test_mnist

#!/bin/bash

INFO=$'
Usage:

\ttest_mnist <saved.model>
\ttest_mnist <saved.model> capture
\ttest_mnist <saved.model> <images dir>
'

if [ ! -d "$1" ]; then
    echo "$INFO" 1>&2
    exit 1
fi

SCRIPT_DIR=$(readlink -m "$0/..") || exit 1
export PATH="$SCRIPT_DIR:$PATH"

TEST_PY='
import sys
import numpy as np
import tensorflow.examples.tutorials.mnist as tf_mnist

import loadsave
from utils import *

mnist = tf_mnist.input_data.read_data_sets("__mnist__")

with loadsave.load(sys.argv[1]) as s:
    data = np.reshape(mnist.test.images, [-1, 28, 28]), mnist.test.labels
    total_ok = 0

    for x, t in random_batches(data, 1000):
        y = s.run(s.y, feed_dict = {s.x: x})
        mistakes = np.not_equal(np.argmax(y, 1), t)
        total_ok += len(x) - np.sum(mistakes)
        write_array(sys.stdout.buffer, x[mistakes])

    print_info("Accuracy: %.2f%% (%d/%d)" % (
        100.0 * (total_ok / len(data[0])),
        total_ok,
        len(data[0])
    ))
'

function print_mistakes() {
    echo "$TEST_PY" | python3 - "$@"
}

if [ "x$2" = xcapture ]; then
    image_capture 200,200 | \
        image_write 200,200 -resize 28x28 GRAY:- | \
        model_run "$1" x:0 y:0 | \
        image_show 10,1 || exit 1
    exit 0
fi

if [ "x$2" != x -a -d "$2" ]; then
    N=$(ls "$2" | wc -l)
    if ! [ "$N" -ge 1 ]; then
        echo "Error: no images found in '$2'" 1>&2
        exit 1
    fi
    image_read 28,28 "$2" | \
        model_run "$1" x:0 y:0 | \
        od -v -An -t u1 -w10 || exit 1
    exit 0
fi

if [ "x$2" != x ]; then
    mkdir "$2" || exit 1
    print_mistakes "$1" | \
        image_write 28,28 "$2/%05d.png" || exit 1
    exit 0
fi

print_mistakes "$1" >/dev/null || exit 1
exit 0