_confusion)
return None
tf.app.flags.DEFINE_string('rnn_type', default_value='uni-dir',
docstring='''uni-dir or bi-dir.''')
tf.app.flags.DEFINE_float('initial_lr', default_value=0.00001,
docstring='''Initial learning rate for training.''')
tf.app.flags.DEFINE_integer('num_filters', default_value=64,
docstring='''Number of convolutional filters.''')
tf.app.flags.DEFINE_float('moving_avg_decay', default_value=0.9999,
docstring='''Decay to use for the moving average of weights.''')
tf.app.flags.DEFINE_integer('num_epochs_per_decay', default_value=5,
docstring='''Epochs after which learning rate decays.''')
tf.app.flags.DEFINE_float('lr_decay_factor', default_value=0.9,
docstring='''Learning rate decay factor.''')
logger.info('Running model for the following paramethers:')
train_data = DataGenerator(FLAGS.data_dir, FLAGS.batch_size)
valid_data = DataGenerator('C:\\Users\\user\\Desktop\\tfVirtEnv\\deepBrain\\data\\json\\Patient1\\valid.json', FLAGS.batch_size)
num_channels = 32 # compute from data
def inference_graph(inputs, seq_len, batch_size, prob, graph):
'''
Function to create the model graph for training and evaluation.
:inputs: ECoG ndarray of shape (batch_size, T, CH)
:seq_len: list (shape = (batch_size)) of ints which hold the lengh of each sequence
:train: bool that indicates if the graph is used for train or evaluation
If false deactivate the dropout layer.
:returns: tf.graph, logits
'''
num_channels = 32 # compute from data
self.pool_of_ack.remove(ack)
def get_all_peers(self):
peers = []
with open('peers_info.json') as f:
peers_json = json.load(f)
for port, name in peers_json.items():
peers.append([name, port])
return peers
def broadcast_data(self, data, type):
all_nodes = self.get_all_peers()
for node in all_nodes:
self.send_data(node, data, type)
def run(self):
receiving_thread = threading.Thread(target=self.receive_data)
receiving_thread.start()
ack_thread = threading.Thread(target=self.send_ack_to_peers)
ack_thread.start()
sync_thread = threading.Thread(target=self.sync_with_peers())
sync_thread.start()
p1 = Node("peer1")
p2 = Node("peer2")
data_generator = DataGenerator()
# node1.send_data(['127.0.0.1', node2.port], "data")
# n2.broadcast_data("Im node 2 and msg is broadcasted")
def main(trainF, testF, validF, modelF):
batch_size = 32
num_classes = 2
epochs = 100
patience = 5#int(epochs/20)
targetIdx = range(3, 5)
currentStateIdx = range(1, 3) # Includes number of time steps
# input image dimensions
img_rows, img_cols = 128, 96
# Parameters
params = {'dim': (img_rows, img_cols, 3),
'batch_size': batch_size,
'n_channels': 3,
'preTrain': preTrain,
'shuffle': True}
# File name
partition = readInp(trainF, testF, validF)
trainF = partition['train'][:, 0]
testF = partition['test'][:, 0]
validF = partition['valid'][:, 0]
trainTarget = partition['train'][:, targetIdx].astype(float)
testTraget = partition['test'][:, targetIdx].astype(float)
validTarget = partition['valid'][:, targetIdx].astype(float)
# Current state position
trainCurrState = partition['train'][:, currentStateIdx].astype(float)
testCurrState = partition['test'][:, currentStateIdx].astype(float)
validCurrState = partition['valid'][:, currentStateIdx].astype(float)
training_generator = DataGenerator(trainF, trainCurrState, trainTarget, **params)
valid_generator = DataGenerator(validF, validCurrState, validTarget, **params)
#
#testparams = {'dim': (img_rows, img_cols, 3),
# 'batch_size': 64,
# 'n_channels': 3,
# 'shuffle': False}
#test_generator = DataGenerator(testF, testTraget, **testparams)
# Input that stores the previous step
model = spatialModel(img_rows, img_cols)
model.compile(loss=keras.losses.mse,
optimizer=keras.optimizers.Adam(lr=0.001),
metrics=['mse', 'mae'])
#pre_dict = evaluate(model, modelF + '.h5', valid_generator)
#truedict = dict(zip(validF, validTarget))
#for i in range(36):
# true = np.zeros((19, 2), dtype=float)
# pred = np.zeros((19, 2), dtype=float)
# for j in range(1, 19):
# key = './output//valid//input%d.lp_%d.png'%(i+80, j-1)
# true[j, :] = truedict[key]
# pred[j, :] = pre_dict[key]
#
# pred = np.cumsum(pred, axis=0)
# true = np.cumsum(true, axis=0)
# pred = np.array(pred)
# true = np.array(true)
#
# import matplotlib.pyplot as plt
# plt.scatter(pred[:, 0], pred[:, 1], c='r')
# plt.scatter(true[:, 0], true[:, 1], c='g')
# plt.show()
early_stopping = EarlyStopping(monitor='mean_squared_error', patience=patience)
#validation_data=valid_generator,
#callbacks=[early_stopping,],
history = model.fit_generator(generator=training_generator,
use_multiprocessing=False,
epochs=epochs, verbose=1, workers=1)
model.save(modelF)
del model
print('Done saving model: %s'%modelF)
return
mse_x = ((p3d_out_17x3[:, :, 0] - p3d_gt_17x3[:, :, 0]) * p3d_std)**2
mse_y = ((p3d_out_17x3[:, :, 1] - p3d_gt_17x3[:, :, 1]) * p3d_std)**2
mse_z = ((p3d_out_17x3[:, :, 2] - p3d_gt_17x3[:, :, 2]) * p3d_std)**2
mean_mse_xy = tf.reduce_mean(tf.sqrt(mse_x + mse_y))
mean_mse_xz = tf.reduce_mean(tf.sqrt(mse_x + mse_z))
mean_mse_yz = tf.reduce_mean(tf.sqrt(mse_y + mse_z))
return p3d_out_17x3[0, 2, 0]
tf.enable_eager_execution()
with tf.Session(config=config) as sess:
trainGenerator = DataGenerator(DATA_PATH,
batch_size=BATCH_SIZE,
val_fraction=VAL_FRACTION,
name="train")
valGenerator = DataGenerator(DATA_PATH,
batch_size=BATCH_SIZE,
val_fraction=VAL_FRACTION,
name="val")
# define model
model = get_model(batch_size=BATCH_SIZE)
if LOAD:
latest = tf.train.latest_checkpoint("./log/test8/saved_models/retrain")
print("loading weight from", latest)
model.load_weights(latest)
# optimizer
optimizer = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE)
# saver
pass
print("Found " + str(len(labels)) + " labels.")
outputAges = []
for label in list(labels.values()):
for i in range(len(ageClassLabels)):
if label[0] >= ageClassLabels[i][0] and label[0] <= ageClassLabels[
i][1]:
outputAges.append(oneHotEncode(i, len(ageClassLabels)))
break
outputAges = np.array(outputAges, dtype=float)
print("Binned Ages into classes. Shape: " + str(outputAges.shape))
imageIDs = np.array(list(labels.keys()))
trainData, testData = DataGenerator(np.array(list(labels.keys())),
outputAges, 100).split(validationSplit)
ConvTemplate = functools.partial(
tf.keras.layers.Conv2D,
bias_initializer=tf.keras.initializers.lecun_normal(),
kernel_initializer=tf.keras.initializers.lecun_normal(),
activation=tf.keras.activations.relu)
ConvTemplate.__name__ = 'ConvTemplate'
def LayerTemplate(filters, kernel_size):
def Layer(x):
return tf.keras.layers.BatchNormalization()(ConvTemplate(
filters=filters, kernel_size=kernel_size))
return Layer
def main():
########################################################
# Initialize
########################################################
global keysPressed
dataIndex = 0
EpisodicMemCap = 100
dataIndex_episode = pickle.load(
open("trainingData/episodeData/dataIndex_episode.p", "rb"))
# Load model
WIDTH = HEIGHT = 224
FILENUM = pickle.load(
open("trainingData/Unbalanced_rgb_299/dataIndex_{}.p".format('both'),
"rb"))
#FILENUM_both = 458
model = get_Keras_model('both', WIDTH, HEIGHT, 8, "AdamReg", FILENUM)
# Setup for screenGrab
last_time = time.time()
W, H = 575, 525
dsp = display.Display()
screen = dsp.screen()
root = dsp.screen().root
countDown(5)
while True:
########################################################
#Reinforcement Loop
########################################################
isSolved = False
trainingData = []
while not isSolved:
"""
During the puzzle, this loop will play till it gets determined that
the puzzle has been solved.
"""
############################
# Get Image
############################
image = grabscreen(root, W, H)
image_rs = cv2.resize(image, (WIDTH, HEIGHT))
############################
# Make Predictions
############################
body_moves = [0, 0, 0, 0, 0]
head_moves = [0, 0, 0, 0, 0]
p_body, p_head = model.predict(
[image_rs.reshape(1, WIDTH, HEIGHT, 3)])
# This method ensures that each frame, the agent will pick the most likely option.
body_m_index = np.argmax(p_body[0])
head_m_index = np.argmax(p_head[0])
body_moves[body_m_index] = 1
head_moves[head_m_index] = 1
print(body_moves)
print(head_moves)
############################
# See if Supervisor overrides
############################
trainingData = SupervisorCheck(cv2.resize(image,
(299, 299)), body_moves,
head_moves, trainingData)
print('loop took {:0.3f} seconds'.format(time.time() - last_time))
last_time = time.time()
############################
# Store episodic memory to file.
############################
if len(trainingData) == 100:
# Sets of 100 so I don't need to preprocess.
print(dataIndex)
shuffle(trainingData) # Just randomize it now.
dataIndex += 1
#Write to new datafile. Formatted to fit with current dataGenerator code.
np.save(
"trainingData/episodeData/episodeData/data_{}".format(
dataIndex), trainingData)
# Restarts array for next batch
trainingData = []
if 'm' in keysPressed or 'M' in keysPressed:
# My manual way to break the loop
isSolved = True
"""
Needs to be there for the async hook and this synced loop
to keep up with each other
"""
time.sleep(0.01)
if cv2.waitKey(25) & 0xFF == ord('q'):
cv2.destroyAllWindows()
break
print("Number of frames in Episodic Memory: {}".format(dataIndex))
#Stop the agent
move_body([0, 0, 0, 0, 1])
move_head([0, 0, 0, 0, 1])
if dataIndex > EpisodicMemCap:
########################################################
# Retrain the Agent's Policy
########################################################
for i, layer in enumerate(model.layers):
print(i, layer.name)
# Only train the last dense layers that I personally added to VGG16.
Layer_End = 19
for layer in model.layers[:Layer_End]:
layer.trainable = False
for layer in model.layers[Layer_End:]:
layer.trainable = True
# Training Setup
EPOCHS = 4
DTYPE = DATA_TYPE = 'episodeData'
TrainLen = dataIndex - 10
Indicies = np.array(range(1, dataIndex + 1))
shuffle(Indicies)
TrainIndex = Indicies[:TrainLen]
ValidIndex = Indicies[TrainLen:]
params = {
'WIDTH': WIDTH,
'HEIGHT': HEIGHT,
'DTYPE': DTYPE,
'DATA_TYPE': DATA_TYPE,
'isConcat': True,
'batch_size': 1,
'shuffle': True
}
training_generator = DataGenerator(Indicies, **params)
validation_generator = DataGenerator(ValidIndex, **params)
# Defining my callbacks:
filepath = "models/SupervisedReinforcement/best_weights.h5"
checkpoint = ModelCheckpoint(filepath,
monitor='val_dense_5_acc',
verbose=1,
save_best_only=True,
mode='max')
callbacks_list = [checkpoint]
# compile the model (should be done *after* setting layers to non-trainable)
model.compile(optimizer=SGD(lr=1e-4, momentum=0.9),
loss='categorical_crossentropy',
metrics=['acc'])
#model.compile(optimizer=Adam(lr=0.0001), loss='categorical_crossentropy', metrics=['acc'])
# Train the model on the new data for a few epochs
model.fit_generator(generator=training_generator,
validation_data=validation_generator,
callbacks=callbacks_list,
use_multiprocessing=True,
workers=6,
epochs=EPOCHS,
steps_per_epoch=dataIndex)
#Reload the best weights from that update session
model.load_weights(
"models/SupervisedReinforcement/best_weights.h5")
# Save data rather than delete because it is like 10k data samples
dataIndex_episode += 1
shutil.move(
"trainingData/episodeData/episodeData",
"trainingData/episodeData/episodeDataArchived/E{}".format(
dataIndex_episode))
subprocess.call(['mkdir', 'trainingData/episodeData/episodeData'])
pickle.dump(
dataIndex_episode,
open("trainingData/episodeData/dataIndex_episode.p", "wb"))
dataIndex = 0
else:
# wait till I tell it to start testing again.
testStart = False
while not testStart:
if 'n' in keysPressed or 'N' in keysPressed:
testStart = True
import h5py
#%% train generator
paramsTrain = {
'trackName': 'skank',
'encoding': 'one-hot',
'batch_size': 32,
'maxLength': 16,
'timesteps': 8,
'seqOverlap': 'max',
'shuffle': True,
'dataAugmentation': True
}
paramsTrain['datasetPath'] = './dataset/'
trainGenerator = DataGenerator(**paramsTrain)
#%% model
# hyperparameters
timesteps = trainGenerator.timesteps
nFeat = trainGenerator.nFeatures
lstmUnits = 64
lstmAct = 'tanh'
outputUnits = nFeat
dropout = 0.2
nEpochs = 1
lr = 10e-3
# optimizer
Adam = tf.keras.optimizers.Adam(lr)
image_type = gen_param['imgType'],
label_type = gen_param["labelType"])
train_ids["train"] = train_ids["train"][:args.n_samples]
# Test set remains the same
test_ids = generatePartitionTrainAndValidFromFolderRandomly(imagePathFolder = val_param['imgPath'],
_validProportion = val_param['validprop'],
shuffle_train_val = val_param['shuffletrain'],
threshold=20,
labelPathFolder = val_param['labelPath'],
image_type = val_param['imgType'],
label_type = val_param["labelType"])
# Generate data generator objects.
training_generator = DataGenerator(train_ids['train'], gen_param['imgPath'], gen_param['labelPath'], norm_param, gen_param["augment"], aug_param, **misc_param_gen)
validation_generator = DataGenerator(test_ids['validation'], val_param['imgPath'], val_param['labelPath'], norm_param, val_param["augment"], aug_param, **misc_param_val)
# Generate the model.
model = load_model(args.path + "model.h5", compile=False)
model.summary()
# Set up metrics.
metrics = [dice_coef_background, dice_coef_first_label]
# Set up loss function
lossfunc = create_loss_func(mp)
opt = Adam(lr=mp["learningrate"])
model.compile(optimizer=opt, loss=lossfunc, metrics=metrics)
# Print report prior to fit of the model.
print_report(p)
preProc.saveArticles()
# Initialising the RNN
regressor = Sequential()
# Adding the input layerand the LSTM layer
regressor.add(
LSTM(units=32,
activation='relu',
input_shape=(preProc.seq_length, preProc.data_dim)))
# Adding the output layer
regressor.add(Dense(units=1))
# Compiling the RNN
regressor.compile(optimizer='adam', loss='mean_squared_error')
list_IDs = range(preProc.num_ts * preProc.num_articles)
batch_size = 32
# Initializing the generator
training_generator = DataGenerator(batch_size=batch_size).generate(list_IDs)
# Fitting the RNN to the Training set
regressor.fit_generator(generator=training_generator,
steps_per_epoch=len(list_IDs) // batch_size,
epochs=10)
regressor.save('model_full_data_7ts_32u_lstm_adam_32b_10e_sh')
import pickle
from dataGenerator import DataGenerator
from utils import visualize_2D_trip
last = 19000
i = 19100
seed_ = 19053
with open("supervised_data/batch_seed_" + str(last) + "_" + str(i) + ".pkl",
'rb') as handle:
b = pickle.load(handle)
dataset = DataGenerator()
input_batch, dist_batch = dataset.create_batch(256, 50, 2, seed=seed_)
visualize_2D_trip(input_batch[100], b[seed_][100], 'niente', 1)
def train(in_dir,
hdf5_dir,
hdf5_name,
dim_patch=224,
pre_trained='',
rs_rate=4,
balancing=False):
"""
Training pipeline
Performing training and evaluation of the tiramisu model on the given dataset
The dataset should be loaded into a hdf5 file
"""
# ------------------------------------------------------------------------------------------------------------------- #
# Read the loaded HDF5 file
hdf5_file = h5py.File(hdf5_dir + hdf5_name + '.hdf5', 'r')
# ------------------------------------------------------------------------------------------------------------------- #
# Generate batch sample
# settings
batch_size = 10
num_fusion = 0 # the number of previous/following patch to be fused in the input
params_train = {
'hdf5_file': hdf5_file,
'dim_x': dim_patch,
'dim_y': dim_patch,
'dim_z': 3,
'batch_size': batch_size,
'num_fusion': num_fusion,
'tag': 'Train',
'aug': rs_rate,
'balancing': balancing
}
params_val = {
'hdf5_file': hdf5_file,
'dim_x': dim_patch,
'dim_y': dim_patch,
'dim_z': 3,
'batch_size': batch_size,
'num_fusion': num_fusion,
'tag': 'Val'
}
train_setting = DataGenerator(**params_train)
val_setting = DataGenerator(**params_val)
# calculate class weighting
class_weighting = train_setting.class_weight
num_train = len(train_setting.crop_list)
num_val = len(val_setting.crop_list)
# Generators
training_generator = train_setting.generate()
validation_generator = val_setting.generate()
# ------------------------------------------------------------------------------------------------------------------- #
# record the settings,training process,results in a file
import time
curr_date = time.strftime("%d/%m/%Y")
curr_time = time.strftime("%H:%M:%S")
file_id = curr_date[0:2] + curr_date[3:5] + '_' + curr_time[0:2]
# file_id = 'test'
# create a new folder
newPath = '../Result/' + file_id
try:
os.makedirs(newPath)
except OSError:
if not os.path.isdir(newPath):
raise
np.save(newPath + 'mean_train', train_setting.mean)
text_file = open(newPath + 'Log.txt', "w")
text_file.write("Date: %s\n" % curr_date)
text_file.write("Start time: %s\n\n" % curr_time)
# ------------------------------------------------------------------------------------------------------------------- #
# load the model (and weights):
model_id = hdf5_name
# Tiramisu_3D().create([dim_patch, dim_patch], [4, 6, 8], 10, [8, 6, 4], 12, 0.0001, 0.5, 5, model_id)
Tiramisu().create([dim_patch, dim_patch], [4, 6, 8], 10, [8, 6, 4], 12,
0.0001, 0.5, 5, model_id)
load_model_name = '../Model/tiramisu_fc_dense_' + model_id + '.json'
with open(load_model_name) as model_file:
tiramisu = models.model_from_json(model_file.read())
if len(pre_trained) > 0:
tiramisu.load_weights(pre_trained, by_name=True)
# specify optimizer
optimizer = Nadam(lr=0.0005)
# number of epochs
nb_epoch = 20
# metrics using accuracy or IoU
tiramisu.compile(loss="categorical_crossentropy",
optimizer=optimizer,
metrics=["accuracy"])
# checkpoint 278
TensorBoard = callbacks.TensorBoard(log_dir='./logs',
histogram_freq=5,
write_graph=True,
write_images=True)
filePath = newPath + 'prop_tiramisu_weights_' + model_id + '.best.hdf5'
checkpoint = ModelCheckpoint(filePath,
monitor='val_acc',
verbose=2,
save_best_only=True,
save_weights_only=False,
mode='max')
# the training stops when the validation accuracy stops improving for 5 epochs
earlyStopping = EarlyStopping(monitor='val_acc',
patience=4,
verbose=1,
mode='max')
callbacks_list = [checkpoint, earlyStopping]
# ------------------------------------------------------------------------------------------------------------------- #
# record the settings in a text file
with open('../Model/Setting.txt', 'r') as model_setting_log:
model_setting = model_setting_log.read()
text_file.write(model_setting)
text_file.write("\nData set: C1_17\n")
text_file.write('\nBalancing: %s' % str(train_setting.balancing))
text_file.write('\nSampling rate: %d' % train_setting.aug)
text_file.write("\nclass weights:\n")
for w_item in class_weighting:
text_file.write("%f\n" % w_item)
text_file.write("\n# training data: %d\n" % num_train)
text_file.write("# validation data: %d\n" % num_val)
text_file.write("model: %s\n" % load_model_name)
# text_file.write("loaded the weights: %s\n\n"%"1303_19/prop_tiramisu_weights_69.best.hdf5")
text_file.write("optimizer = Nadam(lr=0.0005)\n")
text_file.write("loss function: categorical_crossentropy\n\n")
text_file.write("# epoch: %d\n batch size: %d\n" % (nb_epoch, batch_size))
text_file.write("weights stored in: %s\n" % filePath)
text_file.write("number of stacks: %d\n" % (num_fusion * 2 + 1))
text_file.write("no floating ice in training set")
text_file.close()
# ------------------------------------------------------------------------------------------------------------------- #
# Fit the model
history = tiramisu.fit_generator(
generator=training_generator, # gen(train_setting),
steps_per_epoch=math.ceil(num_train / float(batch_size)),
validation_data=validation_generator, # gen(val_setting),
validation_steps=math.ceil(num_val / float(batch_size)),
epochs=nb_epoch,
verbose=1,
callbacks=callbacks_list)
# This save the trained model weights to this file with number of epochs
tiramisu.save_weights(newPath + 'prop_tiramisu_weights_' + model_id +
'_{}.hdf5'.format(nb_epoch))
# ------------------------------------------------------------------------------------------------------------------- #
# plot and save training history
import matplotlib.pyplot as plt
# list all data in history
print(history.history.keys())
# summarize history for accuracy
plt.figure(1)
plt.plot(history.history['acc'])
plt.plot(history.history['val_acc'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper left')
plt.savefig(newPath + 'acc.png')
# summarize history for loss
plt.figure(2)
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper left')
plt.savefig(newPath + 'loss.png')
# save history to numpy array
log_history = [[]] * 4
log_history[0] = np.array(history.history['loss'])
log_history[1] = np.array(history.history['val_loss'])
log_history[2] = np.array(history.history['acc'])
log_history[3] = np.array(history.history['val_acc'])
np.save(newPath + 'Train_history', log_history)
hdf5_file.close()
RADIUS = 750 # space around the subject
xroot, yroot, zroot = vals[0, 0], vals[0, 1], vals[0, 2]
ax.set_xlim3d([-RADIUS + xroot, RADIUS + xroot])
ax.set_zlim3d([-RADIUS + zroot, RADIUS + zroot])
ax.set_ylim3d([-RADIUS + yroot, RADIUS + yroot])
ax.view_init(elev=-90., azim=-90)
ax.get_xaxis().set_ticklabels([])
ax.get_yaxis().set_ticklabels([])
ax.set_zticklabels([])
ax.set_aspect('equal')
Generator = DataGenerator(DATA_PATH, batch_size=BATCH_SIZE, name="train")
images = Generator.__getitem__(0)[0]
#annotations = unnormalize_pose_numpy(p3d,0,2*1100)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = "0"
with tf.Session(config=config) as sess:
if PREDICT:
model = get_model(batch_size=BATCH_SIZE)
# Load weights into the new model
latest = tf.train.latest_checkpoint(SAVE_PATH)
print("weights: ", latest)
model.load_weights(latest)
p3d_out, _ = model.predict(x=images, verbose=1)
p3d_out = unnormalize_pose_numpy(p3d_out, 0, 1100 * 2)
def train_network(name, the_model, config, train, test, LABELS, num_cores):
PREDICTION_FOLDER = 'models_for_top_five_cat' #"predictions_1d_conv_output_2"
MODEL_PATH = PREDICTION_FOLDER + '/' + name + '/'
if not os.path.exists(PREDICTION_FOLDER):
os.mkdir(PREDICTION_FOLDER)
if not os.path.exists(PREDICTION_FOLDER + '/' + name):
os.mkdir(PREDICTION_FOLDER + '/' + name)
if os.path.exists('../logs/' + PREDICTION_FOLDER + name + '/'):
shutil.rmtree('../logs/' + PREDICTION_FOLDER + name + '/')
skf = StratifiedKFold(n_splits=config.n_folds)
splits = skf.split(train.label, train.actual_label_idx)
fold_acc = []
for i, (train_split, val_split) in enumerate(splits):
train_set = train.iloc[train_split]
val_set = train.iloc[val_split]
checkpoint = ModelCheckpoint(MODEL_PATH + 'best_%d.h5' % i,
monitor='val_loss',
verbose=1,
save_best_only=True)
early = EarlyStopping(monitor="val_loss", mode="min", patience=3)
tb = TensorBoard(log_dir='./logs/' + name + '/' + '/fold_%d' % i,
write_graph=True)
callbacks_list = [checkpoint, early, tb]
print("Fold: ", i)
print("#" * 50)
model = the_model(config)
train_generator = DataGenerator(config,
'audio_train/',
train_set.index,
train_set.actual_label_idx,
batch_size=64,
preprocessing_fn=audio_norm)
val_generator = DataGenerator(config,
'audio_train/',
val_set.index,
val_set.actual_label_idx,
batch_size=64,
preprocessing_fn=audio_norm)
history = model.fit_generator(train_generator,
callbacks=callbacks_list,
validation_data=val_generator,
epochs=config.max_epochs,
use_multiprocessing=True,
workers=num_cores,
max_queue_size=20)
model.load_weights(MODEL_PATH + 'best_%d.h5' % i)
# Save train predictions
train_generator = DataGenerator(config,
'audio_train/',
train.index,
batch_size=128,
preprocessing_fn=audio_norm)
predictions = model.predict_generator(train_generator,
use_multiprocessing=True,
workers=num_cores,
max_queue_size=20,
verbose=1)
np.save(MODEL_PATH + "train_predictions_%d.npy" % i, predictions)
# Save test predictions
test_generator = DataGenerator(config,
'audio_train/',
test.index,
batch_size=128,
preprocessing_fn=audio_norm)
predictions = model.predict_generator(test_generator,
use_multiprocessing=True,
workers=num_cores,
max_queue_size=20,
verbose=1)
np.save(MODEL_PATH + "test_predictions_%d.npy" % i, predictions)
# Make a submission file
top_3 = np.array(LABELS)[np.argsort(-predictions, axis=1)[:, :1]]
predicted_labels = [' '.join(list(x)) for x in top_3]
test['predicted_label'] = predicted_labels
num_wrong = 0
for entry in test.values:
if (entry[2] != entry[0]):
num_wrong += 1
fold_acc.append((1 - (num_wrong / test.shape[0])) * 100)
# test[['pred_label']].to_csv(PREDICTION_FOLDER + "/predictions_%d.csv"%i)
f = open(MODEL_PATH + 'fold_accuracies.csv', 'a+')
for i in range(len(fold_acc)):
val = "Fold_" + str(i) + ", " + str(fold_acc[i]) + '\n'
f.write(val)
val = "Avg_accuracy, " + str(sum(fold_acc) / len(fold_acc))
f.write(val)
f.close()
f = open(MODEL_PATH + 'fold_accuracies.csv', 'a+')
for i in range(len(fold_acc)):
val = "Fold_" + str(i) + ", " + str(fold_acc[i]) + '\n'
f.write(val)
val = "Avg_accuracy, " + str(sum(fold_acc) / len(fold_acc))
f.write(val)
f.close()
for cv_index in range(len(chunks)):
trainset = []
for i in range(len(chunks)):
if not i == cv_index:
trainset = trainset + chunks[i]
valset = chunks[cv_index]
partition = dict()
partition['train'] = trainset
partition['validation'] = valset
# Generate data generator objects.
training_generator = DataGenerator(partition['train'],
gen_param['imgPath'],
gen_param['labelPath'], norm_param,
gen_param['augment'], aug_param,
**misc_param)
validation_generator = DataGenerator(partition['validation'],
gen_param['imgPath'],
gen_param['labelPath'],
norm_param, gen_param['augment'],
aug_param, **misc_param)
# Generate the model.
n_rows, n_cols, n_slices = mp['x_end'] - mp['x_start'], mp[
'y_end'] - mp['y_start'], mp['z_end'] - mp['z_start']
model = UNet_3D(input_shape=(n_rows, n_cols, n_slices, mp['channels']),
nb_labels=mp['labels'],
filters=mp['features'],
depth=mp['depth'],
SPACE = " "
SPECIAL_CHARS = "?!,."
ALPHANUMERIC = string.printable[:62]
CHARS = ALPHANUMERIC + SPECIAL_CHARS + SPACE
INPUT_SOURCE_NAME = "iam_word"
BATCH_SIZE = 16
MAX_TEXT_LENGTH = 128
CHARSET_BASE = CHARS
dtgen = DataGenerator(
source=f"/home/kuadmin01/terng/Dataset/dataset_filter.hdf5",
batch_size=BATCH_SIZE,
charset=CHARSET_BASE,
max_text_length=MAX_TEXT_LENGTH,
predict=False
)
INPUT_SHAPE = (1024, 128, 1)
OUTPUT_SHAPE = dtgen.tokenizer.vocab_size + 1
inputs, outputs = ResFlor(input_shape=INPUT_SHAPE, output_shape=OUTPUT_SHAPE)
model = keras.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer=keras.optimizers.RMSprop(
learning_rate=5e-4), loss=ctc_loss)
model.summary()
callbacks = Callback(INPUT_SOURCE_NAME, 'Flor')
def pool_operation(input_b, it):
from concorde.tsp import TSPSolver
import os
# Concorde
solver = TSPSolver.from_data(input_b[it, :, 0] * 1000,
input_b[it, :, 1] * 1000,
norm="EUC_2D")
# Find tour
solution = solver.solve()
return solution.tour
config = get_config()
dataset = DataGenerator()
real_tour_concorde = {}
pool = Pool(processes=8)
last = 10000
if not os.path.exists("supervised_data"):
os.mkdir("supervised_data")
for i in tqdm(range(10000, 30000)): # test instance
seed_ = 1 + i
reward, tour = [], []
input_batch, dist_batch = dataset.create_batch(config.batch_size,
config.graph_dimension,
config.dimension,
seed=seed_)
sys.stdout = open(os.devnull, "w")
params = {
'WIDTH': WIDTH,
'HEIGHT': HEIGHT,
'DTYPE': DTYPE,
'DATA_TYPE': DATA_TYPE,
'isConcat': True,
'batch_size': 1,
'shuffle': True
}
MODEL_NAME = 'pytalos_{}_{}_{}_{}_files_{}_epocs_{}_{}.h5'.format(
DTYPE, ARCH, OPTIMIZER, FILENUM, EPOCHS_1, DATA_TYPE, LR)
model_path = "models/{}/{}".format(DTYPE, MODEL_NAME)
training_generator = DataGenerator(TrainIndex, **params)
validation_generator = DataGenerator(ValidIndex, **params)
# Define the model w/ Keras from their documentation on applications
base_model = VGG16(weights='imagenet',
include_top=False,
input_shape=(WIDTH, HEIGHT, 3))
# Add a global spatial average pooling layer
x = base_model.output
x = GlobalAveragePooling2D()(x)
# Add two fully connected layers
x = Dense(1024, activation='relu', kernel_regularizer=regularizers.l2(0.01))(x)
# Branch out to both modalities, body and head movement.
print(dir_actor)
print(config.version)
tf.reset_default_graph()
actor = Actor(config) # Build graph
# Save & restore all the variables.
variables_to_save_all = [
v for v in tf.global_variables() if 'Adam' not in v.name
]
saver_all = tf.train.Saver(var_list=variables_to_save_all,
keep_checkpoint_every_n_hours=1.0)
########################################## TRAIN #################################
dataset = DataGenerator()
tf.set_random_seed(123)
tf.random.set_random_seed(123)
temperature = 1.0
last = 10000
save_path = "saveSL/" + dir_actor
if not os.path.exists(save_path):
os.makedirs(save_path)
restore_path = "saveSL/" + dir_actor + "/actor.ckpt"
with tf.Session() as sess: # start session+ "next"
count_no_weights()
sess.run(tf.global_variables_initializer()) # run initialize op