def sample_vae(vae, dirs, count):
z_size = vae.z_dim
batch_size = count
z = np.random.normal(size=(batch_size, z_size))
samples = vae.decode(z)
input_dim = samples.shape[1:]
n = batch_size
plt.figure(figsize=(20, 6), tight_layout=False)
plt.title('VAE samples')
for i in range(n):
ax = plt.subplot(3, n, i + 1)
plt.imshow(samples[i].reshape(input_dim[0], input_dim[1],
input_dim[2]))
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
if 0 == i:
ax.set_title("Random")
data = {}
orig = []
opts = {'cfg': cfg, 'categorical': False}
records = gather_records(cfg, dirs, opts, verbose=True)
collate_records(records, data, opts)
keys = list(data.keys())
keys = shuffle(keys)[0:batch_size]
for key in keys:
filename = data[key]['image_path']
img_arr = load_scaled_image_arr(filename, cfg)
orig.append(img_arr)
orig = np.array(orig)
recon = vae.decode(vae.encode(orig))
for i in range(n):
ax = plt.subplot(3, n, n + i + 1)
plt.imshow(orig[i].reshape(input_dim[0], input_dim[1], input_dim[2]))
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
if 0 == i:
ax.set_title("Real")
ax = plt.subplot(3, n, (2 * n) + i + 1)
plt.imshow(recon[i].reshape(input_dim[0], input_dim[1], input_dim[2]))
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
if 0 == i:
ax.set_title("Reconstructed")
plt.savefig("samples_vae.png")
plt.show()
def load_tubs(cfg, tub_names, kl, aux_name=None, pilot=False):
opts = { 'cfg' : cfg}
opts['categorical'] = False
opts['keras_pilot'] = kl # This will be needed for generator
opts['continuous'] = False
gen_records = {}
records = gather_records(cfg, tub_names, verbose=True)
collate_records(records, gen_records, opts)
# These options should be part of the KerasPilot class
if type(kl.model.output) is list:
opts['model_out_shape'] = (2, 1)
else:
opts['model_out_shape'] = kl.model.output.shape
if type(kl.model.input) is list:
opts['model_in_shape'] = (2, 1)
else:
opts['model_in_shape'] = kl.model.input.shape
opts['has_imu'] = type(kl) is KerasIMU
opts['has_bvh'] = type(kl) is KerasBehavioral
opts['img_out'] = type(kl) is KerasLatent
opts['vae_out'] = type(kl) is KerasVAE
train_gen = vae_generator(cfg, gen_records, cfg.BATCH_SIZE, isTrainSet=True, aug=False, aux=aux_name, pilot=pilot)
val_gen = vae_generator(cfg, gen_records, cfg.BATCH_SIZE, isTrainSet=False, aug=False, aux=aux_name, pilot=pilot)
num_train = 0
num_val = 0
for key, _record in gen_records.items():
if _record['train'] == True:
num_train += 1
else:
num_val += 1
steps = num_train // cfg.BATCH_SIZE
val_steps = num_val // cfg.BATCH_SIZE
print( "Num/Steps: {} {}".format( num_train, steps ) )
print( "Val/Steps: {} {}".format( num_val, val_steps ) )
return train_gen, val_gen, steps, val_steps
def make_generator(cfg, tub_names, verbose=True):
opts = {'cfg': cfg}
opts['categorical'] = False
opts['continuous'] = False
gen_records = {}
records = gather_records(cfg, tub_names, verbose=True)
collate_records(records, gen_records, opts)
train_gen = vae_generator(cfg,
gen_records,
cfg.BATCH_SIZE,
isTrainSet=True,
aug=False)
val_gen = vae_generator(cfg,
gen_records,
cfg.BATCH_SIZE,
isTrainSet=False,
aug=False)
return train_gen, val_gen, gen_records
def test_train_TrainTestSplit_continuous(tub_path):
# Check whether the Train-Test splitting is working correctly when a dataset is extended.
initial_records = 100
# Setup the test data
t = create_sample_tub(tub_path, records=initial_records)
assert t is not None
# Import the configuration
import donkeycar.templates.cfg_complete as cfg
# Initial Setup
gen_records = {}
opts = {'categorical': False}
opts['cfg'] = cfg
# Perform the initial split
print()
print("Initial split of {} records to {} Test-Train split...".format(
initial_records, opts['cfg'].TRAIN_TEST_SPLIT))
print()
records = gather_records(cfg, tub_path, opts, verbose=True)
assert len(records) == initial_records
collate_records(records, gen_records, opts)
ratio = calculate_TrainTestSplit(gen_records)
assert ratio == cfg.TRAIN_TEST_SPLIT
# Add some more records and recheck the ratio (only the NEW records should be added)
additional_records = 200
print()
print(
"Added an extra {} records, aiming for overall {} Test-Train split...".
format(additional_records, opts['cfg'].TRAIN_TEST_SPLIT))
print()
create_sample_tub(tub_path, records=additional_records)
records = gather_records(cfg, tub_path, opts, verbose=True)
assert len(records) == (initial_records + additional_records)
collate_records(records, gen_records, opts)
ratio = calculate_TrainTestSplit(gen_records)
assert ratio == cfg.TRAIN_TEST_SPLIT
def test_train_TrainTestSplit_simple(tub_path):
# Check whether the Train-Test splitting is working correctly on a dataset.
initial_records = 100
# Setup the test data
t = create_sample_tub(tub_path, records=initial_records)
assert t is not None
# Import the configuration
import donkeycar.templates.cfg_complete as cfg
# Initial Setup
opts = {'categorical': False}
opts['cfg'] = cfg
orig_TRAIN_TEST_SPLIT = cfg.TRAIN_TEST_SPLIT
records = gather_records(cfg, tub_path, opts, verbose=True)
assert len(records) == initial_records
# Attempt a 50:50 split
gen_records = {}
cfg.TRAIN_TEST_SPLIT = 0.5
print()
print("Testing a {} Test-Train split...".format(
opts['cfg'].TRAIN_TEST_SPLIT))
print()
collate_records(records, gen_records, opts)
ratio = calculate_TrainTestSplit(gen_records)
assert ratio == cfg.TRAIN_TEST_SPLIT
# Attempt a split based on config file (reset of the record set)
gen_records = {}
cfg.TRAIN_TEST_SPLIT = orig_TRAIN_TEST_SPLIT
print()
print("Testing a {} Test-Train split...".format(
opts['cfg'].TRAIN_TEST_SPLIT))
print()
collate_records(records, gen_records, opts)
ratio = calculate_TrainTestSplit(gen_records)
assert ratio == cfg.TRAIN_TEST_SPLIT
except FileNotFoundError:
cfg = dk.load_config("config.py") # retry in the current directory
tub_names = preprocessFileList(args.file)
input_shape = (cfg.IMAGE_W, cfg.IMAGE_H, cfg.IMAGE_DEPTH)
# Code for multiple inputs: http://digital-thinking.de/deep-learning-combining-numerical-and-text-features-in-deep-neural-networks/
aux_out = 0
if args.aux is not None:
aux_out = 7 # need to get number of aux outputs from data
opts = {'cfg': cfg}
opts['categorical'] = False
opts['continuous'] = False
gen_records = {}
records = gather_records(cfg, tub_names, verbose=True)
collate_records(records, gen_records, opts)
train_gen = vae_generator(cfg,
gen_records,
cfg.BATCH_SIZE,
isTrainSet=True,
aug=False,
aux=args.aux,
pilot=True)
for X, y in train_gen:
print("X {} {}".format(type(X[0]), X[0].shape))
img = y['main_output'][0]
print("main {} min/max/avg: {}/{}/{}".format(img.shape, np.min(img),
np.max(img),
np.mean(img)))
if 'aux_output' in y:
