def trainLSTM():
#path_to_checkpoint = poseDataset.base_dir + '/checkpoints_LSTM_no_Trans_no_rot_lr_{0}_mu_{1}_gclip_{2}_batch_size_{3}_truncate_gradient_{4}/'.format(args.initial_lr,args.momentum,args.g_clip,args.batch_size,args.truncate_gradient)
path_to_checkpoint = poseDataset.base_dir + '/{0}/'.format(args.checkpoint_path)
if not os.path.exists(path_to_checkpoint):
os.mkdir(path_to_checkpoint)
saveNormalizationStats(path_to_checkpoint)
trX,trY = poseDataset.getMalikFeatures()
trX_validation,trY_validation = poseDataset.getMalikValidationFeatures()
trX_forecasting,trY_forecasting = poseDataset.getMalikTrajectoryForecasting()
saveForecastedMotion(trY_forecasting,path_to_checkpoint)
saveForecastedMotion(trX_forecasting,path_to_checkpoint,'motionprefix_N_')
print 'X forecasting ',trX_forecasting.shape
print 'Y forecasting ',trY_forecasting.shape
inputDim = trX.shape[2]
print inputDim
rnn = []
if args.use_pretrained == 1:
rnn = load(path_to_checkpoint+'checkpoint.'+str(args.iter_to_load))
else:
args.iter_to_load = 0
rnn = LSTMRegression(inputDim)
rnn.fitModel(trX, trY, snapshot_rate=args.snapshot_rate, path=path_to_checkpoint, epochs=args.epochs, batch_size=args.batch_size,
decay_after=args.decay_after, learning_rate=args.initial_lr, learning_rate_decay=args.learning_rate_decay, trX_validation=trX_validation,
trY_validation=trY_validation, trX_forecasting=trX_forecasting, trY_forecasting=trY_forecasting, iter_start=args.iter_to_load,
decay_type=args.decay_type, decay_schedule=args.decay_schedule, decay_rate_schedule=args.decay_rate_schedule,
use_noise=args.use_noise, noise_schedule=args.noise_schedule, noise_rate_schedule=args.noise_rate_schedule,maxiter=args.maxiter,poseDataset=poseDataset,unNormalizeData=unNormalizeData)
def trainLSTM():
#path_to_checkpoint = poseDataset.base_dir + '/checkpoints_LSTM_no_Trans_no_rot_lr_{0}_mu_{1}_gclip_{2}_batch_size_{3}_truncate_gradient_{4}/'.format(args.initial_lr,args.momentum,args.g_clip,args.batch_size,args.truncate_gradient)
path_to_checkpoint = poseDataset.base_dir + '/{0}/'.format(
args.checkpoint_path)
if not os.path.exists(path_to_checkpoint):
os.mkdir(path_to_checkpoint)
saveNormalizationStats(path_to_checkpoint)
trX, trY = poseDataset.getMalikFeatures()
trX_validation, trY_validation = poseDataset.getMalikValidationFeatures()
trX_forecasting, trY_forecasting = poseDataset.getMalikTrajectoryForecasting(
)
saveForecastedMotion(trY_forecasting, path_to_checkpoint)
saveForecastedMotion(trX_forecasting, path_to_checkpoint,
'motionprefix_N_')
print 'X forecasting ', trX_forecasting.shape
print 'Y forecasting ', trY_forecasting.shape
inputDim = trX.shape[2]
print inputDim
rnn = []
if args.use_pretrained == 1:
rnn = load(path_to_checkpoint + 'checkpoint.' + str(args.iter_to_load))
else:
args.iter_to_load = 0
rnn = LSTMRegression(inputDim)
rnn.fitModel(trX,
trY,
snapshot_rate=args.snapshot_rate,
path=path_to_checkpoint,
epochs=args.epochs,
batch_size=args.batch_size,
decay_after=args.decay_after,
learning_rate=args.initial_lr,
learning_rate_decay=args.learning_rate_decay,
trX_validation=trX_validation,
trY_validation=trY_validation,
trX_forecasting=trX_forecasting,
trY_forecasting=trY_forecasting,
iter_start=args.iter_to_load,
decay_type=args.decay_type,
decay_schedule=args.decay_schedule,
decay_rate_schedule=args.decay_rate_schedule,
use_noise=args.use_noise,
noise_schedule=args.noise_schedule,
noise_rate_schedule=args.noise_rate_schedule,
maxiter=args.maxiter,
poseDataset=poseDataset,
unNormalizeData=unNormalizeData)
fname = 'motionprefixlong'
model.saveForecastedMotion(trX_forecast_nodeFeatures_,path,fname)
cellstate = model.predict_cell(trX_forecasting,trX_forecast_nodeFeatures,sequence_length=trY_forecasting.shape[0],poseDataset=poseDataset,graph=graph)
fname = 'forecast_celllong_{0}'.format(iteration)
model.saveCellState(cellstate,path,fname)
t1 = time.time()
del model
elif args.forecast == 'lstm' or args.forecast == 'malik':
path_to_checkpoint = '{0}checkpoint.{1}'.format(path,iteration)
if os.path.exists(path_to_checkpoint):
model = load(path_to_checkpoint)
print 'Loaded LSTM/Malik: ',path_to_checkpoint
trX_forecasting,trY_forecasting = poseDataset.getMalikTrajectoryForecasting()
fname = 'ground_truth_forecast'
model.saveForecastedMotion(trY_forecasting,path,fname)
fname = 'motionprefix'
model.saveForecastedMotion(trX_forecasting,path,fname)
forecasted_motion = model.predict_sequence(trX_forecasting,sequence_length=trY_forecasting.shape[0])
fname = 'forecast_iteration_{0}'.format(iteration)
model.saveForecastedMotion(forecasted_motion,path,fname)
'''
skel_err = np.mean(np.sqrt(np.sum(np.square((forecasted_motion - trY_forecasting)),axis=2)),axis=1)
err_per_dof = skel_err / trY_forecasting.shape[2]
fname = 'forecast_error_iteration_{0}'.format(iteration)
trX_forecast_nodeFeatures,
sequence_length=trY_forecasting.shape[0],
poseDataset=poseDataset,
graph=graph)
fname = 'forecast_celllong_{0}'.format(iteration)
model.saveCellState(cellstate, path, fname)
t1 = time.time()
del model
elif args.forecast == 'lstm' or args.forecast == 'malik':
path_to_checkpoint = '{0}checkpoint.{1}'.format(path, iteration)
if os.path.exists(path_to_checkpoint):
model = load(path_to_checkpoint)
print 'Loaded LSTM/Malik: ', path_to_checkpoint
trX_forecasting, trY_forecasting = poseDataset.getMalikTrajectoryForecasting(
)
fname = 'ground_truth_forecast'
model.saveForecastedMotion(trY_forecasting, path, fname)
fname = 'motionprefix'
model.saveForecastedMotion(trX_forecasting, path, fname)
forecasted_motion = model.predict_sequence(
trX_forecasting, sequence_length=trY_forecasting.shape[0])
fname = 'forecast_iteration_{0}'.format(iteration)
model.saveForecastedMotion(forecasted_motion, path, fname)
'''
skel_err = np.mean(np.sqrt(np.sum(np.square((forecasted_motion - trY_forecasting)),axis=2)),axis=1)
err_per_dof = skel_err / trY_forecasting.shape[2]
fname = 'forecast_error_iteration_{0}'.format(iteration)
def model():
if args.forecast == 'srnn':
path_to_checkpoint = '{0}checkpoint.{1}'.format(path, args.iteration)
print "Using checkpoint at: ",path_to_checkpoint
if os.path.exists(path_to_checkpoint):
[nodeNames,
nodeList,
nodeFeatureLength,
nodeConnections,
edgeList,
edgeListComplete,
edgeFeatures,
nodeToEdgeConnections,
trX,trY,
trX_validation,
trY_validation,
trX_forecasting,
trY_forecasting,
trX_forecast_nodeFeatures] = graph.readCRFgraph(poseDataset)
print trX_forecast_nodeFeatures.keys()
print 'Loading the model (this takes long, can take upto 25 minutes)'
model = loadDRA(path_to_checkpoint)
print model
print 'Loaded S-RNN from ', path_to_checkpoint
t0 = time.time()
save_full = True
trY_forecasting = model.convertToSingleVec(trY_forecasting,new_idx,featureRange) # 8x100x54
print trY_forecasting.shape
trX_forecast_nodeFeatures_ = model.convertToSingleVec(trX_forecast_nodeFeatures, new_idx, featureRange)
# normalize_data_for_srnn()
trY_forecasting_tem = trY_forecasting
trX_forecasting_tem = trX_forecast_nodeFeatures_
filterList = [10, 11,
16, 17, 18, 19, 20,
25, 26,
31, 32, 33, 34, 35,
48, 49, 50,
58, 59,
63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74,
82, 83, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98]
for i in filterList:
trY_forecasting_tem = np.insert(trY_forecasting_tem, i, 0, axis=2)
trX_forecasting_tem = np.insert(trX_forecasting_tem, i, 0, axis=2)
trY_forecasting_full = trY_forecasting_tem
trX_forecasting_full = trX_forecasting_tem
print(trY_forecasting_full.shape)
print(trX_forecasting_full.shape)
fname = 'ground_truth_forecast'
if(save_full):
# unnormalize_data()
trY_forecasting_full_orig = unnormalize_data(trY_forecasting_full, poseDataset.data_mean, poseDataset.data_std, [])
model.saveForecastedMotion(trY_forecasting_full_orig, path, fname)
else:
model.saveForecastedMotion(trY_forecasting, path, fname)
fname = 'motionprefix'
if(save_full):
trX_forecast_nodeFeatures_full_orig = unnormalize_data(trX_forecasting_full, poseDataset.data_mean, poseDataset.data_std, [])
model.saveForecastedMotion(trX_forecast_nodeFeatures_full_orig, path, fname)
else:
model.saveForecastedMotion(trX_forecasting, path, fname)
forecasted_motion = model.predict_sequence(trX_forecasting, trX_forecast_nodeFeatures,sequence_length=trY_forecasting.shape[0],poseDataset=poseDataset,graph=graph)
forecasted_motion = model.convertToSingleVec(forecasted_motion, new_idx, featureRange)
fname = 'forecast'
if(save_full):
filter_list = []
for x in range(trY_forecasting_full.shape[2]):
if x in filterList:
continue
filter_list.append(x)
forecasted_motion_full = trY_forecasting_full
forecasted_motion_full[:, :, filter_list] = forecasted_motion
# unnormalize_data()
forecasted_motion_full_orig = unnormalize_data(forecasted_motion_full, poseDataset.data_mean, poseDataset.data_std, [])
model.saveForecastedMotion(forecasted_motion_full_orig, path, fname)
else:
model.saveForecastedMotion(forecasted_motion, path, fname)
skel_err = np.mean(np.sqrt(np.sum(np.square((forecasted_motion - trY_forecasting)),axis=2)),axis=1)
err_per_dof = skel_err / trY_forecasting.shape[2]
fname = 'forecast_error'
model.saveForecastError(skel_err,err_per_dof,path,fname)
t1 = time.time()
print t1-t0
del model
else:
print "the path is not found!!!"
elif args.forecast == 'lstm3lr' or args.forecast == 'erd':
path_to_checkpoint = '{0}checkpoint.{1}'.format(path, args.iteration)
if os.path.exists(path_to_checkpoint):
print "Loading the model {0} (this may take sometime)".format(args.forecast)
model = load(path_to_checkpoint)
print 'Loaded the model from ',path_to_checkpoint
save_full = True
# more colum than srnn1 8x100x99
trX_forecasting_full, trY_forecasting_full = poseDataset.getMalikTrajectoryForecastingFull()
# same mat as srnn1 8x100x54
trX_forecasting, trY_forecasting = poseDataset.getMalikTrajectoryForecasting()
fname = 'ground_truth_forecast'
if(save_full):
# unnormalize_data()
# print(poseDataset.dimensions_to_ignore)
trY_forecasting_full_orig = unnormalize_data(trY_forecasting_full, poseDataset.data_mean,
poseDataset.data_std, [])
model.saveForecastedMotion(trY_forecasting_full_orig, path, fname)
else:
model.saveForecastedMotion(trY_forecasting, path, fname)
fname = 'motionprefix'
if(save_full):
# unnormalize_data()
trX_forecasting_full_orig = unnormalize_data(trX_forecasting_full, poseDataset.data_mean,
poseDataset.data_std, [])
model.saveForecastedMotion(trX_forecasting_full_orig, path, fname)
else:
model.saveForecastedMotion(trX_forecasting, path, fname)
forecasted_motion = model.predict_sequence(trX_forecasting,sequence_length=trY_forecasting.shape[0])
fname = 'forecast'
if(save_full):
filter_list = []
for x in range(trY_forecasting_full.shape[2]):
if x in poseDataset.dimensions_to_ignore:
continue
filter_list.append(x)
forecasted_motion_full = trY_forecasting_full
forecasted_motion_full[:, :, filter_list] = forecasted_motion
# unnormalize_data()
forecasted_motion_full_orig = unnormalize_data(forecasted_motion_full, poseDataset.data_mean,
poseDataset.data_std, [])
model.saveForecastedMotion(forecasted_motion_full_orig, path, fname)
else:
model.saveForecastedMotion(forecasted_motion,path,fname)
skel_err = np.mean(np.sqrt(np.sum(np.square((forecasted_motion - trY_forecasting)),axis=2)),axis=1)
err_per_dof = skel_err / trY_forecasting.shape[2]
fname = 'forecast_error'
model.saveForecastError(skel_err,err_per_dof,path,fname)
# print(trX_forecasting_full.shape)
# print(trX_forecasting.shape)
del model
else:
print "the path is not found!!!"
elif args.forecast == 'dracell':
path_to_checkpoint = '{0}checkpoint.{1}'.format(path, args.iteration)
if os.path.exists(path_to_checkpoint):
[nodeNames,nodeList,nodeFeatureLength,nodeConnections,edgeList,edgeListComplete,edgeFeatures,nodeToEdgeConnections,trX,trY,trX_validation,trY_validation,trX_forecasting,trY_forecasting,trX_forecast_nodeFeatures] = graph.readCRFgraph(poseDataset,noise=0.7,forecast_on_noisy_features=True)
print trX_forecast_nodeFeatures.keys()
print 'Loading the model'
model = loadDRA(path_to_checkpoint)
print 'Loaded DRA: ',path_to_checkpoint
t0 = time.time()
trY_forecasting = model.convertToSingleVec(trY_forecasting,new_idx,featureRange)
trX_forecast_nodeFeatures_ = model.convertToSingleVec(trX_forecast_nodeFeatures,new_idx,featureRange)
fname = 'motionprefixlong'
model.saveForecastedMotion(trX_forecast_nodeFeatures_,path,fname)
cellstate = model.predict_cell(trX_forecasting,trX_forecast_nodeFeatures,sequence_length=trY_forecasting.shape[0],poseDataset=poseDataset,graph=graph)
fname = 'forecast_celllong_{0}'.format(args.iteration)
model.saveCellState(cellstate,path,fname)
t1 = time.time()
del model
print 'ending program'
