def __init__(self, name=None, autoconnect=None, **kwargs):
Model.__init__(self)
Base.__init__(self)
self.metadata = Base.metadata
self.name = name
self.autoconnect = autoconnect
def __init__(self, train, column, choices=[], beta0=None, name='Enum model', bounds=None): Model.__init__(self, train) self.choices = choices self.column = column self.name = name self._beta0 = list(beta0) self._bounds = list(bounds)
def train(logdir1='logdir/default/train1', logdir2='logdir/default/train2', queue=True):
model = Model(mode="train2", batch_size=hp.Train2.batch_size, queue=queue)
# Loss
loss_op = model.loss_net2()
# Training Scheme
global_step = tf.Variable(0, name='global_step', trainable=False)
optimizer = tf.train.AdamOptimizer(learning_rate=hp.Train2.lr)
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, 'net/net2')
train_op = optimizer.minimize(loss_op, global_step=global_step, var_list=var_list)
# Summary
summ_op = summaries(loss_op)
session_conf = tf.ConfigProto(
gpu_options=tf.GPUOptions(
allow_growth=True,
per_process_gpu_memory_fraction=0.6,
),
)
# Training
with tf.Session(config=session_conf) as sess:
# Load trained model
sess.run(tf.global_variables_initializer())
model.load(sess, mode='train2', logdir=logdir1, logdir2=logdir2)
writer = tf.summary.FileWriter(logdir2, sess.graph)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
for epoch in range(1, hp.Train2.num_epochs + 1):
for step in tqdm(range(model.num_batch), total=model.num_batch, ncols=70, leave=False, unit='b'):
if queue:
sess.run(train_op)
else:
mfcc, spec, mel = get_batch(model.mode, model.batch_size)
sess.run(train_op, feed_dict={model.x_mfcc: mfcc, model.y_spec: spec, model.y_mel: mel})
# Write checkpoint files at every epoch
summ, gs = sess.run([summ_op, global_step])
if epoch % hp.Train2.save_per_epoch == 0:
tf.train.Saver().save(sess, '{}/epoch_{}_step_{}'.format(logdir2, epoch, gs))
# Eval at every n epochs
with tf.Graph().as_default():
eval2.eval(logdir2, queue=False)
# Convert at every n epochs
with tf.Graph().as_default():
convert.convert(logdir2, queue=False)
writer.add_summary(summ, global_step=gs)
writer.close()
coord.request_stop()
coord.join(threads)
def __init__(self, **kwargs):
Model.__init__(self)
Base.__init__(self)
self.metadata = Base.metadata
for name, val in kwargs.iteritems():
self.__setattr__(name, val)
def get_friends_from_ml100k():
dst_file = "./src/dataset/movie100k/ml100kfriend.dat"
if path.isfile(dst_file):
friend_data = pickle.load(open(dst_file, "rb"))
return friend_data
raw_data = get_moive100k(True);
raw_model = Model(raw_data);
cosine_sim = CosineSimilarity(raw_model)
friend_data = {}
for user_id in raw_model.get_user_ids():
neighbors = cosine_sim.get_similarities(user_id)[:250]
user_ids, x = zip(*neighbors)
user_ids = list(user_ids)
shuffle(user_ids)
# note:
# Randomly choose 150 out of 250 neighbors as friends.
# In such case, systems is able to (possiblly) choose strangers which
# are in top-250 similar users, but with a probability slightly
# smaller than friends selection.
friend_data[user_id] = user_ids[:150]
pickle.dump(friend_data,open(dst_file, "w"),protocol=2)
return friend_data
def tearDown(self):
for m in (Model, ModelFK, ModelM2M):
try:
moderator.unregister(m)
except:
pass
Model.add_to_class('objects', self.default_manager)
def initialize_table(self):
self.session.add_all([
Server(network_id=1, address="irc.gamesurge.net", port=6667,
nickname="ppbot", alt_nickname="ppbot",
realname="Powered by Dong Energy"),
])
Model.initialize_table(self)
def __init__(self, token):
self.token = token
self._initialize()
self.schedule(self.__cleanup, after=timedelta(minutes=1),
recurring=True)
Model.bind_bot(self)
Hook.bind_bot(self)
def test_can_filter_model_after_register(self):
moderator.register(Model)
model = Model()
model.save()
self.assertTrue(model)
models = Model.objects.unmoderated()
self.assertIn(model, models)
def test_can_create_model_after_register(self):
moderator.register(Model)
model = Model()
model.save()
self.assertTrue(model)
model = Model.objects.create()
self.assertTrue(model)
def __init__(self, lower=[0.01, 50, 100], upper = [0.2, 150, 1000], algo_model=dtlz.dtlz1, algo_num_obj=1, algo_num_decs=1):
Model.__init__(self)
#define upper and lower bounds for dec variables: Mutation rate, no of candidates and no of gens
self.name = 'Tuner Model'
self.lower_bounds = lower
self.upper_bounds = upper
self.no_of_decisions = len(lower)
self.algo_model = algo_model
self.algo_num_obj = algo_num_obj
self.algo_num_decs = algo_num_decs
def modelsPage():
page = Model(dsn = app.config['dsn'])
if request.method == 'GET':
return page.list()
elif 'addModel' in request.form:
name = request.form['name']
rider = request.form['rider']
constructor = request.form['constructor']
return page.addModel(name, rider, constructor)
elif 'dbynameModel' in request.form:
name = request.form['name']
return page.deletebyName(name)
elif 'dbyidModel' in request.form:
ID = request.form['ID']
return page.deletebyId(ID)
elif 'updateModel' in request.form:
ID = request.form['ID']
name = request.form['name']
rider = request.form['rider']
constructor = request.form['constructor']
return page.update(ID,name, rider, constructor)
elif 'deleteAllModels' in request.form:
return page.deleteAll()
elif 'AutoFillModels' in request.form:
return page.autoFill()
elif 'searchbyName' in request.form:
name = request.form['name']
return page.find(name)
else:
return redirect(url_for('home_page'))
def __init__(self, name=None, password=None, access=None, **kwargs):
Model.__init__(self)
Base.__init__(self)
self.metadata = Base.metadata
self.name = name
self.password = password
self.access = access
for name, val in kwargs.iteritems():
self.__setattr__(name, val)
def __init__(self, train):
Model.__init__(self, train)
self.bymonth = bymonth = train.groupby('month')
# First of each month and last of last month
dayages = bymonth.first()['dayage'].values
self.dayages = numpy.concatenate([dayages, train.iloc[-1:]['dayage'].values])
self.means = bymonth['count'].mean()
self.nsegments = len(bymonth)
self.nparams = self.nsegments + 1
def test_diff_creates_changes_record(self):
objects = Model.objects
Model.add_to_class('objects', ModeratorManager())
m = Model.objects.create(name='model first version')
me = ModeratorEntry.objects.create(
content_type=ContentType.objects.get_for_model(Model),
object_id=m.pk
)
m.name = 'model second version'
m.save()
me.diff()
self.assertTrue(me.changes.all())
Model.add_to_class('objects', objects)
def eval(logdir='logdir/default/train1', queue=False):
# Load graph
model = Model(mode="test1", batch_size=hp.Test1.batch_size, queue=queue)
# Accuracy
acc_op = model.acc_net1()
# Loss
loss_op = model.loss_net1()
# Summary
summ_op = summaries(acc_op, loss_op)
session_conf = tf.ConfigProto(
allow_soft_placement=True,
device_count={'CPU': 1, 'GPU': 0},
)
with tf.Session(config=session_conf) as sess:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
writer = tf.summary.FileWriter(logdir, sess.graph)
# Load trained model
sess.run(tf.global_variables_initializer())
model.load(sess, 'train1', logdir=logdir)
if queue:
summ, acc, loss = sess.run([summ_op, acc_op, loss_op])
else:
mfcc, ppg = get_batch(model.mode, model.batch_size)
summ, acc, loss = sess.run([summ_op, acc_op, loss_op], feed_dict={model.x_mfcc: mfcc, model.y_ppgs: ppg})
writer.add_summary(summ)
print("acc:", acc)
print("loss:", loss)
print('\n')
writer.close()
coord.request_stop()
coord.join(threads)
def calculate_time_by_hours_global(rw_data):
print len(rw_data.keys())
model = Model(rw_data)
iplot = {}
print len(model.get_user_ids())
for u_id in model.get_user_ids():
rates = model.get_rates(u_id)
for i_id in rates.keys():
ra, ts = rates[i_id]
timestamp = float(ts)
value = datetime.datetime.fromtimestamp(timestamp)
hr = value.hour
if hr in iplot:
iplot[hr] = iplot[hr] + 1
else:
iplot[hr] = 1
return iplot
def __init__( self, feature_engineer=None, model=None ):
# configure logging
self.logger = logging.getLogger("Pipeline")
handler = logging.FileHandler(settings.LOG_PATH)
handler.setFormatter(logging.Formatter(
'%(asctime)s %(levelname)s %(name)s: %(message)s'))
self.logger.addHandler(handler)
self.logger.setLevel(logging.DEBUG)
self.settings = Settings()
self.feature_engineer = FeatureEngineer( self.settings )
self.model = Model().get_model( Settings() )
print("testing time:", time.time() - start_time)
# for ref, pre in zip(references, predictions):
# print(ref)
# print("-"*20)
# print(pre)
# print("*"*100)
if __name__ == '__main__':
args = set_parser()
config = Config(args)
if args.mode == 'train':
train(config)
else:
test_set = get_dataset(config.test_path,
config.w2i_vocabs,
config,
is_train=False)
model = Model(n_emb=args.n_emb,
n_hidden=args.n_hidden,
vocab_size=args.vocab_size,
dropout=args.dropout,
d_ff=args.d_ff,
n_head=args.n_head,
n_block=args.n_block)
model_dict = torch.load(args.restore)
model.load_state_dict(model_dict)
model.to(device)
test(test_set, model)
for i in range(1, args.nlayer):
v[..., i * int(shape[-1] / args.nlayer):] = vp_i[i] # Bottom velocity
v0 = ndimage.gaussian_filter(v, sigma=10)
v0[v < 1.51] = v[v < 1.51]
v0 = ndimage.gaussian_filter(v0, sigma=3)
rho0 = (v0 + .5) / 2
dm = v0**(-2) - v**(-2)
dm[:, -1] = 0.
# Set up model structures
if is_tti:
model = Model(shape=shape,
origin=origin,
spacing=spacing,
vp=v0,
epsilon=.045 * (v0 - 1.5),
delta=.03 * (v0 - 1.5),
rho=rho0,
theta=.1 * (v0 - 1.5),
dm=dm,
space_order=so)
else:
model = Model(shape=shape,
origin=origin,
spacing=spacing,
vp=v0,
rho=rho0,
dm=dm,
space_order=so)
# Time axis
t0 = 0.
import sys
sys.path.append(".")
import numpy as np
from sklearn.cross_validation import KFold, LabelKFold
from utils import weighted_precision_recall, read_training_data, f1_score
from models import Model
if __name__ == "__main__":
csv_path = sys.argv[1]
image_prefix = sys.argv[2]
(X, Y, categories, sample_weights) = read_training_data(csv_path, image_prefix = image_prefix, category = True, sample_weight = True)
print "#T = %d, #F = %d, #N = %d"%(sum(1 for y in Y if y == 0), sum(1 for y in Y if y == 1), sum(1 for y in Y if y == 2))
print "X =", X.shape, ", Y =",Y.shape
# Machine Learning models
model = Model()
kf = LabelKFold(categories, n_folds = 5) # ***shuffle internally***
#kf = KFold(len(X), n_folds = 5, shuffle = True, random_state = 123)
#cross validation
S = []
for train_index, test_index in kf:
trainX, testX = X[train_index], X[test_index]
trainY, testY = Y[train_index], Y[test_index]
K = None
#fit data
model.fit(trainX[:K], trainY[:K], sample_weight = sample_weights[train_index][:K])
#make prediction
predY = model.predict(testX)
#evaluation
def train(logdir='logdir/default/train1', queue=True):
model = Model(mode="train1", batch_size=hp.Train1.batch_size, queue=queue)
# Loss
loss_op = model.loss_net1()
# Accuracy
acc_op = model.acc_net1()
# Training Scheme
global_step = tf.Variable(0, name='global_step', trainable=False)
optimizer = tf.train.AdamOptimizer(learning_rate=hp.Train1.lr)
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
var_list = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, 'net/net1')
train_op = optimizer.minimize(loss_op, global_step=global_step, var_list=var_list)
# Summary
for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, 'net/net1'):
tf.summary.histogram(v.name, v)
tf.summary.scalar('net1/train/loss', loss_op)
tf.summary.scalar('net1/train/acc', acc_op)
summ_op = tf.summary.merge_all()
session_conf = tf.ConfigProto(
gpu_options=tf.GPUOptions(
allow_growth=True,
),
)
# Training
with tf.Session(config=session_conf) as sess:
# Load trained model
sess.run(tf.global_variables_initializer())
model.load(sess, 'train1', logdir=logdir)
writer = tf.summary.FileWriter(logdir, sess.graph)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
for epoch in range(1, hp.Train1.num_epochs + 1):
for step in tqdm(range(model.num_batch), total=model.num_batch, ncols=70, leave=False, unit='b'):
if queue:
sess.run(train_op)
else:
mfcc, ppg = get_batch(model.mode, model.batch_size)
#print("MFCC shape: {}".format(mfcc.shape))
#print("types: {} and {}".format(mfcc.dtype, ppg.dtype))
#print("PPG shape: {}".format(ppg.shape))
sess.run(train_op, feed_dict={model.x_mfcc: mfcc, model.y_ppgs: ppg})
# Write checkpoint files at every epoch
summ, gs = sess.run([summ_op, global_step], feed_dict={model.x_mfcc: mfcc, model.y_ppgs: ppg})
# There was a problem where in certain environments placeholder must be fed for these ops
if epoch % hp.Train1.save_per_epoch == 0:
tf.train.Saver().save(sess, '{}/epoch_{}_step_{}'.format(logdir, epoch, gs))
# Write eval accuracy at every epoch
with tf.Graph().as_default():
eval1.eval(logdir=logdir, queue=False)
writer.add_summary(summ, global_step=gs)
writer.close()
coord.request_stop()
coord.join(threads)
parser.add_argument('--gpu', default=0, type=int, help='gpu id')
parser.add_argument('--log_path',
default='./logs',
type=str,
help='path to log')
args = parser.parse_args()
train_dataset = Dataset(args.data_path, args.dataset, True)
train_loader = data.DataLoader(train_dataset, batch_size=args.bs, shuffle=True)
test_dataset = Dataset(args.data_path, args.dataset, False)
test_loader = data.DataLoader(test_dataset, batch_size=1, shuffle=False)
device = torch.device('cuda:' + str(args.gpu))
model = Model().to(device)
writer = SummaryWriter(args.log_path)
mseloss = nn.MSELoss(reduction='sum').to(device)
bceloss = nn.BCELoss(reduction='sum').to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-4)
if args.load:
checkpoint = torch.load(
os.path.join(args.save_path, 'checkpoint_latest.pth'))
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
best_mae = torch.load(os.path.join(args.save_path,
'checkpoint_best.pth'))['mae']
def create_model(self):
return Model(3, 6)
from flask import Flask, Response, request, render_template
from models import Model
app = Flask(__name__)
greeter = Model("bolt://103.89.6.76:7778", "proglan", "kelas2020")
@app.route('/')
def getdata_kurasi():
getdata = greeter.get_kurasi()
return render_template("index.html", data=getdata)
if __name__ == "__main__":
app.run(port=80, debug=True)
def __init__(self, train, name='Unnamed'): self.name = name Model.__init__(self, train)
def initialize_table(self):
self.session.add_all([
Network('gamesurge')
])
Model.initialize_table(self)
def train(X_train,
y_train,
batch_size,
num_features,
num_targets,
learning_rate,
weight_decay,
hidden_size,
epochs,
fold_id,
X_valid=None,
y_valid=None,
filter_p=False):
device = checkDevice()
train_dataset = TrainDataset(X_train, y_train)
if ((X_valid is not None) and (y_valid is not None)):
valid_dataset = TrainDataset(X_valid, y_valid)
trainloader = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True)
if ((X_valid is not None) and (y_valid is not None)):
validloader = torch.utils.data.DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=False)
model = Model(num_features=num_features,
num_targets=num_targets,
hidden_size=hidden_size,
weights=getClassWeights(y_train))
model.to(device)
optimizer = torch.optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
scheduler = optim.lr_scheduler.OneCycleLR(optimizer=optimizer,
pct_start=0.1,
div_factor=1e3,
max_lr=1e-2,
epochs=epochs,
steps_per_epoch=len(trainloader))
loss_fn = model.loss_fn
for epoch in range(epochs):
start = time.time()
train_loss = train_epoch(model, optimizer, scheduler, loss_fn,
trainloader)
end = time.time()
print("Epoch: {} | Train Loss: {:.5f} | Time Elapsed: {:.2f} |".
format(epoch, train_loss, end - start))
if ((X_valid is not None) and (y_valid is not None)):
if filter_p:
valid_loss = filtered_valid_BCE(model, validloader)
else:
valid_loss, _ = valid_fn(model, model.valid_fn, validloader)
print("Epoch: {} | Test Score: {:.5f} |".format(
epoch, valid_loss))
if ((X_valid is None) or (y_valid is None)):
torch.save(model.state_dict(), f"./kaggle/working/AllDataset.pth")
return model
lambda size: (BORDER, BORDER))
dirty.append(label_rect)
# Render the widgets.
dirty += render_widgets(surface, widgets, surf_w, surf_h, index, \
label_rect)
return render_frame, len(frame_map)
if __name__ == "__main__":
import os.path
localpath = os.path.dirname(__file__)
# Create a Models.
small = Model(os.path.join(localpath, "gis/SmallPatches"), \
os.path.join(localpath, "csv/small"))
# Create the values. We also include the transformation, for later use.
values = [(Values(small, "SWTotal"), "None"),
(Values(small, "NO3Total"), "None")]
# Create the graphs.
graphs = [Graph([Graphable(Values(small, "SWTotal"), \
"SWTotal (min, mean, max)", \
statistics = ['min', 'mean', 'max'])]), \
Graph([Graphable(Values(small, "NO3Total", \
transforms = [lambda v: patch_filter(v, patch_set)]), \
"Field #{}".format(i), statistics = ['mean']) \
for i, patch_set in small.get_patch_fields().items()])
]
# Create the description strings...
descriptions = []
for value, transform in values:
def addParamsClick(self):
self.model.listOfParams.addToList(self.paramsEdt.text())
# print(self.model.listoffeatures.list)
l = len(self.model.listOfParams.list) - 1
self.paramsListWidget.insertItem(l, self.model.listOfParams.list[-1])
print(self.model.listOfParams.list)
def findParamsClick(self):
self.model.set.findParams(self.model.listOfParams.list)
for i, a in enumerate(self.model.set.listofResults):
self.findParamsListWidget.insertItem(i, str(a))
def resetClick(self):
self.model.resetmodel()
self.clearLayout(self.tab_layout)
def clearLayout(self, layout):
self.featuresListWidget.clear()
self.setWidget.clear()
self.paramsListWidget.clear()
self.findParamsListWidget.clear()
if __name__ == '__main__':
app = QApplication(sys.argv)
model = Model()
window = StartWindow(model)
window.show()
app.exit(app.exec_())
def MNISTNet1():
print("BESTNET")
conv1_params={
'kernel_h': 3,
'kernel_w': 3,
'pad': 0,
'stride': 1,
'in_channel': 1,
'out_channel': 32
}
conv2_params={
'kernel_h': 3,
'kernel_w': 3,
'pad': 0,
'stride': 1,
'in_channel': 32,
'out_channel': 32
}
pool1_params={
'pool_type': 'max',
'pool_height': 2,
'pool_width': 2,
'stride': 2,
'pad': 0
}
pool2_params={
'pool_type': 'max',
'pool_height': 3,
'pool_width': 3,
'stride': 2,
'pad': 0
}
model = Model()
# fan_in = in_channel * kernel_h * kernel_w
# fan_out = num_filters * kernel_h * kernel_w / Pooling_layer_area
model.add(Convolution(conv1_params, name='conv1', initializer=Guassian(std=0.001)))
model.add(ReLU(name='relu1'))
model.add(Pooling(pool1_params, name='pooling1'))
model.add(Convolution(conv2_params, name='conv2', initializer=Guassian(std=0.001)))
model.add(ReLU(name='relu2'))
model.add(Pooling(pool2_params, name='pooling2'))
model.add(Dropout(ratio=0.25, name='dropout1'))
model.add(Flatten(name='flatten'))
model.add(FCLayer(800, 256, name='fclayer1', initializer=Guassian(std=0.01)))
model.add(ReLU(name='relu3'))
model.add(FCLayer(256, 10, name='fclayer2', initializer=Guassian(std=0.01)))
return model
def MNISTNet():
conv1_params={
'kernel_h': 3,
'kernel_w': 3,
'pad': 0,
'stride': 1,
'in_channel': 1,
'out_channel': 6
}
conv2_params={
'kernel_h': 3,
'kernel_w': 3,
'pad': 0,
'stride': 1,
'in_channel': 6,
'out_channel': 16
}
pool1_params={
'pool_type': 'max',
'pool_height': 2,
'pool_width': 2,
'stride': 2,
'pad': 0
}
pool2_params={
'pool_type': 'max',
'pool_height': 3,
'pool_width': 3,
'stride': 2,
'pad': 0
}
model = Model()
model.add(Convolution(conv1_params, name='conv1', initializer=Guassian(std=0.001)))
model.add(ReLU(name='relu1'))
model.add(Pooling(pool1_params, name='pooling1'))
model.add(Convolution(conv2_params, name='conv2', initializer=Guassian(std=0.001)))
model.add(ReLU(name='relu2'))
model.add(Pooling(pool2_params, name='pooling2'))
# model.add(Dropout(ratio=0.25, name='dropout1'))
model.add(Flatten(name='flatten'))
model.add(FCLayer(400, 256, name='fclayer1', initializer=Guassian(std=0.01)))
model.add(ReLU(name='relu3'))
# model.add(Dropout(ratio=0.5))
model.add(FCLayer(256, 10, name='fclayer2', initializer=Guassian(std=0.01)))
return model
import os
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import tensorflow as tf
import os
import shutil
from keras.backend.tensorflow_backend import set_session
from config import Config
from models import Model
import matplotlib
matplotlib.use('Agg')
if os.path.isdir("output/"):
shutil.rmtree("output/")
os.makedirs("output/")
config = Config(epochs = 30, gpu = "1", sampling_size_train = 40, sampling_size_val = 40, batch_size = 1 ,lr = 1e-4, val_size = 0.25)
session_config = tf.ConfigProto()
session_config.gpu_options.visible_device_list = config.gpu
session_config.gpu_options.allow_growth = True
set_session(tf.Session(config=session_config))
model = Model(config)
model.train_predict()
#y_scores, y_preds = model.train_predict()
#model.get_metrics(y_scores, y_preds)
#model.plot_ROCs(y_scores)
#model.plot_PRs(y_scores)
def dqnmain(project_name,
do_boltzman_exploration=False,
test=False,
chkpt=None,
hypeparams=hyperparameter_defaults,
steps=1000,
device='cuda'):
image_arr = []
if (not test):
wdbrun = wandb.init(project=project_name,
config=hypeparams,
name=hypeparams['run_name'],
reinit=True,
monitor_gym=False)
# run.save("*.pth")
config = wdbrun.config
max_reward = config.max_reward
max_steps = config.max_steps
memory_size = config.memory_size
min_rb_size = config.min_rb_size
sample_size = config.sample_size
env_steps_before_train = config.env_steps_before_train
tgt_model_update = config.tgt_model_update
reward_scaler = config.reward_scaler
eps_min = config.eps_min
eps_decay = config.eps_decay
gamma = config.gamma
learning_rate = config.learning_rate
else:
max_reward = hypeparams['max_reward']
max_steps = steps
memory_size = hypeparams['memory_size']
min_rb_size = hypeparams['min_rb_size']
sample_size = hypeparams['sample_size']
env_steps_before_train = hypeparams['env_steps_before_train']
tgt_model_update = hypeparams['tgt_model_update']
reward_scaler = hypeparams['reward_scaler']
eps_min = hypeparams['eps_min']
eps_decay = hypeparams['eps_decay']
gamma = hypeparams['gamma']
learning_rate = hypeparams['learning_rate']
env = gym.make(hypeparams['env_name'])
if hypeparams['env_name'] == 'Breakout-v0':
#TODO
env = FrameStackingAndResizingEnv(env, 84, 84,
4) # change stack size here
env._max_episode_steps = 4000
test_env = gym.make(hypeparams['env_name'])
if hypeparams['env_name'] == 'Breakout-v0':
#TODO
test_env = FrameStackingAndResizingEnv(test_env, 84, 84,
4) # change stack size here
test_env._max_episode_steps = 4000
last_observation = env.reset()
if hypeparams['env_name'] == 'Breakout-v0':
m = ConvModel(env.observation_space.shape, env.action_space.n,
learning_rate).to(device)
else:
m = Model(env.observation_space.shape, env.action_space.n,
learning_rate).to(device)
if chkpt is not None:
m.load_state_dict(torch.load(chkpt))
if hypeparams['env_name'] == 'Breakout-v0':
tgt = ConvModel(env.observation_space.shape,
env.action_space.n).to(device)
else:
tgt = Model(env.observation_space.shape, env.action_space.n).to(
device) # target model, gets update fewer times
update_tgt_model(m, tgt)
rb = ReplayBuffer(memory_size)
steps_since_train = 0
epochs_since_tgt = 0
step_num = -1 * min_rb_size
i = 0
episode_rewards = []
rolling_reward = 0
solved = False
try:
while (not solved) and step_num < max_steps:
if test:
screen = env.render('rgb_array')
image_arr.append(screen)
eps = 0
else:
eps = eps_decay**(step_num)
if do_boltzman_exploration:
if hypeparams['env_name'] == 'Breakout-v0':
logits = m(
torch.Tensor(last_observation).unsqueeze(0).to(
device))[0]
action = torch.distributions.Categorical(
logits=logits).sample().item()
else:
logits = m(torch.Tensor(last_observation).to(device))[0]
action = torch.distributions.Categorical(
logits=logits).sample().item()
else:
if random.random() < eps:
action = env.action_space.sample()
else:
if hypeparams['env_name'] == 'Breakout-v0':
action = m(
torch.Tensor(last_observation).unsqueeze(0).to(
device)).max(-1)[-1].item()
else:
action = m(torch.Tensor(last_observation).to(
device)).max(-1)[-1].item()
observation, reward, done, info = env.step(action)
rolling_reward += reward
reward = reward / reward_scaler
rb.insert(SARS(last_observation, action, reward, done,
observation))
last_observation = observation
if done:
episode_rewards.append(rolling_reward)
if test:
print(rolling_reward)
rolling_reward = 0
observation = env.reset()
steps_since_train += 1
i += 1
step_num += 1
if (
not test
) and rb.idx > min_rb_size and steps_since_train > env_steps_before_train:
loss = train_step(m, rb.sample(sample_size), tgt,
env.action_space.n, gamma, device)
ave_reward = np.mean(episode_rewards)
wdbrun.log(
{
'loss': loss.detach().cpu().item(),
'epsilon': eps,
'avg_reward': ave_reward
},
step=step_num)
if ave_reward >= max_reward:
solved = True
episode_rewards = []
epochs_since_tgt += 1
# print(step_num, loss.detach().item())
if epochs_since_tgt > tgt_model_update:
# print('updating target model')
update_tgt_model(m, tgt)
rew, frames = run_test_episode(m, test_env, device)
# frames.shape == (T, H, W, C)
# wandb.log({'test_reward': rew, 'test_video': wandb.Video(frames.transpose(0, 3, 1, 2), str(rew), fps=25, format='mp4')})
wandb.log({'test_reward': rew})
epochs_since_tgt = 0
torch.save(
tgt.state_dict(),
f"{wandb.run.dir}/{hypeparams['run_name']}_{step_num}.pth"
)
steps_since_train = 0
if ave_reward >= max_reward:
solved = True
wandb.join()
env.close()
except KeyboardInterrupt:
sys.exit()
def __init__(self, path):
self.config = Configuration.construct(path)
self.env = Environment(self.config)
self.memory = ReplayMemory(self.config)
self.model = Model(self.config)
self.ep = None
def train(config):
# train_path:train-context.json
args = config.args
train_set = get_dataset(config.train_path,
config.w2i_vocabs,
config,
is_train=True)
dev_set = get_dataset(config.dev_path,
config.w2i_vocabs,
config,
is_train=False)
# X:img,torch.stack;
train_batch = get_dataloader(train_set, args.batch_size, is_train=True)
model = Model(n_emb=args.n_emb,
n_hidden=args.n_hidden,
vocab_size=args.vocab_size,
dropout=args.dropout,
d_ff=args.d_ff,
n_head=args.n_head,
n_block=args.n_block)
if args.restore != '':
model_dict = torch.load(args.restore)
model.load_state_dict(model_dict)
model.to(device)
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=args.lr)
best_score = -1000000
for i in range(args.epoch):
model.train()
report_loss, start_time, n_samples = 0, time.time(), 0
count, total = 0, len(train_set) // args.batch_size + 1
for batch in train_batch:
Y, T = batch
Y = Y.to(device)
T = T.to(device)
optimizer.zero_grad()
loss = model(Y, T)
loss.backward()
optimizer.step()
report_loss += loss.item()
#break
n_samples += len(Y.data)
count += 1
if count % args.report == 0 or count == total:
print('%d/%d, epoch: %d, report_loss: %.3f, time: %.2f' %
(count, total, i + 1, report_loss / n_samples,
time.time() - start_time))
score = eval(model, dev_set, args.batch_size)
model.train()
if score > best_score:
best_score = score
save_model(os.path.join(args.dir, 'best_checkpoint.pt'),
model)
else:
save_model(os.path.join(args.dir, 'checkpoint.pt'), model)
report_loss, start_time, n_samples = 0, time.time(), 0
return model
import sys
sys.path.append(".")
import cPickle
import numpy as np
from models import Model
from utils import weighted_precision_recall, read_training_data
if __name__ == "__main__":
csv_path = sys.argv[1]
image_prefix = sys.argv[2]
model_path = sys.argv[3]
(X, Y, categories, sample_weights) = read_training_data(csv_path, image_prefix = image_prefix, category = True, sample_weight = True)
print "#T = %d, #F = %d, #N = %d"%(sum(1 for y in Y if y == 0), sum(1 for y in Y if y == 1), sum(1 for y in Y if y == 2))
print "X =", X.shape, ", Y =",Y.shape
# Machine Learning models
model = Model()
#fit data
model.fit(X, Y, sample_weight = sample_weights)
#make prediction
predY = model.predict(X)
#evaluation
scores = weighted_precision_recall(Y, predY, sample_weight = sample_weights)
print "IN SAMPLE -- (Precision/Recall) T: (%.3f,%.3f) / F: (%.3f,%.3f) / N:(%.3f,%.3f)" %(scores[0], scores[1], scores[2], scores[3], scores[4], scores[5])
#save model
with open(model_path, "wb") as f:
cPickle.dump(model, f, cPickle.HIGHEST_PROTOCOL)
vocab = build_vocab(train_dataset + valid_dataset)
train_dataset.index_with(vocab)
valid_dataset.index_with(vocab)
data_loader_params = params.pop('data_loader')
batch_size = data_loader_params['batch_size']
train_loader = DataLoader.from_params(dataset=train_dataset,
params=data_loader_params)
valid_loader = DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=False)
# 2. setting up model and training details
# model = build_model(vocab)
model = Model.from_params(vocab=vocab, params=params["model"])
model.cuda()
trainer = Trainer.from_params(
model=model,
serialization_dir=serialization_dir,
data_loader=train_loader,
validation_data_loader=valid_loader,
params=params['trainer'],
)
trainer.optimizer.weight_decay = 0.00001
# 3. perform training with early stopping
print("Starting training")
trainer.train()
print("Finished training")
seed = 0
if opt.seed == 0:
seed = random.randint(1, 10000)
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(args.seed)
else:
seed = opt.seed
torch.manual_seed(opt.seed)
torch.cuda.manual_seed(opt.seed)
torch.backends.cudnn.benchmark = True
dictionary = Dictionary.load_from_file('data/dictionary.pkl')
opt.ntokens = dictionary.ntoken
model = Model(opt)
model.apply(weights_init_kn)
model = nn.DataParallel(model).cuda()
train_dset = SelfCriticalDataset(opt.split, dictionary, opt)
train_loader = DataLoader(train_dset,
opt.batch_size,
shuffle=True,
num_workers=0)
opt.use_all = 1
eval_dset = SelfCriticalDataset(opt.split_test, dictionary, opt)
eval_loader = DataLoader(eval_dset,
opt.batch_size,
shuffle=False,
num_workers=0)
args = gen_args()
set_seed(args.random_seed)
os.system('mkdir -p ' + args.evalf)
os.system('mkdir -p ' + os.path.join(args.evalf, 'render'))
tee = Tee(os.path.join(args.evalf, 'eval.log'), 'w')
### evaluating
data_names = args.data_names
use_gpu = torch.cuda.is_available()
# create model and load weights
model = Model(args, use_gpu)
print("model_kp #params: %d" % count_parameters(model))
if args.eval_epoch < 0:
model_name = 'net_best.pth'
else:
model_name = 'net_epoch_%d_iter_%d.pth' % (args.eval_epoch, args.eval_iter)
model_path = os.path.join(args.outf, model_name)
print("Loading network from %s" % model_path)
if args.stage == 'dy':
pretrained_dict = torch.load(model_path)
model_dict = model.state_dict()
# only load parameters in dynamics_predictor
pretrained_dict = {
class Base:
def __init__(self, config, loaders):
self.config = config
self.loaders = loaders
# Model Configuration
self.part_num = config.part_num
self.pid_num = config.pid_num
# Logger Configuration
self.max_save_model_num = config.max_save_model_num
self.output_path = config.output_path
self.model_path = os.path.join(self.output_path, 'models/')
self.log_path = os.path.join(self.output_path, 'logs/')
# Train Configuration
self.base_learning_rate = config.base_learning_rate
self.milestones = config.milestones
# init model
self._init_device()
self._init_model()
self._init_creiteron()
self._init_optimizer()
def _init_device(self):
self.device = torch.device('cuda')
def _init_model(self):
self.model = Model(part_num=self.part_num, class_num=self.config.pid_num)
self.model = nn.DataParallel(self.model).to(self.device)
def _init_creiteron(self):
self.ide_creiteron = nn.CrossEntropyLoss()
## compute average ide loss of all outputs
def compute_ide_loss(self, logits_list, pids):
avg_ide_loss = 0
avg_logits = 0
for i in xrange(self.part_num):
logits_i = logits_list[i]
avg_logits += 1.0 / float(self.part_num) * logits_i
ide_loss_i = self.ide_creiteron(logits_i, pids)
avg_ide_loss += 1.0 / float(self.part_num) * ide_loss_i
return avg_ide_loss, avg_logits
## init optimizer and lr_lr_scheduler
def _init_optimizer(self):
params = [{'params': self.model.module.resnet_conv.parameters(), 'lr': 0.1*self.base_learning_rate}]
for i in xrange(self.part_num):
params.append({'params': getattr(self.model.module, 'classifier' + str(i)).parameters(), 'lr': self.base_learning_rate})
for i in xrange(self.part_num):
params.append({'params': getattr(self.model.module, 'embedder' + str(i)).parameters(), 'lr': self.base_learning_rate})
self.optimizer = optim.SGD(params=params, weight_decay=5e-4, momentum=0.9, nesterov=True)
self.lr_scheduler = optim.lr_scheduler.MultiStepLR(self.optimizer, self.milestones, gamma=0.1)
## save model as save_epoch
def save_model(self, save_epoch):
# save model
file_path = os.path.join(self.model_path, 'model_{}.pkl'.format(save_epoch))
torch.save(self.model.state_dict(), file_path)
# if saved model is more than max num, delete the model with smallest iter
if self.max_save_model_num > 0:
root, _, files = os_walk(self.model_path)
if len(files) > self.max_save_model_num:
file_iters = sorted([int(file.replace('.pkl', '').split('_')[1]) for file in files], reverse=False)
file_path = os.path.join(root, 'model_{}.pkl'.format(file_iters[0]))
os.remove(file_path)
## resume model from resume_epoch
def resume_model(self, resume_epoch):
model_path = os.path.join(self.model_path, 'model_{}.pkl'.format(resume_epoch))
self.model.load_state_dict(torch.load(model_path))
print('successfully resume model from {}'.format(model_path))
## set model as train mode
def set_train(self):
self.model = self.model.train()
## set model as eval mode
def set_eval(self):
self.model = self.model.eval()
def train(dataset='A'):
# 训练数据集路径
img_root_dir = './data/ShanghaiTech/part_' + dataset + '_final/train_data/images/'
gt_root_dir = './data/ShanghaiTech/part_' + dataset + '_final/train_data/ground_truth/'
# 测试数据集路径
test_img_dir = './data/ShanghaiTech/part_' + dataset + '_final/test_data/images/'
test_gt_dir = './data/ShanghaiTech/part_' + dataset + '_final/test_data/ground_truth/'
img_file_list = os.listdir(img_root_dir)
# gt_file_list = os.listdir(gt_root_dir)
test_img_list = os.listdir(test_img_dir)
# test_gt_list = os.listdir(test_gt_dir)
model_name = 'mcdcnn_SHTech_' + dataset
cfg = Config(model_name)
cfg.lr = 1e-4
tool = Tool()
model = Model()
network = model.mcdcnn()
INPUT, GT_DMP, GT_CNT, EST_DMP, EST_CNT = network['INPUT'], network['GT_DMP'], network['GT_CNT'], \
network['EST_DMP'], network['EST_CNT']
loss = model.losses(GT_DMP, GT_CNT, EST_DMP, EST_CNT)
# 学习率衰变设置
global_step = tf.Variable(0, trainable=False)
lr = tf.train.exponential_decay(cfg.lr,
global_step=global_step,
decay_steps=cfg.lr_decay_step,
decay_rate=cfg.lr_decay_rate,
staircase=True)
# 优化器
optimizer = tf.train.AdamOptimizer(lr).minimize(loss, global_step)
# 模型保存设置
saver = tf.train.Saver(max_to_keep=cfg.max_ckpt_keep)
ckpt = tf.train.get_checkpoint_state(cfg.ckpt_router)
# 创建会话
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# 加载模型
if ckpt and ckpt.model_checkpoint_path:
print('load model')
saver.restore(sess, ckpt.model_checkpoint_path)
# 训练日志文件路径设置
if not os.path.exists(cfg.log_router):
os.makedirs(cfg.log_router)
log = open(cfg.log_router + 'train' + r'.logs',
mode='a+',
encoding='utf-8')
# 迭代训练
num = len(img_file_list)
# 开始训练
for i in tqdm(range(cfg.iter_num)):
# 随机打乱
np.random.shuffle(img_file_list)
for j in tqdm(range(num)):
img_path = img_root_dir + img_file_list[j]
gt_path = gt_root_dir + 'GT_' + img_file_list[j].split(r'.')[0]
img, dmp, cnt = tool.read_train_data(img_path,
gt_path,
use_knn=True)
feed_dict = {INPUT: img, GT_DMP: dmp, GT_CNT: cnt}
_, est_dmp, est_cnt, train_loss = sess.run(
[optimizer, EST_DMP, EST_CNT, loss], feed_dict=feed_dict)
format_time = str(
time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
format_str = 'step %d, loss = %.3f, gt = %d, inference = %.3f'
log_line = format_time, img_file_list[j], format_str % (
(i * num + j), train_loss, cnt, est_cnt)
print(log_line)
log.writelines(str(log_line) + '\n')
sess.run(lr, feed_dict={global_step: i * num + j})
if i % cfg.snap == 0:
# 保存模型
saver.save(sess, cfg.ckpt_router + '/v1', global_step=i)
# 进行测试
print('testing', i, '> > > > > > > > > > > > > > > > > > >')
total_mae = 0.0
total_mse = 0.0
num = len(test_img_list)
log_line = ''
for k in tqdm(range(num - 2, num)):
img_path = test_img_dir + test_img_list[k]
gt_path = test_gt_dir + 'GT_' + test_img_list[k].split(
r'.')[0]
img, dmp, cnt = tool.read_test_data(img_path,
gt_path,
use_knn=True)
feed_dict = {INPUT: img, GT_DMP: dmp, GT_CNT: cnt}
est_cnt, test_loss = sess.run([EST_CNT, loss],
feed_dict=feed_dict)
format_time = str(
time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
format_str = 'step %d, joint loss = %.3f, gt = %d, inference = %.3f'
line = str(format_time + ' ' + test_img_list[k] + ' ' +
format_str % (k, test_loss, cnt, est_cnt) +
'\n')
log_line += line
total_mae += tool.mae_metrix(cnt, est_cnt)
total_mse += tool.mse_metrix(cnt, est_cnt)
print(line)
avg_mae = total_mae / num
avg_mse = pow(total_mse / num, 0.5)
result = str('_MAE_%.5f_MSE_%.5f' % (avg_mae, avg_mse))
test_log = open(cfg.log_router + 'test_' + str(i) + result +
r'.logs',
mode='a+',
encoding='utf-8')
test_log.writelines(result + '\n')
test_log.writelines(str(log_line) + '\n')
test_log.close()
log.close()
# defining hyperparameters
NUM_EPOCHS = 1
LEARNING_RATE = 0.01
GRADIENT_NORM = 5
EMBEDDING_DIM = 32
TOP_K = 5
HIDDEN_DIM = 32
BATCH_SIZE = 16
CHUNK_SIZE = 32
TRAIN_PATH = "data/french.txt"
# instantiating dataset, model, loss function, and optimizer
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dataset = CorpusDataset(TRAIN_PATH, CHUNK_SIZE, BATCH_SIZE)
model = Model(EMBEDDING_DIM, HIDDEN_DIM, len(dataset.vocabulary), device)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE)
iteration = 0
for i in range(NUM_EPOCHS):
hidden_state, cell_state = model.init_hidden(BATCH_SIZE)
for text, target in dataset:
model.train()
optimizer.zero_grad()
output, (hidden_state, cell_state) = model(text,
(hidden_state, cell_state))
loss = criterion(output.transpose(1, 2), target).to(device)
def eval_robustness(ARGS, verbose=True):
#############################################
# Load pre-trained model
#############################################
if verbose:
print('\n- Loading pre-trained model...')
# Build evaluation graph
eval_graph = tf.Graph()
config = tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(graph=eval_graph, config=config)
# Define input TF placeholder
with eval_graph.as_default():
with tf.device('/gpu:0'):
# Define placeholders
with tf.name_scope('Placeholders'):
x = tf.placeholder(dtype=tf.float32,
shape=input_shape,
name='inputs')
y = tf.placeholder(dtype=tf.float32,
shape=(None, n_classes),
name='labels')
is_training = tf.placeholder_with_default(False,
shape=(),
name='is-training')
# Define model
with tf.name_scope('Model'):
model = Model(nb_classes=n_classes,
input_shape=input_shape,
is_training=is_training)
# Define forward-pass
with tf.name_scope('Logits'):
logits = model.get_logits(x)
with tf.name_scope('Probs'):
preds = tf.nn.softmax(logits)
# Restore the pre-trained model
with sess.as_default():
saver = tf.train.Saver()
saver.restore(sess, ARGS.restore_path + '/model.ckpt')
# Define accuracy ops
with tf.name_scope('Accuracy'):
ground_truth = tf.argmax(y, axis=1)
predicted_label = tf.argmax(preds, axis=1)
correct_prediction = tf.equal(predicted_label, ground_truth)
clean_acc = tf.reduce_mean(tf.to_float(correct_prediction),
name='accuracy')
# Define PGD adversary
if ARGS.attack == 'PGD':
if verbose:
print('\n- Building {:s} attack graph...'.format(
ARGS.attack))
with tf.name_scope('PGD-Attacker'):
pgd_params = {
'ord': np.inf,
'y': y,
'eps': ARGS.eps / 255,
'eps_iter': ARGS.eps_iter / 255,
'nb_iter': ARGS.nb_iter,
'rand_init': ARGS.rand_init,
'rand_minmax': ARGS.eps / 255,
'clip_min': 0.,
'clip_max': 1.,
'sanity_checks': True
}
pgd = ProjectedGradientDescent(model, sess=None)
adv_x = pgd.generate(x, **pgd_params)
# Define SPSA adversary
elif ARGS.attack == 'SPSA':
if verbose:
print('\n- Building {:s} attack graph...'.format(
ARGS.attack))
with tf.name_scope('PGD-Attacker'):
spsa_params = {
'y': y,
'eps': ARGS.eps / 255,
'nb_iter': ARGS.nb_iter,
'spsa_samples': ARGS.spsa_samples,
'spsa_iters': ARGS.spsa_iters,
'clip_min': 0.,
'clip_max': 1.,
'learning_rate': ARGS.spsa_lr,
'delta': ARGS.spsa_delta
}
spsa = SPSA(model, sess=sess)
adv_x = spsa.generate(x, **spsa_params)
else:
raise NotImplementedError
with tf.name_scope('Logits'):
adv_logits = model.get_logits(adv_x)
with tf.name_scope('Probs'):
adv_preds = tf.nn.softmax(adv_logits)
adv_loss = tf.nn.softmax_cross_entropy_with_logits(
logits=adv_logits, labels=y)
adv_predicted_label = tf.argmax(adv_preds, axis=1)
correct_prediction = tf.equal(adv_predicted_label, ground_truth)
adv_accuracy = tf.reduce_mean(tf.to_float(correct_prediction),
name='adv-accuracy')
is_adv_example = tf.not_equal(ground_truth, adv_predicted_label)
#############################################
# Run evaluation
#############################################
if verbose:
print('\n- Running robustness evaluation against {:s} attacker...\n'.
format(ARGS.attack))
if ARGS.attack == 'PGD':
clean, adv_mean, adv_worstcase = run_pgd_eval(x,
y,
is_training,
sess,
adv_testloader,
clean_acc,
adv_accuracy,
adv_loss,
is_adv_example,
ARGS,
save_loss_dist=False,
verbose=verbose)
elif ARGS.attack == 'SPSA':
clean, adv_mean = run_spsa_eval(x,
y,
is_training,
sess,
adv_testloader,
clean_acc,
adv_accuracy,
adv_loss,
is_adv_example,
ARGS,
save_loss_dist=False,
verbose=verbose)
adv_worstcase = adv_mean
else:
raise NotImplementedError
return clean, adv_mean, adv_worstcase
def __init__(self, access=None, **kwargs):
Model.__init__(self)
Base.__init__(self)
self.access = access
USE_MQTT = False # To be able to test without MQTT
DEFAULT_CONFIDENCE = 0.5
DEFAULT_LOGFILE = "logs/infer_and_publish.log"
DEFAULT_LOGLEVEL = "DEBUG"
DEFAULT_DEVICE = "MYRIAD" #CPU
if DEFAULT_DEVICE == "CPU":
DEFAULT_PREC = 32
else:
DEFAULT_PREC = 16
DEFAULT_INPUT = 'resources/Pedestrian_Detect_2_1_1.mp4'
isImage = False
allow_print = False #if --dev: print() else log.debug()
model = Model().get_default()
ini_time = time.perf_counter()
def build_argparser():
"""
Parse command line arguments.
:return: command line arguments
"""
parser = ArgumentParser()
parser.add_argument("-m",
"--model",
required=False,
type=str,
default=model['path'],
def populate(self): new_model = Model() return new_model.save()
batch_inputs = torch.zeros((batch_size, max_doc_len, max_sent_len),
dtype=torch.int64)
batch_masks = torch.zeros((batch_size, max_doc_len, max_sent_len),
dtype=torch.int64)
batch_labels = torch.LongTensor(doc_labels)
for b in range(batch_size):
for sent_idx in range(doc_lens[b]):
sent_data = batch_data[b][2][sent_idx] # 表示一个句子
for word_idx in range(sent_data[0]):
batch_inputs[b, sent_idx,
word_idx] = sent_data[1][word_idx]
batch_masks[b, sent_idx, word_idx] = 1
if use_cuda:
batch_inputs = batch_inputs.to(device)
batch_masks = batch_masks.to(device)
batch_labels = batch_labels.to(device)
return (batch_inputs, batch_masks), batch_labels
if __name__ == '__main__':
train_data, dev_data, test_data = get_data(fold_num, dev_fold)
vocab = Vocab(train_data)
model = Model(vocab)
trainer = Trainer(model, vocab, train_data, dev_data, test_data)
trainer.train()
print(vocab._label2id)
class Pipeline():
def __init__( self, feature_engineer=None, model=None ):
# configure logging
self.logger = logging.getLogger("Pipeline")
handler = logging.FileHandler(settings.LOG_PATH)
handler.setFormatter(logging.Formatter(
'%(asctime)s %(levelname)s %(name)s: %(message)s'))
self.logger.addHandler(handler)
self.logger.setLevel(logging.DEBUG)
self.settings = Settings()
self.feature_engineer = FeatureEngineer( self.settings )
self.model = Model().get_model( Settings() )
def read( self, file_path ):
self.settings.read( file_path )
self.read_feature_engineer( settings.feature_engineer_path )
self.read_model( settings.model_path )
def read_feature_engineer( self, file_path ):
print('implement this')
def read_model( self, file_path ):
self.model = joblib.load( file_path+'_main.pkl' )
self.model.load_model(file_path)
def write( self, file_path ):
self.settings.write( file_path )
self.write_feature_engineer( settings.feature_engineer_path )
self.write_model( settings.model_path )
def write_feature_engineer( self, file_path ):
print('implement this')
def write_model( self, file_path ):
joblib.dump( self.model, file_path+'_main.pkl' )
self.model.store_model(file_path)
def set_feature_engineer( self, feature_engineer ):
self.feature_engineer = feature_engineer
def set_model( self, model ):
self.model = model
def train( self ):
# Create a "local" train and testset
initial_preparation.create_test_and_train_set()
# Feature engineer ( Transform features and create new ones )
self.logger.info('Start Feature Engineering')
train_data = self.feature_engineer.engineer_features(settings.LOKAL_TRAIN, settings.PROCESSED_TRAIN)
self.feature_engineer.train_state = False
test_data = self.feature_engineer.engineer_features (settings.LOKAL_TEST, settings.PROCESSED_TEST )
# Train model and evaluate
self.logger.info('Train and Evaluate Model')
score = self.model.train_model()
self.model.store_model()
self.logger.info('Model Score is %f', score)
def predict( self ):
# Feature engineer ( Transform features and create new ones )
self.logger.info('Start Feature Engineering')
self.feature_engineer.train_state = False
test_data = self.feature_engineer.engineer_features( settings.GLOBAL_TEST, settings.PROCESSED_GLOBAL )
#test_data.drop( settings.ID_FIELD, inplace=True, axis=1 )
# Train model and evaluate
self.logger.info('Train and Evaluate Model')
self.model.load_model()
ids, prediction = self.model.predict_submission()
submission_writer.create_submission( ids, prediction )
#
if __name__ == '__main__':
n_burn_in = 10
n_samples = 100
pot = Ring_Potential(scale=5, bias=-1)
x0 = np.asarray([[0.], [0.]])
rng = np.random.RandomState()
model = Model(x0,
rng=rng,
pot=pot,
n_steps=50,
step_size=0.1,
verbose=True,
rand_steps=True)
# adjust for nice plotting
print "Running Model: {}".format(__file__)
model.run(n_samples=n_samples,
n_burn_in=n_burn_in,
mixing_angle=np.pi / 6.)
print "Finished Running Model: {}".format(__file__)
# change shape from (n, 2) -> (2, n)
samples = model.samples
burn_in = model.burn_in
def __init__(self, dbpath): Model.__init__(self, dbpath, "proxy_users")
def main(opts):
if not opts.do_train and not opts.do_eval:
raise ValueError(
'At least one of `do_train` or `do_eval` must be True.')
if os.path.exists(opts.output_dir) and os.listdir(
opts.output_dir) and opts.do_train:
raise ValueError(
'Output directory ({}) already exists and is not empty.'.format(
opts.output_dir))
# Create the output directory (if not exists)
create_dir_if_not_exists(opts.output_dir)
# Device device type
opts.device = torch.device(
'cuda' if torch.cuda.is_available() and not opts.no_cuda else 'cpu')
opts.n_gpus = torch.cuda.device_count() if str(
opts.device) == 'cuda' else 0
print('Device Type: {} | Number of GPUs: {}'.format(
opts.device, opts.n_gpus),
flush=True)
# Load Datasets and Ontology
if opts.language == 'cn':
dataset, ontology = load_dataset_cn(opts.data_dir)
else:
dataset, ontology = load_dataset(opts.data_dir)
print('Loaded Datasets and Ontology', flush=True)
print('Number of Train Dialogues: {}'.format(len(dataset['train'])),
flush=True)
print('Number of Dev Dialogues: {}'.format(len(dataset['dev'])),
flush=True)
print('Number of Test Dialogues: {}'.format(len(dataset['test'])),
flush=True)
if opts.do_train:
# Load model from scratch
model = Model.from_scratch(opts.bert_model)
model.move_to_device(opts)
print('Number of model parameters is: {}'.format(get_n_params(model)))
# Start Training
print('Start Training', flush=True)
model.run_train(dataset, ontology, opts)
# Free up all memory pytorch is taken from gpu memory
del model
torch.cuda.empty_cache()
if opts.do_eval:
if not (os.path.exists(opts.output_dir)
and os.listdir(opts.output_dir)):
raise ValueError(
'Output directory ({}) is empty. Cannot do evaluation'.format(
opts.output_dir))
# Load trained model
model = Model.from_model_path(opts.output_dir)
model.move_to_device(opts)
print('Number of model parameters is: {}'.format(get_n_params(model)))
# Start evaluating
print('Start Evaluating', flush=True)
print(model.run_dev(dataset, ontology, opts), flush=True)
print(model.run_test(dataset, ontology, opts), flush=True)
def __init__(self, train): Model.__init__(self, train) dayage = train['dayage'].values self.begin = dayage[0] self.end = dayage[-1] self.mean = train['count'].mean()
class Evaluation(object):
def __init__(self, hparams, model=None):
self.hparams = hparams
self.model = model
self._logger = logging.getLogger(__name__)
self.device = (torch.device("cuda", self.hparams.gpu_ids[0]) if
self.hparams.gpu_ids[0] >= 0 else torch.device("cpu"))
self.split = hparams.evaluate_data_type
print("Evaluation Split :", self.split)
do_valid, do_test = False, False
if self.split == "dev":
do_valid = True
else:
do_test = True
self._build_dataloader(do_valid=do_valid, do_test=do_test)
self._dataloader = self.valid_dataloader if self.split == 'dev' else self.test_dataloader
if model is None:
print("No pre-defined model!")
self._build_model()
def _build_dataloader(self, do_valid=False, do_test=False):
if do_valid:
self.valid_dataset = ResponseSelectionDataset(
self.hparams,
split="dev",
)
self.valid_dataloader = DataLoader(
self.valid_dataset,
batch_size=self.hparams.eval_batch_size,
num_workers=self.hparams.cpu_workers,
drop_last=False,
)
if do_test:
self.test_dataset = ResponseSelectionDataset(
self.hparams,
split="test",
)
self.test_dataloader = DataLoader(
self.test_dataset,
batch_size=self.hparams.eval_batch_size,
num_workers=self.hparams.cpu_workers,
drop_last=False,
)
def _build_model(self):
self.model = Model(self.hparams)
self.model = self.model.to(self.device)
# Use Multi-GPUs
if -1 not in self.hparams.gpu_ids and len(self.hparams.gpu_ids) > 1:
self.model = nn.DataParallel(self.model, self.hparams.gpu_ids)
def run_evaluate(self, evaluation_path):
self._logger.info("Evaluation")
model_state_dict, optimizer_state_dict = load_checkpoint(
evaluation_path)
print(evaluation_path)
if isinstance(self.model, nn.DataParallel):
self.model.module.load_state_dict(model_state_dict)
else:
self.model.load_state_dict(model_state_dict)
k_list = self.hparams.recall_k_list
total_mrr, total_prec_at_one, total_map = 0, 0, 0
total_examples, total_correct = 0, 0
self.model.eval()
with torch.no_grad():
for batch_idx, batch in enumerate(tqdm(self._dataloader)):
buffer_batch = batch.copy()
for task_key in batch:
for key in buffer_batch[task_key]:
buffer_batch[task_key][key] = buffer_batch[task_key][
key].to(self.device)
# for key in buffer_batch["res_sel"]:
# buffer_batch["res_sel"][key] = buffer_batch["res_sel"][key].to(self.device)
logits, loss = self.model(buffer_batch)
pred = torch.sigmoid(logits).to("cpu").tolist()
rank_by_pred, pos_index, stack_scores = \
calculate_candidates_ranking(np.array(pred), np.array(buffer_batch["res_sel"]["label"].to("cpu").tolist()),
self.hparams.evaluate_candidates_num)
num_correct = logits_recall_at_k(pos_index, k_list)
if self.hparams.task_name in ["douban", "kakao"]:
total_prec_at_one += precision_at_one(rank_by_pred)
total_map += mean_average_precision(pos_index)
for pred in rank_by_pred:
if sum(pred) == 0:
total_examples -= 1
total_mrr += logits_mrr(pos_index)
total_correct = np.add(total_correct, num_correct)
total_examples += math.ceil(
buffer_batch["res_sel"]["label"].size()[0] /
self.hparams.evaluate_candidates_num)
recall_result = ""
if (batch_idx + 1) % self.hparams.evaluate_print_step == 0:
for i in range(len(k_list)):
recall_result += "Recall@%s : " % k_list[
i] + "%.2f%% | " % (
(total_correct[i] / total_examples) * 100)
else:
print(
"%d[th] | %s | MRR : %.3f | P@1 : %.3f | MAP : %.3f"
% (batch_idx + 1, recall_result,
float(total_mrr / total_examples),
float(total_prec_at_one / total_examples),
float(total_map / total_examples)))
self._logger.info(
"%d[th] | %s | MRR : %.3f | P@1 : %.3f | MAP : %.3f" %
(batch_idx + 1, recall_result,
float(total_mrr / total_examples),
float(total_prec_at_one / total_examples),
float(total_map / total_examples)))
avg_mrr = float(total_mrr / total_examples)
avg_prec_at_one = float(total_prec_at_one / total_examples)
avg_map = float(total_map / total_examples)
recall_result = ""
for i in range(len(k_list)):
recall_result += "Recall@%s : " % k_list[i] + "%.2f%% | " % (
(total_correct[i] / total_examples) * 100)
print(recall_result)
print("MRR: %.4f" % avg_mrr)
print("P@1: %.4f" % avg_prec_at_one)
print("MAP: %.4f" % avg_map)
self._logger.info(recall_result)
self._logger.info("MRR: %.4f" % avg_mrr)
self._logger.info("P@1: %.4f" % avg_prec_at_one)
self._logger.info("MAP: %.4f" % avg_map)
class Execute:
def __init__(self, path):
self.config = Configuration.construct(path)
self.env = Environment(self.config)
self.memory = ReplayMemory(self.config)
self.model = Model(self.config)
self.ep = None
def get_epsilon(self, is_play):
if is_play:
return self.config.play.ep
ep_start = self.config.train.ep.start
ep_final = self.config.train.ep.final
ep_num_frames = self.config.train.ep.num_frames
decay = (ep_start - ep_final) / ep_num_frames
if self.ep is None:
self.ep = ep_start
self.ep = max(self.ep - decay, ep_final)
return self.ep
def log(self, **kawrgs):
log = ""
for name, value in kawrgs.items():
log += f"{name}: {value}, "
print(log)
def run_episode(self, episode=1, steps=0, is_play=True, debug=False):
config = self.config
self.env.reset()
action = 1
_, _, curr_state, is_done = self.env.step(action)
total_reward = 0
update_net = 0; C = config.train.network_update_freq
t = 0; T = config.max_episode_length
while not is_done and t < T:
if t % config.action_repeat == 0:
ep = self.get_epsilon(is_play)
action = self.model.choose_action(curr_state, ep)
prev_state, reward, curr_state, is_done = self.env.step(action)
total_reward += reward
t += 1
if is_play:
self.env.render("human")
if debug and t % config.play.debug.time == 0:
self.log(ftype=self.env.get_frame_type(), action=action, reward=total_reward)
continue
self.memory.add((prev_state, action, reward, curr_state, is_done))
if self.memory.get_size() > config.train.replay_start_size:
for i in range(config.train.batch_run):
batch = self.memory.sample()
self.model.optimize(batch)
steps = (steps + 1) % C
if steps % C == 0:
self.model.update_qhat()
update_net += 1
if not is_play and debug and episode % config.train.debug.time == 0:
self.log(ftype=self.env.get_frame_type(), total_reward=total_reward, network_update_steps=update_net, episode_time=t, ep=ep)
return total_reward, steps
def load_model(self):
ftype = self.env.get_frame_type()
in_size = self.env.get_in_size()
num_actions = self.env.get_num_actions()
self.model.load_model(ftype, in_size, num_actions)
def play(self, debug=False):
self.load_model()
for ep in range(1):
self.run_episode(is_play=True, debug=debug)
def train(self, debug=False):
self.load_model()
optimize_steps = 0
episodes = self.config.train.episodes
for episode in range(1, episodes+1):
reward, steps = self.run_episode(episode=episode, steps=optimize_steps, is_play=False, debug=debug)
optimize_steps += steps
if episode % self.config.train.save_model_episode == 0:
self.model.save_model()
self.model.update_qhat()
self.model.save_model()
def close(self):
self.env.close()
self.memory.close()
def _build_model(self):
self.model = Model(self.hparams)
self.model = self.model.to(self.device)
# Use Multi-GPUs
if -1 not in self.hparams.gpu_ids and len(self.hparams.gpu_ids) > 1:
self.model = nn.DataParallel(self.model, self.hparams.gpu_ids)
def __init__(self, dbpath): Model.__init__(self, dbpath, "downloads") self.Status = Status
class Agent:
def __init__(self, action_dims, model_version=Config.model_version,
q_version=Config.q_version,
target_version=Config.target_version,
loss_version=Config.loss_version):
# self.action_num = action_num
self.action_dims = action_dims
self.epsilon = Config.initial_epsilon
self.epsilon_final = Config.epsilon_final
self.epsilon_start = Config.epsilon_start
self.epsilon_decay = Config.epsilon_decay
self.model_version = model_version
self.q_version = q_version
self.target_version = target_version
self.loss_version = loss_version
self.build_network()
def build_network(self):
self.Q_network = Model(self.action_dims, self.model_version, self.loss_version)
self.target_network = Model(self.action_dims, self.model_version, self.loss_version)
# Change learning rate for commen net !!!! Start from here
if self.loss_version == 0:
self.optimizers = optim.Adam(self.Q_network.parameters(), lr=Config.lr)
elif self.loss_version == 1:
if self.model_version == 1:
self.optimizers = [optim.Adam([
{'params': self.Q_network.multi_output_1.parameters(), 'lr':Config.lr},
{'params': self.Q_network.fc2.parameters()},
{'params': self.Q_network.fc1.parameters()},
{'params': self.Q_network.lstm1.parameters()}
], lr=0.5*Config.lr),
optim.Adam([
{'params': self.Q_network.multi_output_2.parameters(), 'lr':Config.lr},
{'params': self.Q_network.fc2.parameters()},
{'params': self.Q_network.fc1.parameters()},
{'params': self.Q_network.lstm1.parameters()}
], lr=0.5*Config.lr)]
elif self.model_version == 2:
self.optimizers = [optim.Adam([
{'params': self.Q_network.multi_output_1.parameters(), 'lr':Config.lr},
{'params': self.Q_network.fc2.parameters()},
{'params': self.Q_network.fc1.parameters()},
{'params': self.Q_network.lstm1.parameters()},
{'params': self.Q_network.dueling.parameters()}
], lr=0.5*Config.lr),
optim.Adam([
{'params': self.Q_network.multi_output_2.parameters(), 'lr':Config.lr},
{'params': self.Q_network.fc2.parameters()},
{'params': self.Q_network.fc1.parameters()},
{'params': self.Q_network.lstm1.parameters()},
{'params': self.Q_network.dueling.parameters()}
], lr=0.5*Config.lr)]
def update_target_network(self):
# copy current_network to target network
self.target_network.load_state_dict(self.Q_network.state_dict())
def update_Q_network_v0(self, state, action, reward, state_new, terminal):
state = torch.from_numpy(state).float()
action = torch.from_numpy(action).float()
state_new = torch.from_numpy(state_new).float()
terminal = torch.from_numpy(terminal).float()
reward = torch.from_numpy(reward).float()
state = Variable(state)
action = Variable(action) # shape (batch, 6*7)
state_new = Variable(state_new)
terminal = Variable(terminal)
reward = Variable(reward)
self.Q_network.eval()
self.target_network.eval()
# use current network to evaluate action argmax_a' Q_current(s', a')_
actions_new = self.Q_network.forward(state_new).max(dim=1)[1].cpu().data.view(-1, 1)
actions_new_onehot = torch.zeros(Config.sampling_batch_size, self.action_dims[0]*self.action_dims[1])
actions_new_onehot = Variable(actions_new_onehot.scatter_(1, actions_new, 1.0))
# Different loss and object
# use target network to evaluate value y = r + discount_factor * Q_tar(s', a')
y = reward + torch.mul(((self.target_network.forward(state_new)*actions_new_onehot).sum(dim=1)*terminal),Config.discount_factor)
self.Q_network.train()
Q = (self.Q_network.forward(state)*action).sum(dim=1)
losses = []
# y = reward + torch.mul(((self.target_network.forward(state_new)[action_idx]*actions_new_onehot[action_idx]).sum(dim=1)*terminal), Config.discount_factor)
# Q = (self.Q_network.forward(state)[action_idx]*actions[action_idx]).sum(dim=1)
loss = mse_loss(input=Q, target=y.detach())
self.optimizers.zero_grad()
loss.backward()
self.optimizers.step()
losses.append(loss.item())
# print(losses)
return losses
def update_Q_network_v1(self, state, action_1, action_2, reward, state_new, terminal):
state = torch.from_numpy(state).float()
action_1 = torch.from_numpy(action_1).float()
action_2 = torch.from_numpy(action_2).float()
state_new = torch.from_numpy(state_new).float()
terminal = torch.from_numpy(terminal).float()
reward = torch.from_numpy(reward).float()
state = Variable(state)
action_1 = Variable(action_1)
action_2 = Variable(action_2)
state_new = Variable(state_new)
terminal = Variable(terminal)
reward = Variable(reward)
self.Q_network.eval()
self.target_network.eval()
# use current network to evaluate action argmax_a' Q_current(s', a')_
new_q_values = self.Q_network.forward(state_new)
actions_new = [torch.max(q_value, 1)[1].cpu().data.view(-1, 1) for q_value in new_q_values]
actions_new_onehot = [torch.zeros(Config.sampling_batch_size, action_dim) for action_dim in self.action_dims]
actions_new_onehot = [Variable(actions_new_onehot[action_idx].scatter_(1, actions_new[action_idx], 1.0)) for action_idx in range(len(self.action_dims))]
# Different loss and object
# use target network to evaluate value y = r + discount_factor * Q_tar(s', a')
actions = [action_1, action_2]
raw_y = self.target_network.forward(state_new)
y = []
for new_q_idx in range(len(raw_y)):
y.append(reward + torch.mul(((raw_y[new_q_idx]*actions_new_onehot[new_q_idx]).sum(dim=1)*terminal),Config.discount_factor))
self.Q_network.train()
raw_Q = self.Q_network.forward(state)
Q = []
for q_idx in range(len(raw_Q)):
Q.append((raw_Q[q_idx]*actions[q_idx]).sum(dim=1))
# Calculate loss
losses = []
if self.loss_version == 0:
loss = 0.0
for action_idx in range(len(self.action_dims)):
loss += mse_loss(input=Q[action_idx], target=y[action_idx].detach())
loss /= len(self.action_dims)
self.optimizers.zero_grad()
loss.backward()
self.optimizers.step()
losses.append(loss.item())
elif self.loss_version == 1:
for action_idx in range(len(self.action_dims)):
loss = mse_loss(input=Q[action_idx], target=y[action_idx].detach())
self.optimizers[action_idx].zero_grad()
if action_idx < len(self.action_dims)-1:
loss.backward(retain_graph=True)
else:
loss.backward()
self.optimizers[action_idx].step()
losses.append(loss.item())
# print(losses)
return losses
def update_Q_network_v2(self, state, action_1, action_2, reward, state_new, terminal):
state = torch.from_numpy(state).float()
action_1 = torch.from_numpy(action_1).float()
action_2 = torch.from_numpy(action_2).float()
state_new = torch.from_numpy(state_new).float()
terminal = torch.from_numpy(terminal).float()
reward = torch.from_numpy(reward).float()
state = Variable(state)
action_1 = Variable(action_1)
action_2 = Variable(action_2)
state_new = Variable(state_new)
terminal = Variable(terminal)
reward = Variable(reward)
self.Q_network.eval()
self.target_network.eval()
# use current network to evaluate action argmax_a' Q_current(s', a')_
# Get argmax action does not need aggregation
new_q_values = self.Q_network.forward(state_new)
# print(torch.max(new_q_values[0],1)[0], torch.max(new_q_values[0],1)[0].data)
# print(new_q_values[0], new_q_values[0] + new_q_values[2], new_q_values[2])
actions_new = [torch.max(q_value, 1)[1].cpu().data.view(-1, 1) for q_value in new_q_values[:2]]
actions_new_onehot = [torch.zeros(Config.sampling_batch_size, action_dim) for action_dim in self.action_dims]
actions_new_onehot = [Variable(actions_new_onehot[action_idx].scatter_(1, actions_new[action_idx], 1.0)) for action_idx in range(len(self.action_dims))]
actions = [action_1, action_2]
raw_y = self.target_network.forward(state_new)
if self.q_version == 0:
# Q = V + A
# Get Q(s',a')
aggregation = [raw_y[0]+ raw_y[2], raw_y[1]+raw_y[2]]
# Get Q(s,a)
self.Q_network.train()
raw_Q = self.Q_network.forward(state)
Q = []
for q_idx in range(len(raw_Q[:2])):
Q.append(((raw_Q[q_idx]+raw_Q[2])*actions[q_idx]).sum(dim=1))
elif self.q_version == 1:
# Q = V + Ad-max(Ad)
# print(raw_y[0])
# print(raw_y[1])
# print(raw_y[2])
# print(torch.max(raw_y[0],1, keepdim=True)[0])
branch_0 = raw_y[0]-torch.max(raw_y[0], 1, keepdim=True)[0].data + raw_y[2]
branch_1 = raw_y[1]-torch.max(raw_y[1], 1, keepdim=True)[0].data + raw_y[2]
aggregation = [branch_0, branch_1]
print(aggregation)
# Get Q(s,a)
self.Q_network.train()
raw_Q = self.Q_network.forward(state)
Q = []
for q_idx in range(len(raw_Q[:2])):
Q.append(((raw_Q[q_idx]-torch.max(raw_y[0], 1, keepdim=True)[0].data+raw_Q[2])*actions[q_idx]).sum(dim=1))
elif self.q_version == 2:
# Q = V + Ad - mean(Ad)
branch_0 = raw_y[0]-torch.mean(raw_y[0], 1, keepdim=True)[0].data + raw_y[2]
branch_1 = raw_y[1]-torch.mean(raw_y[1], 1, keepdim=True)[0].data + raw_y[2]
aggregation = [branch_0, branch_1]
# Get Q(s,a)
self.Q_network.train()
raw_Q = self.Q_network.forward(state)
Q = []
for q_idx in range(len(raw_Q[:2])):
Q.append(((raw_Q[q_idx]-torch.mean(raw_y[0], 1, keepdim=True)[0].data+raw_Q[2])*actions[q_idx]).sum(dim=1))
# Calculate target indep or global and then do loss (v0 or v1)
losses = []
if self.target_version == 0:
# Independent target, 2 branches
y = []
for new_q_idx in range(len(aggregation)):
y.append(reward + torch.mul(((aggregation[new_q_idx]*actions_new_onehot[new_q_idx]).sum(dim=1)*terminal),Config.discount_factor))
# print(y)
# Calculate loss
if self.loss_version == 0:
loss = 0.0
for action_idx in range(len(self.action_dims)):
loss += mse_loss(input=Q[action_idx], target=y[action_idx].detach())
loss/=len(self.action_dims)
self.optimizers.zero_grad()
loss.backward()
self.optimizers.step()
losses.append(loss.item())
elif self.loss_version == 1:
for action_idx in range(len(self.action_dims)):
loss = mse_loss(input=Q[action_idx], target=y[action_idx].detach())
self.optimizers[action_idx].zero_grad()
if action_idx < len(self.action_dims)-1:
loss.backward(retain_graph=True)
else:
loss.backward()
self.optimizers[action_idx].step()
losses.append(loss.item())
elif self.target_version == 1:
# Global target using max of branches, one y
# Torch.max support find max over two 300*6 and 300*7 tensors
local_max_0 = torch.mul(((aggregation[0]*actions_new_onehot[0]).sum(dim=1)*terminal),Config.discount_factor)
local_max_1 = torch.mul(((aggregation[1]*actions_new_onehot[1]).sum(dim=1)*terminal),Config.discount_factor)
# print(local_max_0, local_max_1)
y = reward + torch.max(local_max_0, local_max_1)
# print(y)
# Calculate loss
if self.loss_version == 0:
loss = 0.0
for action_idx in range(len(self.action_dims)):
loss += mse_loss(input=Q[action_idx], target=y.detach())
loss/=len(self.action_dims)
self.optimizers.zero_grad()
loss.backward()
self.optimizers.step()
losses.append(loss.item())
elif self.loss_version == 1:
for action_idx in range(len(self.action_dims)):
loss = mse_loss(input=Q[action_idx], target=y.detach())
self.optimizers[action_idx].zero_grad()
if action_idx < len(self.action_dims)-1:
loss.backward(retain_graph=True)
else:
loss.backward()
self.optimizers[action_idx].step()
losses.append(loss.item())
# Global target using global average
elif self.target_version == 2:
local_max_0 = torch.mul(((aggregation[0]*actions_new_onehot[0]).sum(dim=1)*terminal),Config.discount_factor)
local_max_1 = torch.mul(((aggregation[1]*actions_new_onehot[1]).sum(dim=1)*terminal),Config.discount_factor)
# print(local_max_0)
# print(local_max_1)
# print(torch.mean(torch.stack((local_max_0, local_max_1)),0))
y = reward + torch.mean(torch.stack((local_max_0, local_max_1)),dim=0)
# Calculate loss
if self.loss_version == 0:
loss = 0.0
for action_idx in range(len(self.action_dims)):
loss += mse_loss(input=Q[action_idx], target=y.detach())
loss/=len(self.action_dims)
self.optimizers.zero_grad()
loss.backward()
self.optimizers.step()
losses.append(loss.item())
elif self.loss_version == 1:
for action_idx in range(len(self.action_dims)):
loss = mse_loss(input=Q[action_idx], target=y.detach())
self.optimizers[action_idx].zero_grad()
if action_idx < len(self.action_dims)-1:
loss.backward(retain_graph=True)
else:
loss.backward()
self.optimizers[action_idx].step()
losses.append(loss.item())
# print(losses)
return losses
def take_action(self, state):
state = torch.from_numpy(state).float()
state = Variable(state)
self.Q_network.eval()
if self.model_version == 0:
estimate = torch.max(self.Q_network.forward(state), 1)[1].data[0]
if np.random.random() < self.epsilon:
return np.random.randint(0, self.action_dims[0]*self.action_dims[1])
else:
return estimate
elif self.model_version == 1:
outputs = self.Q_network.forward(state)
estimate = [torch.max(q_value, 1)[1].data[0] for q_value in outputs]
# with epsilon prob to choose random action else choose argmax Q estimate action
if np.random.random() < self.epsilon:
return [np.random.randint(0, self.action_dims[action_idx]) for action_idx in range(len(self.action_dims))]
else:
return estimate
elif self.model_version == 2:
outputs = self.Q_network.forward(state)[:2]
estimate = [torch.max(q_value, 1)[1].data[0] for q_value in outputs]
# with epsilon prob to choose random action else choose argmax Q estimate action
if np.random.random() < self.epsilon:
return [np.random.randint(0, self.action_dims[action_idx]) for action_idx in range(len(self.action_dims))]
else:
return estimate
def testing_take_action(self, state):
state = torch.from_numpy(state).float()
state = Variable(state)
self.Q_network.eval()
if self.model_version == 0:
estimate = torch.max(self.Q_network.forward(state), 1)[1].data[0]
return estimate
elif self.model_version == 1:
q_values = self.Q_network.forward(state)
estimate = [torch.max(q_value, 1)[1].data[0] for q_value in q_values]
# with epsilon prob to choose random action else choose argmax Q estimate action
return estimate
elif self.model_version == 2:
q_values = self.Q_network.forward(state)[:2]
estimate = [torch.max(q_value, 1)[1].data[0] for q_value in q_values]
# with epsilon prob to choose random action else choose argmax Q estimate action
return estimate
def update_epsilon_by_epoch(self, epoch):
self.epsilon = self.epsilon_final+(self.epsilon_start - self.epsilon_final) * math.exp(-1.*epoch/self.epsilon_decay)
def save(self, step, logs_path):
os.makedirs(logs_path, exist_ok=True)
model_list = glob.glob(os.path.join(logs_path, '*.pth'))
if len(model_list) > Config.maximum_model - 1 :
min_step = min([int(li.split('/')[-1][6:-4]) for li in model_list])
os.remove(os.path.join(logs_path, 'model-{}.pth' .format(min_step)))
logs_path = os.path.join(logs_path, 'model-{}.pth' .format(step))
self.Q_network.save(logs_path, step=step, optimizers=self.optimizers)
print('=> Save {}' .format(logs_path))
def train_restore(self, logs_path):
model_list = glob.glob(os.path.join(logs_path, '*.pth'))
max_step = max([int(li.split('/')[-1][6:-4]) for li in model_list])
model_path = os.path.join(logs_path, 'model-{}.pth' .format(max_step))
self.Q_network.load(model_path, self.optimizers)
self.target_network.load(model_path, self.optimizers)
print('=> Restore {}' .format(model_path))
return max_step + 1
def restore(self, logs_path):
if self.loss_version == 0:
self.Q_network.load(logs_path, self.optimizers)
self.target_network.load(logs_path, self.optimizers)
print('=> Restore {}' .format(logs_path))
# self.optimizers = optim.Adam(self.Q_network.parameters(), lr=Config.lr)
elif self.loss_version == 1:
self.Q_network.load(logs_path, self.optimizers)
self.target_network.load(logs_path, self.optimizers)
print('=> Restore {}' .format(logs_path))
def __init__(self, env, alpha, gamma, eps):
super().__init__(Model(), alpha, gamma)
self.env = env
self.eps = eps
self.reset()
mask = K.cast(K.greater(targ, 0),
dtype='float32') #cast bool-tensor to float
n_correct = K.sum(mask * correct) #
n_total = K.sum(mask)
return n_correct / n_total
#custom loss because we dont want to consider padding
def loss(y_true, y_pred):
# both are of shape ( _, Ty, VOCAB_SIZE )
mask = K.cast(y_true > 0, dtype='float32')
out = mask * y_true * K.log(y_pred) #cross entopy loss
return -K.sum(out) / K.sum(mask)
models = Model(VOCAB_SIZE, MAX_LEN, embedding_matrix)
train_model = models.model_train_teacher_forcing()
train_model.compile(optimizer="adam", loss=loss, metrics=[acc])
print("Training Started")
all_metrics = []
BATCH_SIZE = 224
TRAIN_BATCHES, VAL_BATCHES = len(x_train) // BATCH_SIZE, len(
x_val) // BATCH_SIZE
for epoch in range(len(all_metrics) + 1, 15):
tloss, tacc, ti, t0 = 0, 0, 0, time.time()
for x, y in batch_generator(x_train,
y_train,
ytf_train,
batch_size=BATCH_SIZE):
metrics_train = train_model.train_on_batch(x, y)