def deep_iv_fit(x, z, t, y, epochs=100, hidden=[128, 64, 32]):
from deepiv.models import Treatment, Response
import deepiv.architectures as architectures
import deepiv.densities as densities
from keras.layers import Input, Dense
from keras.models import Model
from keras.layers.merge import Concatenate
n = z.shape[0]
dropout_rate = min(1000. / (1000. + n), 0.5)
batch_size = 100
images = False
act = "relu"
n_components = 10
instruments = Input(shape=(z.shape[1], ), name="instruments")
features = Input(shape=(x.shape[1], ), name="features")
treatment_input = Concatenate(axis=1)([instruments, features])
est_treat = architectures.feed_forward_net(
treatment_input,
lambda x: densities.mixture_of_gaussian_output(x, n_components),
hidden_layers=hidden,
dropout_rate=dropout_rate,
l2=0.0001,
activations=act)
treatment_model = Treatment(inputs=[instruments, features],
outputs=est_treat)
treatment_model.compile('adam',
loss="mixture_of_gaussians",
n_components=n_components)
treatment_model.fit([z, x], t, epochs=epochs, batch_size=batch_size)
# Build and fit response model
treatment = Input(shape=(t.shape[1], ), name="treatment")
response_input = Concatenate(axis=1)([features, treatment])
est_response = architectures.feed_forward_net(response_input,
Dense(1),
activations=act,
hidden_layers=hidden,
l2=0.001,
dropout_rate=dropout_rate)
response_model = Response(treatment=treatment_model,
inputs=[features, treatment],
outputs=est_response)
response_model.compile('adam', loss='mse')
response_model.fit([z, x],
y,
epochs=epochs,
verbose=1,
batch_size=batch_size,
samples_per_batch=2)
return response_model
def conv_embedding(images,
output,
other_features=[],
dropout_rate=0.1,
embedding_dropout=0.1,
embedding_l2=0.05,
constrain_norm=True):
print("Building conv net")
x_embedding = architectures.convnet(images,
Dense(64, activation='linear'),
dropout_rate=embedding_dropout,
activations='relu',
l2_rate=embedding_l2,
constrain_norm=constrain_norm)
if len(other_features) > 0:
embedd = Concatenate(axis=1)([x_embedding] + other_features)
else:
embedd = x_embedding
out = architectures.feed_forward_net(embedd,
output,
hidden_layers=[32],
dropout_rate=dropout_rate,
activations='relu',
constrain_norm=constrain_norm)
return out
Response:{y}".format(**{'x':x.shape, 'z':z.shape,
't':t.shape, 'y':y.shape}))
# Build and fit treatment model
instruments = Input(shape=(z.shape[1],), name="instruments")
features = Input(shape=(x.shape[1],), name="features")
treatment_input = Concatenate(axis=1)([instruments, features])
hidden = [128, 64, 32]
act = "relu"
n_components = 10
est_treat = architectures.feed_forward_net(treatment_input, lambda x: densities.mixture_of_gaussian_output(x, n_components),
hidden_layers=hidden,
dropout_rate=dropout_rate, l2=0.0001,
activations=act)
treatment_model = Treatment(inputs=[instruments, features], outputs=est_treat)
treatment_model.compile('adam',
loss="mixture_of_gaussians",
n_components=n_components)
treatment_model.fit([z, x], t, epochs=epochs, batch_size=batch_size)
# Build and fit response model
treatment = Input(shape=(t.shape[1],), name="treatment")
response_input = Concatenate(axis=1)([features, treatment])
est_response = architectures.feed_forward_net(response_input, Dense(1),