def run_model(n,
train_steps=50 * 1000,
verbose=2,
lr=0.001,
anneal=True,
plot=False,
batch_size=32,
epochs=1):
model = student.Model([20, n, n, 1],
softmax=False,
loss='power',
power=power)
loc = 0
bs = 200 #batch_size
tt = 200000 #training time (steps)
lr = 0.01
history = model.train(teach.data_matrix_x[loc:loc + bs * tt],
teach.data_matrix_y[loc:loc + bs * tt],
learning_rate=lr,
batch_size=bs,
verbose=verbose,
epochs=1)
loc += bs * tt
bs = 1000 #batch_size
tt = 20000 #training time (steps)
lr = 0.01
history = model.train(teach.data_matrix_x[loc:loc + bs * tt],
teach.data_matrix_y[loc:loc + bs * tt],
learning_rate=lr,
batch_size=bs,
verbose=verbose,
epochs=1)
loc += bs * tt
bs = 400 #batch_size
tt = 20000 #training time (steps)
lr = 0.001
history = model.train(teach.data_matrix_x[loc:loc + bs * tt],
teach.data_matrix_y[loc:loc + bs * tt],
learning_rate=lr,
batch_size=bs,
verbose=verbose,
epochs=1)
loc += bs * tt
x1, y1 = teach.predict_from_matrix(batch_size=10000)
loss = model.evaluate(x1, y1)
if plot:
student.graph_sample(parent, model, folder='plot')
#return [loss,model,model.count_params()-n-1]
return loss, model
def run_model(n,
train_steps=50 * 1000,
verbose=2,
lr=0.001,
anneal=True,
plot=False,
batch_size=32,
epochs=1):
model = student.Model([9, n, n, 1], softmax=False, loss='power')
dm_x = np.concatenate(
(teach1.data_matrix_x[:, 0:3], teach2.data_matrix_x[:, 3:6],
teach3.data_matrix_x[:, 6:9]),
axis=1)
dm_y = teach1.data_matrix_y + teach2.data_matrix_y + teach3.data_matrix_y
loc = 0
bs = 200 #batch_size
tt = 200000 #training time (steps)
lr = 0.01
history = model.train(dm_x[loc:loc + bs * tt],
dm_y[loc:loc + bs * tt],
learning_rate=lr,
batch_size=bs,
verbose=verbose,
epochs=1)
loc += bs * tt
bs = 1000 #batch_size
tt = 20000 #training time (steps)
lr = 0.01
history = model.train(dm_x[loc:loc + bs * tt],
dm_y[loc:loc + bs * tt],
learning_rate=lr,
batch_size=bs,
verbose=verbose,
epochs=1)
loc += bs * tt
bs = 4000 #batch_size
tt = 20000 #training time (steps)
lr = 0.001
history = model.train(dm_x[loc:loc + bs * tt],
dm_y[loc:loc + bs * tt],
learning_rate=lr,
batch_size=bs,
verbose=verbose,
epochs=1)
loc += bs * tt
x1, y1 = teach1.predict_from_matrix(batch_size=10000)
x2, y2 = teach2.predict_from_matrix(batch_size=10000)
x3, y3 = teach3.predict_from_matrix(batch_size=10000)
x1 = np.concatenate((x1[:, 0:3], x2[:, 3:6], x3[:, 6:9]), axis=1)
y1 = y1 + y2 + y3
loss = model.evaluate(x1, y1)
if plot:
student.graph_sample(parent, model, folder='plot')
#return [loss,model,model.count_params()-n-1]
return loss, model
