def resnet_exp(model_name, data_name, channel_id, stages, blocks, filter_list,
nb_epoch, batch_size, frame_cut):
"""Perform ResNet experiment."""
model_path = os.path.join(spiker.SPIKER_EXPS, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
else:
raise ValueError("[MESSAGE] This experiment has been done before."
" Create a new config model if you need.")
model_pic = os.path.join(model_path, model_name+"-model-pic.png")
model_file_base = os.path.join(model_path, model_name)
# print model info
print("[MESSAGE] Model Name: %s" % (model_name))
print("[MESSAGE] Number of epochs: %d" % (nb_epoch))
print("[MESSAGE] Batch Size: %d" % (batch_size))
print("[MESSAGE] Number of stages: %d" % (stages))
print("[MESSAGE] Number of blocks: %d" % (blocks))
# load data
data_path = os.path.join(spiker.SPIKER_DATA, "ddd17",
data_name)
if not os.path.isfile(data_path):
raise ValueError("This dataset does not exist at %s" % (data_path))
print("[MESSAGE] Dataset %s" % (data_path))
assert len(frame_cut) == 2
print("[MESSAGE] Frame cut: first at %d, last at %d"
% (frame_cut[0], -frame_cut[1]))
frames, steering = ddd17.prepare_train_data(
data_path, y_name="accel",
frame_cut=frame_cut)
frames /= 255.
frames -= np.mean(frames, keepdims=True)
num_samples = frames.shape[0]
num_train = int(num_samples*0.7)
X_train = frames[:num_train]
Y_train = steering[:num_train]
X_test = frames[num_train:]
Y_test = steering[num_train:]
del frames, steering
if channel_id != 2:
X_train = X_train[:, :, :, channel_id][..., np.newaxis]
X_test = X_test[:, :, :, channel_id][..., np.newaxis]
print("[MESSAGE] Number of samples %d" % (num_samples))
print("[MESSAGE] Number of train samples %d" % (X_train.shape[0]))
print("[MESSAGE] Number of test samples %d" % (X_test.shape[0]))
# setup image shape
input_shape = (X_train.shape[1], X_train.shape[2], X_train.shape[3])
# Build model
model = resnet.resnet_builder(
model_name=model_name, input_shape=input_shape,
batch_size=batch_size,
filter_list=filter_list, kernel_size=(3, 3),
output_dim=1, stages=stages, blocks=blocks,
bottleneck=False, network_type="regress")
model.summary()
plot_model(model, to_file=model_pic, show_shapes=True,
show_layer_names=True)
# configure optimizer
# def step_decay(epoch):
# "step decay callback."""
# if epoch >= 80 and epoch < 120:
# return float(0.01)
# elif epoch >= 120:
# return float(0.001)
# else:
# return float(0.1)
# sgd = optimizers.SGD(lr=0.0, momentum=0.9, nesterov=True)
model.compile(loss='mean_squared_error',
optimizer="adam",
metrics=["mse"])
print ("[MESSAGE] Model is compiled.")
model_file = model_file_base + \
"-{epoch:02d}-{val_mean_squared_error:.2f}.hdf5"
checkpoint = ModelCheckpoint(model_file,
monitor='val_mean_squared_error',
verbose=1,
save_best_only=True,
mode='min')
# scheduler = LearningRateScheduler(step_decay)
csv_his_log = os.path.join(model_path, "csv_history.log")
csv_logger = CSVLogger(csv_his_log, append=True)
callbacks_list = [checkpoint, csv_logger]
# training
model.fit(
x=X_train, y=Y_train,
batch_size=batch_size,
epochs=nb_epoch,
validation_data=(X_test, Y_test),
callbacks=callbacks_list)
[200, 2200],
[500, 500],
[200, 600],
[500, 1500]]
# load steering and speed
all_train_steering = np.zeros((0,))
for data_idx in xrange(30):
# get path
data_path = os.path.join(data_root_path, data_list[data_idx])
print("Reading %s" % (data_path))
# get steering
train_steering = ddd17.prepare_train_data(
data_path, y_name="steering", only_y=True,
frame_cut=[frame_cuts[data_idx][0]*2, frame_cuts[data_idx][1]*2],
data_portion="train",
speed_threshold=15, data_type="float32")
print ("Number of instance: %d" % (train_steering.shape[0]))
all_train_steering = np.append(all_train_steering, train_steering, axis=0)
print ("Total number of instances: %d" % (all_train_steering.shape[0]))
steering_mean = np.mean(all_train_steering)
three_sigma = np.std(all_train_steering)*3
high_threshold = steering_mean+three_sigma
low_threshold = steering_mean-three_sigma
# filter high degrees
filter_inds = np.logical_and(
def resnet_exp(model_name, data_name, channel_id, stages, blocks, filter_list,
nb_epoch, batch_size, frame_cut):
"""Perform ResNet experiment."""
model_path = os.path.join(spiker.HOME, "data", "exps", "ral-exps",
model_name)
model_file_base = os.path.join(model_path, model_name)
# print model info
print("[MESSAGE] Model Name: %s" % (model_name))
print("[MESSAGE] Number of epochs: %d" % (nb_epoch))
print("[MESSAGE] Batch Size: %d" % (batch_size))
print("[MESSAGE] Number of stages: %d" % (stages))
print("[MESSAGE] Number of blocks: %d" % (blocks))
# load data
data_path = os.path.join(spiker.HOME, "data", "exps", "data", "ddd17",
data_name)
if not os.path.isfile(data_path):
raise ValueError("This dataset does not exist at %s" % (data_path))
print("[MESSAGE] Dataset %s" % (data_path))
assert len(frame_cut) == 2
print("[MESSAGE] Frame cut: first at %d, last at %d" %
(frame_cut[0] * 2, -frame_cut[1] * 2))
frames, steering = ddd17.prepare_train_data(
data_path,
y_name="steering",
frame_cut=[frame_cut[0] * 2, frame_cut[1] * 2],
speed_threshold=15.)
num_samples = frames.shape[0]
num_train = int(num_samples * 0.7)
# save the training files
save_path = os.path.join(spiker.HOME, "data", "exps", "data", "ddd17")
train_file_name = os.path.join(save_path, model_name + ".hdf5")
# only save once
if not os.path.isfile(train_file_name):
save_data = h5py.File(train_file_name, "w")
train_group = save_data.create_group("train")
test_group = save_data.create_group("test")
train_group.create_dataset(name="frame",
data=frames[:num_train],
dtype=np.uint8)
train_group.create_dataset(name="steering",
data=steering[:num_train],
dtype=np.float32)
test_group.create_dataset(name="frame",
data=frames[num_train:],
dtype=np.uint8)
test_group.create_dataset(name="steering",
data=steering[num_train:],
dtype=np.float32)
save_data.flush()
save_data.close()
frames /= 255.
frames -= np.mean(frames, keepdims=True)
X_test = frames[num_train:]
Y_test = steering[num_train:]
del frames, steering
if channel_id != 2:
X_test = X_test[:, :, :, channel_id][..., np.newaxis]
print("[MESSAGE] Number of samples %d" % (num_samples))
print("[MESSAGE] Number of test samples %d" % (X_test.shape[0]))
model_file = model_file_base + "-best.hdf5"
model = load_model(model_file)
Y_predict = model.predict(X_test)
with open(model_file_base + "-prediction.pkl", "wb") as f:
pickle.dump([Y_test, Y_predict], f)
from __future__ import print_function
import os
import numpy as np
import matplotlib.pyplot as plt
import spiker
from spiker.models import utils
from spiker.data import ddd17
# load and process data
data_path = os.path.join(spiker.SPIKER_DATA, "ddd17",
"aug09/rec1502336427-export.hdf5")
frame_cut = [100, 400]
frames, steering = ddd17.prepare_train_data(
data_path, y_name="steering",
frame_cut=frame_cut)
frames /= 255.
frames -= np.mean(frames, keepdims=True)
num_samples = frames.shape[0]
num_train = int(num_samples*0.7)
X_train = frames[:num_train]
Y_train = steering[:num_train]
X_test = frames[num_train:]
Y_test = steering[num_train:]
del frames, steering
model_name_base = "steering-night-6-"
# load model
model_path = os.path.join(
spiker.SPIKER_EXPS, model_name_base+"full",
def resnet_exp(model_name, data_name, channel_id, stages, blocks, filter_list,
nb_epoch, batch_size, frame_cut):
"""Perform ResNet experiment."""
model_path = os.path.join(spiker.HOME, "data", "exps", "ral-exps",
model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
else:
raise ValueError("[MESSAGE] This experiment has been done before."
" Create a new config model if you need.")
model_file_base = os.path.join(model_path, model_name)
# print model info
print("[MESSAGE] Model Name: %s" % (model_name))
print("[MESSAGE] Number of epochs: %d" % (nb_epoch))
print("[MESSAGE] Batch Size: %d" % (batch_size))
print("[MESSAGE] Number of stages: %d" % (stages))
print("[MESSAGE] Number of blocks: %d" % (blocks))
# load data
data_path = os.path.join(spiker.HOME, "data", "exps", "data", "ddd17",
data_name)
if not os.path.isfile(data_path):
raise ValueError("This dataset does not exist at %s" % (data_path))
print("[MESSAGE] Dataset %s" % (data_path))
assert len(frame_cut) == 2
print("[MESSAGE] Frame cut: first at %d, last at %d" %
(frame_cut[0] * 2, -frame_cut[1] * 2))
frames, steering = ddd17.prepare_train_data(
data_path,
y_name="steering",
frame_cut=[frame_cut[0] * 2, frame_cut[1] * 2],
speed_threshold=15.)
frames /= 255.
frames -= np.mean(frames, keepdims=True)
num_samples = frames.shape[0]
num_train = int(num_samples * 0.7)
X_train = frames[:num_train]
Y_train = steering[:num_train]
X_test = frames[num_train:]
Y_test = steering[num_train:]
del frames, steering
if channel_id != 2:
X_train = X_train[:, :, :, channel_id][..., np.newaxis]
X_test = X_test[:, :, :, channel_id][..., np.newaxis]
print("[MESSAGE] Number of samples %d" % (num_samples))
print("[MESSAGE] Number of train samples %d" % (X_train.shape[0]))
print("[MESSAGE] Number of test samples %d" % (X_test.shape[0]))
# setup image shape
input_shape = (X_train.shape[1], X_train.shape[2], X_train.shape[3])
# Build model
model = resnet.resnet_builder(model_name=model_name,
input_shape=input_shape,
batch_size=batch_size,
filter_list=filter_list,
kernel_size=(3, 3),
output_dim=1,
stages=stages,
blocks=blocks,
bottleneck=False,
network_type="regress")
model.summary()
model.compile(loss='mean_squared_error', optimizer="adam", metrics=["mse"])
print("[MESSAGE] Model is compiled.")
model_file = model_file_base + "-best.hdf5"
checkpoint = ModelCheckpoint(model_file,
monitor='val_mean_squared_error',
verbose=1,
save_best_only=True,
mode='min')
csv_his_log = os.path.join(model_path, "csv_history.log")
csv_logger = CSVLogger(csv_his_log, append=True)
callbacks_list = [checkpoint, csv_logger]
# training
model.fit(x=X_train,
y=Y_train,
batch_size=batch_size,
epochs=nb_epoch,
validation_data=(X_test, Y_test),
callbacks=callbacks_list)
model_base = "-day-4-"
exp_type = "steering"
sensor_type = ["full", "dvs", "aps"]
load_prediction = os.path.join(spiker.SPIKER_EXTRA,
"pred" + model_base + "result-run-3")
if os.path.isfile(load_prediction):
print("[MESSAGE] Prediction available")
with open(load_prediction, "r") as f:
(steer_full, steer_dvs, steer_aps) = pickle.load(f)
f.close()
else:
# export ground truth
test_frames, _ = ddd17.prepare_train_data(data_path,
y_name="steering",
frame_cut=frame_cut)
test_frames /= 255.
test_frames -= np.mean(test_frames, keepdims=True)
num_samples = test_frames.shape[0]
num_train = int(num_samples * 0.7)
X_test = test_frames[num_train:]
del test_frames
# steering full
steer_full = get_prediction(X_test, exp_type, model_base, sensor_type[0],
"steering-day-4-full-103-0.02.hdf5")
print("[MESSAGE] Steering Full")
# steering dvs
steer_dvs = get_prediction(X_test[:, :, :, 0][..., np.newaxis], exp_type,
model_base, sensor_type[1],