def load_data():
"""Loads data"""
x_train = preprocessing.get_x("data/train")
y_train = preprocessing.get_y1("data/train")
x_test = preprocessing.get_x("data/test")
y_test = preprocessing.get_y1("data/test")
return x_train, y_train, x_test, y_test
def accuracy_calculate(x, y, L):
x = preprocessing.get_x("data/train")
y = preprocessing.get_y1("data/train")
x = x[y[:, 0] == L]
y = y[y[:, 0] == L]
indices = np.random.choice(len(x), 2000, replace=False)
examples = x[indices]
answers = y[indices]
model = Model()
with tf.Session(graph=model.graph) as session:
# Restore model with 82% accuracy on test data
model.saver.restore(session, "./try2.ckpt")
log_1, log_2, log_3, log_4, log_5 = session.run([
model.logits_1, model.logits_2, model.logits_3, model.logits_4,
model.logits_5
],
feed_dict={
model.x: examples,
model.keep_prob:
4.0
})
# Make predictions
predictions = predictions(log_1, log_2, log_3, log_4, log_5)
acc = 0.0
r1 = 0.0
r2 = 0.0
for i, el in enumerate(indices):
number = predictions[i][predictions[i] < 10]
ans = answers[i][answers[i] < 10]
if len(ans[1:]) != len(number):
r1 += 1.0
continue
if not (ans[1:] - number).any():
acc += 1.0
else:
r2 += 1.0
try:
ans = int(''.join(ans[1:]))
number = int(''.join(number))
if ans == number:
acc += 1.0
except:
pass
fac = len(x) / 100.0
print('acc:%f' % (acc / fac))
print('r1:%f' % (r1 / fac))
print('r2:%f' % (r2 / fac))
"""Try to predict 7 random examples from test data"""
import scipy.misc
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import preprocessing
from model import Model
from train import predictions
x = preprocessing.get_x("data/test")
indices = np.random.choice(len(x), 7)
examples = x[indices]
model = Model()
with tf.Session(graph=model.graph) as session:
# Restore model with 82% accuracy on test data
model.saver.restore(session, "./try2.ckpt")
log_1, log_2, log_3, log_4, log_5 = session.run([
model.logits_1, model.logits_2, model.logits_3, model.logits_4,
model.logits_5
],
feed_dict={
model.x: examples,
model.keep_prob: 1.0
})
# Make predictions on 7 random examples from test data
def ckpt_accuracy(L, ckpt, cut=True):
x = preprocessing.get_x("data/test")
y = preprocessing.get_y1("data/test")
x = x[y[:, 0] == L]
y = y[y[:, 0] == L]
if len(x) > 2000:
indices = np.random.choice(len(x), 2000, replace=False)
examples = x[indices]
answers = y[indices]
else:
examples = x
answers = y
model = Model()
with tf.Session(graph=model.graph) as session:
# Restore model with filename as ckpt
model.saver.restore(session, "./" + ckpt)
log_1, log_2, log_3, log_4, log_5 = session.run([
model.logits_1, model.logits_2, model.logits_3, model.logits_4,
model.logits_5
],
feed_dict={
model.x: examples,
model.keep_prob:
4.0
})
# Make predictions
preds = predictions(log_1, log_2, log_3, log_4, log_5)
#cut the digit longer than L
if cut:
preds = preds[:, :L]
#r1 is first type error that predict wrong length
#r2 is second type error that predict wrong number
acc = 0.0
r1 = 0.0
r2 = 0.0
for i, el in enumerate(indices):
number = preds[i][preds[i] < 10]
ans = answers[i][answers[i] < 10]
if len(ans[1:]) != len(number):
r1 += 1.0
continue
if not (ans[1:] - number).any():
acc += 1.0
else:
r2 += 1.0
try:
ans = int(''.join(ans[1:]))
number = int(''.join(number))
if ans == number:
acc += 1.0
except:
pass
fac = len(examples) / 100.0
print('ckpt_acc:%f' % (acc / fac))
print('r1:%f' % (r1 / fac))
print('r2:%f' % (r2 / fac))
return acc, r1, r2