def preprocess(self, tokenizer_name, var_length=False):
# types of csv columns
time_start = time.time()
tokenizer = self.choose_tokenizer(tokenizer_name)
self.text = data.Field(batch_first=True,
tokenize=tokenizer,
include_lengths=var_length)
self.qid = data.Field()
self.target = data.Field(sequential=False,
use_vocab=False,
is_target=True)
# read and tokenize data
print('read and tokenize data...')
self.train = MyTabularDataset(path=self.train_csv,
format='csv',
fields={
'qid': ('qid', self.qid),
'question_text': ('text', self.text),
'target': ('target', self.target)
})
self.test = MyTabularDataset(path=self.test_csv,
format='csv',
fields={
'qid': ('qid', self.qid),
'question_text': ('text', self.text)
})
print_duration(time_start, 'time to read and tokenize data: ')
self.text.build_vocab(self.train, self.test, min_freq=1)
self.qid.build_vocab(self.train, self.test)
print_duration(time_start, 'time to read, tokenize and build vocab: ')
def k_split_indices(randperm, cut_idxs, k, i, is_test):
time_start = time.time()
# val index
val_start_idx = cut_idxs[i]
val_end_idx = cut_idxs[i + 1]
val_index = randperm[val_start_idx:val_end_idx]
# test index
if is_test:
if i <= k - 2:
test_start_idx = cut_idxs[i + 1]
test_end_idx = cut_idxs[i + 2]
else:
test_start_idx = cut_idxs[0]
test_end_idx = cut_idxs[1]
test_index = randperm[test_start_idx:test_end_idx]
else:
test_index = []
val_test_index = set(val_index + test_index)
# train index
print_duration(time_start, message='k_split_indices time')
train_index = [idx for idx in randperm if idx not in val_test_index]
print_duration(time_start, message='k_split_indices time')
return train_index, val_index, test_index
def read_embedding(self, embeddings, unk_std, max_vectors, to_cache):
time_start = time.time()
unk_init = partial(normal_init, std=unk_std)
for emb in embeddings:
self.vectors.append(
MyVectors(emb,
cache=self.cache,
to_cache=to_cache,
unk_init=unk_init,
max_vectors=max_vectors))
print_duration(time_start, 'time to read embedding: ')
def main():
# load up the SVM stored from prior training
try:
svm = cv2.ml.SVM_load(params.HOG_SVM_PATH_SAVED)
except:
print("Missing files SVM")
print("-- have you performed training to produce this file ?")
exit()
# load ** testing ** data sets in the same class order as training
# (here we perform patch sampling only from the centre of the +ve
# class and only a single sample is taken
# hence [0,0] sample sizes and [False,True] centre weighting flags)
print("Loading test data as a batch ...")
paths = [params.DATA_testing_path_neg, params.DATA_testing_path_pos]
use_centre_weighting = [False, True]
class_names = params.DATA_CLASS_NAMES
imgs_data = utils.load_images(paths, class_names, [0, 0],
use_centre_weighting)
print("Computing HOG descriptors...") # for each testing image
start = cv2.getTickCount()
[img_data.compute_hog_descriptor() for img_data in imgs_data]
utils.print_duration(start)
# get the example/sample HOG descriptors and class labels
samples, class_labels = utils.get_hog_descriptors(
imgs_data), utils.get_class_labels(imgs_data)
# perform batch SVM classification over the whole set
print("Performing batch SVM classification over all data ...")
results = svm.predict(samples)
output = results[1].ravel()
# compute and report the error over the whole set
error = ((np.absolute(class_labels.ravel() - output).sum()) /
float(output.shape[0]))
print("Successfully trained SVM with {}% testing set error".format(
round(error * 100, 2)))
print(
"-- meaining the SVM got {}% of the testing examples correct!".format(
round((1.0 - error) * 100, 2)))
def main():
############################################################################
# load our training data set of images examples
program_start = cv2.getTickCount()
print("Loading images...")
start = cv2.getTickCount()
# N.B. specify data path names in same order as class names (neg, pos)
paths = [params.DATA_training_path_neg, params.DATA_training_path_pos]
# build a list of class names automatically from our dictionary of class (name,number) pairs
class_names = [utils.get_class_name(class_number) for class_number in range(len(params.DATA_CLASS_NAMES))]
# specify number of sub-window samples to take from each positive and negative
# example image in the data set
# N.B. specify in same order as class names (neg, pos) - again
sampling_sizes = [params.DATA_training_sample_count_neg, params.DATA_training_sample_count_pos]
# do we want to take samples only centric to the example image or ramdonly?
# No - for background -ve images (first class)
# Yes - for object samples +ve images (second class)
sample_from_centre = [False, True];
# perform image loading
imgs_data = utils.load_images(paths, class_names, sampling_sizes, sample_from_centre,
params.DATA_WINDOW_OFFSET_FOR_TRAINING_SAMPLES, params.DATA_WINDOW_SIZE);
print(("Loaded {} image(s)".format(len(imgs_data))))
utils.print_duration(start)
############################################################################
# perform HOG feature extraction
print("Computing HOG descriptors...") # for each training image
start = cv2.getTickCount()
#each HoG descriptor is stored in its respective img_data instance
[img_data.compute_hog_descriptor() for img_data in imgs_data]
utils.print_duration(start)
############################################################################
# train an SVM based on these norm_features
print("Training SVM...")
start = cv2.getTickCount()
# define SVM parameters
svm = cv2.ml.SVM_create()
svm.setType(cv2.ml.SVM_C_SVC) # set SVM type
svm.setKernel(params.HOG_SVM_kernel) # use specific kernel type
# get hog descriptor for each image and store in single global array
samples = utils.get_hog_descriptors(imgs_data)
# get class label for each training image (i.e. 0 for other, 1 for pedestrian... can extend)
class_labels = utils.get_class_labels(imgs_data);
# specify the termination criteria for the SVM training
svm.setTermCriteria((cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_COUNT, params.HOG_SVM_max_training_iterations, 1.e-06))
# perform auto training for the SVM which will essentially perform grid
# search over the set of parameters for the chosen kernel and the penalty
# cost term, C (N.B. trainAuto() syntax is correct as of OpenCV 3.4.x)
svm.trainAuto(samples, cv2.ml.ROW_SAMPLE, class_labels, kFold = 10, balanced = True);
# save the trained SVM to file so that we can load it again for testing / detection
svm.save(params.HOG_SVM_PATH_TRAIN)
############################################################################
# measure performance of the SVM trained on the bag of visual word features
# perform prediction over the set of examples we trained over
output = svm.predict(samples)[1].ravel()
error = (np.absolute(class_labels.ravel() - output).sum()) / float(output.shape[0])
# we are succesful if our prediction > than random
# e.g. for 2 class labels this would be 1/2 = 0.5 (i.e. 50%)
if error < (1.0 / len(params.DATA_CLASS_NAMES)):
print("Trained SVM obtained {}% training set error".format(round(error * 100,2)))
print("-- meaining the SVM got {}% of the training examples correct!".format(round((1.0 - error) * 100,2)))
else:
print("Failed to train SVM. {}% error".format(round(error * 100,2)))
utils.print_duration(start)
print(("Finished training HoG detector. {}".format(format_time(get_elapsed_time(program_start)))))
def fit(self, epoch, n_eval, tresh, early_stop, warmup_epoch, clip):
step = 0
min_loss = 1e5
max_f1 = -1
max_test_f1 = -1
no_improve_epoch = 0
no_improve_in_previous_epoch = False
fine_tuning = False
losses = []
best_test_info = None
torch.backends.cudnn.benchmark = False
eval_every = int(len(list(iter(self.train_dl))) / n_eval)
time_start = time.time()
print(self.model)
for e in range(epoch):
self.scheduler.step()
if e >= warmup_epoch:
if no_improve_in_previous_epoch:
no_improve_epoch += 1
if no_improve_epoch >= early_stop:
e = e - 1
break
else:
no_improve_epoch = 0
no_improve_in_previous_epoch = True
if not fine_tuning and e >= warmup_epoch:
self.model.embedding.weight.requires_grad = True
fine_tuning = True
self.train_dl.init_epoch()
for train_batch in iter(self.train_dl):
step += 1
self.model.zero_grad()
self.model.train()
model_input = self.to_cuda(train_batch.text)
y = train_batch.target.type(torch.Tensor).cuda()
pred = self.model.forward(*model_input).view(-1)
loss = self.loss_func(pred, y)
self.recorder.tr_record.append(
{'tr_loss': loss.cpu().data.numpy()})
loss.backward()
total_norm = nn.utils.clip_grad_norm_(self.model.parameters(),
clip)
# parameters = list(filter(lambda p: p.grad is not None, self.model.parameters()))
# total_norm = 0
# for p in parameters:
# param_norm = p.grad.data.norm(2)
# total_norm += param_norm.item() ** 2
# total_norm = total_norm ** (1. / 2)
self.recorder.norm_record.append({'grad_norm': total_norm})
self.optimizer.step()
if step % eval_every == 0:
# self.scheduler.step()
with torch.no_grad():
# train evaluation
losses.append(loss.cpu().data.numpy())
train_loss = np.mean(losses)
# val evaluation
val_loss, val_f1, val_ids, val_prob, val_true = self.evaluate(
self.val_dl, tresh)
pred_to_csv(
val_ids, val_prob, val_true,
f'tmp/val_probs_{int(step / eval_every) - 1}.csv')
self.recorder.val_record.append({
'step': step,
'loss': val_loss,
'f1': val_f1
})
info = {
'best_ep': e,
'step': step,
'train_loss': train_loss,
'val_loss': val_loss,
'val_f1': val_f1
}
self.recorder.save_step(info, message=True)
if val_f1 > max_f1:
self.recorder.save(self.model, info)
max_f1 = val_f1
no_improve_in_previous_epoch = False
if val_loss < min_loss:
min_loss = val_loss
# test evaluation
# if self.args.test:
# test_loss, test_f1 = self.evaluate(self.test_dl, tresh)
# test_info = {'test_ep': e, 'test_step': step, 'test_loss': test_loss, 'test_f1': test_f1}
# self.recorder.test_record.append({'step': step, 'loss': test_loss, 'f1': test_f1})
# print('epoch {:02} - step {:06} - test_loss {:.4f} - test_f1 {:.4f}'.format(*list(test_info.values())))
# if test_f1 >= max_test_f1:
# max_test_f1 = test_f1
# best_test_info = test_info
tr_time = print_duration(time_start, 'training time: ')
self.recorder.append_info({'ep_time': tr_time / (e + 1)})
#if self.args.test:
# self.recorder.append_info(best_test_info, message='Best results for test:')
self.recorder.append_info({'min_loss': min_loss}, 'min val loss: ')
self.model, info = self.recorder.load(message='best model:')
# final train evaluation
train_loss, train_f1, _, _, _ = self.evaluate(self.train_dl, tresh)
tr_info = {'train_loss': train_loss, 'train_f1': train_f1}
self.recorder.append_info(tr_info, message='train loss and f1:')
def embedding_lookup(self):
print('embedding lookup...')
time_start = time.time()
self.text.vocab.load_vectors(self.vectors)
print_duration(time_start, 'time for embedding lookup: ')
return
feed_dict={x: testX, y_: testY, keep_prob: 1.0})
exam_cross_entropy_avg = utils.get_moving_avg(exam_cross_entropy_avg, test_cross_entropy)
exam_accuracy_avg = utils.get_moving_avg(exam_accuracy_avg, test_accuracy)
exam_writer.add_summary(summary, i)
if i % DISPLAY_STEP == 0:
print("step %d, exam accuracy %f cross_entropy %f" % (i, exam_accuracy_avg, exam_cross_entropy_avg))
print("=====> test result training accuracy %f cross_entropy %f" % (exam_accuracy_avg, exam_cross_entropy_avg))
exam_writer.close()
if PUBLISH:
# start publish:
test_data_source = Input(TEST_PATH, TRAIN_FILE, shuffle=False,
loop=False, temp_file_path=TEMP_FILE,
n_mfcc=FLATTEN_SIZE_W,
fixed_sample=FLATTEN_SIZE_H)
result_frame = pd.DataFrame(np.zeros([test_data_source.get_total_files(), 2]), columns=['fname', 'label'])
for i in range(0, test_data_source.get_total_files()):
testX, _ = test_data_source.next(1)
# testX = np.pad(testX, ((0, BATCH_SIZE - 1), (0, 0)), 'constant')
logits = sess.run(y_conv,
feed_dict={x: testX, keep_prob: 1.0})
result_frame.iloc[i, 0] = test_data_source.get_last_read_file_name()
result_frame.iloc[i, 1] = utils.trans_index_to_label(test_data_source, utils.top_n_value(tf.squeeze(logits).eval(session=sess), 3))
result_frame.to_csv(path_or_buf=SUBMIT_FILE)
print("=====> publish done")
sess.close()
utils.print_duration(start_time)
print("===== job finish =====")
# scale the disparity to 8-bit for viewing
disparity_scaled = (disparity / 16.).astype(np.uint8)
# crop area not seen by *both* cameras and and area with car bonnet
disparity_scaled = utils.crop_image(disparity_scaled, 0, 390, 135, width)
return disparity_scaled
# </section>End of Functions Section
# <section>~~~~~~~~~~~~~~~~~~~~~~~~~~~~Main~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Colors = utils.gen_N_colors(
81) #get N different colors for the N possible classes
utils.print_duration(time_to_setup) #print how long it took to set up
# cycle through the images
for filename_left in left_file_list:
# <section>---------------Directory Checks---------------
# skipping if requested
if check_skip(skip_forward_file_pattern, filename_left):
continue
else:
skip_forward_file_pattern = ""
# from the left image filename get the correspondoning right image
filename_right = filename_left.replace("_L", "_R")
full_path_filenames = join_paths_both_sides(full_path_directory_left,
filename_left,
full_path_directory_right,
