# load data
file_name = os.path.join(rlvision.RLVISION_DATA, "chain_data", "grid16_po.pkl")
model_path = os.path.join(rlvision.RLVISION_MODEL, "grid16-po")
if not os.path.isdir(model_path):
os.makedirs(model_path)
# parameters
batch_size = 256
nb_epochs = 80
print('# Minibatch-size: {}'.format(batch_size))
print('# epoch: {}'.format(nb_epochs))
print('')
train, test, _ = utils.process_map_data(file_name)
model = vin_model(l_s=train[0].shape[2], k=20)
model.compile(optimizer="adam",
loss='categorical_crossentropy',
metrics=['accuracy'])
model_file = os.path.join(model_path,
"vin-model-po-16-{epoch:02d}-{acc:.2f}.h5")
checkpoint = ModelCheckpoint(model_file,
monitor='acc',
verbose=1,
save_best_only=True,
mode='max',
save_weights_only=True)
history = model.fit([
train[0].transpose(
reward = get_layer_output(model, 'reward', im_ary)
value = get_layer_output(model, 'value{}'.format(k), im_ary)
reward = np.reshape(reward, im.shape[1:])
value = np.reshape(value, im.shape[1:])
return action, reward, value
file_name = os.path.join(rlvision.RLVISION_DATA, "chain_data",
"grid28_with_idx.pkl")
model_file = os.path.join(rlvision.RLVISION_MODEL, "grid28-po",
"vin-model-po-28-77-0.89.h5")
im_data, state_data, label_data, sample_idx = process_map_data(
file_name, return_full=True)
model = vin_model(l_s=im_data.shape[2], k=20)
model.load_weights(model_file)
sampler = GridSampler(im_data, state_data, label_data, sample_idx, (28, 28))
gt_collector = []
po_collector = []
diff_collector = []
grid, state, label, goal = sampler.get_grid(77)
gt_collector.append(state)
step_map = np.zeros((2, 28, 28))
step_map[0] = np.ones((28, 28))
step_map[1] = grid[1]
pos = [state[0, 0], state[0, 1]]
reward = get_layer_output(model, 'reward', im_ary)
value = get_layer_output(model, 'value{}'.format(k), im_ary)
reward = np.reshape(reward, im.shape[1:])
value = np.reshape(value, im.shape[1:])
return action, reward, value
# load data
file_name = os.path.join(rlvision.RLVISION_DATA, "chain_data",
"grid16_with_idx.pkl")
model_file = os.path.join(rlvision.RLVISION_MODEL, "vin_model_po_16.h5")
k = 20
_, test, _ = process_map_data(file_name)
model = vin_model(l_s=test[0].shape[2], k=k)
model.load_weights(model_file)
for d in zip(*test):
im = d[0]
print im.shape
pos = d[1]
print pos
action, reward, value = predict(im, pos, model, k)
path = [tuple(pos)]
for _ in range(30):
if im[1][pos[1], pos[0]] == 1:
break
action, _, _ = predict(im, pos, model, k)
dx, dy = get_action(action)