def visualize_similarity_smallset(in_dir, out_dir, model_path):
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
# load model
model_state_dict = torch.load(model_path)
model = network.resnet_50_embedding()
model.load_state_dict(model_state_dict['model_state_dict'])
model = model.cuda()
list_dir = os.listdir(in_dir)
for d in list_dir:
in_path = os.path.join(in_dir, d)
if not os.path.isdir(in_path):
continue
out_path = os.path.join(out_dir, d)
if not os.path.isdir(out_path):
os.mkdir(out_path)
dataset = OPENFusionStereoDatasetPath(in_path, transform=image_transform)
# Run the images through the network, get last conv features
fvecs, maps, labels, image_path_list = embed(dataset,model)
maps = np.moveaxis(maps,1,3)
#fvecs = np.array(fvecs)
#labels = np.array(labels)
# Compute the spatial similarity maps (returns a heatmap that's the size of the last conv layer)
for i in range(len(image_path_list)):
for j in range(len(image_path_list)):
if abs(i-j) > 2:
continue
if i == j:
continue
conv1 = maps[i]
conv2 = maps[j]
conv1 = conv1.reshape(-1,conv1.shape[-1])
conv2 = conv2.reshape(-1,conv2.shape[-1])
heatmap1, heatmap2 = compute_spatial_similarity(conv1, conv2)
im1_path = image_path_list[i][0]
im2_path = image_path_list[j][0]
# Combine the images with the (interpolated) similarity heatmaps.
im1_with_similarity = combine_image_and_heatmap_noninter(load_and_resize(im1_path),heatmap1)
im2_with_similarity = combine_image_and_heatmap_noninter(load_and_resize_grayscale(im2_path),heatmap2)
save_img = combine_horz([im1_with_similarity,im2_with_similarity])
save_img.save(os.path.join(out_path,'b{}-{}_t{}-{}.png'.format(os.path.basename(im1_path)[:-4], i, os.path.basename(im2_path)[:-4], j)))
#Image.fromarray(im2_with_similarity).save(os.path.join(out_path,'b{}_t{}.png'.format(labels[j],labels[i])))
return
def grid_match_jointdata_scan_whole_image(rgb_img_path, thermal_img_path, out_dir, model_path):
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
imgSize = 400
gridSize = 256
stepSize = 32
# init grid pool, return: vec of box coordinates
grid_pool = init_grid_pool_largeGrid(imgSize, gridSize, stepSize)
# init image set base on grid pool, both sensor's image
rgb_grid_data = GridMatchDatasetPath(rgb_img_path, grid_pool, imgSize, image_transform)
thermal_grid_data = GridMatchDatasetPath(thermal_img_path, grid_pool, imgSize, image_transform)
# load model
model_state_dict = torch.load(model_path)
model = network.resnet_50_embedding()
# remove `module.` for parallel
new_state_dict = OrderedDict()
for k, v in model_state_dict['model_state_dict'].items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
# load params
model.load_state_dict(new_state_dict)
model.cuda()
model.eval()
output_number = 0
for base_grid in grid_pool:
# select base roi from RGB
base_roi_image = init_base_rgb_roi_image(rgb_img_path, base_grid, out_dir)
# get base embedding
base_embedding = embed_single_image(base_roi_image, model)
# loop thermal data set and get pair embeddings+++
pair_embeddings, labels = embed_pair_grid(thermal_grid_data, model)
# get a score for the grid
scores = get_scores_for_grids(base_embedding, pair_embeddings, labels, base_grid)
# output grid images sorted by score
output_grid_images_by_scores_topN(thermal_img_path, scores, labels, out_dir, output_number, base_grid)
# draw a map base on all the grid score
#generate_score_map(imgSize, scores, labels)
# output a scored heat map
return
def loadModelParallel(model_path):
# load model
model_state_dict = torch.load(model_path)
model = network.resnet_50_embedding()
# remove `module.` for parallel
new_state_dict = OrderedDict()
for k, v in model_state_dict['model_state_dict'].items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
# load params
model.load_state_dict(new_state_dict)
model.cuda()
model.eval()
return model
def grid_match_jointdata_single_image(rgb_img_path, thermal_img_path, out_dir, model_path):
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
imgSize = 224
gridSize = 128
stepSize = 8
# init grid pool, return: vec of box coordinates
grid_pool = init_grid_pool_largeGrid(imgSize, gridSize, stepSize)
# init image set base on grid pool, both sensor's image
rgb_grid_data = GridMatchDatasetPath(rgb_img_path, grid_pool, imgSize, image_transform)
thermal_grid_data = GridMatchDatasetPath(thermal_img_path, grid_pool, imgSize, image_transform)
# load model
model_state_dict = torch.load(model_path)
model = network.resnet_50_embedding()
model.load_state_dict(model_state_dict['model_state_dict'])
model = model.cuda()
model.eval()
#base_grid = (64,40,128,104)
base_grid = (64,40,64+128,40+128)
# select base roi from RGB, human selected for now
base_roi_image = init_base_rgb_roi_image(rgb_img_path, base_grid, out_dir)
# get base embedding
base_embedding = embed_single_image(base_roi_image, model)
# loop thermal data set and get pair embeddings+++
pair_embeddings, labels = embed_pair_grid(thermal_grid_data, model)
# get a score for the grid
scores = get_scores_for_grids(base_embedding, pair_embeddings, labels, base_grid)
# output grid images sorted by score
output_grid_images_by_scores(thermal_img_path, scores, labels, out_dir)
# draw a map base on all the grid score
#generate_score_map(imgSize, scores, labels)
# output a scored heat map
return
def loss_plot():
dataset = OPENFusionJointDataset(
'/media/zli/Seagate Backup Plus Drive/OPEN/ua-mac/Level_3/joint_training_data_test',
transform=image_transform)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
shuffle=False,
num_workers=1)
model_state_dict = torch.load(
'/media/zli/Seagate Backup Plus Drive/trained_models/pytorch-py3/checkpoints/NearDateJointTripletMarginLoss/20200915_221501874/epoch_11.pth'
)
model = network.resnet_50_embedding()
model.load_state_dict(model_state_dict['model_state_dict'])
loss_func = MyNCALoss()
# inference
#model.load_state_dict(model_state_dict['model_state_dict'])
model.cuda()
model.eval()
vector_list = []
label_list = []
loss_list = []
with torch.no_grad():
for data, target in tqdm(dataloader, total=len(dataloader)):
data = data.cuda()
model_output = model(data)
loss = loss_func(model_output, target)
loss = loss.item()
vector_list.append(model_output)
loss_list.append(loss)
label_list.append(target)
vectors = torch.cat(vector_list, 0)
labels = {}
for l in label_list:
for key, val in l.items():
if key not in labels:
labels[key] = []
labels[key].extend(val)
return
def visualize_sim_pairs_in_folder(input_folder, model_path):
# load model
model_state_dict = torch.load(model_path)
model = network.resnet_50_embedding()
model.load_state_dict(model_state_dict['model_state_dict'])
model = model.cuda()
list_dir = os.listdir(input_folder)
for d in list_dir:
sub_dir = os.path.join(input_folder, d)
if not os.path.isdir(sub_dir):
continue
png_suffix = os.path.join(sub_dir, '*.png')
pngs = glob(png_suffix)
im1_path = pngs[0]
im2_path = pngs[1]
image1 = Image.open(im1_path)
image2 = Image.open(im2_path)
image1 = image1.crop((40,40,360,360))
image2 = image2.crop((40,40,360,360))
#image1.show()
fv1, fm1 = embed_single_image(image1, model)
fv2, fm2 = embed_single_image(image2, model)
fm1 = np.moveaxis(fm1,1,3)
fm2 = np.moveaxis(fm2,1,3)
conv1 = fm1.reshape(-1,fm1.shape[-1])
conv2 = fm2.reshape(-1,fm2.shape[-1])
#heatmap1, heatmap2 = compute_spatial_similarity(conv1, conv2)
heatmap1, heatmap2 = combine_SVD(conv1,conv2)
# Combine the images with the (interpolated) similarity heatmaps.
im1_with_similarity = combine_image_and_heatmap_noninter(load_and_resize_grayscale(im1_path),heatmap1)
im2_with_similarity = combine_image_and_heatmap_noninter(load_and_resize_grayscale(im2_path),heatmap2)
save_img = combine_horz([im1_with_similarity,im2_with_similarity])
save_img.save(os.path.join(sub_dir,'{}_simVis.png'.format(os.path.basename(im1_path)[:-4])))
return
def output_image_and_activations(in_dir, out_dir, model_path):
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
# load model
model_state_dict = torch.load(model_path)
model = network.resnet_50_embedding()
model.load_state_dict(model_state_dict['model_state_dict'])
model = model.cuda()
list_dir = os.listdir(in_dir)
for d in list_dir:
in_path = os.path.join(in_dir, d)
if not os.path.isdir(in_path):
continue
out_path = os.path.join(out_dir, d)
if not os.path.isdir(out_path):
os.mkdir(out_path)
dataset = OPENFusionStereoDatasetPath(in_path, transform=image_transform)
# Run the images through the network, get last conv features
fvecs, maps, labels, image_path_list = embed(dataset,model)
maps = np.moveaxis(maps,1,3)
for i in range(len(image_path_list)):
conv1 = maps[i]
im1_path = image_path_list[i][0]
basename = os.path.basename(im1_path)[:-4]
conv_file_path = os.path.join(out_path, '{}.pkl'.format(basename))
mat_file_path = os.path.join(out_path, '{}.mat'.format(basename))
img_path = os.path.join(out_path, '{}.png'.format(basename))
sio.savemat(mat_file_path, {"conv": conv1})
conv_df = pd.DataFrame({"conv": [conv1]})
shutil.copyfile(im1_path, img_path)
conv_df.to_pickle(conv_file_path)
return
'/media/zli/Seagate Backup Plus Drive/OPEN/ua-mac/Level_3/joint_training_data_validation_double/',
transform=val_image_transform)
val_fusion_dateset = TripletNearDateJointDataset(
val_dataset,
class_by='plot',
date_by='scan_date',
neighbor_distance=neighbor_distance,
collate_fn=None,
transform=val_image_transform)
val_dataloader = torch.utils.data.DataLoader(val_fusion_dateset,
batch_size=val_batch_size,
shuffle=True,
num_workers=0)
print('Init model')
model = resnet_50_embedding()
#loss_func = TripletMarginLoss()
loss_func = MyNCALoss()
model.cuda()
optimizer = optim.SGD(model.parameters(), lr=lr)
scheduler = lr_scheduler.StepLR(optimizer, 15, gamma=0.1, last_epoch=-1)
print('Train 1 epoch')
train(model,
dataloader,
loss_func,
optimizer,
scheduler,
n_epochs,
resume_dict=resume_dict,
tb_log_dir=tb_log_dir,
ckp_dir=ckp_dir,
def get_output():
dataset = OPENFusionStereoDatasetPath(
'/media/zli/Seagate Backup Plus Drive/OPEN/ua-mac/Level_3/joint_training_data_test_double',
transform=image_transform)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
shuffle=False,
num_workers=4)
#model_state_dict = torch.load('/media/zli/Seagate Backup Plus Drive/trained_models/pytorch-py3/checkpoints/NearDateJointTripletMarginLoss_RotateSameWay30WithFLIP_320_temp3.5/20210106_180416201/epoch_22.pth')
#model_state_dict = torch.load('/media/zli/Seagate Backup Plus Drive/trained_models/pytorch-py3/checkpoints/NearDateJointTripletMarginLoss_RotateDifferentWay30WithFLIP_320_temp3.5/20210110_161216957/epoch_21.pth')
model_state_dict = torch.load(
'/media/zli/Seagate Backup Plus Drive/trained_models/pytorch-py3/checkpoints/NearDateJointTripletMarginLoss_RotateDiffWay33_temp3.5_linear/20210122_192953695/epoch_33.pth'
)
model = network.resnet_50_embedding()
# remove `module.` for parallel
new_state_dict = OrderedDict()
for k, v in model_state_dict['model_state_dict'].items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
# load params
model.load_state_dict(new_state_dict)
#model.load_state_dict(model_state_dict['model_state_dict'])
# inference
#model.load_state_dict(model_state_dict['model_state_dict'])
model.cuda()
model.eval()
vector_list = []
label_list = []
image_path_list = []
with torch.no_grad():
for data, target, image_path in tqdm(dataloader,
total=len(dataloader)):
data = data.cuda()
vector_list.append(model(data))
label_list.append(target)
image_path_list.append(image_path)
#vectors = torch.cat(vector_list, 0)
labels = {}
for l in label_list:
for key, val in l.items():
if key not in labels:
labels[key] = []
labels[key].extend(val)
image_paths = np.concatenate(image_path_list)
vectors_norm = [
torch.nn.functional.normalize(d, p=2, dim=1) for d in vector_list
]
vectors = torch.cat(vectors_norm, 0).cpu().numpy()
torch.save(
{
'vectors': vectors,
'labels': labels,
'image_paths': image_paths
}, './results/{}_ep_{}.pth'.format('RotateDiffWay33_temp3.5_linear',
33))
return
def pca_vis_by_score(out_dir, model_path):
# load model
model_state_dict = torch.load(model_path)
model = network.resnet_50_embedding()
'''
model.load_state_dict(model_state_dict['model_state_dict'])
model = model.cuda()
'''
# remove `module.` for parallel
new_state_dict = OrderedDict()
for k, v in model_state_dict['model_state_dict'].items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
# load params
model.load_state_dict(new_state_dict)
model.cuda()
model.eval()
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
task_name = 'RotateDiffWay30WithFLIP_320_temp3.5'
load_result_ep = 21
result_dict = torch.load('/home/zli/WorkSpace/PyWork/pytorch-py3/reverse-pheno-master/results/{}_ep_{}.pth'.format(task_name, load_result_ep))
vectors = result_dict['vectors']
labels = (result_dict['labels'])
image_paths = result_dict['image_paths']
total_nums = len(vectors)
imsize = (320,320)
for i in range(15,total_nums):
print(i)
base_vec = vectors[i]
base_label = labels['plot'][i]
base_sensor = labels['sensor'][i]
dist_rec = np.zeros(total_nums)
for j in range(total_nums):
pair_sensor = labels['sensor'][j]
loop_vec = vectors[j]
if base_sensor is pair_sensor:
dist_rec[j] = float('inf')
else:
dist_rec[j] = np.linalg.norm(loop_vec-base_vec)
# dot product
#dist_rec[j] = abs(np.matmul(loop_vec, base_vec))
min_index = dist_rec.argsort()[:4]
mid_ind = int(total_nums/2)
start_ind = mid_ind - 5
max_index = dist_rec.argsort()[start_ind:mid_ind]
out_path = os.path.join(out_dir, str(i))
if not os.path.isdir(out_path):
os.mkdir(out_path)
src_path_base = image_paths[i]
image1 = Image.open(src_path_base)
image1 = image1.crop((40,40,360,360))
fv1, fm1 = embed_single_image(image1, model)
fm1 = np.moveaxis(fm1,1,3)
conv1 = fm1.reshape(-1,fm1.shape[-1])
for ind in range(len(min_index)):
val = min_index[ind]
src_path_pair = image_paths[val]
if image_paths[val] == image_paths[i]:
continue
image2 = Image.open(src_path_pair)
image2 = image2.crop((40,40,360,360))
#image1.show()
fv2, fm2 = embed_single_image(image2, model)
fm2 = np.moveaxis(fm2,1,3)
conv2 = fm2.reshape(-1,fm2.shape[-1])
heatmap1, heatmap2 = combine_SVD(conv1,conv2)
featPCAIm1 = cv2.resize(heatmap1, (imsize[0], imsize[1]), interpolation=cv2.INTER_NEAREST)
featPCAIm1 = featPCAIm1.astype('uint8')
featPCAIm2 = cv2.resize(heatmap2, (imsize[0], imsize[1]), interpolation=cv2.INTER_NEAREST)
featPCAIm2 = featPCAIm2.astype('uint8')
# Combine the images with the (interpolated) similarity heatmaps.
im1 = cv2.imread(src_path_base, 1)
im1 = im1[40:360, 40:360]
im1_with_similarity = cv2.addWeighted(im1, 0.5, featPCAIm1, 0.5, 0.0)
im2 = cv2.imread(src_path_pair, 0)
im2 = im2[40:360, 40:360]
im2 = cv2.merge((im2,im2,im2))
im2_with_similarity = cv2.addWeighted(im2, 0.5, featPCAIm2, 0.5, 0.0)
save_img = combine_horz([im1_with_similarity,im2_with_similarity])
save_path = os.path.join(out_path, 'close_{}.png'.format(ind))
save_img.save(save_path)
for c in range(len(max_index)):
val = max_index[c]
src_path_pair = image_paths[val]
if image_paths[val] == image_paths[i]:
continue
image2 = Image.open(src_path_pair)
image2 = image2.crop((40,40,360,360))
#image1.show()
fv2, fm2 = embed_single_image(image2, model)
fm2 = np.moveaxis(fm2,1,3)
conv2 = fm2.reshape(-1,fm2.shape[-1])
heatmap1, heatmap2 = combine_SVD(conv1,conv2)
featPCAIm1 = cv2.resize(heatmap1, (imsize[0], imsize[1]), interpolation=cv2.INTER_NEAREST)
featPCAIm1 = featPCAIm1.astype('uint8')
featPCAIm2 = cv2.resize(heatmap2, (imsize[0], imsize[1]), interpolation=cv2.INTER_NEAREST)
featPCAIm2 = featPCAIm2.astype('uint8')
# Combine the images with the (interpolated) similarity heatmaps.
im1 = cv2.imread(src_path_base, 1)
im1 = im1[40:360, 40:360]
im1_with_similarity = cv2.addWeighted(im1, 0.7, featPCAIm1, 0.5, 0.0)
im2 = cv2.imread(src_path_pair, 0)
im2 = im2[40:360, 40:360]
im2 = cv2.merge((im2,im2,im2))
im2_with_similarity = cv2.addWeighted(im2, 0.7, featPCAIm2, 0.5, 0.0)
save_img = combine_horz([im1_with_similarity,im2_with_similarity])
save_path = os.path.join(out_path, 'far_{}.png'.format(c))
save_img.save(save_path)
return
def visualize_pca_jointdata(in_dir, out_dir, model_path):
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
# load model
model_state_dict = torch.load(model_path)
model = network.resnet_50_embedding()
# remove `module.` for parallel
new_state_dict = OrderedDict()
for k, v in model_state_dict['model_state_dict'].items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
# load params
model.load_state_dict(new_state_dict)
model.cuda()
model.eval()
rgb_dir = os.path.join(in_dir, 'rgb')
thermal_dir = os.path.join(in_dir, 'laser')
list_dir = os.listdir(rgb_dir)
len_dir = len(list_dir)
for i in range(0,len_dir):
print(i)
d = list_dir[i]
print(d)
rgb_path = os.path.join(rgb_dir, d)
thermal_path = os.path.join(thermal_dir, d)
if not os.path.isdir(rgb_path):
continue
if not os.path.isdir(thermal_path):
continue
out_path = os.path.join(out_dir, d)
if not os.path.isdir(out_path):
os.mkdir(out_path)
dataset_rgb = OPENFusionStereoDatasetPath(rgb_path, transform=image_transform)
dataset_thermal = OPENFusionStereoDatasetPath(thermal_path, transform=image_transform)
# Run the images through the network, get last conv features
fvecs, maps_rgb, labels, image_path_list_rgb = embed(dataset_rgb,model)
fvecs, maps_thermal, labels, image_path_list_thermal = embed(dataset_thermal,model)
maps_rgb = np.moveaxis(maps_rgb,1,3)
maps_thermal = np.moveaxis(maps_thermal,1,3)
#fvecs = np.array(fvecs)
#labels = np.array(labels)
imsize = (320,320)
# Compute the spatial similarity maps (returns a heatmap that's the size of the last conv layer)
for i in range(len(image_path_list_rgb)):
im1_path = image_path_list_rgb[i][0]
rgb_time = os.path.basename(im1_path)[:-4]
rgb_time_int = datetime.strptime(rgb_time, '%Y-%m-%d').date().toordinal()
for j in range(len(image_path_list_thermal)):
im2_path = image_path_list_thermal[j][0]
thermal_time = os.path.basename(im2_path)[:-4]
thermal_time_int = datetime.strptime(thermal_time, '%Y-%m-%d').date().toordinal()
if abs(rgb_time_int-thermal_time_int) > 2:
continue
conv1 = maps_rgb[i]
conv2 = maps_thermal[j]
heatmap1, heatmap2 = combine_SVD(conv1,conv2)
featPCAIm1 = cv2.resize(heatmap1, (imsize[0], imsize[1]), interpolation=cv2.INTER_NEAREST)
featPCAIm1 = featPCAIm1.astype('uint8')
featPCAIm2 = cv2.resize(heatmap2, (imsize[0], imsize[1]), interpolation=cv2.INTER_NEAREST)
featPCAIm2 = featPCAIm2.astype('uint8')
# Combine the images with the (interpolated) similarity heatmaps.
im1 = cv2.imread(im1_path, 1)
#im1 = cv2.flip(im1, 0)
#im1 = cv2.rotate(im1, cv2.ROTATE_90_CLOCKWISE)
im1 = im1[40:360, 40:360]
im1 = cv2.resize(im1, (imsize[0], imsize[1]))
im1_with_similarity = cv2.addWeighted(im1, 0.5, featPCAIm1, 0.5, 0.0)
im2 = cv2.imread(im2_path, 0)
im2 = im2[40:360, 40:360]
im2 = cv2.merge((im2,im2,im2))
im2 = cv2.resize(im2, (imsize[0], imsize[1]))
im2_with_similarity = cv2.addWeighted(im2, 0.5, featPCAIm2, 0.5, 0.0)
save_img = combine_horz([im1_with_similarity,im2_with_similarity])
save_img.save(os.path.join(out_path,'b{}-{}_t{}-{}.png'.format(os.path.basename(im1_path)[:-4], i, os.path.basename(im2_path)[:-4], j)))
#Image.fromarray(im2_with_similarity).save(os.path.join(out_path,'b{}_t{}.png'.format(labels[j],labels[i])))
return
def pca_optimization(in_dir, out_dir, model_path):
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
# load model
model_state_dict = torch.load(model_path)
model = network.resnet_50_embedding()
# remove `module.` for parallel
new_state_dict = OrderedDict()
for k, v in model_state_dict['model_state_dict'].items():
name = k[7:] # remove `module.`
new_state_dict[name] = v
# load params
model.load_state_dict(new_state_dict)
model.cuda()
model.eval()
imsize = (320,320)
rgb_dir = os.path.join(in_dir, 'rgb')
thermal_dir = os.path.join(in_dir, 'laser')
list_dir = os.listdir(rgb_dir)
len_dir = len(list_dir)
for i in range(0,len_dir):
print(i)
d = list_dir[i]
rgb_path = os.path.join(rgb_dir, d)
thermal_path = os.path.join(thermal_dir, d)
if not os.path.isdir(rgb_path):
continue
if not os.path.isdir(thermal_path):
continue
list_rgb_files = os.walk(rgb_path)
for root, dirs, files in list_rgb_files:
for f in files:
base_file = os.path.join(rgb_path, f)
if os.path.exists(base_file):
pair_file = os.path.join(thermal_path, f)
if os.path.exists(pair_file):
out_path = os.path.join(out_dir, d)
if not os.path.isdir(out_path):
os.mkdir(out_path)
image1 = Image.open(base_file)
image1 = image1.crop((40,40,360,360))
fv1, fm1 = embed_single_image(image1, model)
fm1 = np.moveaxis(fm1,1,3)
conv1 = fm1.reshape(-1,fm1.shape[-1])
image2 = Image.open(pair_file)
image2 = image2.crop((40,40,360,360))
fv2, fm2 = embed_single_image(image2, model)
fm2 = np.moveaxis(fm2,1,3)
conv2 = fm2.reshape(-1,fm2.shape[-1])
heatmap1, heatmap2 = combine_SVD(conv1,conv2)
featPCAIm1 = cv2.resize(heatmap1, (imsize[0], imsize[1]), interpolation=cv2.INTER_LINEAR)
featPCAIm2 = cv2.resize(heatmap2, (imsize[0], imsize[1]), interpolation=cv2.INTER_LINEAR)
# Combine the images with the (interpolated) similarity heatmaps.
im1 = cv2.imread(base_file, 1)
im1 = im1[40:360, 40:360]
im1 = cv2.resize(im1, (imsize[0], imsize[1]))
im2 = cv2.imread(pair_file, 1)
im2 = im2[40:360, 40:360]
#im2 = cv2.merge((im2,im2,im2))
im2 = cv2.resize(im2, (imsize[0], imsize[1]))
#featPCAIm1 = featPCAIm1.astype('uint8')
#featPCAIm2 = featPCAIm2.astype('uint8')
#im1_with_similarity = cv2.addWeighted(im1, 0.5, featPCAIm1, 0.5, 0.0)
#im2_with_similarity = cv2.addWeighted(im2, 0.5, featPCAIm2, 0.5, 0.0)
#save_img = combine_horz([im1_with_similarity,im2_with_similarity])
#save_img.save(os.path.join(out_path,'{}.png'.format(os.path.basename(base_file)[:-4])))
save_path = os.path.join(out_path,'{}_fusion.png'.format(os.path.basename(base_file)[:-4]))
'''
PCA = torch.from_numpy(featPCAIm1)
PCAT = torch.from_numpy(featPCAIm2)
Im1 = torch.from_numpy(im1)
Im2 = torch.from_numpy(im2)
pytorch_optimization_frame(PCA, PCAT, Im1, Im2)
'''
a, b, _ = pca_template_matching(featPCAIm1, featPCAIm2)
save_fusion_img(im1, im2, a, b, save_path)
'''
im_t = cv2.imread(pair_file, 0)
im_t = im2[40:360, 40:360]
im_t = cv2.merge((im_t,im_t,im_t))
im_t = cv2.resize(im_t, (imsize[0], imsize[1]))
x0 = [a,b,1.0]
res = minimize(scale_translation_optimization_fun, x0, args=(featPCAIm1, featPCAIm2, im1, im_t), method='trust-constr', constraints=cons, options={'gtol': 1e-10, 'disp': True})
print(res)
'''
return
def visualize_similarity_jointdata(in_dir, out_dir, model_path):
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
# load model
model_state_dict = torch.load(model_path)
model = network.resnet_50_embedding()
model.load_state_dict(model_state_dict['model_state_dict'])
model = model.cuda()
rgb_dir = os.path.join(in_dir, 'rgb')
thermal_dir = os.path.join(in_dir, 'thermal')
list_dir = os.listdir(rgb_dir)
for d in list_dir:
rgb_path = os.path.join(rgb_dir, d)
thermal_path = os.path.join(thermal_dir, d)
if not os.path.isdir(rgb_path):
continue
if not os.path.isdir(thermal_path):
continue
out_path = os.path.join(out_dir, d)
if not os.path.isdir(out_path):
os.mkdir(out_path)
dataset_rgb = OPENFusionStereoDatasetPath(rgb_path, transform=image_transform)
dataset_thermal = OPENFusionStereoDatasetPath(thermal_path, transform=image_transform)
# Run the images through the network, get last conv features
fvecs, maps_rgb, labels, image_path_list_rgb = embed(dataset_rgb,model)
fvecs, maps_thermal, labels, image_path_list_thermal = embed(dataset_thermal,model)
maps_rgb = np.moveaxis(maps_rgb,1,3)
maps_thermal = np.moveaxis(maps_thermal,1,3)
#fvecs = np.array(fvecs)
#labels = np.array(labels)
# Compute the spatial similarity maps (returns a heatmap that's the size of the last conv layer)
for i in range(len(image_path_list_rgb)):
im1_path = image_path_list_rgb[i][0]
rgb_time = os.path.basename(im1_path)[:-4]
rgb_time_int = datetime.strptime(rgb_time, '%Y-%m-%d').date().toordinal()
for j in range(len(image_path_list_thermal)):
im2_path = image_path_list_thermal[j][0]
thermal_time = os.path.basename(im2_path)[:-4]
thermal_time_int = datetime.strptime(thermal_time, '%Y-%m-%d').date().toordinal()
if abs(rgb_time_int-thermal_time_int) > 2:
continue
conv1 = maps_rgb[i]
conv2 = maps_thermal[j]
conv1 = conv1.reshape(-1,conv1.shape[-1])
conv2 = conv2.reshape(-1,conv2.shape[-1])
heatmap1, heatmap2 = compute_spatial_similarity(conv1, conv2)
# Combine the images with the (interpolated) similarity heatmaps.
im1_with_similarity = combine_image_and_heatmap_noninter(load_and_resize(im1_path),heatmap1)
im2_with_similarity = combine_image_and_heatmap_noninter(load_and_resize_grayscale(im2_path),heatmap2)
save_img = combine_horz([im1_with_similarity,im2_with_similarity])
save_img.save(os.path.join(out_path,'b{}-{}_t{}-{}.png'.format(os.path.basename(im1_path)[:-4], i, os.path.basename(im2_path)[:-4], j)))
#Image.fromarray(im2_with_similarity).save(os.path.join(out_path,'b{}_t{}.png'.format(labels[j],labels[i])))
return
dataloader = torch.utils.data.DataLoader(torch_u_data.ConcatDataset(list_of_datasets), batch_size=batch_size,
shuffle=False, num_workers=4)
'''
train_dataset = OPENFusionStereoDatasetPath(
'/media/zli/Seagate Backup Plus Drive/OPEN/ua-mac/Level_3/joint_training_data_test',
transform=image_transform)
#fusion_dateset = TripletNearDateJointDataset(train_dataset, class_by='plot', date_by='scan_date', neighbor_distance=neighbor_distance, collate_fn=None)
dataloader = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
model_state_dict = torch.load(
'/media/zli/Seagate Backup Plus Drive/trained_models/pytorch-py3/checkpoints/NearDateJointTripletMarginLoss/20200915_221501874/epoch_11.pth'
)
model = network.resnet_50_embedding()
#model.load_state_dict(model_state_dict['model_state_dict'])
# inference
model.load_state_dict(model_state_dict['model_state_dict'])
model.cuda()
model.eval()
vector_list = []
label_list = []
image_path_list = []
with torch.no_grad():
for data, target, image_path in tqdm(dataloader, total=len(dataloader)):
data = data.cuda()
vector_list.append(model(data))
label_list.append(target)
image_path_list.append(image_path)
