def main():
backend.set_floatx("float16")
backend.set_epsilon(1e-4)
data = get_dataset(
block_interval=8,
block_size=INPUT_COUNT,
file_count=40,
output_size=OUTPUT_COUNT,
)
train_data = data.train_data.reshape(len(data.train_data), INPUT_COUNT, 1)
test_data = data.test_data.reshape(len(data.test_data), INPUT_COUNT, 1)
model = ExperimentalModel()
model.load()
# model.train(train_data, data.train_out, test_data, data.test_out)
for i in range(min(len(data.files), min(len(data.files), 10))):
result = data.files[i][40]
inp = result.reshape(INPUT_COUNT, 1)
result = list(result)
generate_steps = 4000
iterations = int(generate_steps / OUTPUT_COUNT)
for j in range(iterations):
if j % 1000 == 0:
print(f"Progress: {j} / {iterations}")
out = model.predict_output(inp.reshape(1, *inp.shape))
result.extend(out[0])
# inp = out.reshape(32, 1)
inp = np.concatenate([inp, out.reshape(OUTPUT_COUNT,
1)])[-INPUT_COUNT:]
data.write_wav(f"output-lstm-{MODEL_ID}-{i}.wav", np.array(result))
def _enable_float32(self):
dtype = 'float32'
K.set_floatx(dtype)
K.set_epsilon(1e-7)
return dtype
def main():
backend.set_floatx("float16")
backend.set_epsilon(1e-4)
data = get_dataset(block_interval=int(INPUT_COUNT / 4), block_size=INPUT_COUNT, file_count=10)
# train_data = data.train_data.reshape(len(data.train_data), INPUT_COUNT, 1)
# test_data = data.test_data.reshape(len(data.test_data), INPUT_COUNT, 1)
train_data = data.train_data
test_data = data.test_data
# plt.plot(train_data.flatten())
# plt.show()
# plot_data = np.bincount((train_data.flatten() * 255).astype(int))
# plt.plot(plot_data)
# plt.show()
# print(plot_data.shape)
model = AutoEncoder()
model.load()
model.train(train_data, test_data)
for i in range(min(len(data.files), 10)):
output = model.predict_output(data.files[i])
data.write_wav(f"output-conv-{CONV_ID}-{i}.wav", output)
def _enable_float16(self):
dtype = 'float16'
K.set_floatx(dtype)
K.set_epsilon(1e-4)
return dtype
def F1(y_true, y_pred):
K.set_epsilon(1e-12)
predicted_positives = pre_t(y_true, y_pred) # 预测为正的个数
possible_positives = real_t(y_true, y_pred) # 实际为正的个数
intersect = inter(y_true, y_pred) # 为1的交集个数
P = intersect / (predicted_positives + K.epsilon())
R = intersect / (possible_positives + K.epsilon())
return 2*P*R/(P+R+K.epsilon())
def main():
backend.set_floatx("float16")
backend.set_epsilon(1e-4)
data = get_dataset(block_interval=1000, block_size=2000, file_count=10)
# train_data = data.train_data.reshape(len(data.train_data), 2000, 1)
# test_data = data.test_data.reshape(len(data.test_data), 2000, 1)
model = AutoEncoder()
model.load()
# model.train(train_data, test_data)
for i in range(10):
inp = data.files[i].reshape(len(data.files[i]), 2000, 1)
output = model.predict_output(inp).reshape(len(data.files[i]), 2000)
data.write_wav(f"output-conv{i}.wav", output)
def pre_encoder(self, X_train, param_reg):
"""
First part of the stacked AE.
Train the AE on the ROI input images.
:param X_train: ROI input image
:param get_history: boolean to return the loss history
:param loss: if "customized_loss" -> customized_loss
:return: encoded ROI image
"""
K.set_epsilon(0.1)
rm = run_model.run_model('pre')
#define callbacks
callbacks = rm.callbacks_define(self.monitor, self.weight_name)
autoencoder_0 = Sequential()
dim = self.pretrain_window * self.pretrain_window
#print(dim)
encoder_0 = Dense(input_dim=dim,
units=self.units_ed,
kernel_regularizer=regularizers.l2(param_reg))
decoder_0 = Dense(input_dim=self.units_ed,
units=dim,
kernel_regularizer=regularizers.l2(param_reg))
autoencoder_0.add(encoder_0)
autoencoder_0.add(decoder_0)
rm = run_model.run_model('roi')
loss = rm.load_loss(self.loss_pre)
autoencoder_0.compile(loss=loss,
optimizer=self.roi_optimizer,
metrics=['accuracy'])
h = autoencoder_0.fit(X_train,
X_train,
epochs=self.epoch_pre,
verbose=self.verbose,
callbacks=callbacks)
autoencoder_0.load_weights(self.store_model + "/weights_%s_%s.hdf5" %
('pre', self.weight_name))
weights = autoencoder_0.layers[0].get_weights()
temp_0 = Sequential()
temp_0.add(encoder_0)
temp_0.compile(loss=loss,
optimizer=self.roi_optimizer,
metrics=['accuracy'])
encoded_X = temp_0.predict(X_train, verbose=self.verbose)
if self.get_history:
return h.history['loss'], encoded_X, encoder_0, weights
else:
return encoded_X, encoder_0, weights
def main():
backend.set_floatx("float16")
backend.set_epsilon(1e-4)
data = get_dataset(block_interval=1, block_size=INPUT_COUNT, file_count=1)
train_data = data.train_data.reshape(len(data.train_data), INPUT_COUNT, 1)
test_data = data.test_data.reshape(len(data.test_data), INPUT_COUNT, 1)
model = ExperimentalModel()
model.load()
# model.train(train_data, test_data)
for i in range(min(len(data.files), min(len(data.files), 10))):
inp = data.files[i].reshape(len(data.files[i]), INPUT_COUNT, 1)
output = model.predict_output(inp).reshape(len(data.files[i]),
INPUT_COUNT)
data.write_wav(f"output-lstm-{MODEL_ID}-{i}.wav", output)
def build_model(model, learning_rate):
# 1e-4
# 1e-5 cnn 6-1-6 mape-5.86
# 1e-6 cnn-6-1-6 mape-6.49
K.set_epsilon(1e-7)
def rmse(y_true, y_pred):
return losses.mean_squared_error(y_true, y_pred) ** 0.5
def mape(y_true, y_pred):
return losses.mae(y_true, y_pred)
adam = Adam(lr=learning_rate)
model.compile(loss='mse', optimizer=adam, metrics=[rmse, mape])
# model.summary()
return model
def init_tensorflow(seed, use_float16=False):
if use_float16:
import keras.backend as K
dtype = 'float16'
K.set_floatx(dtype)
K.set_epsilon(1e-4)
# Solves an issue with regard to the use of newest CUDA versions
from tensorflow.compat.v1 import ConfigProto
from tensorflow.compat.v1 import InteractiveSession
config = ConfigProto()
config.gpu_options.allow_growth = True
_ = InteractiveSession(config=config)
tf.random.set_seed(seed)
def fm_model(x, y, p):
model = Sequential()
model.add(
Dense(p['l1'],
input_dim=3,
kernel_initializer=p['kernel_init'],
activation=p['ac1']))
model.add(Dropout(p['d1']))
for i in range(4):
layerstr = "l" + str(i + 2)
dropoutstr = "d" + str(i + 2)
activationstr = "ac" + str(i + 2)
model.add(Dense(p[layerstr], activation=p[activationstr]))
model.add(Dropout(p[dropoutstr]))
model.add(Dense(3, activation='linear'))
K.set_epsilon(1)
model.compile(optimizer=tf.train.AdamOptimizer(p['lr']),
loss=[p['loss_func']],
metrics=[p['metrics']])
history = model.fit(x=x,
y=y,
validation_split=0.2,
epochs=p['epoch_size'],
batch_size=p['batch_size'],
shuffle=True,
verbose=0,
callbacks=[
EarlyStopping(monitor='val_mean_absolute_error',
min_delta=0.1,
patience=30,
verbose=0,
mode='auto')
])
return history, model
def __init__(self, debug=False):
# Limit GPU memory(VRAM) usage in TensorFlow 2.0
# https://github.com/tensorflow/tensorflow/issues/34355
# https://medium.com/@starriet87/tensorflow-2-0-wanna-limit-gpu-memory-10ad474e2528
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
try:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
print("Limit GPU memory ...")
except RuntimeError as e:
print(e)
# lower float precision when RTX optimizer is switched on
if self.rtx_optimizer == True:
print("RTX optimizer is on ...")
K.set_epsilon(1e-4)
self.debug = debug
if self.debug:
print("ada_conv initiated...")
def worker(self,
path,
data,
task_sequence,
render_que,
new_weights,
use_gpu=True):
import os
if not use_gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
import sys
from keras import backend as K
K.set_epsilon(1e-9)
gpu_options = tf.GPUOptions(allow_growth=True)
K.set_session(
tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)))
policy_net = create_policy_net('policy_model.h5')
value_net = create_policy_net('value_model.h5')
while True:
item = task_sequence.get()
render = False
if render_que.qsize() > 0:
render = render_que.get()
if new_weights.qsize() > 0:
if new_weights.get():
try:
policy_net.load_weights('policy_model.h5')
value_net.load_weights('value_model.h5')
print('new weights are loaded by process', os.getpid())
except:
print('new weights are not loaded by process',
os.getpid())
#if item == 0:
# render = True
state_memory, reward_memory, action_memory, old_memory, value_memory = self.run_episode(
item, policy_net, value_net, render)
advantage_memory, target_values = self.compute_advantages(
reward_memory, action_memory, value_memory)
data.put((state_memory[:-1], advantage_memory, target_values,
old_memory, sum(reward_memory)))
def __init__(self,
summary_writer=tf.summary.FileWriter('./train'),
start_point=0,
seed=123456,
decay=1e-12,
value_loss_coeff=0.1,
entropy_coeff=1e-1,
batch_size=16,
learning_rate=1e-4,
data_format='channels_first'):
self.model = None
self.target_model = None
self.value_loss_coeff = value_loss_coeff
self.batch_size = batch_size
self.entropy_coeff = entropy_coeff
self.learning_rate = learning_rate
self.iterations = start_point
self.seed = seed
self.data_format = data_format
self.channel_axis = -1 if self.data_format == 'channels_last' else -3
self.decay = decay
# summary
self.writer = summary_writer
###################################
# TensorFlow wizardry
config = tf.ConfigProto()
# Don't pre-allocate memory; allocate as-needed
config.gpu_options.allow_growth = True
# Only allow a total of half the GPU memory to be allocated
# config.gpu_options.per_process_gpu_memory_fraction = 0.5
# Create a session with the above options specified.
k.clear_session()
k.set_session(tf.Session(config=config))
k.set_image_data_format(self.data_format)
k.set_epsilon(1e-12)
def main():
backend.set_floatx("float16")
backend.set_epsilon(1e-4)
data = get_dataset(
block_interval=1,
block_size=INPUT_COUNT,
file_count=1,
output_size=OUTPUT_COUNT,
shuffle=True,
)
train_data = data.train_data.reshape(len(data.train_data), INPUT_COUNT, 1)
test_data = data.test_data.reshape(len(data.test_data), INPUT_COUNT, 1)
model = ExperimentalModel()
model.load()
# model.train(train_data, train_data, test_data, test_data)
for i in range(1):
inp = data.files[i].reshape(len(data.files[i]), INPUT_COUNT, 1)
output = model.predict_output(inp).reshape(len(data.files[i]),
INPUT_COUNT)
data.write_wav(f"output-{NAME}-{MODEL_ID}-{i}.wav", output)
print(f"output-{NAME}-{MODEL_ID}-{i}.wav created")
from keras.models import Model
from keras.optimizers import RMSprop
'''**************************************Parameters******************************************************************'''
# data import options
class_list_filename = "siamese_trainingSet_classes.txt"
# specify root directory where training / validation folders are located
root_dir = "data/product"
n_classes = 26
# training hyperparameters
base_id = 1 # 0=Inception-v3, 1=MobileNet, 2=InceptionResNet-v2, 3=ResNet50
n_epochs = 100
batch_size = 32
K.set_epsilon(1e-07)
epsilon = K.epsilon()
# save parameters
base_name = ' ' # base network name corresponding to base_ids 0-3
date_dir = ' ' # specify save directory for results using YYYY-MM-DD format
similarity_type = 'cosine' # l1_sum, l2_sum, cosine
enable_model_save = True
'''******************************************************************************************************************'''
'''**************************************Functions*******************************************************************'''
# create a list of training image directory locations,class labels and image file formats
def create_class_list(list_filename):
with open(list_filename, 'r') as f:
import keras
from keras.callbacks import EarlyStopping
from keras.models import Input, Model, Sequential
from keras.layers import Dense, Conv2D, Flatten, MaxPooling2D, BatchNormalization
from keras.layers import *
from keras.models import load_model
from keras import initializers
from keras import optimizers
from keras.utils.vis_utils import plot_model
import keras.backend as K
dtype = 'float32'
K.set_floatx(dtype)
# default is 1e-7 which is too small for float16. Without adjusting the epsilon, we will get NaN predictions because of divide by zero problems
K.set_epsilon(1e-4)
from gym_drones.classes.map import WorldMap
from gym_drones.envs.drones_env import DronesDiscreteEnv
def testing_network(
world_size: [int], target_num: int, drone_goal: str, drone_num: int,
extra_drone_num: int, world_gain_peak_range: [float],
world_gain_var_range: [float], world_evolution_speed: [float],
drone_comm: float, drone_view: float, drone_memory: int,
drone_battery: float, action_step: int, max_age: int,
lookahead_step: int, malus: float, final_bonus: float, malus_sm: float,
random_episode: bool, alpha: float, alpha_dec: float, epsilon: float,
epsilon_dec: float, temperature: float, temperature_dec: float,
state_MR: int, limit_MR: bool, three_MR: bool, prioritized: bool,
theano.config.nvcc.fastmath = True
theano.config.nvcc.flags = '-D_FORCE_INLINES'
theano.config.cxx = os.environ['CONDA_PREFIX'] + '/bin/g++'
# import and check Keras version
import keras
import keras.backend as K
from keras import __version__ as keras_version
from pkg_resources import parse_version
if (parse_version(keras_version) >= parse_version('2.0')):
raise RuntimeError('DeepCell requires Keras 1 to run')
# configure Keras, to avoid using file ~/.keras/keras.json
K.set_image_dim_ordering('th') # theano's image format (required by DeepCell)
K.set_floatx('float32')
K.set_epsilon('1e-07')
##
###############################################################################
import cytometer.deepcell as deepcell
import cytometer.deepcell_models as deepcell_models
###############################################################################
## Check response changes with step size changes
###############################################################################
# create test images with a corner step
im = np.ones(shape=(1, 2, 200, 200))
im[:, :, 0:50, 0:50] = 2
from keras.callbacks import Callback, TerminateOnNaN, TensorBoard, ModelCheckpoint, CSVLogger, EarlyStopping
import tensorflow as tf
from tensorflow.python.framework import ops
from tensorflow.python.ops import math_ops, control_flow_ops
from tensorflow.python.training import optimizer
from heat.utils import hyperboloid_to_poincare_ball, load_data, load_embedding
from heat.utils import perform_walks, determine_positive_and_negative_samples
from heat.losses import hyperbolic_softmax_loss, hyperbolic_sigmoid_loss, hyperbolic_hypersphere_loss
from heat.generators import TrainingDataGenerator
from heat.visualise import draw_graph, plot_degree_dist
from heat.callbacks import Checkpointer
K.set_floatx("float64")
K.set_epsilon(1e-15)
np.set_printoptions(suppress=True)
# TensorFlow wizardry
config = tf.ConfigProto()
# Don't pre-allocate memory; allocate as-needed
config.gpu_options.allow_growth = True
# Only allow a total of half the GPU memory to be allocated
config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.log_device_placement=False
config.allow_soft_placement=True
# MULTI_CPU = True
EXPERIMENTS_DIR = "D:\\!DA-20092018\\AVCexperimentsData"
MULTI_CPU = False
# SET OUTPUT_SAVES OUTSIDE THE DOCKER CONTAINER
OUTPUT_SAVES = "./"
# EXTEND_ACTIONS = True
USE_SLIDINGWINDOW = True
# USE_SCALER = False
CNN_TRAINABLE = True
FRAMES_THRESHOLD = 13
COEFF_SLIDINGWINDOW = 0.8
COEFF_DROPOUT = 0.6
COEFF_REGULARIZATION_L2 = 0.015
CNN_BATCH_SIZE = 50
RNN_BATCH_SIZE = 16
k.set_epsilon(1e-06)
ITERATIONS = 1
NUM_EPOCHS = 1
AUGMENTATIONS = [
'none',
# "scale_shift",
# 'scale',
# 'shift_gauss_xy',
# 'noise',
# 'subsample',
# 'interpol',
# 'translate',
# 'scale_translate'
]
OPTIMIZER = [
from keras.layers import Lambda
from keras.layers.wrappers import TimeDistributed
from keras.layers.normalization import BatchNormalization
from keras import regularizers
from keras import callbacks
from keras.optimizers import RMSprop,adam
from keras.models import load_model
import wtte.weibull as weibull
import wtte.wtte as wtte
from wtte.wtte import WeightWatcher
# from phased_lstm_keras.PhasedLSTM import PhasedLSTM as PLSTM
np.random.seed(1)
K.set_epsilon(1e-8)
## Callbacks
checkpointer = callbacks.ModelCheckpoint('./tensorboard_log/wtte-rnn/model_checkpoint.h5',
monitor='loss',
verbose=1,
save_best_only=True,
save_weights_only=True,
mode='auto', period=5)
reduce_lr = callbacks.ReduceLROnPlateau(monitor='loss',
factor =0.5,
patience=15,
verbose=1,
mode='auto',
epsilon=0.0001,
def dsc_crossentropy(y_true, y_pred, smooth=0.00000001):
K.set_epsilon(1e-8)
num = K.binary_crossentropy(y_true, y_pred)
den = ( K.sum( K.abs(y_true), axis=(1,2,3)) + smooth ) * K.ones_like(y_true)
return 0.5 * 256.0*256.0 * num/den
from keras import backend as K
from keras.callbacks import ModelCheckpoint, TerminateOnNaN, EarlyStopping
import tensorflow as tf
from hednet.utils import hyperboloid_to_poincare_ball, load_data
from hednet.utils import determine_positive_and_negative_samples
from hednet.generators import TrainingDataGenerator
from hednet.visualise import draw_graph, plot_degree_dist
from hednet.callbacks import Checkpointer, elu
from hednet.models import build_hyperboloid_asym_model
from hednet.optimizers import ExponentialMappingOptimizer
from hednet.losses import asym_hyperbolic_loss
K.set_floatx("float64")
K.set_epsilon(np.float64(1e-15))
np.set_printoptions(suppress=True)
# TensorFlow wizardry
config = tf.ConfigProto()
# Don't pre-allocate memory; allocate as-needed
config.gpu_options.allow_growth = True
# Only allow a total of half the GPU memory to be allocated
config.gpu_options.per_process_gpu_memory_fraction = 0.5
config.log_device_placement = False
config.allow_soft_placement = True
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Thu Jun 8 11:57:25 2017
@author: shubham
"""
from default import *
from keras import backend as K
from keras import metrics
K.set_epsilon(1e-7)
np.set_printoptions(threshold=np.inf)
#%% Handle dim ordering
assert (BACKEND == K.backend())
if BACKEND == 'theano':
K.theano.config.__setattr__('exception_verbosity', 'high')
K.set_image_data_format('channels_first')
elif BACKEND == 'tensorflow':
K.set_image_data_format('channels_last')
print("Image_ordering:", K.image_dim_ordering())
print("Backend:", K.backend())
#%%LOSS FUNCTIONS
#%%1.DICE LOSS
smooth = 1.0
w_dice = 0.5
def old_dice_coef(y_true, y_pred):
import numpy as np
np.random.seed(123) # for reproducibility
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten
from keras.layers import Convolution2D, MaxPooling2D
from keras.utils import np_utils
from keras.datasets import mnist
from keras import backend as K
K.set_epsilon(1e-6)
# 4. Load pre-shuffled MNIST data into train and test sets
(X_train, y_train), (X_test, y_test) = mnist.load_data()
# 5. Preprocess input data
X_train = X_train.reshape(X_train.shape[0], 1, 28, 28)
X_test = X_test.reshape(X_test.shape[0], 1, 28, 28)
X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255
X_test /= 255
# 6. Preprocess class labels
Y_train = np_utils.to_categorical(y_train, 10)
Y_test = np_utils.to_categorical(y_test, 10)
# 7. Define model architecture
model = Sequential()
model.add(Convolution2D(32, (3, 3), activation='relu',
callbacks=[schedule, checkpoint1, checkpoint2])
pd.DataFrame(result.history).to_csv('models\\GRU_fit_history.csv')
#get the highest validation accuracy of the training epochs
validation_loss = np.amin(result.history['val_loss'])
print('validation loss of epoch:', validation_loss)
return model
if __name__ == '__main__':
X_train, y_train, X_test, y_test, look_back, num_features = data()
"""
GTX 960m and GTX 970 support FP32
"""
from keras import backend as K
float_type = 'float32' # Change this to float16 to use FP16
K.set_floatx(float_type)
K.set_epsilon(1e-4) #default is 1e-7
X_train = X_train.astype(float_type)
y_train = y_train.astype(float_type)
X_test = X_test.astype(float_type)
y_test = y_test.astype(float_type)
model = create_model(X_train, y_train, X_test, y_test, look_back,
num_features)
#print("Evalutation of best performing model:")
#print(model.evaluate(X_test, y_test))
def train_two_phrase(train_file_names, test_file_names, params):
K.set_epsilon(1e-4)
# form vocabs and data
numpy.set_printoptions(precision=3)
print 'loading data.'
old_std = sys.stdout
f_print = open('debug_print_phr1.txt','w')
f_test = open('debug_print_test.txt','w')
sys.stdout = f_print
lines = []
for f_name in train_file_names:
f = open(f_name, 'r')
lines.extend(f.readlines())
f.close()
lines_test = []
for f_name in test_file_names:
f = open(f_name, 'r')
lines_test.extend(f.readlines())
f.close()
vocab_text = {}
inv_vocab_text = {}
data = parse_stories(lines, vocab_text, inv_vocab_text, 1)
data_test = parse_stories(lines_test, vocab_text, inv_vocab_text, 0)
l_train = len(data)
l_test = len(data_test)
data = data[:min(params['train_set_size'], l_train)]
data_test = data_test[:min(params['test_set_size'], l_test)]
story_int, story_mask, q_int, q_mask, ans_int, ans_mask, sizes = int_stories(data, vocab_text)
story_mask = repeatTensor(story_mask, params['dim_emb_story'])
q_mask = repeatTensor(q_mask, params['dim_emb_story'])
story_int_test, story_mask_test, q_int_test, q_mask_test, ans_int_test, ans_mask_test, sizes_test = int_stories(data_test, vocab_text)
story_mask_test = repeatTensor(story_mask_test, params['dim_emb_story'])
q_mask_test = repeatTensor(q_mask_test, params['dim_emb_story'])
inv_vocab_text['0'] = 'dududu'
params['max_story_len'] = sizes[0]
params['max_sent_len'] = max(sizes[1], sizes[3])
params['max_q_num'] = sizes[2]
params['max_story_len_test'] = sizes_test[0]
params['max_q_num_test'] = sizes_test[2]
params['vocab_size'] = len(vocab_text)
params['vocab_size_ans'] = len(vocab_text)
n_samples = len(story_int)
params['ent_range'] = 1 # normal order and inverse order.
print 'params:'
print params
print 'n_samples:'
print n_samples
print 'vocab_text:'
print vocab_text
sys.stdout = old_std
# initialize the env embeddings, actions taken, rewards for the whole train data of an epoch
story_train = numpy.zeros((n_samples * params['story_sims_per_epoch'], params['max_story_len'], params['max_sent_len']))
story_mask_train = numpy.zeros((n_samples * params['story_sims_per_epoch'], params['max_story_len'], params['max_sent_len'], params['dim_emb_story']))
q_train = numpy.zeros((n_samples * params['story_sims_per_epoch'], params['max_q_num'], params['max_sent_len']))
q_mask_train = numpy.zeros((n_samples * params['story_sims_per_epoch'], params['max_q_num'], params['max_sent_len'], params['dim_emb_story']))
act_selected_train = numpy.zeros((n_samples * params['story_sims_per_epoch'], params['max_story_len'], (params['ent_range']) * 1 * params['relation_num'])) # one hot vector
reward_train = numpy.ones((n_samples * params['story_sims_per_epoch'], params['max_story_len'], (params['ent_range']) * 1 * params['relation_num'])) # real number reward signal
reward_immediate_train = numpy.zeros((n_samples * params['story_sims_per_epoch'], params['max_story_len']))
ans_train = numpy.zeros((n_samples * params['story_sims_per_epoch'], params['max_q_num'], params['vocab_size']))
ans_mask_train = numpy.zeros((n_samples * params['story_sims_per_epoch'], params['max_q_num']))# maybe used as mul input at final answer position
fin_one_hot_train = numpy.zeros((n_samples * params['story_sims_per_epoch'], params['max_q_num'], (params['ent_range']) * 1 * params['relation_num'])) # one hot arg1 arg2 r_pred
reward_fin_train = numpy.ones((n_samples * params['story_sims_per_epoch'], params['max_q_num'], (params['ent_range']) * 1 * params['relation_num']))
print 'building model.'
# simulate and generate final train data, do batch train and policy gd
# build model
reasoner, simulator = DRL_Reasoner(params)
test_simulator = DRL_Reasoner_Test(params)
working_memory = working_environment(params['enti_num'], params['relation_num_expand'])
work_space_table = varTable()
# for every single story-question group, take simulations, compute the env embeddings, actions taken, and total rewards.
# take a certain number of simulations for each group, and use the pool as total train data, do sgd and batch training.
# loop that progress for many epoches.
mask_sim = numpy.zeros((1, params['max_story_len'], params['dim_emb_story'] * params['max_sent_len']))
all_sim_number = 0
print 'two phrase training started.'
for epoch_id in range(params['epoch_num']):
train_trace_id = 0
avg_reward_q_epoch = 0.
avg_reward_action_epoch = numpy.zeros(params['max_story_len'])
print 'epoch %d phrase 1' %epoch_id
# phrase 1: simulate and train data generating
epoch_precision_rate = 0.
sample_rate = numpy.zeros(n_samples)
for story_id in range(n_samples):
story_precision = 0.
for sim_round in range(params['story_sims_per_epoch']):
# write story, q, ans, data
# print story_train[train_trace_id,:,:], story_int[story_id]
sys.stdout = f_print
story_train[train_trace_id,:len(story_int[0]),:len(story_int[0][0])] = story_int[story_id]
story_mask_train[train_trace_id,:len(story_int[0]),:len(story_int[0][0]),:] = story_mask[story_id]
q_train[train_trace_id,:len(q_int[0]),:len(q_int[0][0])] = q_int[story_id]
q_mask_train[train_trace_id,:len(q_int[0]),:len(q_int[0][0]),:] = q_mask[story_id]
ans_train[train_trace_id,:len(ans_int[0]),:len(ans_int[0][0])] = ans_int[story_id]
arg_cached = []
for time_slice in range(params['max_story_len']):
mask_sim[0][time_slice][:] = numpy.ones(params['dim_emb_story'] * params['max_sent_len'])
action_probs, retrieve_probs = simulator.predict([story_train[numpy.newaxis,train_trace_id],
story_mask_train[numpy.newaxis,train_trace_id],
q_train[numpy.newaxis,train_trace_id],
q_mask_train[numpy.newaxis,train_trace_id],
mask_sim[:],
reward_train[numpy.newaxis,train_trace_id],
reward_fin_train[numpy.newaxis,train_trace_id]])
for time_slice in range(params['max_story_len']):
action_selected, action_one_hot = select_action(action_probs[:,time_slice,:], params['epsilon'])
act_selected_train[train_trace_id, time_slice,:] = action_one_hot
arg_1_ptr = action_selected // ((1) * params['relation_num']) # start from 0 (empty number)
arg_2_ptr = (action_selected - arg_1_ptr * (1) * params['relation_num']) // params['relation_num']
arg_r_ptr = (action_selected - arg_1_ptr * (1) * params['relation_num']) % params['relation_num'] + 1
arg_r = arg_r_ptr
ents_in_sent = []
if time_slice < len(story_int[story_id]):
for w_id in story_int[story_id][time_slice]:
if inv_vocab_text[str(int(w_id))] in ent_list.values():
ents_in_sent.append(w_id)
if len(ents_in_sent) > 0:
arg_1 = int(ents_in_sent[0])
if arg_1 == vocab_text['they']:
arg_1 = arg_cached[0]
arg_2 = arg_cached[1]
arg_3 = int(ents_in_sent[1])
else:
arg_2 = int(ents_in_sent[1])
arg_cached = [arg_1, arg_2]
arg_3 = int(ents_in_sent[2])
# if len(ents_in_sent) == 3:
# slice_reward = 0
# for tt in range(time_slice+1):
# reward_immediate_train[train_trace_id][tt] += slice_reward * (params['reward_decay'] ** (time_slice - tt))
if arg_1 > 0 and arg_2 > 0 and story_mask_train[train_trace_id, time_slice, 0, 0] > 0:
arg_1_int = work_space_table.retr_insert(arg_1, inv_vocab_text)
arg_2_int = work_space_table.retr_insert(arg_2, inv_vocab_text)
arg_3_int = work_space_table.retr_insert(arg_3, inv_vocab_text)
if arg_r <= 2:
working_memory.modifyEnv((arg_1_int, arg_r, arg_3_int))
working_memory.modifyEnv((arg_2_int, arg_r, arg_3_int))
# compute fin_train and fin_train_pred
retrieved_relation_list = []
reward_temp_list = [] # reward for every question based on arg1/2 in q or not.
for q_idx in range(len(retrieve_probs[0])):
retr_idx, action_one_hot_retr = select_action(retrieve_probs[0,q_idx,:], params['epsilon'])
arg1_retr_ptr = retr_idx // ((1) * params['relation_num'])
arg2_retr_ptr = (retr_idx - arg1_retr_ptr * (1) * params['relation_num']) // params['relation_num']
relation_pred = (retr_idx - arg1_retr_ptr * (1) * params['relation_num']) % params['relation_num'] + 1
# convert from ptr to id:
flag = 0
for q_w in q_int[story_id, 0]:
if inv_vocab_text[str(int(q_w))] in ent_list.values() and flag == 0:
arg1_retr = int(q_w)
flag = 1
elif inv_vocab_text[str(int(q_w))] in ent_list.values() and flag == 1:
arg2_retr = int(q_w)
retrieve_reward_pre = 0
reward_temp_list.append(retrieve_reward_pre)
arg1_retr_int = work_space_table.retr(arg1_retr, inv_vocab_text)
# arg2_retr_int = work_space_table.retr(arg2_retr, inv_vocab_text)
search_result = working_memory.retrieveArg(arg1_retr_int)
arg_retr_str = work_space_table.inv_retr(search_result)
retrieved_relation_list.append(arg_retr_str)
one_hot_single = numpy.zeros(((params['ent_range']) * 1 * params['relation_num']))
one_hot_single[retr_idx] = 1
fin_one_hot_train[train_trace_id, q_idx, :] = one_hot_single
reward_q_total = 0.
for q_idx in range(len(retrieve_probs[0])):
ans_word_int = numpy.argmax(ans_train[train_trace_id][0]) + 1
ans_word = inv_vocab_text[str(ans_word_int)]
reward_scalar_q, ans_q, label_q = compute_single_reward(retrieved_relation_list[q_idx],ans_word)
if reward_scalar_q > 0 and q_idx == 0:
epoch_precision_rate += 1.
story_precision += 1.
reward_scalar_q *= q_mask_train[train_trace_id, q_idx, 0, 0]
reward_q_total += reward_scalar_q
reward_q_total += reward_temp_list[0]
for time_slice in range(params['max_story_len']):
reward_immediate_train[train_trace_id, time_slice] += reward_q_total * (params['reward_decay'] ** (params['max_story_len'] - time_slice))
for q_idx in range(len(retrieve_probs[0])):
# pass
if q_idx == 0:
mask_reward = 1.
else:
mask_reward = 0.
reward_fin_train[train_trace_id, q_idx, :] *= (reward_q_total * mask_reward) # used as input at the last softmax
for time_slice in range(params['max_story_len']):
# pass
reward_train[train_trace_id, time_slice, :] *= (reward_immediate_train[train_trace_id, time_slice] * story_mask_train[train_trace_id, time_slice, 0, 0]) # used as input at the last softmax
train_trace_id += 1
all_sim_number += 1
mask_sim *= 0
working_memory.resetEnv()
work_space_table.reset()
sample_rate[story_id] = story_precision / params['story_sims_per_epoch']
epoch_precision_rate = epoch_precision_rate / (n_samples * params['story_sims_per_epoch'])
print 'the total answer precision of this epoch:'
print epoch_precision_rate
sys.stdout = old_std
print 'precision of this epoch: %f' %epoch_precision_rate
print 'epoch %d phrase 2' %(epoch_id)
print 'sample_rate:'
print sample_rate
# phrase2: go batch train on the trace pool.
mask_sim_2 = numpy.ones((n_samples * params['story_sims_per_epoch'], params['max_story_len'], params['dim_emb_story'] * params['max_sent_len']))
reasoner.fit([story_train, story_mask_train, q_train, q_mask_train, mask_sim_2, reward_train, reward_fin_train],
{'action_probs_re': act_selected_train, 'retrieve_probs_re': fin_one_hot_train},
batch_size=params['batch_size_phrase2'], nb_epoch=params['nb_epoch_phrase2'], verbose=2)
sys.stdout = old_std
# test the model
weights_train = [i for i in simulator.layers if len(i.get_weights()) != 0]
weights_test = [i for i in test_simulator.layers if len(i.get_weights()) != 0]
for (l1,l2) in zip(weights_train,weights_test):
l2.set_weights(l1.get_weights())
if (epoch_id + 1) % params['test_epoch_period'] == 0:
test_model(test_simulator, story_int_test, story_mask_test, q_int_test, q_mask_test, ans_int_test, ans_mask_test, params, old_std, f_test, vocab_text, inv_vocab_text)
sys.stdout = old_std
story_train *= 0
story_mask_train *= 0
q_train *= 0
q_mask_train *= 0
act_selected_train *= 0
reward_train = numpy.ones((n_samples * params['story_sims_per_epoch'], params['max_story_len'], (params['ent_range']) * 1 * params['relation_num'])) # real number reward signal
ans_train *= 0
ans_mask_train *= 0
reward_fin_train = numpy.ones((n_samples * params['story_sims_per_epoch'], params['max_q_num'], params['ent_range'] * 1 * params['relation_num']))
fin_one_hot_train *= 0
reward_immediate_train *= 0
f_print.close()
f_test.close()
comp_count] = G_temp_component
G_temp_count += comp_count
return Gfunc_data
##################################
###### BPNN models ######
##################################
import os
#os.environ["THEANO_FLAGS"] = "mode=FAST_RUN,device=cuda,floatX=float64,dnn.enabled=True"
os.environ["KERAS_BACKEND"] = "tensorflow"
from keras.models import load_model, model_from_json
#===== setting for float64 for keras ======
from keras import backend as K
K.set_floatx('float64')
K.set_epsilon(1e-14)
def bpnn_2b_model(base_path, at_idx_map, Gfunc_data, s):
"""calculate sum of all atomic energies from subnetworks for dimers
# Arguments
base_path: path to bpnn-2b model
at_idx_map: a mapping between atom types and atom indexes; dictionary
Gfunc_data: symmetry function values;
dictionary with 1st layer keys = atom types,
2nd layer keys = atom indexes,
values = 2D arrays with shape=(nb_samples, nb_sym_functions)
s: switching function
# Returns
bpnn 2B-energy
"""
def main():
backend.set_floatx("float16")
backend.set_epsilon(1e-4)
just_files = False
data = get_dataset(
block_interval=50,
block_size=INPUT_COUNT,
file_count=1,
output_size=0,
shuffle=True,
just_files=just_files,
)
if not just_files:
train_data = data.train_data.reshape(len(data.train_data), INPUT_COUNT, 1)
test_data = data.test_data.reshape(len(data.test_data), INPUT_COUNT, 1)
level1 = Level1Autoencoder()
level1.train(train_data, train_data, test_data, test_data)
# Prepare data by running it through our first level autoencoder
data = level1.encode(data.files[0].reshape(len(data.files[0]), INPUT_COUNT, 1))
plotdata = data.reshape(len(data), 128)[:1000]
plt.subplot(2, 1, 1)
plt.plot(plotdata)
data = data[:int(len(data) / LEVEL_2_INPUT_COUNT) * LEVEL_2_INPUT_COUNT]
data = np.array(np.split(data, len(data) / LEVEL_2_INPUT_COUNT))
data = data.reshape(len(data), LEVEL_2_INPUT_COUNT, 128)
# Unload level 1 model
del level1
backend.clear_session()
level2 = Level2Autoencoder()
level2.train(data, data, data, data)
output = level2.predict_output(data)
print(output.shape)
plotdata = output.reshape(output.shape[0] * output.shape[1], 128)[:1000]
plt.subplot(2, 1, 2)
plt.plot(plotdata)
plt.show()
print(output.shape)
output = output.reshape(output.shape[0] * output.shape[1], 1, 128)
print(output.shape)
del level2
backend.clear_session()
level1 = Level1Autoencoder()
output = level1.decode(output).flatten()
data = Dataset()
data.write_wav(f"output-{NAME}-{MODEL_ID}-level2.wav", output)
for i in range(min(len(data.files), 2)):
inp = data.files[i].reshape(len(data.files[i]), INPUT_COUNT, 1)
output = level1.decode(level1.encode(inp)).reshape(len(data.files[i]), INPUT_COUNT)
# output = level1.predict_output(inp).reshape(len(data.files[i]), INPUT_COUNT)
data.write_wav(f"output-{NAME}-{MODEL_ID}-level1-{i}.wav", output)
print(f"output-{NAME}-{MODEL_ID}-{i}.wav created")
plt.subplot(2, 2)
plt.plot(inp.flatten()[2000:8000])
plt.subplot(2, 2)
plt.plot(output.flatten()[2000:8000])
plt.show()
#!/usr/bin/env python2 # -*- coding: utf-8 -*- """ Created on Tue Mar 21 19:44:05 2017 @author: hh """ from keras import backend as K import numpy as np #def Solve(): K.set_epsilon(1e-11) x = K.placeholder(shape=(None, 1)) y = K.placeholder(shape=(None, 1)) label = K.placeholder(shape=(None, 1)) #x= K.variable(value = X_train) #y = K.variable(value = Y_train) #label = K.variable(value = L_train) W = K.variable(value=0.5) #np.random.rand(1)) b = K.variable(value=1.0) #np.random.rand(1)) U = K.square(W * x + b) lamda = K.variable(value=1.0) #np.random.rand(1)) U2 = y - lamda * U #U2 = K.concatenate([U, 1-U], axis = -1)
