File size: 10,689 Bytes
1403119 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 | import numpy as np
import argparse
import os
import time
import torch
import torch.backends.cudnn as cudnn
import torch.nn.init as init
import torch.optim as optim
import torch.utils.data
from glob import glob
torch.autograd.set_detect_anomaly(True)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(device)
parser = argparse.ArgumentParser()
parser.add_argument('--train', help='number of the train set to load', required=True)
parser.add_argument('--timestamps', help='evaluate test data every x iterations', required=True)
parser.add_argument('--beta', help='multiplication fact for rotation loss', required=True)
parser.add_argument('--iterations', help='iterations to train for', required=True)
parser.add_argument('--batchsize', help='batchsize', required=True)
parser.add_argument('--size', help='sequence length', required=True)
parser.add_argument('--rnn_size', help='network hidden size', required=False, default=10)
parser.add_argument('--rot', help='0: keep rel as is, 1: rotate rel with abs', required=False, default=0)
parser.add_argument('--m', help='1: APR + PRP, 2: APR[s] + PRP[s], 3: SfM + PRP, 4: APR + PRP[s], 5: APR[s] + PRP, 6: SfM + PRP[s]', required=False, default=1)
parser.add_argument('--net', help='Network to use', required=False, default="lstm")
parser.add_argument('--stack', help='Number of layers', required=False, default="2")
parser.add_argument('--run_num', help='Id of the run', required=False, default="0")
args = parser.parse_args()
train_num = int(args.train)
TIMESTAMP = int(args.timestamps)
BETA = float(args.beta)
ITERATIONS = int(args.iterations)
batchSize = int(args.batchsize)
time_size = int(args.size)
rnn_size = int(args.rnn_size)
backrot = int(args.rot)
mode = int(args.m)
net_name = args.net
stacks = int(args.stack)
run_num = int(args.run_num)
all_folders = glob("...") # Collects all previous trainings (the folder names start with the training number)
all_folders = np.array([int(x.split("/")[-1].split("_")[0]) for x in all_folders])
if run_num in all_folders:
# if the number already exists, then skipp it
print("Skipped: ", run_num, train_num, TIMESTAMP, BETA, ITERATIONS, batchSize, time_size,
rnn_size, backrot, mode, net_name, stacks)
exit()
print("Running: ", run_num, train_num, TIMESTAMP, BETA, ITERATIONS, batchSize, time_size, rnn_size,
backrot, mode, net_name, stacks)
# Choose and create the network specified in the arguments
if net_name == "lstm":
from lstm import *
model = LSTM_stack(14, rnn_size, device, stacks)
model.train()
model = model.to(device)
elif net_name == "cfn":
from cfn import *
model = CFN_stack(14, rnn_size, device, stacks)
model.train()
model.to(device)
elif net_name == "gru":
from gru import *
model = GRU_stack(14, rnn_size, device, stacks)
model.train()
model.to(device)
elif net_name == "mgu":
from mgu import *
model = MGU_stack(14, rnn_size, device, stacks)
model.train()
model.to(device)
elif net_name == "qrnn":
from qrnn import *
model = fakeQRNN_stack(14, rnn_size, device, stacks)
model.train()
model.to(device)
elif net_name == "ran":
from ran import *
model = RAN_stack(14, rnn_size, device, stacks)
model.train()
model.to(device)
elif net_name == "sru":
from sru import *
model = SRU_stack(14, rnn_size, device, stacks)
model.train()
model.to(device)
elif net_name == "trnn":
from trnn import *
model = TRNN_stack(14, rnn_size, device, stacks)
model.train()
model.to(device)
import utilities_comb as utilities
# The following deals with paths and folders for the data. It has to be adjusted to your setup. -------------------- #
# In the end 4 file paths are defined: train_dataset_abs/train_dataset_rel, test_dataset_abs/test_dataset_rel
# Folders with the train/test data for the different methods
# Each is a dictionary with an entry for every train number
abs_datasets_sfm = {1: ..., 2: ...}
abs_datasets = {1: ..., 2: ...}
rel_datasets = {1: ..., 2: ...}
abs_synt_post_datasets = {1: ..., 2: ...}
rel_synt_post_datasets = {1: ..., 2: ...}
# Depending on the mode, the absolute and relative data is combined from different methods
if mode == 1:
train_base_abs = abs_datasets[train_num]
train_base_rel = rel_datasets[train_num]
train_dataset_abs = train_base_abs + "..."
train_dataset_rel = train_base_rel + "..."
test_dataset_abs = train_base_abs + "..."
test_dataset_rel = train_base_rel + "..."
elif mode == 2:
train_base_abs = abs_synt_post_datasets[train_num]
train_base_rel = rel_synt_post_datasets[train_num]
train_dataset_abs = train_base_abs + "..."
train_dataset_rel = train_base_rel + "..."
test_dataset_abs = train_base_abs + "..."
test_dataset_rel = train_base_rel + "..."
elif mode == 3:
train_base_abs = abs_datasets_sfm[train_num]
train_base_rel = rel_datasets[train_num]
train_dataset_abs = train_base_abs + "..."
train_dataset_rel = train_base_rel + "..."
test_dataset_abs = train_base_abs + "..."
test_dataset_rel = train_base_rel + "..."
elif mode == 4:
train_base_abs = abs_datasets[train_num]
train_base_rel = rel_synt_post_datasets[train_num]
train_dataset_abs = train_base_abs + "..."
train_dataset_rel = train_base_rel + "..."
test_dataset_abs = train_base_abs + "..."
test_dataset_rel = train_base_rel + "..."
elif mode == 5:
train_base_abs = abs_synt_post_datasets[train_num]
train_base_rel = rel_datasets[train_num]
train_dataset_abs = train_base_abs + "..."
train_dataset_rel = train_base_rel + "..."
test_dataset_abs = train_base_abs + "..."
test_dataset_rel = train_base_rel + "..."
elif mode == 6:
train_base_abs = abs_datasets_sfm[train_num]
train_base_rel = rel_synt_post_datasets[train_num]
train_dataset_abs = train_base_abs + "..."
train_dataset_rel = train_base_rel + "..."
test_dataset_abs = train_base_abs + "..."
test_dataset_rel = train_base_rel + "..."
# ------------------------------------------------------------------------------------------------------------------- #
# Creates the folders in which the results are saved
currentDate = time.strftime('%Y-%m-%d_%H-%M-%S', time.localtime())
os.mkdir("../results/" + "%05d_" % (run_num) + str(currentDate))
os.mkdir("../results/" + "%05d_" % (run_num) + str(currentDate) + "/weights/")
# writes the current parameters in a file
parameters = open("../results/" + "%05d_" % (run_num) + str(currentDate) + "/parameters.txt", "w")
parameters.write("Modelname: flownet" + "\n" +
"Dataset: " + str(train_num) + "\n" +
"Timestamp: " + str(TIMESTAMP) + "\n" +
"Iterations: " + str(ITERATIONS) + "\n" +
"Batch size: " + str(batchSize) + "\n" +
"Time_size: " + str(time_size) + "\n" +
"Rnn_size: " + str(rnn_size) + "\n" +
"Back_rotate: " + str(backrot) + "\n" +
"Mode: " + str(mode) + "\n" +
"Net name: " + str(net_name) + "\n" +
"Stacks: " + str(stacks) + "\n")
parameters.close()
outputWeightspath = "../results/" + "%05d_" % (run_num) + str(currentDate) + "/weights/" + str(train_num) + '_BS_' + str(batchSize) + '_'
log = open("../results/" + "%05d_" % (run_num) + str(currentDate) + "/log_" + str(train_num) + '_BS_' + str(batchSize) + '.txt', "w")
def getError(posx, posq, actualx, actualq):
""" Calculates the position and rotation error """
q1 = actualq / np.linalg.norm(actualq)
q2 = posq / np.linalg.norm(posq)
d = abs(np.sum(np.multiply(q1, q2)))
theta = 2 * np.arccos(d) * 180 / np.pi
errx = np.linalg.norm(actualx - posx)
return errx, theta
def rmse_p(y_true, y_pred):
return torch.sqrt(torch.sum(torch.square(y_pred-y_true), axis=-1))
def rmse_q(y_true, y_pred):
return BETA * torch.sqrt(torch.sum(torch.square(y_true-y_pred), axis=-1))
class rmse_wrap(torch.nn.Module):
""" Wrapper for dealing with the position loss and rotation loss. The rotation loss is multiplied with BETA. """
def __init__(self):
super(rmse_wrap, self).__init__()
def __call__(self, output, target):
loss1 = rmse_p(output[:, :3], target[:, :3])
loss2 = rmse_q(output[:, 3:], target[:, 3:])
return loss1 + loss2
# For keeping track of the total run time
s_time = time.time()
# Loads the training/test data
pred_abs, pred_rel, gt_abs, gt_rel = utilities.get_training_data(train_dataset_abs, train_dataset_rel, backrot)
pred_abs_test, pred_rel_test, gt_abs_test, gt_rel_test = utilities.get_training_data(test_dataset_abs, test_dataset_rel, backrot)
# Generator for creating batches of size batchSize
data_gen = utilities.gen_data_batch(pred_abs, pred_rel, gt_abs, gt_rel, batchSize, time_size)
optimizer = optim.Adam(model.parameters())
criterion = rmse_wrap().to(device)
# train for ITERATIONS
for i in range(ITERATIONS):
print(i)
# Load the next batch and put in on the device
X_batch, Y_batch = next(data_gen)
X_batch = X_batch.transpose((1, 0, 2))
X_batch = torch.tensor(X_batch)
Y_batch = torch.from_numpy(Y_batch)
inputs = X_batch.to(device)
labels = Y_batch.to(device)
# predict on the current input data
preds = model(inputs).float()
# Loss function
cost = criterion(preds, labels).mean()
optimizer.zero_grad()
cost.backward(retain_graph=True)
optimizer.step()
model.zero_grad()
model.detach_hidden()
print(cost)
if (i+1) % int(TIMESTAMP) == 0: # every TIMESTAMP steps: evaluate the network on the test data
model.eval()
hidden1_copy = model.cell1.hidden
model.cell1.hidden = None
if stacks == 2:
hidden2_copy = model.cell2.hidden
model.cell2.hidden = None
# create a generator for the test data (Batchsize = 1)
datagen = utilities.limited_gen_data(pred_abs_test, pred_rel_test, gt_abs_test, gt_rel_test, time_size)
posxs = []
posqs = []
for ims, xs, qs in datagen: # Iterate over the test data and evaluate it
inputs = np.zeros((1, time_size, 14)).astype("float32")
inputs[0, :] = ims
inputs = inputs.transpose((1, 0, 2))
inputs = torch.tensor(inputs).to(device)
out = model(inputs).float()
posx = out[0][:3].detach().cpu().numpy()
posq = out[0][3:].detach().cpu().numpy()
actualx = xs
actualq = qs
errx, theta = getError(posx, posq, actualx, actualq)
posxs.append(errx)
posqs.append(theta)
# Write the results into a file
print('Median error ' + str(np.median(posxs)) + ' m and ' + str(np.nanmedian(posqs)) + ' degrees')
log.write("Median error %f m and %f degrees \n" % (np.median(posxs), np.nanmedian(posqs)))
log.flush()
torch.save(model.state_dict(), outputWeightspath + str(i) + '.h5')
model.cell1.hidden = hidden1_copy
if stacks == 2:
model.cell2.hidden = hidden2_copy
model.train()
print("Total time: ", time.time() - s_time)
|