def dnn_concat(data,
configfile,
layers=[16, 8],
activation='tanhlecun',
initrange=1e-3,
bn=True,
keep_prob=.95,
concat_size=24,
uembed=32,
iembed=32,
learnrate=.00001,
verbose=True,
epochs=10,
maxbadcount=20,
mb=2000,
eval_rate=500):
with tf.name_scope('ant_graph'):
ant = config.AntGraph(configfile,
data=data.dev.features,
marker='-',
graph_name='dnn_concat',
variable_bindings={
'layers': layers,
'activation': activation,
'initrange': initrange,
'bn': bn,
'keep_prob': keep_prob,
'concat_size': concat_size,
'uembed': uembed,
'iembed': iembed,
})
y = ant.tensor_out
y_ = tf.placeholder("float", [None, None], name='Target')
ant.placeholderdict[
'ratings'] = y_ # put the new placeholder in the graph for training
with tf.name_scope('objective'):
objective = tf.reduce_sum(tf.square(y_ - y))
with tf.name_scope('dev_rmse'):
dev_rmse = tf.sqrt(
tf.div(tf.reduce_sum(tf.square(y - y_)), data.dev.num_examples))
with tf.name_scope('training'):
model = generic_model.Model(objective,
ant.placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=dev_rmse,
predictions=y,
model_name='res_concat',
make_histograms=False,
save=False,
tensorboard=False)
model.train(data.train, dev=data.dev, eval_schedule=eval_rate)
return model
def mf(data, configfile, lamb=0.001,
kfactors=20,
learnrate=0.01,
verbose=True,
epochs=1000,
maxbadcount=20,
mb=500,
initrange=1,
eval_rate=500,
random_seed=None,
develop=False,
train_dev_eval_factor=3):
with tf.name_scope('ant_graph'):
ant = config.AntGraph(configfile,
data=data.dev.features,
marker='-',
graph_name='basic_mf',
develop=develop,
variable_bindings={'kfactors': kfactors,
'initrange': initrange})
y = ant.tensor_out
y_ = tf.placeholder("float", [None, None], name='Target')
ant.placeholderdict['ratings'] = y_
with tf.name_scope('objective'):
objective = (tf.reduce_sum(tf.square(y_ - y)))
objective += (lamb*tf.reduce_sum(tf.square(ant.tensordict['huser'])) +
lamb*tf.reduce_sum(tf.square(ant.tensordict['hitem'])) +
lamb*tf.reduce_sum(tf.square(ant.tensordict['ubias'])) +
lamb*tf.reduce_sum(tf.square(ant.tensordict['ibias'])))
with tf.name_scope('dev_rmse'):
_rmse = node_ops.rmse(y_, y)
mae = node_ops.mae(y_, y)
model = generic_model.Model(objective, ant.placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=_rmse,
train_evaluate=_rmse,
predictions=y,
model_name='mf',
random_seed=random_seed,
save_tensors={'mae': mae})
model.train(data.train, dev=data.dev, eval_schedule=eval_rate,train_dev_eval_factor= train_dev_eval_factor)
return model
def tensorfactor(data,
configfile,
lamb=.001,
kfactors=1000,
learnrate=0.01,
verbose=True,
epochs=1000,
maxbadcount=20,
mb=500,
initrange=1,
eval_rate=500,
random_seed=None):
with tf.name_scope('ant_graph'):
ant = config.AntGraph(configfile,
data=data.dev.features,
marker='-',
graph_name='basic_mf',
variable_bindings={
'kfactors': kfactors,
'initrange': initrange,
'lamb': lamb
})
y = ant.tensor_out
y_ = tf.placeholder("float", [None, None], name='Target')
ant.placeholderdict['ratings'] = y_
with tf.name_scope('objective'):
objective = (tf.reduce_sum(tf.square(y_ - y)))
with tf.name_scope('dev_rmse'):
dev_rmse = node_ops.rmse(y_, y)
model = generic_model.Model(objective,
ant.placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=dev_rmse,
predictions=y,
model_name='mf',
random_seed=random_seed)
model.train(data.train, dev=data.dev, eval_schedule=200)
return model
def svdplus(data,
lamb_bias=0.005,
lambfactor=0.015,
kfactors=20,
learnrate=0.01,
verbose=True,
epochs=1000,
maxbadcount=20,
mb=500,
initrange=1,
eval_rate=500,
random_seed=None,
develop=False):
data = loader.read_data_sets(data,
folders=['train', 'dev', 'item'],
hashlist=['user', 'item', 'ratings'])
data.train.labels['ratings'] = loader.center(data.train.labels['ratings'])
data.dev.labels['ratings'] = loader.center(data.dev.labels['ratings'])
utility_matrix = sps.csr_matrix(
(numpy.ones(data.train.features['user'].vec.shape[0]),
(data.train.features['user'].vec, data.train.features['item'].vec)),
shape=(data.train.features['user'].dim,
data.train.features['item'].dim))
data.item.features['util'] = utility_matrix
xuser = tf.placeholder(tf.int32, [None])
xitem = tf.placeholder(tf.int32, [None])
xutil = tf.placeholder(tf.float32, [None, None])
wuser = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['user'].dim, kfactors]))
witem = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['item'].dim, kfactors]))
wplus = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['item'].dim, kfactors]))
ubias = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['user'].dim]))
ibias = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['item'].dim]))
i_bias = tf.nn.embedding_lookup(ibias, xitem)
u_bias = tf.nn.embedding_lookup(ubias, xuser)
huser = tf.nn.embedding_lookup(wuser, xuser)
hitem = tf.nn.embedding_lookup(witem, xitem)
hplus = tf.nn.embedding_lookup(xutil, xuser)
plus = tf.mul(
tf.matmul(hplus, wplus, a_is_sparse=True),
tf.rsqrt(tf.reduce_sum(hplus, reduction_indices=1, keep_dims=True)))
huserplus = huser + plus
y = node_ops.x_dot_y([huserplus, hitem, i_bias, u_bias])
y_ = tf.placeholder("float", [None, None], name='Target')
with tf.name_scope('objective'):
objective = (tf.reduce_sum(tf.square(y_ - y)) +
lambfactor * tf.reduce_sum(tf.square(huser)) +
lambfactor * tf.reduce_sum(tf.square(hitem)) +
lambfactor * tf.reduce_sum(tf.square(wplus)) +
lamb_bias * tf.reduce_sum(tf.square(i_bias)) +
lamb_bias * tf.reduce_sum(tf.square(u_bias)))
placeholderdict = {
'ratings': y_,
'util': xutil,
'user': xuser,
'item': xitem
}
mae = node_ops.mae(y_, y)
with tf.name_scope('dev_rmse'):
dev_rmse = node_ops.rmse(y_, y)
model = generic_model.Model(objective,
placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=dev_rmse,
predictions=y,
model_name='svdplus',
random_seed=random_seed,
decay=(500, 0.999),
save_tensors={'mae': mae})
model.train(data.train,
dev=data.dev,
supplement=data.item.features,
eval_schedule=eval_rate)
return model
def dssm(data,
configfile,
layers=[10, 10, 10],
bn=True,
keep_prob=.95,
act='tanhlecun',
initrange=1,
kfactors=10,
lamb=.1,
mb=500,
learnrate=0.0001,
verbose=True,
maxbadcount=10,
epochs=100,
model_name='dssm',
random_seed=500,
eval_rate=500):
datadict = data.user.features.copy()
datadict.update(data.item.features)
configdatadict = data.dev.features.copy()
configdatadict.update(datadict)
with tf.name_scope('ant_graph'):
ant = config.AntGraph(configfile,
data=configdatadict,
marker='-',
graph_name='basic_mf',
variable_bindings={
'initrange': initrange,
'kfactors': kfactors
})
y_ = tf.placeholder("float", [None, None], name='Target')
ant.placeholderdict['ratings'] = y_
with tf.name_scope('objective'):
if type(ant.tensor_out) is list:
objective = tf.reduce_sum(tf.square(y_ - ant.tensor_out[0]))
for i in range(1, len(ant.tensor_out)):
objective += tf.reduce_sum(tf.square(y_ - ant.tensor_out[i]))
objective += (
lamb * tf.reduce_sum(tf.square(ant.tensordict['huser'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['hitem'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['ubias'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['ibias'])))
with tf.name_scope('dev_rmse'):
dev_rmse = node_ops.rmse(ant.tensor_out[0], y_)
model = generic_model.Model(objective,
ant.placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=dev_rmse,
predictions=ant.tensor_out[0],
model_name='dssm',
random_seed=random_seed)
model.train(data.train,
dev=data.dev,
supplement=datadict,
eval_schedule=eval_rate)
return model
def tensorfactor(data,
context_key='occ',
lamb=0.01,
learnrate=0.0001,
verbose=True,
epochs=5,
maxbadcount=20,
mb=500,
initrange=0.0001,
eval_rate=10000,
random_seed=None,
uembed=50,
iembed=50,
cembed=50):
data = loader.read_data_sets(data,
folders=('train', 'dev', 'item', 'user'),
hashlist=('user', 'item', context_key,
'ratings'))
data.train.labels['ratings'] = loader.center(data.train.labels['ratings'])
data.dev.labels['ratings'] = loader.center(data.dev.labels['ratings'])
if context_key in data.item.features:
data.train.features[context_key] = data.item.features[context_key][
data.train.features['item']]
data.dev.features[context_key] = data.item.features[context_key][
data.dev.features['item']]
del data.item.features[context_key]
elif context_key in data.user.features:
data.train.features[context_key] = data.user.features[context_key][
data.train.features['user']]
data.dev.features[context_key] = data.user.features[context_key][
data.dev.features['user']]
del data.user.features[context_key]
data.show()
item = tf.placeholder(tf.int32, [None])
user = tf.placeholder(tf.int32, [None])
context = tf.placeholder(tf.int32, [None])
wuser = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['user'].shape[1], uembed]))
witem = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['item'].shape[1], iembed]))
wcontext = initrange * tf.Variable(
tf.truncated_normal([data.dev.features[context_key].shape[1], cembed]))
xuser = tf.nn.embedding_lookup(wuser, user)
xitem = tf.nn.embedding_lookup(witem, item)
xcontext = tf.nn.embedding_lookup(wcontext, context)
ibias = tf.Variable(
tf.truncated_normal([data.dev.features['item'].shape[1]]))
ubias = tf.Variable(
tf.truncated_normal([data.dev.features['user'].shape[1]]))
cbias = tf.Variable(
tf.truncated_normal([data.dev.features[context_key].shape[1]]))
i_bias = tf.nn.embedding_lookup(ibias, item)
u_bias = tf.nn.embedding_lookup(ubias, user)
c_bias = tf.nn.embedding_lookup(cbias, context)
y = node_ops.ternary_tensor_combine([xuser, xitem, xcontext],
initrange=initrange,
l2=lamb) + i_bias + u_bias
y_ = tf.placeholder("float", [None, None], name='Target')
placeholderdict = {
'user': user,
'item': item,
context_key: context,
'ratings': y_
}
with tf.name_scope('objective'):
objective = (tf.reduce_sum(tf.square(y_ - y)) +
lamb * tf.reduce_sum(tf.square(wcontext)) +
lamb * tf.reduce_sum(tf.square(xuser)) +
lamb * tf.reduce_sum(tf.square(xitem)) +
lamb * tf.reduce_sum(tf.square(i_bias)) +
lamb * tf.reduce_sum(tf.square(u_bias)) +
lamb * tf.reduce_sum(tf.square(c_bias)))
with tf.name_scope('dev_rmse'):
dev_rmse = node_ops.rmse(y_, y)
model = generic_model.Model(objective,
placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=dev_rmse,
predictions=y,
model_name='tensorfactor',
random_seed=random_seed)
model.train(data.train, dev=data.dev, eval_schedule=eval_rate)
return model
def tree(data,
configfile,
lamb=0.001,
kfactors=20,
learnrate=0.0001,
verbose=True,
maxbadcount=20,
mb=500,
initrange=0.00001,
epochs=10,
random_seed=None,
eval_rate=500,
keep_prob=0.95,
act='tanh'):
datadict = data.user.features.copy()
datadict.update(data.item.features)
configdatadict = data.dev.features.copy()
configdatadict.update(datadict)
with tf.name_scope('ant_graph'):
ant = config.AntGraph(configfile,
data=configdatadict,
marker='-',
variable_bindings={
'kfactors': kfactors,
'initrange': initrange,
'keep_prob': keep_prob,
'act': act
},
graph_name='tree')
y = ant.tensor_out
y_ = tf.placeholder("float", [None, None], name='Target')
ant.placeholderdict[
'ratings'] = y_ # put the new placeholder in the graph for training
with tf.name_scope('objective'):
objective = (tf.reduce_sum(tf.square(y_ - y)) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['huser'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['hitem'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['ubias'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['ibias'])))
with tf.name_scope('dev_rmse'):
dev_rmse = tf.sqrt(
tf.div(tf.reduce_sum(tf.square(y - y_)), data.dev.num_examples))
with tf.name_scope('training'):
model = generic_model.Model(objective,
ant.placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=dev_rmse,
predictions=y,
model_name='tree')
model.train(data.train,
dev=data.dev,
supplement=datadict,
eval_schedule=eval_rate)
return model
def mf(data,
configfile,
lamb=0.001,
kfactors=20,
learnrate=0.01,
verbose=True,
epochs=1000,
maxbadcount=20,
mb=500,
initrange=1,
eval_rate=500,
random_seed=None,
develop=False):
data = loader.read_data_sets(data,
hashlist=['item', 'user', 'ratings'],
folders=['dev', 'train', 'item'])
data.train.labels['ratings'] = loader.center(data.train.labels['ratings'])
data.dev.labels['ratings'] = loader.center(data.dev.labels['ratings'])
with tf.name_scope('ant_graph'):
ant = config.AntGraph(configfile,
data=data.dev.features,
marker='-',
graph_name='basic_mf',
develop=develop,
variable_bindings={
'kfactors': kfactors,
'initrange': initrange,
'lamb': lamb
})
print(ant.tensor_out)
y = node_ops.x_dot_y(ant.tensor_out)
y_ = tf.placeholder("float", [None, None], name='Target')
data.item.features['util'] = utility_matrix
xuser = tf.placeholder(tf.int32, [None])
xitem = tf.placeholder(tf.int32, [None])
xutil = tf.placeholder(tf.float32, [None, None])
wuser = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['user'].dim, kfactors]))
witem = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['item'].dim, kfactors]))
wplus = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['item'].dim, kfactors]))
ubias = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['user'].dim]))
ibias = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['item'].dim]))
i_bias = tf.nn.embedding_lookup(ibias, xitem)
u_bias = tf.nn.embedding_lookup(ubias, xuser)
huser = tf.nn.embedding_lookup(wuser, xuser)
hitem = tf.nn.embedding_lookup(witem, xitem)
hplus = tf.nn.embedding_lookup(xutil, xuser)
plus = tf.mul(
tf.matmul(hplus, wplus, a_is_sparse=True),
tf.rsqrt(tf.reduce_sum(hplus, reduction_indices=1,
keep_dims=True)))
huserplus = huser #+ plus
ant.placeholderdict['ratings'] = y_
with tf.name_scope('objective'):
objective = (tf.reduce_sum(tf.square(y_ - y)))
objective += (
lamb * tf.reduce_sum(tf.square(ant.tensordict['huser'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['hitem'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['ubias'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['ibias'])))
with tf.name_scope('dev_rmse'):
dev_rmse = node_ops.rmse(y_, y)
model = generic_model.Model(objective,
ant.placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=dev_rmse,
predictions=y,
model_name='mf',
random_seed=random_seed)
model.train(data.train, dev=data.dev, eval_schedule=eval_rate)
return model
configdatadict = data.dev.features.copy()
configdatadict.update(datadict)
with tf.variable_scope('mfgraph'):
ant = config.AntGraph('tree.config',
data=configdatadict,
marker='-',
variable_bindings={
'kfactors': 100,
'initrange': 0.0001
},
develop=False)
y = ant.tensor_out
y_ = tf.placeholder("float", [None, None], name='Target')
ant.placeholderdict[
'ratings'] = y_ # put the new placeholder in the graph for training
objective = tf.reduce_sum(tf.square(y_ - y))
dev_rmse = node_ops.rmse(y, y_)
model = generic_model.Model(objective,
ant.placeholderdict,
mb=500,
learnrate=0.00001,
verbose=True,
maxbadcount=20,
epochs=100,
evaluate=dev_rmse,
predictions=y)
model.train(data.train, dev=data.dev, supplement=datadict, eval_schedule=2000)
def dsadd(data,
configfile,
initrange=0.1,
kfactors=20,
lamb=.01,
mb=500,
learnrate=0.003,
verbose=True,
maxbadcount=10,
epochs=100,
model_name='dssm',
random_seed=500,
eval_rate=500):
datadict = data.user.features.copy()
datadict.update(data.item.features)
configdatadict = data.dev.features.copy()
configdatadict.update(datadict)
with tf.name_scope('ant_graph'):
ant = config.AntGraph(configfile,
data=configdatadict,
marker='-',
graph_name='basic_mf',
variable_bindings={
'initrange': initrange,
'kfactors': kfactors
})
y_ = tf.placeholder("float", [None, None], name='Target')
ant.placeholderdict['ratings'] = y_
with tf.name_scope('objective'):
if type(ant.tensor_out) is list:
scalars = tf.Variable(
0.001 * tf.truncated_normal([len(ant.tensor_out), 1]))
prediction = tf.mul(ant.tensor_out[0],
tf.slice(scalars, [0, 0], [1, 1]))
for i in range(1, len(ant.tensor_out)):
with tf.variable_scope('predictor%d' % i):
prediction = prediction + tf.mul(
ant.tensor_out[i], tf.slice(scalars, [i, 0], [1, 1]))
prediction = tf.square(y_ - prediction)
objective = (
tf.reduce_sum(prediction) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['huser'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['hitem'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['ubias'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['ibias'])) +
lamb * tf.reduce_sum(tf.square(tf.concat(1, ant.tensor_out))))
with tf.name_scope('dev_rmse'):
dev_rmse = tf.sqrt(
tf.div(tf.reduce_sum(prediction), data.dev.num_examples))
model = generic_model.Model(objective,
ant.placeholderdict,
mb=500,
learnrate=0.000001,
verbose=True,
maxbadcount=10,
epochs=100,
evaluate=dev_rmse,
predictions=ant.tensor_out[0],
model_name='dssm',
random_seed=500)
model.train(data.train,
dev=data.dev,
supplement=datadict,
eval_schedule=eval_rate)
return model
y_ = tf.placeholder(tf.float32, [None, 19])
objective = node_ops.other_cross_entropy(y, y_)
detection = node_ops.detection(y, args.threshold)
# recall percentage of actual genres detected
recall = node_ops.recall(y, y_, detects=detection)
# precision: percentage of the genres we predicted that were correct
precision = node_ops.precision(y, y_, detects=detection)
fscore = node_ops.fscore(precisions=precision, recalls=recall)
placeholderdict = {'words': x, 'genre': y_}
model = generic_model.Model(objective,
placeholderdict,
mb=args.mb,
learnrate=args.learnrate,
verbose=args.verbose,
maxbadcount=args.maxbadcount,
epochs=100,
evaluate=objective,
predictions=detection,
model_name='genrepred',
random_seed=500,
save_tensors={
'fscore': fscore,
'precision': precision,
'recall': recall,
'dist':
tf.get_collection('activation_layers')[0]
})
z = model.train(data.train, dev=data.dev, eval_schedule=args.eval_rate)
with tf.variable_scope('mfgraph'):
ant = config.AntGraph('mf2.config',
data=data.dev.features,
marker='-',
variable_bindings={
'kfactors': 100,
'initrange': 0.001,
'l2': 0.1
})
y = ant.tensor_out
y_ = tf.placeholder("float", [None, None], name='Target')
ant.placeholderdict[
'ratings'] = y_ # put the new placeholder in the graph for training
objective = tf.reduce_sum(tf.square(y_ - y))
dev_rmse = node_ops.rmse(y, y_)
dev_mae = node_ops.mae(y, y_)
model = generic_model.Model(objective,
ant.placeholderdict,
mb=500,
learnrate=0.01,
verbose=True,
maxbadcount=10,
epochs=100,
evaluate=dev_rmse,
predictions=y,
save_tensors={'dev_mae': dev_mae})
model.train(data.train, dev=data.dev)
y_ = tf.placeholder("float", [None, None], name='Target')
ant.placeholderdict['targets'] = y_
if type == 'r':
objective = node_ops.mse(y, y_)
if type == 'c':
objective = node_ops.accuracy(y, y_)
with tf.name_scope('objective'):
objective = (tf.reduce_sum(tf.square(y_ - y)) +
lamb*tf.reduce_sum(tf.square(ant.tensordict['huser'])) +
lamb*tf.reduce_sum(tf.square(ant.tensordict['hitem'])) +
lamb*tf.reduce_sum(tf.square(ant.tensordict['ubias'])) +
lamb*tf.reduce_sum(tf.square(ant.tensordict['ibias'])))
with tf.name_scope('dev_rmse'):
dev_rmse = node_ops.rmse(y_, y)
#dev_rmse = tf.sqrt(tf.div(tf.reduce_sum(tf.square(y - y_)), data.dev.num_examples))
model = generic_model.Model(objective, ant.placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=dev_rmse,
predictions=y,
model_name='mf',
random_seed=500)
model.train(data.train, dev=data.dev, eval_schedule=200)
return model
def tensorfactor(data, lamb=0.01,
learnrate=0.0001,
verbose=True,
epochs=100,
maxbadcount=20,
mb=500,
initrange=0.0001,
eval_rate=10000,
random_seed=None,
uembed=50,
iembed=50,
gembed=50):
data = loader.read_data_sets(data, folders=('train', 'dev', 'item'),
hashlist=('user', 'item', 'genres', 'ratings'))
data.train.labels['ratings'] = loader.center(data.train.labels['ratings'])
data.dev.labels['ratings'] = loader.center(data.dev.labels['ratings'])
data.train.features['genre'] = data.item.features['genres'][data.train.features['item'].vec, :]
data.dev.features['genre'] = data.item.features['genres'][data.dev.features['item'].vec, :]
data.show()
item = tf.placeholder(tf.int32, [None])
user = tf.placeholder(tf.int32, [None])
genre = tf.placeholder(tf.float32, [None, data.dev.features['genre'].shape[1]])
wuser = initrange*tf.Variable(tf.truncated_normal([data.dev.features['user'].shape[1], uembed]))
witem = initrange*tf.Variable(tf.truncated_normal([data.dev.features['item'].shape[1], iembed]))
wgenre = initrange*tf.Variable(tf.truncated_normal([data.dev.features['genre'].shape[1], gembed]))
xuser = tf.nn.embedding_lookup(wuser, user)
xitem = tf.nn.embedding_lookup(witem, item)
xgenre = tf.matmul(genre, wgenre, a_is_sparse=True)
ibias = tf.Variable(tf.truncated_normal([data.dev.features['item'].shape[1]]))
ubias = tf.Variable(tf.truncated_normal([data.dev.features['user'].shape[1]]))
gbias = tf.Variable(tf.truncated_normal([data.dev.features['genre'].shape[1], 1]))
i_bias = tf.nn.embedding_lookup(ibias, item)
u_bias = tf.nn.embedding_lookup(ubias, user)
g_bias = tf.matmul(genre, gbias, a_is_sparse=True)
y = node_ops.ternary_tensor_combine([xuser, xitem, xgenre],
initrange=initrange,
l2=lamb) + i_bias + u_bias
y_ = tf.placeholder("float", [None, None], name='Target')
placeholderdict = {'user': user, 'item': item, 'genre': genre, 'ratings': y_}
with tf.name_scope('objective'):
objective = (tf.reduce_sum(tf.square(y_ - y)) +
lamb*tf.reduce_sum(tf.square(wgenre)) +
lamb*tf.reduce_sum(tf.square(xuser)) +
lamb*tf.reduce_sum(tf.square(xitem)) +
lamb*tf.reduce_sum(tf.square(i_bias)) +
lamb*tf.reduce_sum(tf.square(u_bias)))
with tf.name_scope('dev_rmse'):
dev_rmse = node_ops.rmse(y_, y)
model = generic_model.Model(objective, placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=dev_rmse,
predictions=y,
model_name='tensorfactor',
random_seed=random_seed)
model.train(data.train, dev=data.dev, eval_schedule=eval_rate)
return model
def tree(data,
configfile,
lamb=0.001,
kfactors=50,
learnrate=0.00001,
verbose=True,
maxbadcount=20,
mb=500,
initrange=0.00001,
epochs=10,
random_seed=None,
eval_rate=500,
keep_prob=0.95,
act='tanh'):
datadict = data.user.features.copy()
datadict.update(data.item.features)
configdatadict = data.dev.features.copy()
configdatadict.update(datadict)
with tf.name_scope('ant_graph'):
ant = config.AntGraph(configfile,
data=configdatadict,
marker='-',
variable_bindings={
'kfactors': kfactors,
'initrange': initrange,
'keep_prob': keep_prob,
'act': act
},
graph_name='tree')
# ant.display_graph()
alpha = node_ops.weights('tnorm', [1, 1], l2=1.0)
beta = node_ops.weights('tnorm', [1, 1], l2=1.0)
print(ant.tensor_out)
ubias = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['user'].dim]))
ibias = initrange * tf.Variable(
tf.truncated_normal([data.dev.features['item'].dim]))
i_bias = tf.nn.embedding_lookup(ibias, ant.placeholderdict['item'])
u_bias = tf.nn.embedding_lookup(ubias, ant.placeholderdict['user'])
y = alpha * ant.tensor_out[0] + beta * ant.tensor_out[1] + u_bias + i_bias
y_ = tf.placeholder("float", [None, None], name='Target')
ant.placeholderdict[
'ratings'] = y_ # put the new placeholder in the graph for training
with tf.name_scope('objective'):
objective = (
tf.reduce_sum(tf.square(y_ - y)) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['huser'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['hitem'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['huser2'])) +
lamb * tf.reduce_sum(tf.square(ant.tensordict['hitem2'])) +
lamb * tf.reduce_sum(tf.square(u_bias)) +
lamb * tf.reduce_sum(tf.square(i_bias)))
with tf.name_scope('dev_rmse'):
dev_rmse = node_ops.rmse(y_, y)
with tf.name_scope('training'):
model = generic_model.Model(objective,
ant.placeholderdict,
mb=mb,
learnrate=learnrate,
verbose=verbose,
maxbadcount=maxbadcount,
epochs=epochs,
evaluate=dev_rmse,
predictions=y,
model_name='tree',
train_evaluate=dev_rmse)
model.train(data.train,
dev=data.dev,
supplement=datadict,
eval_schedule=eval_rate,
train_dev_eval_factor=5)
return model
x = tf.placeholder(tf.float32, [None, 12734])
y = node_ops.dnn(x, [19], initrange=args.initrange, activation='sigmoid')
y_ = tf.placeholder(tf.float32, [None, 19])
objective = node_ops.other_cross_entropy(y, y_)
detection = node_ops.detection(y, args.threshold)
# recall percentage of actual genres detected
recall = node_ops.recall(y_, detects=detection)
# precision: percentage of the genres we predicted that were correct
precision = node_ops.precision(y_, detects=detection)
fscore = node_ops.fscore(precisions=precision, recalls=recall)
placeholderdict = {'words': x, 'genre': y_}
model = generic_model.Model(objective,
placeholderdict,
mb=args.mb,
learnrate=args.learnrate,
verbose=args.verbose,
maxbadcount=args.maxbadcount,
epochs=100,
evaluate=objective,
predictions=detection,
model_name='genrepred',
save_tensors={
'precision': precision,
'recall': recall,
'fscore': fscore
},
random_seed=args.random_seed)
z = model.train(data.train, dev=data.dev, eval_schedule=args.eval_rate)
