def get_transform_fun(resized=False):
if resized == True:
transform_function = et.ExtCompose([
et.ExtRandomCrop(size=2048),
et.ExtRandomCrop(scale=0.7, size=None),
et.ExtEnhanceContrast(),
et.ExtRandomCrop(size=2048, pad_if_needed=True),
et.ExtResize(scale=0.5),
et.ExtRandomHorizontalFlip(p=0.5),
et.ExtRandomCrop(size=512),
et.ExtRandomVerticalFlip(p=0.5),
et.ExtToTensor()
])
else:
transform_function = et.ExtCompose([
et.ExtRandomCrop(size=256),
et.ExtRandomHorizontalFlip(p=0.5),
et.ExtRandomVerticalFlip(p=0.5),
et.ExtEnhanceContrast(),
et.ExtToTensor()
])
return transform_function
model.eval()
data_loader = DataLoader(data_path=path_original_data,
metadata_path=path_meta_data)
file_names_val = np.array([
image_name[:-8] for image_name in os.listdir(path_val)
if image_name[-5] != "k"
])
file_names_val = file_names_val[file_names_val != ".DS_S"]
torch.manual_seed(1)
if not resize:
transform_function = et.ExtCompose([
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
else:
transform_function = et.ExtCompose([
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
denorm = Denormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
if binary:
color_dict = data_loader.color_dict_binary
target_dict = data_loader.get_target_dict()
annotations_dict = data_loader.annotations_dict
else:
import matplotlib.pyplot as plt
from object_detect.train_hpc import define_model
from torchvision.models.segmentation import deeplabv3_resnet101
def get_predictions(samples,ids,path_save,file_names):
for (img, m), id in zip(samples, ids):
image = (img[i].detach().cpu().numpy() * 255).transpose(1, 2, 0).astype(np.uint8)
mask = m[i]
for i in range(len(ids)):
Image.fromarray(image.astype(np.uint8)).save(
path_save + '/{}_img.png'.format(file_names[ids[i].numpy()[0]]),format='PNG')
Image.fromarray(mask.astype(np.uint8)).save(
path_save + '/{}_mask.png'.format(file_names[ids[i].numpy()[0]]), format='PNG')
transform_function = et.ExtCompose([et.ExtToTensor()])
HPC=False
splitted_data=True
binary=True
tif=False
if __name__ == '__main__':
random_seed = 1
torch.manual_seed(random_seed)
np.random.seed(random_seed)
random.seed(random_seed)
device = torch.device('cpu')
model_name = 'resnet50'
file_names_train = file_names_train[file_names_train != ".DS_S"]
file_names_val = np.array([
image_name[:-4] for image_name in os.listdir(path_val)
if image_name[-5] != "k"
])
N_files = len(file_names_val)
shuffled_index = np.random.permutation(len(file_names_val))
file_names_val = file_names_val[shuffled_index]
file_names_val = file_names_val[file_names_val != ".DS_S"]
# #FOR EXTENDED DATASET EXPERIMENT
transform_function = et.ExtCompose([
et.ExtRandomHorizontalFlip(p=0.5),
et.ExtRandomCrop(size=SIZE),
et.ExtEnhanceContrast(),
et.ExtRandomVerticalFlip(p=0.5),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
if binary:
color_dict = data_loader.color_dict_binary
target_dict = data_loader.get_target_dict()
annotations_dict = data_loader.annotations_dict
train_dst = LeatherData(path_mask=path_train,
path_img=path_train,
list_of_filenames=file_names_train,
transform=transform_function,
color_dict=color_dict,
target_dict=target_dict)
if image_name[-5] != "k"
])
file_names_train = file_names_train[file_names_train != ".DS_S"]
file_names_val = np.array([
image_name[:-4] for image_name in os.listdir(path_val)
if image_name[-5] != "k"
])
file_names_val = file_names_val[file_names_val != ".DS_S"]
#transform_function = et.ExtCompose([et.ExtEnhanceContrast(), et.ExtRandomCrop((1000, 1000)), et.ExtToTensor(),
# et.ExtNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
transform_function = et.ExtCompose([
et.ExtResize(scale=0.5),
et.ExtRandomCrop(size=512, semantic_evaluation_resize=True, scale=0.7),
et.ExtEnhanceContrast()
])
denorm = Denormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
if binary:
color_dict = data_loader.color_dict_binary
target_dict = data_loader.get_target_dict()
annotations_dict = data_loader.annotations_dict
else:
color_dict = data_loader.color_dict
target_dict = data_loader.get_target_dict(labels)
annotations_dict = data_loader.annotations_dict
train_dst = LeatherData(path_mask=path_train,
device = torch.device('cpu')
cpu_device = torch.device('cpu')
model_name = 'resnet50'
path_original_data = r'C:\Users\johan\OneDrive\Skrivebord\leather_patches'
path_meta_data = r'samples/model_comparison.csv'
optim = "SGD"
tif_path = r'C:\Users\johan\iCloudDrive\DTU\KID\BA\HPC\TIF\good_area1.png'
save_path = r'C:\Users\johan\iCloudDrive\DTU\KID\BA\HPC\last_round\predictions\vda4'
resize = False
if resize:
patch_size = 512
else:
path_size = 256
transform_function = et.ExtCompose([et.ExtEnhanceContrast(),
et.ExtToTensor()])
print("Device: %s" % device)
print("Exp: ", exp)
if brevetti:
print("REDHALF")
else:
print("WALKNAPPA")
data_loader = DataLoader(data_path=path_original_data,
metadata_path=path_meta_data)
array = load_tif_as_numpy_array(tif_path)
print("Shape array: ", np.shape(array))
if resize == True:
resize_function = et.ExtCompose([et.ExtResize(scale=0.5)])
# model=deeplabv3_resnet101(pretrained=True, progress=True,num_classes=21, aux_loss=None)
# model.classifier[-1] = torch.nn.Conv2d(256, n_classes+2, kernel_size=(1, 1), stride=(1, 1)).requires_grad_()
# model.aux_classifier[-1] = torch.nn.Conv2d(256, n_classes+2, kernel_size=(1, 1), stride=(1, 1)).requires_grad_()
#else:
# model=_segm_mobilenet('deeplabv3', 'mobile_net', output_stride=8, num_classes=n_classes+2,pretrained_backbone=True)
#model.load_state_dict(checkpoint['model_state'])
model.to(device)
model.eval()
if resize:
patch_dim = (int(patch_dim[0] * 0.5), int(patch_dim[1] * 0.5))
overlap = int(0.5 * overlap)
transform_function = et.ExtCompose([
et.ExtResize(scale=0.5),
et.ExtEnhanceContrast(),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
else:
transform_function = et.ExtCompose([
et.ExtEnhanceContrast(),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
target_tif = []
label = Image.fromarray(np.zeros(patch_dim, dtype=np.uint8))
for i in range(0, split_x_y[0], step_size):
print(i)
pred_stack = []
import os
from data_import.data_loader import DataLoader
from data_import.data_pipeline import import_data_and_mask
from semantic_segmentation.DeepLabV3.utils import ext_transforms as et
import torch
import object_detect.helper.utils as utils
from semantic_segmentation.DeepLabV3.dataset_class import LeatherData
from data_import.data_loader import DataLoader
from torch.utils import data
import random
import numpy as np
transform_function = et.ExtCompose([
et.ExtRandomCrop(scale=0.7),
et.ExtRandomHorizontalFlip(p=0.5),
et.ExtRandomVerticalFlip(p=0.5),
et.ExtEnhanceContrast(),
et.ExtToTensor()
])
random_seed = 1
torch.manual_seed(random_seed)
np.random.seed(random_seed)
random.seed(random_seed)
device = torch.device('cpu')
lr = 0.01
layers_to_train = 'classifier'
num_epoch = 1
path_original_data = r'C:\Users\johan\OneDrive\Skrivebord\leather_patches'
path_meta_data = r'samples/model_comparison.csv'
from torch.utils import data
from semantic_segmentation.DeepLabV3.metrics import StreamSegMetrics
from semantic_segmentation.DeepLabV3.utils import ext_transforms as et
from semantic_segmentation.DeepLabV3.utils.utils import Denormalize
import torch
import torch.nn as nn
from torchvision.models.segmentation import deeplabv3_resnet101
import os
import PIL
import pickle
import matplotlib.pyplot as plt
transform_function = et.ExtCompose([
et.ExtTransformLabel(),
et.ExtCenterCrop(512),
et.ExtScale(512),
et.ExtEnhanceContrast(),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
num_classes = 2
output_stride = 16
save_val_results = False
total_itrs = 100
lr = 0.01
lr_policy = 'step'
step_size = 10000
batch_size = 16
val_batch_size = 4
loss_type = "cross_entropy"
weight_decay = 1e-4
