def test(self, num_function, num_shots, test_shots, epoch=1, input_range=(-2., 2.)):
fig = plt.figure(figsize=(10,10))
# a = int(np.sqrt(num_function))
for i in range(num_function):
# ax = fig.add_subplot(a,a,i+1)
ax = fig.add_subplot(4,3,i+1)
sampler = self.eval_set.sample(1)[0]
c = [1, 4, 8, 16, 32, 64]
num_shots = c[(i%6)]
X_value, y_value = sampler.sample(num_shots+test_shots)
X_c_value, X_t_value = X_value[:num_shots], X_value[num_shots:]
y_c_value, y_t_value = y_value[:num_shots], y_value[num_shots:]
m = self.parallel_models[0]
X_gt, y_gt = sampler.get_all_samples()
ax.plot(*sort_x(X_gt[:,0], y_gt), "-")
ax.scatter(X_c_value[:,0], y_c_value)
X_eval = np.linspace(self.eval_set.input_range[0], self.eval_set.input_range[1], num=100)[:,None]
# step 1
y_hat = m.predict(self.session, X_c_value, y_c_value, X_eval, step=1)
ax.plot(X_eval[:,0], y_hat, ":", color='gray', alpha=0.3)
# step 5
y_hat = m.predict(self.session, X_c_value, y_c_value, X_eval, step=5)
ax.plot(X_eval[:,0], y_hat, "--", color='gray', alpha=0.3)
# step 10
y_hat = m.predict(self.session, X_c_value, y_c_value, X_eval, step=10)
ax.plot(X_eval[:,0], y_hat, "-", color='gray', alpha=0.3)
fig.savefig("figs/maml-{0}-{1}.pdf".format(self.eval_set.dataset_name, epoch))
plt.close()
def test(self, num_function, num_shots, test_shots, epoch=1):
fig = plt.figure(figsize=(10, 10))
# a = int(np.sqrt(num_function))
for i in range(num_function):
# ax = fig.add_subplot(a,a,i+1)
ax = fig.add_subplot(4, 3, i + 1)
sampler = self.eval_set.sample(1)[0]
#
# total_shots = np.random.randint(low=10, high=50)
# num_shots = np.random.randint(low=total_shots-10, high=total_shots-1)
# test_shots = total_shots - num_shots
#num_shots = np.random.randint(low=1, high=30)
#test_shots = np.random.randint(low=1, high=10)
# num_shots, test_shots = 20, 10
c = [1, 4, 8, 16, 32, 64]
num_shots = c[(i % 6)]
#
X_value, y_value = sampler.sample(num_shots + test_shots)
X_c_value, X_t_value = X_value[:num_shots], X_value[num_shots:]
y_c_value, y_t_value = y_value[:num_shots], y_value[num_shots:]
m = self.parallel_models[0]
X_gt, y_gt = sampler.get_all_samples()
ax.plot(*sort_x(X_gt[:, 0], y_gt), "-")
ax.scatter(X_c_value[:, 0], y_c_value)
#ax.plot(*sort_x(X_value[:,0], y_value), "+")
for k in range(20):
X_eval = np.linspace(self.eval_set.input_range[0],
self.eval_set.input_range[1],
num=100)[:, None]
y_hat = m.predict(self.session, X_c_value, y_c_value, X_eval)
ax.plot(X_eval[:, 0], y_hat, "-", color='gray', alpha=0.3)
#ax.plot(X_value[:,0], y_hat, "-", color='gray', alpha=0.3)
fig.savefig("figs/np-{0}-{1}.pdf".format(self.eval_set.dataset_name,
epoch))
plt.close()
def run_eval(self, num_func, num_shots=1, test_shots=50):
m = self.parallel_models[0]
saver = tf.train.Saver(var_list=self.variables)
ckpt_file = self.save_dir + '/params.ckpt'
print('restoring parameters from', ckpt_file)
saver.restore(self.session, ckpt_file)
evals = []
for _ in range(num_func):
sampler = self.eval_set.sample(1)[0]
X_value, y_value = sampler.sample(num_shots+test_shots)
X_c_value, X_t_value = X_value[:num_shots], X_value[num_shots:]
y_c_value, y_t_value = y_value[:num_shots], y_value[num_shots:]
y_t_hat = m.predict(self.session, X_c_value, y_c_value, X_t_value, step=10)
evals.append(np.mean(np.power(y_t_value - y_t_hat, 2)))
eval = np.nanmean(evals)
print(".......... EVAL : num_func {0} num_shots {1} test_shots {2}............".format(num_func, num_shots, test_shots))
print("\t{0}".format(eval))
fig = plt.figure(figsize=(10,10))
for i in range(4):
# ax = fig.add_subplot(a,a,i+1)
ax = fig.add_subplot(4,1,i+1)
sampler = self.eval_set.sample(1)[0]
c = [5, 10, 15, 20]
num_shots = c[(i%4)]
test_shots = 0
X_value, y_value = sampler.sample(num_shots+test_shots)
X_c_value, X_t_value = X_value[:num_shots], X_value[num_shots:]
y_c_value, y_t_value = y_value[:num_shots], y_value[num_shots:]
m = self.parallel_models[0]
X_gt, y_gt = sampler.get_all_samples()
ax.plot(*sort_x(X_gt[:,0], y_gt), "-")
ax.scatter(X_c_value[:,0], y_c_value)
X_eval = np.linspace(self.eval_set.input_range[0], self.eval_set.input_range[1], num=100)[:,None]
# step 1
y_hat = m.predict(self.session, X_c_value, y_c_value, X_eval, step=1)
ax.plot(X_eval[:,0], y_hat, ":", color='gray', alpha=0.3)
# step 5
y_hat = m.predict(self.session, X_c_value, y_c_value, X_eval, step=5)
ax.plot(X_eval[:,0], y_hat, "--", color='gray', alpha=0.3)
# step 10
y_hat = m.predict(self.session, X_c_value, y_c_value, X_eval, step=10)
ax.plot(X_eval[:,0], y_hat, "-", color='gray', alpha=0.3)
fig.savefig("figs/maml-{0}-{1}.pdf".format(self.eval_set.dataset_name, "eval"))
plt.close()