def train(self, sess, num_epochs=20):
"""
Remember to initialize all variables outside this method
"""
batch_size = self.batch_size
# Placeholder for some more summaries that may be of interest.
merged_summaries = tf.summary.merge([self.loss_summ])
self.writer = tf.summary.FileWriter(addDateTime('./logs/log'))
Ytrain = self.to_one_hot(self.data_dict['Ytrain'], self.yDim)
Xtrain = self.data_dict['Xtrain']
Yvalid = self.to_one_hot(self.data_dict['Yvalid'], self.yDim)
Xvalid = self.data_dict['Xvalid']
for ep in range(num_epochs):
num_batches = len(Xtrain) // batch_size
for idx in range(0, num_batches):
batch_images = Xtrain[idx * batch_size:(idx + 1) * batch_size]
batch_labels = Ytrain[idx * batch_size:(idx + 1) * batch_size]
_, loss, summaries = sess.run(
[self.train_op, self.loss, merged_summaries],
feed_dict={
'FOURIER/X:0': batch_images,
'FOURIER/Y:0': batch_labels
})
if (idx * 10) // num_batches > ((idx - 1) * 10) // num_batches:
print((100 * idx) // num_batches, '% processed...')
self.writer.add_summary(summaries, ep)
print('Ep, cost:', ep, loss)
ypred = sess.run(tf.nn.softmax(self.output),
feed_dict={'FOURIER/X:0': Xtrain[:100]})
ypred = self.nptensorfloat_to_one_hot(ypred)
accuracy = np.sum(ypred * Ytrain[:100]) / 100
print('Accuracy', accuracy)
if ep % 50 == 0:
Ypred_valid = sess.run(self.output,
feed_dict={'FOURIER/X:0': Xvalid[:100]})
Ypred_valid = self.nptensorfloat_to_one_hot(Ypred_valid)
accuracy = np.sum(Ypred_valid * Yvalid[:100]) / 100
print('Accuracy', accuracy)
def main(_):
"""
Launches this whole zingamajinga.
"""
data_path = params.local_data_dir + params.this_data_dir
rlt_dir = params.local_rlt_dir + params.this_data_dir + addDateTime() + '/'
if params.mode == 'generate':
generate_fake_data(lat_mod_class=params.lat_mod_class,
gen_mod_class=params.gen_mod_class,
params=params,
data_path=data_path,
save_data_file=params.save_data_file,
Nsamps=params.genNsamps,
NTbins=params.genNTbins,
write_params_file=True,
draw_quiver=True,
draw_heat_maps=True,
savefigs=True)
if params.mode == 'train':
graph = tf.Graph()
with graph.as_default():
sess = tf.Session(graph=graph)
with sess.as_default():
with open(data_path + params.save_data_file, 'rb+') as f:
# Set encoding='latin1' for python 2 pickled data
datadict = pickle.load(
f,
encoding='latin1') if params.is_py2 else pickle.load(f)
Ytrain = datadict['Ytrain']
Yvalid = datadict['Yvalid']
params.yDim = Ytrain.shape[-1]
if not os.path.exists(rlt_dir): os.makedirs(rlt_dir)
write_option_file(rlt_dir)
opt = Optimizer_TS(params)
sess.run(tf.global_variables_initializer())
opt.train(sess, rlt_dir, Ytrain, Yvalid)
def main(_):
"""
Fly babe!
"""
data_path = params.local_data_dir + params.this_data_dir
rlt_dir = (params.local_rlt_dir + params.this_data_dir + addDateTime() +
'/' if not params.restore_from_ckpt else params.load_ckpt_dir)
if params.mode == 'generate':
generate_fake_data(params,
data_path=data_path,
save_data_file=params.save_data_file)
elif params.mode == 'train':
sess = tf.Session()
with sess.as_default():
train(params, data_path, rlt_dir)
elif params.mode == 'other':
with open(data_path + params.save_data_file, 'rb+') as f:
# Set encoding='latin1' for python 2 pickled data
datadict = pickle.load(
f, encoding='latin1') if params.is_py2 else pickle.load(f)
print(sorted(datadict.keys()))
other(datadict, data_path)
from lasagne.nonlinearities import softmax, linear, softplus, tanh, sigmoid from lasagne.layers import InputLayer, DenseLayer from lasagne.init import Orthogonal, Uniform, Normal # Local imports import vind from code.ObservationModels import GaussianObsTSGM, PoissonObsTSGM from code.LatEvModels import LocallyLinearEvolution from code.RecognitionModels import SmoothingNLDSTimeSeries from code.OptimizerVAEC_TS import VAEC_NLDSOptimizer # from MinibatchIterator import DatasetTrialIterator from code.datetools import addDateTime cur_date = addDateTime() #=== GLOBALS ===# # AS A USER YOU SHOULD BE ABLE TO GET THIS RUNNING JUST BY CHANGING THESE # PLEASE READ THE HELP ON EACH OF THESE OPTIONS IN THE PARSER. ### HARDWARE, META CLUSTER = 0 EXTERNAL_DATA = 1 ### DATA PATHS, ETC ### LOCAL_DATA_ROOT = "data/" THIS_DATA_DIR = "poisson_data_002" # lorenzdata004 IS_DICT = 1 DATA_FILE = 'datadict' TEST_FILE = None
import os
import pickle
import matplotlib
matplotlib.use('Agg')
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
import tensorflow as tf
from code.Optimizer_VAEC import Optimizer_TS
from code.datetools import addDateTime
# DATA_FILE = '/Users/danielhernandez/work/supervind/data/poisson002/datadict'
DATA_FILE = '/Users/danielhernandez/work/vind/data/poisson_data_002/datadict'
RLT_DIR = "/Users/danielhernandez/work/supervind/rslts/poisson002/fit" + addDateTime(
) + '/'
if not os.path.exists(RLT_DIR): os.makedirs(RLT_DIR)
# For information on these parameters, see runner.py
flags = tf.app.flags
flags.DEFINE_integer('yDim', 10, "")
flags.DEFINE_integer('xDim', 2, "")
flags.DEFINE_float('learning_rate', 2e-3, "")
flags.DEFINE_float('initrange_MuX', 0.2, "")
flags.DEFINE_float('initrange_B', 3.0, "")
flags.DEFINE_float('init_Q0', 0.4, "")
flags.DEFINE_float('init_Q', 1.0, "")
flags.DEFINE_float('alpha', 0.2, "")
flags.DEFINE_float('initrange_outY', 3.0, "")
flags.DEFINE_float('initrange_LambdaX', 1.0, "")
params = tf.flags.FLAGS