load_status = transcriber1.load_weights("%s/ckpt" % weights_dir)
load_status.assert_existing_objects_matched()
transcriber2 = transcriberModel()
weights_dir = ("%s/Robot_Rainfall_Rescue/models/ATB2_retuned/" +
"Epoch_%04d") % (
os.getenv("SCRATCH"),
int(args.epoch),
)
if args.epoch < 50:
load_status = transcriber2.load_weights("%s/ckpt" % weights_dir)
load_status.assert_existing_objects_matched()
# Make the probability matrix
testImages = getImageDataset(purpose="test", nImages=args.nimages)
testNumbers = getNumbersDataset(purpose="test", nImages=args.nimages)
testData = tf.data.Dataset.zip((testImages, testNumbers))
# Need the size of the dataset
if args.nimages is None:
nimages = sum(1 for _ in testData)
else:
nimages = args.nimages
count = numpy.zeros(10)
pmatrix = numpy.zeros((10, 10))
dcount = 0
plmatrix = numpy.zeros((436))
origN = None
encN = None
for testCase in testData:
image = testCase[0]
weights_dir = ("%s/Robot_Rainfall_Rescue/models/ATB2_retuned/Epoch_%04d") % (
os.getenv("SCRATCH"),
args.epoch - 1,
)
load_status = transcriber.load_weights("%s/ckpt" % weights_dir)
# Check the load worked
load_status.assert_existing_objects_matched()
# Get test case number args.image
testImage = getImageDataset(subdir="unperturbed",
purpose="test",
nImages=args.image + 1)
testImage = testImage.batch(1)
originalImage = next(itertools.islice(testImage, args.image, args.image + 1))
testNumbers = getNumbersDataset(subdir="unperturbed",
purpose="test",
nImages=args.image + 1)
testNumbers = testNumbers.batch(1)
originalNumbers = next(
itertools.islice(testNumbers, args.image, args.image + 1))
# Run that test image through the transcriber
encoded = transcriber.predict_on_batch(originalImage)
# Plot original image on the left - make an image from the encoded numbers
# on the right
fig = Figure(
figsize=(16.34, 10.56),
dpi=100,
facecolor="white",
edgecolor="black",
# Length of an epoch - if None, same as nTrainingImages
nImagesInEpoch = 1000
if nImagesInEpoch is None:
nImagesInEpoch = nTrainingImages
# Dataset parameters
bufferSize = 100 # Shouldn't make much difference (data is random)
batchSize = 2 # Arbitrary
# Set up the training data
imageData = getImageDataset(subdir="unperturbed",
purpose="training",
nImages=nTrainingImages).repeat()
numbersData = getNumbersDataset(subdir="unperturbed",
purpose="training",
nImages=nTrainingImages).repeat()
trainingData = tf.data.Dataset.zip((imageData, numbersData))
trainingData = trainingData.shuffle(bufferSize).batch(batchSize)
# Set up the test data
testImageData = getImageDataset(subdir="unperturbed",
purpose="test",
nImages=nTestImages).repeat()
testNumbersData = getNumbersDataset(subdir="unperturbed",
purpose="test",
nImages=nTestImages).repeat()
testData = tf.data.Dataset.zip((testImageData, testNumbersData))
testData = testData.batch(batchSize)
# Instantiate the model