h3_t = h3.fprop([[x_t, h2_t], [h3_tm1, h1_tm1, h2_tm1]])
return h1_t, h2_t, h3_t
((h1_temp, h2_temp, h3_temp), updates) = theano.scan(fn=inner_fn,
sequences=[x],
outputs_info=[s1_0, s2_0, s3_0])
ts, _, _ = y.shape
post_scan_shape = ((ts*batch_size, -1))
h1_in = h1_temp.reshape(post_scan_shape)
h2_in = h2_temp.reshape(post_scan_shape)
h3_in = h3_temp.reshape(post_scan_shape)
y_hat_in = output.fprop([h1_in, h2_in, h3_in])
cost = NllMulInd(y.flatten(), y_hat_in)
cost = cost.mean()
cost.name = 'cost'
model.inputs = [x, y]
model._params = params
model.nodes = nodes
model.set_updates(update_list)
optimizer = Adam(
lr=0.001
)
extension = [
GradientClipping(batch_size=batch_size),
EpochCount(100),
# You will fill in a list of nodes
nodes = [h1, output]
# Initalize the nodes
for node in nodes:
node.initialize()
params = flatten([node.get_params().values() for node in nodes])
# Build the Theano computational graph
h1_out = h1.fprop([x])
y_hat = output.fprop([h1_out])
# Compute the cost
cost = NllMulInd(y, y_hat).mean()
err = error(predict(y_hat), y)
cost.name = 'cross_entropy'
err.name = 'error_rate'
model.inputs = [x, y]
model._params = params
model.nodes = nodes
# Define your optimizer: Momentum (Nesterov), RMSProp, Adam
optimizer = RMSProp(lr=0.001)
extension = [
GradientClipping(),
EpochCount(40),
Monitoring(freq=100,
return h1_t, h2_t, h3_t
((h1_temp, h2_temp, h3_temp),
updates) = theano.scan(fn=inner_fn,
sequences=[x],
outputs_info=[s1_0, s2_0, s3_0])
ts, _, _ = y.shape
post_scan_shape = ((ts * batch_size, -1))
h1_in = h1_temp.reshape(post_scan_shape)
h2_in = h2_temp.reshape(post_scan_shape)
h3_in = h3_temp.reshape(post_scan_shape)
y_hat_in = output.fprop([h1_in, h2_in, h3_in])
cost = NllMulInd(y.flatten(), y_hat_in)
cost = cost.mean()
cost.name = 'cost'
model.inputs = [x, y]
model._params = params
model.nodes = nodes
model.set_updates(update_list)
optimizer = Adam(lr=0.001)
extension = [
GradientClipping(batch_size=batch_size),
EpochCount(100),
Monitoring(freq=100, ddout=[cost]),
Picklize(freq=100, path=save_path)
# Initalize the nodes
for node in nodes:
node.initialize()
# Collect parameters
params = flatten([node.get_params().values() for node in nodes])
# Build the Theano computational graph
h1_out = h1.fprop([x])
d1_out = d1.fprop([h1_out])
h2_out = h2.fprop([d1_out])
d2_out = d2.fprop([h2_out])
y_hat = output.fprop([d2_out])
# Compute the cost
cost = NllMulInd(y, y_hat).mean()
err = error(predict(y_hat), y)
cost.name = 'cross_entropy'
err.name = 'error_rate'
d1.set_mode(1)
d2.set_mode(1)
mn_h1_out = h1.fprop([mn_x])
mn_h2_out = h2.fprop([mn_h1_out])
mn_y_hat = output.fprop([mn_h2_out])
mn_cost = NllMulInd(mn_y, mn_y_hat).mean()
mn_err = error(predict(mn_y_hat), mn_y)
mn_cost.name = 'cross_entropy'
mn_err.name = 'error_rate'
# Initalize the nodes
for node in nodes:
node.initialize()
# Collect parameters
params = flatten([node.get_params().values() for node in nodes])
# Build the Theano computational graph
h1_out = h1.fprop([x])
d1_out = d1.fprop([h1_out])
h2_out = h2.fprop([d1_out])
d2_out = d2.fprop([h2_out])
y_hat = output.fprop([d2_out])
# Compute the cost
cost = NllMulInd(y, y_hat).mean()
err = error(predict(y_hat), y)
cost.name = 'cross_entropy'
err.name = 'error_rate'
d1.set_mode(1)
d2.set_mode(1)
mn_h1_out = h1.fprop([mn_x])
mn_h2_out = h2.fprop([mn_h1_out])
mn_y_hat = output.fprop([mn_h2_out])
mn_cost = NllMulInd(mn_y, mn_y_hat).mean()
mn_err = error(predict(mn_y_hat), mn_y)
mn_cost.name = 'cross_entropy'
mn_err.name = 'error_rate'
