def validate_prediction(self, x_train, x_test, y_test): e = self.encoder.predict(x_train) embedding = e[:, self.predict_hierarchies] mean_diff, diff = metrics.get_embedding_diffs(embedding[:,1], embedding[:,0]) add_std = np.std(diff, axis=0) z = self.encoder.predict(x_test)[:,self.predict_hierarchies[0]] z = z + mean_diff y_test_pred = self.decoder.predict(z) y_test_pred = self.unormalize_angle(y_test_pred) return add_std, self.euler_error(y_test, y_test_pred)
def run(self, data_iterator, valid_data):
model_vars = [NAME, self.latent_dim, self.timesteps, self.batch_size]
self.get_ignored_dims()
used_idx = [
6, 7, 8, 9, 12, 13, 14, 15, 21, 22, 23, 24, 27, 28, 29, 30, 36, 37,
38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 51, 52, 53, 54, 55, 56, 57,
60, 61, 62, 75, 76, 77, 78, 79, 80, 81, 84, 85, 86
]
def euler_error(yTrue, yPred):
yPred = self.__recover_parameterization(yPred)[:, :, used_idx]
error = np.square(wrap_angle(yTrue[:, :, used_idx]) - yPred)
error = np.sum(error, -1)
error = np.sqrt(error)
return np.mean(error, 0)
def wrap_angle(rad):
return (rad + np.pi) % (2 * np.pi) - np.pi
if not self.load():
# from keras.utils import plot_model
# plot_model(self.autoencoder, to_file='model.png')
loss = 10000
iter1, iter2 = tee(data_iterator)
for i in range(self.periods):
for x_data, y_data in iter1:
# x_data, y_data = self.__alter_label(x, y)
x_train, x_test, y_train_, y_test_ = cross_validation.train_test_split(
x_data, y_data, test_size=self.cv_splits)
_y_train = self.__alter_parameterization(y_train_)
_y_test = self.__alter_parameterization(y_test_)
y_train = self.__alter_y(_y_train)
y_test = self.__alter_y(_y_test)
# print np.sum(y_train[:,0,-self.label_dim:], axis=0)
history = self.autoencoder.fit(_y_train,
y_train,
shuffle=True,
epochs=self.epochs,
batch_size=self.batch_size,
validation_data=(_y_test,
y_test))
new_loss = np.mean(history.history['loss'])
if new_loss < loss:
print 'Saved model - ', loss
loss = new_loss
# y_test_decoded = self.autoencoder.predict(x_test[:1])
# y_test_decoded = np.reshape(y_test_decoded, (len(self.hierarchies), self.timesteps, -1))
# image.plot_poses(x_test[:1,:,:-self.label_dim], y_test_decoded[:,:,:-self.label_dim])
# image.plot_hierarchies(y_test_orig[:,:,:-self.label_dim], y_test_decoded[:,:,:-self.label_dim])
self.autoencoder.save_weights(self.save_path,
overwrite=True)
rand_idx = np.random.choice(x_test.shape[0],
25,
replace=False)
#metrics.validate(x_test[rand_idx], self, self.log_path, history.history['loss'])
y_test_pred = self.encoder.predict(_y_test[rand_idx])[:,
-1]
y_test_pred = self.decoder.predict(y_test_pred)
mse_ = np.mean(
np.square(y_test[rand_idx, -self.timesteps:] -
y_test_pred))
y_test_gt = y_test_[rand_idx]
mse = euler_error(y_test_gt, y_test_pred)
y_test_pred = self.__recover_parameterization(y_test_pred)
mae = np.mean(
np.abs(
np.arctan2(np.sin(y_test_gt - y_test_pred),
np.cos(y_test_gt - y_test_pred))))
wrap_mae = np.mean(
np.abs(
wrap_angle(y_test_gt[:, :, used_idx]) -
y_test_pred[:, :, used_idx]))
print 'MSE_Sin_Cos', mse_
print 'MAE', mae
print 'Wrap_MAE', wrap_mae
print 'MSE', mse
with open(
'../new_out/%s_t%d_l%d_log.csv' %
(NAME, self.timesteps, self.latent_dim), 'a+') as f:
spamwriter = csv.writer(f)
spamwriter.writerow([
new_loss, mse_, mae, wrap_mae, mse, LEARNING_RATE
])
del x_train, x_test, y_train, y_test, y_train_, y_test_
iter1, iter2 = tee(iter2)
data_iterator = iter2
else:
# load embedding
embedding = []
for _, y in data_iterator:
y = self.__alter_parameterization(y)
e = self.encoder.predict(y)
if len(embedding) == 0:
embedding = e[:, self.hierarchies]
else:
embedding = np.concatenate(
(embedding, e[:, self.hierarchies]), axis=0)
break
embedding = np.array(embedding)
print 'emb', embedding.shape
mean_diff, diff = metrics.get_embedding_diffs(
embedding[:, 1], embedding[:, 0])
load_path = '../human_motion_pred/baselines/euler/'
cut = self.hierarchies[0]
methods = ['closest_partial', 'mean-100', 'add']
_N = 8
pred_n = self.hierarchies[1] - cut
error = {m: np.zeros((15, pred_n)) for m in methods}
a_n = 0
for basename in metric_baselines.iter_actions():
print basename, '================='
cond = np.zeros((_N, self.timesteps, self.input_dim))
cond[:, -cut - 1:] = self.__alter_parameterization(
np.load(load_path + basename + '_cond.npy')[:, -cut - 1:])
# pd = np.load(load_path + basename + '_pred.npy')
gtp = np.load(load_path + basename + '_gt.npy')[:, :pred_n]
enc = self.encoder.predict(cond)[:, cut]
# autoencoding error
autoenc = self.decoder.predict(enc)[:, :cut + 1]
print euler_error(cond[:, :cut + 1], autoenc)
for method in methods:
new_enc = np.zeros(enc.shape)
for i in range(_N):
if method == 'closest_partial':
new_enc[i] = metrics.closest_partial_index(
embedding[:, 0], enc[i])
elif method == 'mean-100':
new_enc[i] = metrics.closest_mean(embedding[:, 1],
enc[i],
n=100)
elif method == 'add':
new_enc[i] = enc[i] + mean_diff
model_pred = self.decoder.predict(new_enc)[:, cut + 1:]
error[method][a_n] = euler_error(gtp, model_pred)
#print method
#print error[method][a_n]
a_n += 1
print 'total ================='
for method in methods:
print np.mean(error[method], 0)
error[method] = error[method].tolist()
with open(
'../new_out/%s_t%d_l%d_compared.json' %
(NAME, self.timesteps, self.latent_dim), 'wb') as result_file:
json.dump(error, result_file)
def run(self, data_iterator, valid_data):
load_path = '../human_motion_pred/baselines/'
# model_vars = [NAME, self.latent_dim, self.timesteps, self.batch_size]
if not self.load():
test_data_y, test_data_x = self.load_validation_data(load_path)
test_data_y = wrap_angle(test_data_y)
print self.loss_opt_str
# from keras.utils import plot_model
# plot_model(self.autoencoder, to_file='model.png')
loss = 10000
iter1, iter2 = tee(data_iterator)
for i in range(self.periods):
for x, _ in iter1:
#image.plot_fk_from_euler(x[:3], title='test')
y, x = self.__alter_parameterization(x)
x_train, x_test, y_train, y_test = cross_validation.train_test_split(
x, x, test_size=self.cv_splits)
y_train = self.__alter_y(y_train)
y_test = self.__alter_y(y_test)
history = self.autoencoder.fit(x_train,
y_train,
shuffle=True,
epochs=self.epochs,
batch_size=self.batch_size,
validation_data=(x_test,
y_test))
# callbacks=[tbCallBack])
print history.history['loss']
new_loss = np.mean(history.history['loss'])
if new_loss < loss:
self.autoencoder.save_weights(self.save_path,
overwrite=True)
loss = new_loss
print 'Saved model - ', loss
rand_idx = np.random.choice(x.shape[0],
5000,
replace=False)
# y_test_pred = self.encoder.predict(x[rand_idx])[:,-1]
# y_test_pred = self.decoder.predict(y_test_pred)
# #y_gt = x_test[rand_idx]
# y_test_pred = self.unormalize_angle(y_test_pred)
y_gt = wrap_angle(y[rand_idx])
# mae = np.mean(np.abs(y_gt-y_test_pred))
# mse = self.euler_error(y_gt, y_test_pred)
mse = self.validate_autoencoding(x[rand_idx], y_gt)
mse_test = self.validate_autoencoding(
test_data_x, test_data_y)
add_std, mse_pred = self.validate_prediction(
x[rand_idx], test_data_x, test_data_y)
# print 'MAE', mae
add_mean_std, add_std_std = np.mean(add_std), np.std(
add_std)
print 'STD', add_mean_std, add_std_std
print 'MSE', np.mean(mse)
print 'MSE TEST', np.mean(mse_test)
print 'MSE PRED', mse_pred[[-9, -7, -3, -1]]
with open(
'../new_out/%s_t%d_l%d_%s_log.csv' %
(NAME, self.timesteps, self.latent_dim,
self.loss_opt_str), 'a+') as f:
spamwriter = csv.writer(f)
spamwriter.writerow([
new_loss, mse, mse_test, mse_pred, add_mean_std,
add_std_std, self.loss_opt_str
])
iter1, iter2 = tee(iter2)
data_iterator = iter2
else:
# load embedding
embedding = []
i = 0
for x, y in data_iterator:
y, norm_y = self.__alter_parameterization(x)
e = self.encoder.predict(norm_y)
#y_test_pred = self.decoder.predict(e[:,-1])
#y_test_pred = self.unormalize_angle(y_test_pred)
#y_gt = wrap_angle(y)
#print self.euler_error(y_gt, y_test_pred)
#np.save('../data/embedding/t40-l1024-euler-nadam-meanAbs-lr0.0001/emb_%d.npy'%i, e[:,self.predict_hierarchies])
#i += 1
#print i
#continue
if len(embedding) == 0:
embedding = e[:, self.predict_hierarchies]
else:
embedding = np.concatenate(
(embedding, e[:, self.predict_hierarchies]), axis=0)
#break
#return
embedding = np.array(embedding)
print 'emb', embedding.shape
mean_diff, diff = metrics.get_embedding_diffs(
embedding[:, 1], embedding[:, 0])
print 'std', np.std(diff)
_N = 8
#methods = ['closest', 'closest_partial', 'mean-5', 'add', 'fn']
methods = ['0v']
nn = NN.Forward_NN({
'input_dim': self.latent_dim,
'output_dim': self.latent_dim,
'mode': 'sample'
})
nn.run(None)
#nn = None
cut_e = self.predict_hierarchies[0]
cut_x = self.hierarchies[0] + 1
pred_n = self.hierarchies[1] - cut_x + 1
error = {
act: {m: {
'euler': None
}
for m in methods}
for act in metric_baselines.iter_actions()
}
# a_n = 0
for basename in metric_baselines.iter_actions():
print basename, '================='
# x, y = valid_data
# x = np.zeros((_N, self.timesteps, 96))
# x[:,:cut_x+1] = np.load(load_path + 'xyz/' + basename + '_cond.npy')[:,-cut_x-1:]
y = np.zeros((_N, self.timesteps, 99))
y[:, :cut_x] = np.load(load_path + 'euler/' + basename +
'_cond.npy')[:, -cut_x:]
# y = np.load(load_path + 'euler/' + basename + '_cond.npy')[:,-25:]
# gtp_x = np.load(load_path + 'xyz/' + basename + '_gt.npy')[:,:pred_n][:,:,self.used_xyz_idx]
gtp_y_orig = np.load(load_path + 'euler/' + basename +
'_gt.npy')[:, :pred_n]
#print np.mean(np.abs(y[:,cut_x-2] - y[:,cut_x-1])), np.mean(np.abs(y[:,cut_x-1] - gtp_y[:,0]))
#y[:,cut_x:] = gtp_y
#image.plot_fk_from_euler(gtp_y[:4])
#image.plot_fk_from_euler( wrap_angle(gtp_y[:4]))
gtp_y = wrap_angle(gtp_y_orig[:, :, self.used_euler_idx])
# rand_idx = np.random.choice(x.shape[0], _N, replace=False)
# x, y, xy = self.__merge_n_reparameterize(x[rand_idx],y[rand_idx], True)
y, norm_y = self.__alter_parameterization(y)
y = wrap_angle(y)
enc = self.encoder.predict(norm_y)
partial_enc = enc[:, cut_e]
#y = wrap_angle(y[:_N])
#partial_enc = e[:_N, cut_e]
# autoencoding error for partial seq
#dec = self.decoder.predict(enc[:,-1])
#dec = self.unormalize_angle(dec)
#print self.euler_error(y, dec)
#image.plot_poses_euler(x[:2,:cut+1], dec[:2,:,:self.euler_start], title='autoencoding', image_dir='../new_out/')
fn_pred = nn.model.predict(partial_enc)
#poses = [None]*(len(methods)+2)
#poses[1] = self.recover(gtp_y[0:1])
#poses[0] = gtp_y_orig[0:1]
#score_name = ''
for k, method in enumerate(methods):
new_enc = np.zeros(partial_enc.shape)
for i in tqdm(range(_N)):
if method == 'closest_partial':
idx = metrics.closest_partial_index(
embedding[:, 0], partial_enc[i])
new_enc[i] = embedding[idx, 1]
elif 'mean' in method:
n = int(method.split('-')[1])
new_enc[i] = metrics.closest_mean(embedding[:, 1],
partial_enc[i],
n=n)
elif method == 'add':
new_enc[i] = partial_enc[i] + mean_diff
elif method == 'closest':
new_enc[i] = metrics.closest(
embedding[:, 1], partial_enc[i])
elif method == 'fn':
new_enc[i] = fn_pred[i]
model_pred = None
if method == '0v':
model_pred = np.zeros(gtp_y.shape)
for i in range(pred_n):
model_pred[:, i] = y[:, cut_x - 1]
else:
model_pred = self.decoder.predict(new_enc)[:, cut_x:]
model_pred = self.unormalize_angle(model_pred)
error[basename][method]['euler'] = self.euler_error(
gtp_y, model_pred) #y[:, -pred_n:], model_pred)
print method
print error[basename][method]['euler'][[1, 3, 7, 9]]
error[basename][method]['euler'] = error[basename][method][
'euler'].tolist()
#score_name = score_name + '%s:%2f_' %(method, error[basename][method]['euler'][-1])
#poses[k+2] = self.recover(model_pred[:1])
# error[method]['z'] = np.mean([np.linalg.norm(new_enc[i] - enc[i,-1]) for i in range(_N)])
# print error[method]['z']
# for i in range(_N):
# pose_err = metrics.pose_seq_error(gtp_x[i], model_pred[i,:,:self.euler_start], cumulative=True)
# error[method]['pose'] = error[method]['pose'] + np.array(pose_err)
# error[method]['pose'] = error[method]['pose']/_N
# print error[method]['pose']
# error[method]['pose'] = error[method]['pose'].tolist()
#poses = np.concatenate(poses, axis=0)
#image.plot_fk_from_euler(poses, title='%s_gt_agt_%s'%(basename, score_name), image_dir='../new_out/')
with open(
'../new_out/zero_velocity_validation-testset-mseMartinez.json',
'wb') as result_file:
json.dump(error, result_file)
def run(self, data_iterator, valid_data):
# model_vars = [NAME, self.latent_dim, self.timesteps, self.batch_size]
if not self.load():
# from keras.utils import plot_model
# plot_model(self.autoencoder, to_file='model.png')
loss = 10000
iter1, iter2 = tee(data_iterator)
for i in range(self.periods):
for x, y in iter1:
x = self.__merge_n_reparameterize(x,y)
x_train, x_test, y_train, y_test = cross_validation.train_test_split(x, x, test_size=self.cv_splits)
y_train = self.__alter_y(y_train)
y_test = self.__alter_y(y_test)
history = self.autoencoder.fit(x_train, y_train,
shuffle=True,
epochs=self.epochs,
batch_size=self.batch_size,
validation_data=(x_test, y_test))
# callbacks=[tbCallBack])
print history.history['loss']
new_loss = np.mean(history.history['loss'])
if new_loss < loss:
self.autoencoder.save_weights(self.save_path, overwrite=True)
loss = new_loss
print 'Saved model - ', loss
rand_idx = np.random.choice(x.shape[0], 25, replace=False)
y_test_pred = self.encoder.predict(x[rand_idx])[:,-1]
y_test_pred = self.decoder.predict(y_test_pred)[:,:,self.euler_start:]
y_test_pred = self.unormalize_angle(y_test_pred)
y_gt = wrap_angle(y[rand_idx][:,:,self.used_euler_idx])
mae = np.mean(np.abs(y_gt-y_test_pred))
mse = self.euler_error(y_gt, y_test_pred)
print 'MAE', mae
print 'MSE', mse
with open('../new_out/%s_t%d_l%d_log.csv'%(NAME, self.timesteps, self.latent_dim), 'a+') as f:
spamwriter = csv.writer(f)
spamwriter.writerow([new_loss, mae, mse, L_RATE])
iter1, iter2 = tee(iter2)
data_iterator = iter2
else:
# load embedding
embedding = []
for x,y in data_iterator:
x = self.__merge_n_reparameterize(x,y)
e = self.encoder.predict(x)
if len(embedding) == 0:
embedding = e[:, self.predict_hierarchies]
else:
embedding = np.concatenate((embedding, e[:,self.predict_hierarchies]), axis=0)
break
embedding = np.array(embedding)
print 'emb', embedding.shape
mean_diff, diff = metrics.get_embedding_diffs(embedding[:,1], embedding[:,0])
_N = 8
methods = ['closest', 'closest_partial', 'add']
cut_e = self.predict_hierarchies[0]
cut_x = self.hierarchies[0]
pred_n = self.hierarchies[1]-cut_x
# a_n = 0
load_path = '../human_motion_pred/baselines/'
for basename in metric_baselines.iter_actions():
print basename, '================='
error = {m: {'euler': None,
# 'z': None
'pose': np.zeros(pred_n)} for m in methods}
# x, y = valid_data
x = np.zeros((_N, self.timesteps, 96))
x[:,:cut_x+1] = np.load(load_path + 'xyz/' + basename + '_cond.npy')[:,-cut_x-1:]
y = np.zeros((_N, self.timesteps, 99))
y[:,:cut_x+1] = np.load(load_path + 'euler/' + basename + '_cond.npy')[:,-cut_x-1:]
gtp_x = np.load(load_path + 'xyz/' + basename + '_gt.npy')[:,:pred_n][:,:,self.used_xyz_idx]
gtp_y = np.load(load_path + 'euler/' + basename + '_gt.npy')[:,:pred_n][:,:,self.used_euler_idx]
# rand_idx = np.random.choice(x.shape[0], _N, replace=False)
# x, y, xy = self.__merge_n_reparameterize(x[rand_idx],y[rand_idx], True)
x, y, xy = self.__merge_n_reparameterize(x,y, True)
# y = unormalize_angle(y)
enc = self.encoder.predict(xy)
partial_enc = enc[:,cut_e]
# autoencoding error for partial seq
dec = self.decoder.predict(partial_enc)[:,:cut_x+1]
dec_euler = unormalize_angle(dec[:,:,self.euler_start:])
print self.euler_error(y[:,:cut_x+1], dec_euler)
#image.plot_poses_euler(x[:2,:cut+1], dec[:2,:,:self.euler_start], title='autoencoding', image_dir='../new_out/')
for method in methods:
new_enc = np.zeros(partial_enc.shape)
for i in tqdm(range(_N)):
if method == 'closest_partial':
idx = metrics.closest_partial_index(embedding[:,0], partial_enc[i])
new_enc[i] = embedding[idx,1]
elif 'mean' in method:
n = int(method.split('-')[1])
new_enc[i] = metrics.closest_mean(embedding[:,1], partial_enc[i], n=n)
elif method == 'add':
new_enc[i] = partial_enc[i]+mean_diff
elif method == 'closest':
new_enc[i] = metrics.closest(embedding[:,1], partial_enc[i])
model_pred = self.decoder.predict(new_enc)[:,cut_x+1:]
model_pred_euler = unormalize_angle(model_pred[:,:,self.euler_start:])
# error[method]['euler'] = self.euler_error(y[:,cut_x+1:], model_pred_euler)
error[method]['euler'] = self.euler_error(gtp_y, model_pred_euler)
print method
print error[method]['euler']
error[method]['euler'] = error[method]['euler'].tolist()
#image.plot_poses_euler(gtp_x[:2], model_pred[:2,:,:self.euler_start], title=method, image_dir='../new_out/')
# error[method]['z'] = np.mean([np.linalg.norm(new_enc[i] - enc[i,-1]) for i in range(_N)])
# print error[method]['z']
for i in range(_N):
pose_err = metrics.pose_seq_error(gtp_x[i], model_pred[i,:,:self.euler_start], cumulative=True)
error[method]['pose'] = error[method]['pose'] + np.array(pose_err)
error[method]['pose'] = error[method]['pose']/_N
print error[method]['pose']
error[method]['pose'] = error[method]['pose'].tolist()
def run(self, data_iterator, valid_data):
# model_vars = [NAME, self.latent_dim, self.timesteps, self.batch_size]
if not self.load():
# from keras.utils import plot_model
# plot_model(self.autoencoder, to_file='model.png')
loss = 10000
iter1, iter2 = tee(data_iterator)
for i in range(self.periods):
for x, y in iter1:
x = self.__merge_n_reparameterize(x, y)
x_train, x_test, y_train, y_test = cross_validation.train_test_split(
x, x, test_size=self.cv_splits)
y_train = self.__alter_y(y_train)
y_test = self.__alter_y(y_test)
history = self.autoencoder.fit(x_train,
y_train,
shuffle=True,
epochs=self.epochs,
batch_size=self.batch_size,
validation_data=(x_test,
y_test))
# callbacks=[tbCallBack])
print history.history['loss']
new_loss = np.mean(history.history['loss'])
if new_loss < loss:
self.autoencoder.save_weights(self.save_path,
overwrite=True)
loss = new_loss
print 'Saved model - ', loss
rand_idx = np.random.choice(x.shape[0], 25, replace=False)
y_test_pred = self.encoder.predict(x[rand_idx])[:, -1]
y_test_pred = self.decoder.predict(
y_test_pred)[:, :, self.euler_start:]
y_test_pred = unormalize_angle(y_test_pred)
y_gt = wrap_angle(y[rand_idx][:, :, self.used_euler_idx])
mae = np.mean(np.abs(y_gt - y_test_pred))
mse = self.euler_error(y_gt, y_test_pred)
print 'MAE', mae
print 'MSE', mse
with open(
'../new_out/%s_t%d_l%d_log.csv' %
(NAME, self.timesteps, self.latent_dim), 'a+') as f:
spamwriter = csv.writer(f)
spamwriter.writerow([new_loss, mae, mse, L_RATE])
iter1, iter2 = tee(iter2)
data_iterator = iter2
else:
# load embedding
embedding = []
for x, y in data_iterator:
x = self.__merge_n_reparameterize(x, y)
e = self.encoder.predict(x)
if len(embedding) == 0:
embedding = e[:, self.hierarchies]
else:
embedding = np.concatenate(
(embedding, e[:, self.hierarchies]), axis=0)
#break
embedding = np.array(embedding)
print 'emb', embedding.shape
mean_diff, diff = metrics.get_embedding_diffs(
embedding[:, 1], embedding[:, 0])
_N = 100
methods = ['closest_partial', 'closest', 'add']
cut = self.hierarchies[0]
pred_n = self.hierarchies[1] - cut
error = {
m: {
'euler': np.zeros(pred_n),
'pose': np.zeros(pred_n)
}
for m in methods
}
x, y = valid_data
rand_idx = np.random.choice(x.shape[0], _N, replace=False)
x, y, valid_data = self.__merge_n_reparameterize(
x[rand_idx], y[rand_idx], True)
y = unormalize_angle(y)
enc = self.encoder.predict(valid_data)
partial_enc = enc[:, cut]
# autoencoding error for partial seq
dec = self.decoder.predict(partial_enc)[:, :cut + 1]
dec_euler = unormalize_angle(dec[:, :, self.euler_start:])
print self.euler_error(y[:, :cut + 1], dec_euler)
#image.plot_poses_euler(x[:2,:cut+1], dec[:2,:,:self.euler_start], title='autoencoding', image_dir='../new_out/')
for method in methods:
new_enc = np.zeros(partial_enc.shape)
for i in tqdm(range(_N)):
if method == 'closest_partial':
idx = metrics.closest_partial_index(
embedding[:, 0], partial_enc[i])
new_enc[i] = embedding[idx, 1]
elif method == 'closest':
new_enc[i] = metrics.closest(embedding[:, 1],
partial_enc[i])
elif method == 'add':
new_enc[i] = partial_enc[i] + mean_diff
model_pred = self.decoder.predict(new_enc)[:, cut + 1:]
model_pred_euler = unormalize_angle(
model_pred[:, :, self.euler_start:])
error[method]['euler'] = self.euler_error(
y[:, cut + 1:], model_pred_euler)
print method
print error[method]['euler']
image.plot_poses_euler(x[:2, cut + 1:],
model_pred[:2, :, :self.euler_start],
title=method,
image_dir='../new_out/')
for i in range(_N):
pose_err = metrics.pose_seq_error(
x[i, cut + 1:],
model_pred[i, :, :self.euler_start],
cumulative=True)
error[method]['pose'] = error[method]['pose'] + np.array(
pose_err)
error[method]['pose'] = error[method]['pose'] / _N
print error[method]['pose']
error[method]['euler'] = error[method]['euler'].tolist()
error[method]['pose'] = error[method]['pose'].tolist()
with open(
'../new_out/%s_t%d_l%d_validation.json' %
(NAME, self.timesteps, self.latent_dim), 'wb') as result_file:
json.dump(error, result_file)