def test_easy(self):
f2vm = Feat2VecModel([
{
"name": "disc1",
"type": "discrete",
"vocab": 10
}, # disc1 has a vocabulary of 10 words
{
"name": "cont1",
"type": "real"
}, #cont1 is a real number
{
"name": "disc100",
"type": "discrete",
"len": 100,
"vocab": 10
}, # disc100 is a sequence of 100 words, with a vocabulary of 10 words
{
"name": "cont100",
"len": 100,
"type": "real"
}, #cont100 is a sequence of 100 numbers
])
keras_model = f2vm.build_model(embedding_dimensions=5)
keras_model.compile(loss='binary_crossentropy',
optimizer=tf.train.AdamOptimizer())
def test_easy_in(self):
import keras.layers
# define custom layer:
input_layer = keras.layers.Input(shape=(1, ))
custom_layer = keras.layers.Embedding(input_dim=1, output_dim=5)(
input_layer) #keeping it simple here
# define feat2vec model
f2vm = Feat2VecModel([
{
"name":
"disc1", # 'disc1 is a (single) discrete feature with a vocabulary of 10 words
"type": "discrete",
"vocab": 10
},
{
"name": "cont1", # 'cont1' is a sequence of 20 real numbers
"type": "real",
"len": 20
},
{
"name": "custom", # 'custom' is a custom-type feature
"type": {
"input": input_layer,
"output": custom_layer
}
}
])
keras_model = f2vm.build_model(
embedding_dimensions=5) # this returns a Keras network
keras_model.compile(loss='binary_crossentropy',
optimizer=tf.train.AdamOptimizer())
def test_rai(self):
dimensions = 10
dup_dimension = 20
EMBEDING_DIM = 5
SEC_EMBEDDING = 5
feature_names = ["principal", "f1", "f2", "f3", "f4", "f5"]
# DEFINE CUSTOM LAYERS
skream_rotator = Dense(units=dimensions,
activation="linear",
use_bias=False,
name="rotator")
# Popularity
f1_input = Input(batch_shape=(None, 1), name="f1")
popularity_embed_inter = Embedding(input_dim=EMBEDING_DIM,
output_dim=dup_dimension,
name="embedding_dup",
mask_zero=True)(f1_input)
popularity_unmasker = Lambda(
lambda x: x, name='unmasker_dup')(popularity_embed_inter)
popularity_embed1 = Dense(
units=dimensions,
activation="linear",
use_bias=False,
name="resized_embedding")(popularity_unmasker)
f1_embed = Reshape((dimensions, ))(popularity_embed1)
# Principal id
principal_input = Input(batch_shape=(None, 1), name="principal")
principal_embedding = Embedding(input_dim=EMBEDING_DIM,
output_dim=SEC_EMBEDDING,
name="embedding_principal",
mask_zero=True)(principal_input)
principal_makser = Lambda(
lambda x: x, name='unmasker_principal')(principal_embedding)
principal_reshape = Reshape((SEC_EMBEDDING, ))(principal_makser)
merged_principal = concatenate([principal_reshape, f1_embed])
rotated_principal = Reshape(
(dimensions, ),
name="rotated_principal")(skream_rotator(merged_principal))
# F2
f2_input = Input(batch_shape=(None, 20), name="f2")
f2_temp = Embedding(input_dim=5,
input_length=20,
output_dim=SEC_EMBEDDING,
mask_zero=True,
name="embedding_f2")(f2_input)
avg_f2 = Reshape((SEC_EMBEDDING, ))(
Lambda(lambda x: K.sum(x, axis=1, keepdims=True),
name="avg_f2_embedding")(f2_temp))
merged_f2 = concatenate([avg_f2, f1_embed])
f2_embed = Reshape((dimensions, ),
name="rotated_f2")(skream_rotator(merged_f2))
# DEFINE FM MACHINE:
feature_specification = []
for feat in feature_names:
if feat == "principal":
feature_specification.append({
"name": "principal",
"type": {
"input": principal_input,
"output": rotated_principal
}
})
elif feat == "f1":
feature_specification.append({
"name": "f1",
"type": {
"input": f1_input,
"output": f1_embed
}
})
elif feat == "f2":
feature_specification.append({
"name": feat,
"type": {
"input": f2_input,
"output": f2_embed
}
})
elif feat == "f3":
feature_specification.append({"name": feat, "type": "real"})
else:
feature_specification.append({
"name": feat,
"type": "discrete",
"len": 10,
"vocab": 100
})
fm = Feat2VecModel(features=feature_specification,
mask_zero=True,
obj='ns')
groups = []
groups.append(
zip(["principal"] * (len(feature_names) - 1), feature_names[1:]))
groups.append(zip(["f1"] * (len(feature_names) - 2),
feature_names[2:]))
groups.append([("f1", "principal")])
print(groups)
keras_model = fm.build_model(dimensions,
deep_out=True,
deep_out_bias=False,
deep_weight_groups=groups,
dropout_layer=0.5,
dropout_input=0.1)
#f5 = keras_model.get_layer("dropout_embedding_f5")
#assert f5.rate > 0.
#f5.rate = 0.
try:
from keras.utils import plot_model
plot_model(keras_model, to_file="rai.png")
except:
pass
def test_rai_faster(self):
dimensions = 10
EMBEDING_DIM = 5
feature_names = ["principal", "f1", "f2", "f3", "f4", "f5"]
# Principal id
principal_input = Input(batch_shape=(None, 1), name="principal")
int = Embedding(input_dim=EMBEDING_DIM,
output_dim=dimensions,
name="embedding_principal")(principal_input)
principal_embedding = int # Reshape((dimensions,1))(int)
# DEFINE FM MACHINE:
feature_specification = []
for feat in feature_names:
if feat == "principal":
feature_specification.append({
"name": "principal",
"type": {
"input": principal_input,
"output": principal_embedding
}
})
elif feat == "f1":
feature_specification.append({"name": feat, "type": "real"})
else:
feature_specification.append({
"name": feat,
"type": "discrete",
"len": 10,
"vocab": 100
})
fm = Feat2VecModel(features=feature_specification,
mask_zero=True,
obj='ns')
groups = []
groups.append([("principal", feature_names)])
groups.append([("f1", feature_names)])
print(groups)
keras_model_collapsed = fm.build_model(dimensions,
collapsed_type=2,
deep_out=True,
deep_out_bias=False,
deep_weight_groups=groups,
dropout_layer=0.5,
dropout_input=0.1)
keras_model_notcollapsed = fm.build_model(dimensions,
collapsed_type=None,
deep_out=True,
deep_out_bias=False,
deep_weight_groups=groups,
dropout_layer=0.5,
dropout_input=0.1)
#f5 = keras_model_notcollapsed.get_layer("dropout_grp_000")
#assert f5.rate > 0.
#f5.rate = 0.
try:
from keras.utils import plot_model
plot_model(keras_model_collapsed,
to_file="rai_faster_collapsed.png")
plot_model(keras_model_notcollapsed,
to_file="rai_faster_notcollapsed.png")
except:
pass