def generate_and_save_heave_files():
root = path_to_echograms()
echs = get_echograms()
for i, ech in enumerate(echs):
if i % 100 == 0:
print(len(echs), i)
# Get vertical pixel resolution
r = ech.range_vector
r_diff = np.median(r[1:] - r[:-1])
# Convert heave value from meters to number of pixels
heave = np.round(ech.heave / r_diff).astype(np.int)
assert heave.size == ech.shape[1]
labels_old = ech.label_numpy()
labels_new = np.zeros_like(labels_old)
# Create new labels: Move each labels column up/down corresponding to heave
for x, h in enumerate(list(heave)):
if h == 0:
labels_new[:, x] = labels_old[:, x]
elif h > 0:
labels_new[:-h, x] = labels_old[h:, x]
else:
labels_new[-h:, x] = labels_old[:h, x]
# Save new labels as new memmap file
path_save = root + ech.name + '/labels_heave'
save_memmap(labels_new, path_save, dtype=labels_new.dtype)
def sampling_echograms_full(args):
path_to_echograms = paths.path_to_echograms()
# path_to_echograms = '/Users/changkyu/Documents/GitHub/save_pts/sampled'
samplers_train = torch.load(
os.path.join(path_to_echograms, 'sampler3_tr.pt'))
supervised_count = int(len(samplers_train[0]) * args.semi_ratio)
samplers_supervised = []
samplers_tr_rest = []
for samplers in samplers_train:
samplers_supervised.append(samplers[:supervised_count])
samplers_tr_rest.append(samplers[supervised_count:])
augmentation = CombineFunctions([add_noise_img, flip_x_axis_img])
data_transform = CombineFunctions([remove_nan_inf_img, db_with_limits_img])
dataset_semi = DatasetImg(samplers_supervised,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
dataset_tr_rest = DatasetImg(samplers_tr_rest,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
return dataset_semi, dataset_tr_rest
def sampling_echograms_full(args):
tr_ratio = [0.97808653, 0.01301181, 0.00890166]
path_to_echograms = paths.path_to_echograms()
########
samplers_train = torch.load(
os.path.join(path_to_echograms, 'sampler3_tr.pt'))
semi_count = int(len(samplers_train[0]) * args.semi_ratio)
samplers_semi = [samplers[:semi_count] for samplers in samplers_train]
augmentation = CombineFunctions([add_noise_img, flip_x_axis_img])
data_transform = CombineFunctions([remove_nan_inf_img, db_with_limits_img])
dataset_cp = DatasetImg(samplers_train,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
dataset_semi = DatasetImg(samplers_semi,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
return dataset_cp, dataset_semi
def sampling_echograms_full_for_comparisonP2(args):
path_to_echograms = paths.path_to_echograms()
data = torch.load(os.path.join(path_to_echograms, 'data_tr_TEST_200.pt'))
label = torch.load(os.path.join(path_to_echograms, 'label_tr_TEST_200.pt'))
data_transform = CombineFunctions(
[remove_nan_inf_for_comparisonP2, db_with_limits_for_comparisonP2])
label_transform = CombineFunctions([
index_0_1_27_for_comparisonP2,
relabel_with_threshold_morph_close_for_comparisonP2,
seabed_checker_for_comparisonP2
])
semi_count = int(len(data) * args.semi_ratio)
dataset_cp = DatasetImg_for_comparisonP2(
data=data,
label=label,
label_transform_function=label_transform,
data_transform_function=data_transform)
dataset_semi = DatasetImg_for_comparisonP2(
data=data[:semi_count],
label=label[:semi_count],
label_transform_function=label_transform,
data_transform_function=data_transform)
return dataset_cp, dataset_semi
def test_and_plot_2019(test_pred_large_2019, test_label_large_2019, epoch, args, idx=2):
path_to_echograms = paths.path_to_echograms()
data_2019, label_2019, patch_loc = torch.load(os.path.join(path_to_echograms, 'data_label_patch_loc_te_2019_%d.pt' % idx))
data_transform = CombineFunctions([remove_nan_inf_for_comparisonP2, db_with_limits_for_comparisonP2])
label_transform = CombineFunctions([index_0_1_27_for_comparisonP2, relabel_with_threshold_morph_close_for_comparisonP2, seabed_checker_for_comparisonP2])
boxsize = 5
plt.figure(figsize=(boxsize * patch_loc[1], boxsize * patch_loc[0] * 4))
for i in range(len(data_2019)):
l = label_2019[i]
d = data_2019[i]
d, l = label_transform(d, l)
d, l = data_transform(d, l)
dim = np.shape(l)
labels_rgb = np.ones((dim[0], dim[1], 3))
sandeel = np.where(l == 1)
other = np.where(l == 2)
labels_rgb[sandeel[0], sandeel[1], 0] = 0
labels_rgb[sandeel[0], sandeel[1], 1] = 0 # sandeel blue
labels_rgb[other[0], other[1], 1] = 0
labels_rgb[other[0], other[1], 2] = 0 # other red
pred = test_pred_large_2019[i]
pred_rgb = np.ones((dim[0], dim[1], 3))
pred_sandeel = np.where(pred == 1)
pred_other = np.where(pred == 2)
pred_rgb[pred_sandeel[0], pred_sandeel[1], 0] = 0
pred_rgb[pred_sandeel[0], pred_sandeel[1], 1] = 0 # sandeel blue
pred_rgb[pred_other[0], pred_other[1], 1] = 0
pred_rgb[pred_other[0], pred_other[1], 2] = 0 # other red
lbb = test_label_large_2019[i]
lbb_rgb = np.ones((dim[0], dim[1], 3))
lbb_sandeel = np.where(lbb == 1)
lbb_other = np.where(lbb == 2)
lbb_rgb[lbb_sandeel[0], lbb_sandeel[1], 0] = 0
lbb_rgb[lbb_sandeel[0], lbb_sandeel[1], 1] = 0 # sandeel blue
lbb_rgb[lbb_other[0], lbb_other[1], 1] = 0
lbb_rgb[lbb_other[0], lbb_other[1], 2] = 0 # other red
plt.subplot(patch_loc[0] * 4, patch_loc[1], i + 1)
plt.imshow(labels_rgb)
plt.subplot(patch_loc[0] * 4, patch_loc[1], patch_loc[1] * patch_loc[0] + i + 1)
plt.imshow(pred_rgb)
plt.subplot(patch_loc[0] * 4, patch_loc[1], patch_loc[1] * patch_loc[0] * 2 + i + 1)
plt.imshow(lbb_rgb)
plt.subplot(patch_loc[0] * 4, patch_loc[1], patch_loc[1] * patch_loc[0] * 3 + i + 1)
plt.imshow(d[-1])
plt.setp(plt.gcf().get_axes(), xticks=[], yticks=[])
plt.tight_layout()
plt.savefig(os.path.join(args.pred_2019, '%d_data_pred_label_2019_%d_patch.pdf' % (epoch, idx)))
plt.close()
return
def sampling_echograms_full(args):
tr_ratio = [0.97808653, 0.01301181, 0.00890166]
path_to_echograms = paths.path_to_echograms()
########
samplers_train = torch.load(
os.path.join(path_to_echograms, 'sampler3_tr.pt'))
samplers_bg = torch.load(
os.path.join(path_to_echograms, 'train_bg_32766.pt'))
supervised_count = int(len(samplers_train[0]) * args.semi_ratio)
total_unsupervised_count = int(
(len(samplers_train[0]) - supervised_count) * args.nmb_category)
unlab_size = [int(ratio * total_unsupervised_count) for ratio in tr_ratio]
if np.sum(unlab_size) != total_unsupervised_count:
unlab_size[0] += total_unsupervised_count - np.sum(unlab_size)
samplers_supervised = []
samplers_unsupervised = []
for samplers in samplers_train:
samplers_supervised.append(samplers[:supervised_count])
samplers_unsupervised.append(samplers[supervised_count:])
samplers_unsupervised[0].extend(samplers_bg)
samplers_unbal_unlab = []
for sampler, size in zip(samplers_unsupervised, unlab_size):
samplers_unbal_unlab.append(sampler[:size])
samplers_semi_unbal_unlab_long = []
for sampler_semi, sampler_unb_unl in zip(samplers_supervised,
samplers_unbal_unlab):
samplers_semi_unbal_unlab_long.extend(
np.concatenate([sampler_semi, sampler_unb_unl]))
list_length = len(samplers_train[0])
num_classes = len(samplers_train)
samplers_cp = [
samplers_semi_unbal_unlab_long[i * list_length:(i + 1) * list_length]
for i in range(num_classes)
]
########
augmentation = CombineFunctions([add_noise_img, flip_x_axis_img])
data_transform = CombineFunctions([remove_nan_inf_img, db_with_limits_img])
dataset_cp = DatasetImg(samplers_cp,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
dataset_semi = DatasetImg(samplers_supervised,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
return dataset_cp, dataset_semi
def save_all_seabeds():
"""
Loop through all echograms and generate seabed-estimates
:return:
"""
path_to_echograms = paths.path_to_echograms()
echogram_names = os.listdir(path_to_echograms)
echograms = [Echogram(path_to_echograms + e) for e in echogram_names]
for e in echograms:
e.get_seabed(save_to_file=True, ignore_saved=True)
def get_echograms_revised(eg_names_full, sample_idx, num_echograms=100):
path_to_echograms = paths.path_to_echograms()
index_list = np.arange(len(eg_names_full)//num_echograms)
if sample_idx not in set(index_list):
sample_idx = 0
print('Reset_sample_idx')
eg_idx = np.arange(num_echograms*sample_idx, num_echograms *(sample_idx+1))
eg_names = list(map(eg_names_full.__getitem__, eg_idx))
echograms = [Echogram(os.path.join(path_to_echograms, e)) for e in eg_names]
sample_idx += 1
return echograms, sample_idx
def sampling_echograms_test(args):
path_to_echograms = paths.path_to_echograms()
samplers_test = torch.load(
os.path.join(path_to_echograms, 'sampler6_te.pt'))
data_transform = CombineFunctions([remove_nan_inf_img, db_with_limits_img])
dataset_test = DatasetImg(samplers_test,
args.sampler_probs,
augmentation_function=None,
data_transform_function=data_transform)
return dataset_test
def sampling_echograms_full(args):
path_to_echograms = paths.path_to_echograms()
samplers_train = torch.load(
os.path.join(path_to_echograms, 'samplers_sb_1024_3.pt'))
augmentation = CombineFunctions([add_noise_img, flip_x_axis_img])
data_transform = CombineFunctions([remove_nan_inf_img, db_with_limits_img])
dataset_cp = DatasetImg(samplers_train,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
return dataset_cp
def test_analysis(predictions, predictions_mat, epoch, args):
path_to_echograms = paths.path_to_echograms()
labels_origin = torch.load(os.path.join(path_to_echograms, 'label_TEST_60_after_transformation.pt'))
if np.shape(predictions_mat) == (60, 256, 256, 3):
predictions_mat = predictions_mat.transpose(0, 3, 1, 2)
keep_test_idx = np.where(labels_origin > -1)
labels_vec = labels_origin[keep_test_idx]
predictions_vec = predictions[keep_test_idx]
predictions_mat_sampled = predictions_mat[keep_test_idx[0], :, keep_test_idx[1], keep_test_idx[2]]
fpr, tpr, roc_auc, roc_auc_macro = roc_curve_macro(labels_vec, predictions_mat_sampled)
prob_mat, mat, f1_score, kappa = conf_mat(ylabel=labels_vec, ypred=predictions_vec, args=args)
acc_bg, acc_se, acc_ot = prob_mat.diagonal()
plot_macro(fpr, tpr, roc_auc, epoch, args)
plot_conf(epoch, prob_mat, mat, f1_score, kappa, args)
return fpr, tpr, roc_auc, roc_auc_macro, prob_mat, mat, f1_score, kappa, acc_bg, acc_se, acc_ot
def sampling_echograms_for_s3vm(args):
path_to_echograms = paths.path_to_echograms()
samplers_bg = torch.load(
os.path.join(path_to_echograms, 'train_bg_32766.pt'))
list_length = len(samplers_bg) // args.nmb_category
samplers_bg = [
samplers_bg[i * list_length:(i + 1) * list_length]
for i in range(args.nmb_category)
]
augmentation = CombineFunctions([add_noise_img, flip_x_axis_img])
data_transform = CombineFunctions([remove_nan_inf_img, db_with_limits_img])
dataset_bg_full = DatasetImg(samplers_bg,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
return dataset_bg_full
def sampling_echograms_test_for_comparisonP2():
path_to_echograms = paths.path_to_echograms()
data = torch.load(os.path.join(path_to_echograms, 'data_te_TEST_60.pt'))
label = torch.load(os.path.join(path_to_echograms, 'label_te_TEST_60.pt'))
data_transform = CombineFunctions(
[remove_nan_inf_for_comparisonP2, db_with_limits_for_comparisonP2])
label_transform = CombineFunctions([
index_0_1_27_for_comparisonP2,
relabel_with_threshold_morph_close_for_comparisonP2,
seabed_checker_for_comparisonP2
])
dataset = DatasetImg_for_comparisonP2(
data=data,
label=label,
label_transform_function=label_transform,
data_transform_function=data_transform)
return dataset
def sampling_echograms_full(window_size, args):
path_to_echograms = paths.path_to_echograms()
with open(os.path.join(path_to_echograms, 'memmap_2014_heave.pkl'),
'rb') as fp:
eg_names_full = pickle.load(fp)
echograms = get_echograms_full(eg_names_full)
echograms_train, echograms_val, echograms_test = cps.partition_data(
echograms,
args.partition,
portion_train_test=0.8,
portion_train_val=0.75)
sampler_bg_train = Background(echograms_train, window_size)
sampler_sh27_train = Shool(echograms_train, window_size, 27)
sampler_sbsh27_train = ShoolSeabed(echograms_train,
window_size,
args.window_dim // 4,
fish_type=27)
sampler_sh01_train = Shool(echograms_train, window_size, 1)
sampler_sbsh01_train = ShoolSeabed(echograms_train,
window_size,
args.window_dim // 4,
fish_type=1)
samplers_train = [
sampler_bg_train, sampler_sh27_train, sampler_sbsh27_train,
sampler_sh01_train, sampler_sbsh01_train
]
augmentation = CombineFunctions([add_noise, flip_x_axis])
label_transform = CombineFunctions(
[index_0_1_27, relabel_with_threshold_morph_close])
data_transform = CombineFunctions([remove_nan_inf, db_with_limits])
dataset_train = Dataset(samplers_train,
window_size,
args.frequencies,
args.batch * args.iteration_train,
args.sampler_probs,
augmentation_function=augmentation,
label_transform_function=label_transform,
data_transform_function=data_transform)
return dataset_train
def sampling_echograms_2019_for_comparisonP2(echogram_idx=2,
path_to_echograms=None):
if path_to_echograms == None:
path_to_echograms = paths.path_to_echograms()
data, label, patch_loc = torch.load(
os.path.join(path_to_echograms,
'data_label_patch_loc_te_2019_%d.pt' % echogram_idx))
data_transform = CombineFunctions(
[remove_nan_inf_for_comparisonP2, db_with_limits_for_comparisonP2])
label_transform = CombineFunctions([
index_0_1_27_for_comparisonP2,
relabel_with_threshold_morph_close_for_comparisonP2,
seabed_checker_for_comparisonP2
])
dataset_2019 = DatasetImg_for_comparisonP2(
data=data,
label=label,
label_transform_function=label_transform,
data_transform_function=data_transform)
return dataset_2019, label, patch_loc
parser.add_argument('--partition', type=str, default='train_only',
help='echogram partition (tr/val/te) by year')
parser.add_argument('--iteration_train', type=int, default=1200,
help='num_tr_iterations per one batch and epoch')
parser.add_argument('--sampler_probs', type=list, default=None,
help='[bg, sh27, sbsh27, sh01, sbsh01], default=[1, 1, 1, 1, 1]')
parser.add_argument('--resume',
default=os.path.join(current_dir, 'checkpoint.pth.tar'), type=str, metavar='PATH',
help='path to checkpoint (default: None)')
parser.add_argument('--exp', type=str,
default=current_dir, help='path to exp folder')
return parser.parse_args(args=[])
args = parse_args()
window_size = [args.window_dim, args.window_dim]
path_to_echograms = paths.path_to_echograms()
with open(os.path.join(path_to_echograms, 'memmap_2014_heave.pkl'), 'rb') as fp:
eg_names_full = pickle.load(fp)
echograms = get_echograms_full(eg_names_full)
echograms_train, echograms_val, echograms_test = cps.partition_data(echograms, args.partition, portion_train_test=0.8,
portion_train_val=0.75)
sampler_bg_train = Background(echograms_train, window_size)
# sampler_sh27_train = Shool(echograms_train, window_size, 27)
# sampler_sh01_train = Shool(echograms_train, window_size, 1)
# sampler_sbsh27_train = ShoolSeabed(echograms_train, window_size, args.window_dim // 4, fish_type=27)
# sampler_sbsh01_train = ShoolSeabed(echograms_train, window_size, args.window_dim // 4, fish_type=1)
data_bg = DatasetSampler(sampler_bg_train, window_size, args.frequencies)
# data_sh27 = DatasetSampler(sampler_sh27_train, window_size, args.frequencies)
def sampling_echograms_full(args):
# idx_2000 = np.random.choice(3000, size=2000, replace=False).tolist()
# sampler_2000 = []
# for i in range(5):
# sampler_2000.append([samplers_train[i][idx] for idx in idx_2000])
# torch.save(sampler_2000, 'samplers_2000.pt')
# path_to_echograms = "/Users/changkyu/Documents/GitHub/echogram/memmap/memmap_set"
# bg = torch.load(os.path.join(path_to_echograms, 'numpy_bg_2999.pt')) + \
# torch.load(os.path.join(path_to_echograms, 'numpy_bg_5999.pt'))
# bg_idx = np.random.choice(np.arange(len(bg)), size=3000, replace=False)
# bg = [bg[idx] for idx in bg_idx]
#
# sbsh01 = torch.load(os.path.join(path_to_echograms, 'numpy_sbsh01_2999.pt')) +\
# torch.load(os.path.join(path_to_echograms, 'numpy_sbsh01_5999.pt')) +\
# torch.load(os.path.join(path_to_echograms, 'numpy_sbsh01_8999.pt')) +\
# torch.load(os.path.join(path_to_echograms, 'numpy_sbsh01_11999.pt')) +\
# torch.load(os.path.join(path_to_echograms, 'numpy_sbsh01_12667.pt'))
# sbsh01_idx = np.random.choice(np.arange(len(sbsh01)), size=3000, replace=False)
# sbsh01 = [sbsh01[idx] for idx in sbsh01_idx]
#
# sbsh27 = torch.load(os.path.join(path_to_echograms, 'numpy_sbsh27_2999.pt'))+\
# torch.load(os.path.join(path_to_echograms, 'numpy_sbsh27_3079.pt'))
# sbsh27_idx = np.random.choice(np.arange(len(sbsh27)), size=3000, replace=False)
# sbsh27 = [sbsh27[idx] for idx in sbsh27_idx]
#
# sh01 = torch.load(os.path.join(path_to_echograms, 'numpy_sh01_2999.pt'))+\
# torch.load(os.path.join(path_to_echograms, 'numpy_sh01_4046.pt'))
# sh01_idx = np.random.choice(np.arange(len(sh01)), size=3000, replace=False)
# sh01 = [sh01[idx] for idx in sh01_idx]
#
# sh27 = torch.load(os.path.join(path_to_echograms, 'numpy_sh27_2999.pt'))+\
# torch.load(os.path.join(path_to_echograms, 'numpy_sh27_3549.pt'))
# sh27_idx = np.random.choice(np.arange(len(sh27)), size=3000, replace=False)
# sh27 = [sh27[idx] for idx in sh27_idx]
# samplers_train = [bg, sh27, sbsh27, sh01, sbsh01]
# torch.save(samplers_train, 'samplers_3000.pt')
# samplers_train = [bg, sh27, sbsh27, sh01, sbsh01]
# def sample_align(samplers):
# num_samples = []
# new_samplers = []
# for i in range(len(samplers)):
# num_samples.append(len(samplers[i]))
# max_num_sample = np.min(num_samples)
# print(max_num_sample)
# for i in range(len(samplers)):
# new_samplers.append(np.random.choice(samplers[i], size=max_num_sample, replace=False))
# return new_samplers
# path_to_echograms = '/Users/changkyu/Documents/GitHub/echogram/memmap/memmap_set/'
path_to_echograms = paths.path_to_echograms()
samplers_train = torch.load(
os.path.join(path_to_echograms, 'samplers_500_three.pt'))
augmentation = CombineFunctions([add_noise_img, flip_x_axis_img])
data_transform = CombineFunctions([remove_nan_inf_img, db_with_limits_img])
dataset_cp = DatasetImg(samplers_train,
1500,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
return dataset_cp
def sampling_echograms_full(args):
path_to_echograms = paths.path_to_echograms()
assert (args.semi_ratio in [0.01, 0.05, 0.1, 0.2]), 'Fix args.semi-ratio in a given range'
samplers_train = torch.load(os.path.join(path_to_echograms, 'sampler6_tr.pt'))
samplers_bg_unb = torch.load(os.path.join(path_to_echograms, 'bg_unb_11000_tr.pt'))
semi_count = int(len(samplers_train[0]) * args.semi_ratio)
if len(samplers_train) == 6:
if args.semi_ratio == 0.01:
unlab_size = [12992, 195, 200, 1344, 90, 29]
elif args.semi_ratio == 0.05:
unlab_size = [12468, 187, 192, 1289, 86, 28]
elif args.semi_ratio == 0.1:
unlab_size = [11811, 177, 182, 1222, 82, 26]
elif args.semi_ratio == 0.2:
unlab_size = [10499, 157, 162, 1086, 73, 23]
elif len(samplers_train) == 3: # combined 3classes case
if args.semi_ratio == 0.01:
unlab_size = [14336, 285, 229]
elif args.semi_ratio == 0.05:
unlab_size = [13757, 273, 220]
elif args.semi_ratio == 0.1:
unlab_size = [13033, 259, 208]
elif args.semi_ratio == 0.2:
unlab_size = [11585, 230, 185]
samplers_semi = []
samplers_rest = []
for samplers in samplers_train:
samplers_semi.append(samplers[:semi_count])
samplers_rest.append(samplers[semi_count:])
samplers_rest[0].extend(samplers_bg_unb)
samplers_unbal_unlab = []
for sampler, size in zip(samplers_rest, unlab_size):
samplers_unbal_unlab.append(sampler[:size])
samplers_semi_unbal_unlab_long = []
for sampler_semi, sampler_unb_unl in zip(samplers_semi, samplers_unbal_unlab):
samplers_semi_unbal_unlab_long.extend(np.concatenate([sampler_semi, sampler_unb_unl]))
list_length = len(samplers_train[0])
num_classes = len(samplers_train)
samplers_cp = [samplers_semi_unbal_unlab_long[i*list_length: (i+1)*list_length] for i in range(num_classes)]
augmentation = CombineFunctions([add_noise_img, flip_x_axis_img])
data_transform = CombineFunctions([remove_nan_inf_img, db_with_limits_img])
dataset_cp = DatasetImg(
samplers_cp,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
dataset_semi = DatasetImg(
samplers_semi,
args.sampler_probs,
augmentation_function=augmentation,
data_transform_function=data_transform)
return dataset_cp, dataset_semi
def sampling_echograms(sample_idx, window_size, args):
path_to_echograms = paths.path_to_echograms()
with open(os.path.join(path_to_echograms, 'memmap_2014_heave.pkl'),
'rb') as fp:
eg_names_full = pickle.load(fp)
print(sample_idx, "IN SAMPLING_ECHOGRAMS")
sample_idx = int(sample_idx)
echograms, sample_idx = get_echograms_revised(
eg_names_full, sample_idx, num_echograms=args.num_echogram)
echograms_train, echograms_val, echograms_test = cps.partition_data(
echograms,
args.partition,
portion_train_test=0.8,
portion_train_val=0.75)
sampler_bg_train = Background(echograms_train, window_size)
# sampler_sb_train = Seabed(echograms_train, window_size)
sampler_sh27_train = Shool(echograms_train, window_size, 27)
sampler_sbsh27_train = ShoolSeabed(echograms_train,
window_size,
args.window_dim // 4,
fish_type=27)
sampler_sh01_train = Shool(echograms_train, window_size, 1)
sampler_sbsh01_train = ShoolSeabed(echograms_train,
window_size,
args.window_dim // 4,
fish_type=1)
# sampler_bg_val = Background(echograms_val, window_size)
# # sampler_sb_val = Seabed(echograms_val, window_size)
# sampler_sh27_val = Shool(echograms_val, window_size, 27)
# sampler_sbsh27_val = ShoolSeabed(echograms_val, window_size, args.window_dim//4, fish_type=27)
# sampler_sh01_val = Shool(echograms_val, window_size, 1)
# sampler_sbsh01_val = ShoolSeabed(echograms_val, window_size, args.window_dim//4, fish_type=1)
samplers_train = [
sampler_bg_train, #sampler_sb_train,
sampler_sh27_train,
sampler_sbsh27_train,
sampler_sh01_train,
sampler_sbsh01_train
]
# samplers_val = [sampler_bg_val, # sampler_sb_val,
# sampler_sh27_val, sampler_sbsh27_val,
# sampler_sh01_val, sampler_sbsh01_val]
augmentation = CombineFunctions([add_noise, flip_x_axis])
label_transform = CombineFunctions(
[index_0_1_27, relabel_with_threshold_morph_close])
data_transform = CombineFunctions([remove_nan_inf, db_with_limits])
dataset_train = Dataset(samplers_train,
window_size,
args.frequencies,
args.batch * args.iteration_train,
args.sampler_probs,
augmentation_function=augmentation,
label_transform_function=label_transform,
data_transform_function=data_transform)
# dataset_val = DatasetVal(
# samplers_val,
# window_size,
# args.frequencies,
# args.batch * args.iteration_val,
# args.sampler_probs,
# augmentation_function=None,
# label_transform_function=label_transform,
# data_transform_function=data_transform)
# val_dataloader = torch.utils.data.DataLoader(dataset_val,
# shuffle=True,
# batch_size=args.batch,
# num_workers=args.workers,
# pin_memory=True)
return dataset_train, int(sample_idx)
