def main():
criterion = nn.CrossEntropyLoss()
criterion = criterion.cuda()
model = Network(args.init_channels, 10, args.layers, criterion, spaces_dict[args.search_space]).cuda()
model.load_state_dict(torch.load("weights-3.pt"))
model.eval()
#for module in model.named_modules():
# print (module[0])
# Open image
raw_image = cv2.imread(args.img_path)
tens = np.load(args.tens_path, allow_pickle=True)
image = torch.from_numpy(tens).unsqueeze(0)#.unsqueeze(0)
image = image.cuda()
print (image.size())
pred = model(image)
print (pred)
# GCAM
gcam = GradCAM(model=model)
predictions = gcam.forward(image)
top_idx = predictions[0][1]
print(predictions, len(predictions), top_idx)
target_layer = "cells.19"
gcam.backward(idx=top_idx)
region = gcam.generate(target_layer=target_layer)
cmap = cm.jet_r(region)[..., :3] * 255.0
cmap = cv2.resize(cmap, (32, 32))
blend = (cmap+raw_image)/2
cv2.imwrite("blend_4.png", blend)
print (region.shape, cmap.shape, raw_image.shape)
def get_data_of_gradcam(gcam: Tensor,
raw_image: Tensor,
paper_cmap: bool = False) -> Tensor:
r"""Returns Grad-CAM data.
Args:
gcam (Tensor): Grad-CAM data.
raw_image (Tensor): raw image data.
paper_cmap (bool, optional): cmap. Defaults to False.
Returns:
Tensor: [description]
"""
np_gcam = gcam.cpu().numpy()
del gcam
cmap = cm.jet_r(np_gcam)[..., :3] * 255.0 # type: ignore
if paper_cmap:
alpha = np_gcam[..., None]
np_gcam = alpha * cmap + ([1] - alpha) * raw_image
# np_gcam = alpha * cmap + ([torch.tensor(1)] - alpha) * raw_image
else:
np_gcam = (cmap.astype(np.float) +
raw_image.clone().cpu().numpy().astype(np.float)) / 2
np_gcam = np.uint8(np_gcam) # type: ignore
return np_gcam
def save_gradcam(gcam, raw_image, paper_cmap=False):
gcam = gcam.cpu().numpy()
cmap = cm.jet_r(gcam)[..., :3] * 255.0
#print("raw_image shape: ", raw_image.shape)
# extract dog from colormap
temp = cmap.astype(np.float)
"""TODO: tune the below parameters"""
rgb_lower1 = np.array([0, 0, 35], dtype='uint8')
rgb_upper1 = np.array([255, 255, 255], dtype='uint8')
mask1 = cv2.inRange(temp, rgb_lower1, rgb_upper1)
contours1, _ = cv2.findContours(mask1.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
contour1 = sorted(contours1, key=cv2.contourArea, reverse=True)[0]
x, y, w, h = cv2.boundingRect(contour1)
dog = [x,y,w,h]
# dog = raw_image[:, y:y+h, x:x+w]
# dog = torch.unsqueeze(dog, 0)
# dog = F.upsample(dog, (224, 224), mode="bilinear", align_corners=False)
rgb_lower2 = np.array([120, 240, 100], dtype='uint8')
rgb_upper2 = np.array([140, 255, 130], dtype='uint8')
mask2 = cv2.inRange(temp, rgb_lower2, rgb_upper2)
if (mask2 is not None):
contours2, _ = cv2.findContours(mask2.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
contour2 = sorted(contours2, key=cv2.contourArea, reverse=True)[0]
x, y, w, h = cv2.boundingRect(contour2)
dog_face = [x, y, w, h]
#dog_face = torch.tensor(dog_face)
####
return dog, dog_face
def main(config, model_path, cuda, crf, camera_id):
# Configuration
CONFIG = Dict(yaml.load(open(config)))
cuda = cuda and torch.cuda.is_available()
if cuda:
current_device = torch.cuda.current_device()
print('Running on', torch.cuda.get_device_name(current_device))
# Label list
with open(CONFIG.LABELS) as f:
classes = {}
for label in f:
label = label.rstrip().split('\t')
classes[int(label[0])] = label[1].split(',')[0]
# Load a model
state_dict = torch.load(model_path)
# Model
model = DeepLabV2_ResNet101_MSC(n_classes=CONFIG.N_CLASSES)
model.load_state_dict(state_dict)
model.eval()
if cuda:
model.cuda()
image_size = (CONFIG.IMAGE.SIZE.TEST, ) * 2
cap = cv2.VideoCapture(camera_id)
cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc(*'YUYV'))
while True:
# Image preprocessing
ret, frame = cap.read()
image = cv2.resize(frame.astype(float), image_size)
raw_image = image.astype(np.uint8)
image -= np.array([
float(CONFIG.IMAGE.MEAN.B),
float(CONFIG.IMAGE.MEAN.G),
float(CONFIG.IMAGE.MEAN.R),
])
image = torch.from_numpy(image.transpose(2, 0, 1)).float().unsqueeze(0)
image = image.cuda() if cuda else image
# Inference
output = model(Variable(image, volatile=True))
output = F.upsample(output, size=image_size, mode='bilinear')
output = F.softmax(output, dim=1)
output = output.data.cpu().numpy()[0]
if crf:
output = dense_crf(raw_image, output)
labelmap = np.argmax(output.transpose(1, 2, 0), axis=2)
labelmap = labelmap.astype(float) / CONFIG.N_CLASSES
labelmap = cm.jet_r(labelmap)[..., :-1] * 255.0
cv2.addWeighted(np.uint8(labelmap), 0.5, raw_image, 0.5, 0.0,
raw_image)
cv2.imshow('DeepLabV2', raw_image)
cv2.waitKey(50)
def massflux_ss(init, end):
"""draw th in average"""
##### retrieve variables #####
xx, zz = np.meshgrid(x, z)
dt = netCDF4.Dataset(dy_files[0])
t = int(np.array(dt['Time']))
density = np.loadtxt('../density.txt')[:len(z)]
dwdzss = np.zeros((end - init, len(z)))
for tidx in range(init, end):
td = netCDF4.Dataset(td_files[tidx])
progressbar(now=tidx - init,
length=end - init,
text='theta in snap shot')
dy = netCDF4.Dataset(dy_files[tidx])
w = np.mean(dy['w'][0, :len(z)] * area, axis=(1, 2))
dwdzss[tidx - init, :] = w * density
figure()
for tidx in range(init, end):
plot(dwdzss[tidx - init, :],
z,
c=cm.jet_r((tidx - init) / (end - init)),
label='%s' % tidx,
alpha=0.5)
plot(np.mean(dwdzss, axis=0), z, color='black', label='average')
legend(fontsize=5)
title(r'Convective Mass Flux $\rho w$')
savefig('dwdzss_ct.jpg', dpi=300)
clf()
def save_gradcam(filename, gcam, paper_cmap=False):
cmap = cm.jet_r(gcam)[..., :3] * 255.0
if paper_cmap:
alpha = gcam[..., None]
gcam = alpha * cmap
else:
gcam = cmap.astype(np.float)
cv2.imwrite(filename, np.uint8(gcam))
def save_gradcam_over_image(filename, gcam, raw_image, paper_cmap=False):
cmap = cm.jet_r(gcam)[..., :3] * 255.0
if paper_cmap:
alpha = gcam[..., None]
gcam = alpha * cmap + (1 - alpha) * raw_image
else:
gcam = (cmap.astype(np.float) + raw_image.astype(np.float)) / 2
cv2.imwrite(filename, np.uint8(gcam))
def getCAM(feature_conv, weight_fc, discard):
_, nc, h, w = feature_conv.shape
cam = weight_fc[discard].dot(feature_conv.reshape((nc, h*w)))
cam = cam.reshape(h, w)
cam = cam - np.min(cam)
cam_img = cam / np.max(cam)
cam_img = cm.jet_r(cam_img)[..., :3] * 255.0
cam_img *= pixel_intensity
return cam_img
def save_gradcam(filename, gcam, raw_image, paper_cmap=False):
gcam = gcam.cpu().numpy()
cmap = cm.jet_r(gcam)[..., :3] * 255.0
if paper_cmap:
alpha = gcam[..., None]
gcam = alpha * cmap + (1 - alpha) * raw_image
else:
gcam = (cmap.astype(np.float) + raw_image.astype(np.float)) / 2
cv2.imwrite(filename, np.uint8(gcam))
def get_gradcam_image(gcam, raw_image, paper_cmap=False):
gcam = gcam.cpu().numpy()
cmap = cm.jet_r(gcam)[..., :3] * 255.0
if paper_cmap:
alpha = gcam[..., None]
gcam = alpha * cmap + (1 - alpha) * raw_image
else:
gcam = (cmap.astype(np.float) + raw_image.astype(np.float)) / 2
return np.uint8(gcam)
def colorize(self, labelmap):
#print(labelmap.shape)
# Assign a unique color to each label
labelmap = labelmap.astype(np.float32) / self.CONFIG.DATASET.N_CLASSES
if self.autumn:
colormap = cm.autumn(labelmap)[..., :-1] * 255.0
else:
colormap = cm.jet_r(labelmap)[..., :-1] * 255.0
return np.uint8(colormap)
def generate_gcam2d(attention_map, raw_input):
assert (len(attention_map.shape) == 2) # No batch dim
assert (isinstance(attention_map, np.ndarray)) # Not a tensor
if raw_input is not None:
attention_map = overlay(raw_input, attention_map)
else:
attention_map = _resize_attention_map(attention_map, MIN_SHAPE)
attention_map = cm.jet_r(attention_map)[..., :3] * 255.0
return np.uint8(attention_map)
def save_gradcam(filename, gcam, raw_image, paper_cmap=False):
gcam = gcam.cpu().numpy()
cmap = cm.jet_r(gcam)[..., :3] * 255.0
if paper_cmap:
alpha = gcam[..., None]
gcam = alpha * cmap + (1 - alpha) * raw_image
else:
gcam = (cmap.astype(np.float) +
255 * raw_image[..., ::-1].astype(np.float)) / 2
cv2.imwrite(filename, np.uint8(np.clip(gcam, 0, 255)))
def save_gradcam(filename, gcam, raw_image, save_as_file=False):
print(f"\t Generating Image : {filename}")
gcam = gcam.cpu().numpy()
cmap = cm.jet_r(gcam)[..., :3] * 255.0
gcam = (cmap.astype(np.float) + raw_image.astype(np.float)) / 2
if save_as_file:
cv2.imwrite(filename, np.uint8(gcam))
return None
return np.uint8(gcam)
def overlay(raw_input, attention_map):
if np.max(raw_input) > 1:
raw_input = raw_input.astype(np.float)
raw_input /= 255
attention_map = cv2.resize(attention_map,
tuple(np.flip(raw_input.shape[:2])))
attention_map = cm.jet_r(attention_map)[..., :3]
attention_map = (attention_map.astype(np.float) +
raw_input.astype(np.float)) / 2
attention_map *= 255
return attention_map
def save_gradcam(filename, gcam, raw_image, paper_cmap=False):
gcam = gcam.cpu().numpy()
cmap = cm.jet_r(gcam)[..., :3] * 255.0
if paper_cmap:
alpha = gcam[..., None]
gcam = alpha * cmap + (1 - alpha) * raw_image
else:
#gcam = (cmap.astype(np.float) + raw_image.astype(np.float)) / 2
gcam = (cmap.astype(np.float) + raw_image.transpose(0, 1).transpose(
1, 2).cpu().detach().numpy()) / 2
cv2.imwrite(filename, np.uint8(gcam))
def _visualize(self, img, res):
for obj in res.objects:
tl_x = obj.region.x_offset
tl_y = obj.region.y_offset
br_x = tl_x + obj.region.width
br_y = tl_y + obj.region.height
color = np.array(cm.jet_r(obj.score)[0:3]) * 255
cv2.rectangle(img, (tl_x, tl_y), (br_x, br_y), color, 3)
cv2.putText(img, obj.class_name, (tl_x, tl_y - 2), cv2.FONT_HERSHEY_COMPLEX, 1.0, color, 2)
cv2.imshow("color", img)
cv2.waitKey(10)
def makePreview(array, filename):
import numpy as np
import matplotlib
import matplotlib.mlab as mlab
import matplotlib.pyplot as plt
import matplotlib.cm as cm
from PIL import Image
import cv2
import scipy.ndimage
x=array.shape[0]
y=array.shape[1]
min= 99999
max=0
for i in range (0,x):
for ii in range (0,y):
if (array[i][ii]!=0 and array[i][ii]<min):
min=array[i][ii]
if (array[i][ii]!=0 and array[i][ii]>max):
max=array[i][ii]
print('%s %s %s %s' %(min,max, x, y))
mask=np.empty([x, y], dtype='float64')
for i in range (0,x):
for ii in range (0,y):
if (array[i][ii]!=0):
mask[i][ii]=255
if (array[i][ii]==0):
mask[i][ii]=0
if (min!=max and max!=0):
array=(1-(array-min)/(max-min))
array = cv2.resize(array, (0,0), fx=5, fy=5)
mask = cv2.resize(mask, (0,0), fx=5, fy=5)
array= scipy.ndimage.median_filter(array, 4)
values = Image.fromarray(np.uint8(cm.jet_r(array)*255)).convert('RGB')
hs_array = Image.fromarray(np.uint8(hillshade(array*255,45, 315, 0.5))).convert('RGB')
new_img = Image.blend(values, hs_array, 0.5).convert('RGBA')
mask = Image.fromarray(np.uint8(mask)).convert('L')
new_img.putalpha(mask)
new_img.save(pathTif+filename+'prev.png')
img = Image.open(pathTif+filename+'prev.png')
img.show()
else:
print('error in reading image')
def colorize_depth(depth_map):
# scale everything to [0, 255]
sorted_depth = np.unique(np.sort(depth_map.flatten()))
min_depth = sorted_depth[0]
max_depth = sorted_depth[len(sorted_depth) - 1]
depth_map = np.asarray(
map(lambda pixel: (pixel - min_depth) * 1.0 / (max_depth - min_depth),
depth_map))
# Apply jet colormap to it
depth_map = np.uint8(cm.jet_r(depth_map) * 255)
return depth_map[:, :, 0:3]
def overlay(raw_input, attention_map):
if isinstance(raw_input, torch.Tensor):
raw_input = raw_input.detach().cpu().numpy()
if raw_input.shape[0] == 1 or raw_input.shape[0] == 3:
raw_input = raw_input.transpose(1, 2, 0)
if np.max(raw_input) > 1:
raw_input = raw_input.astype(np.float)
raw_input /= 255
attention_map = cv2.resize(attention_map, tuple(np.flip(raw_input.shape[:2])))
attention_map = cm.jet_r(attention_map)[..., :3]
attention_map = (attention_map.astype(np.float) + raw_input.astype(np.float)) / 2
attention_map *= 255
return attention_map
def save_gradcam(filename, gcam, raw_image, paper_cmap=False):
gcam = gcam.cpu().numpy()
cmap = cm.jet_r(gcam)[..., :3] * 255.0
if paper_cmap:
alpha = gcam[..., None]
gcam = alpha * cmap + (1 - alpha) * raw_image
else:
#gcam = (cmap.astype(np.float) + raw_image.astype(np.float)) / 2
gcam = cmap.astype(np.float)
gcam = gcam.astype(np.uint8)
cv2.imwrite(filename, gcam)
#cv2.imwrite(filename, np.uint8(gcam))
return gcam
def makePreview(self,array, hillshade_mat, filename):
import matplotlib
import matplotlib.mlab as mlab
import matplotlib.pyplot as plt
import matplotlib.cm as cm
from PIL import Image
import cv2
#import scipy.ndimage
x=array.shape[0]
y=array.shape[1]
min= 99999
max=0
if self.stdClip_check.isChecked():
array=self.stdClip(array, self.stdClip_slider.value()/float(100))
for i in range (0,x):
for ii in range (0,y):
if (array[i][ii]!=0 and array[i][ii]<min):
min=array[i][ii]
if (array[i][ii]!=0 and array[i][ii]>max):
max=array[i][ii]
print('min: %.2f max: %.2f \nsize: %s x %s' %(min,max, x, y))
mask=np.empty([x, y], dtype='float64')
for i in range (0,x):
for ii in range (0,y):
if (array[i][ii]!=0):
mask[i][ii]=255
if (array[i][ii]==0):
mask[i][ii]=0
if (min!=max and max!=0):
array=(1-(array-min)/(max-min))
array = cv2.resize(array, (0,0), fx=5, fy=5)
hillshade_mat = cv2.resize(hillshade_mat, (0,0), fx=5, fy=5)
mask = cv2.resize(mask, (0,0), fx=5, fy=5, interpolation=cv2.INTER_NEAREST)
#array= scipy.ndimage.median_filter(array, 4)
values = Image.fromarray(np.uint8(cm.jet_r(array)*255)).convert('RGB')
hs_array = Image.fromarray(np.uint8(hillshade_mat)).convert('RGB')
new_img = Image.blend(values, hs_array, 0.3).convert('RGBA')
mask = Image.fromarray(np.uint8(mask)).convert('L')
new_img.putalpha(mask)
new_img.save(str(self.pathOutput+filename)+'preview.png')
# self.displayPreview()
else:
print('error in reading image')
def saveGCam(gc, raw_image, opath, paper_cmap=False):
"""
Save the grad-cam overlaid on the image
"""
gc = gc.cpu().numpy()
cmap = cm.jet_r(gc)[..., :3] * 255.0
if paper_cmap:
alpha = gc[..., None]
gc = alpha * cmap + (1 - alpha) * raw_image
else:
gc = (cmap.astype(np.float) + raw_image.astype(np.float)) / 2
cv2.imwrite(opath, np.uint8(gc))
return
def save_gradcam_overlay(filename, gcam_img, input_img, gcam_alpha=0.2):
''' gcam_img and input_img should two tensors with the same shape
'''
gcam_img = gcam_img.numpy()
input_img = input_img.numpy()
assert np.shape(gcam_img) == np.shape(input_img),\
f"gcam_img shape: {np.shape(gcam_img)}; input_img shape: {np.shape(input_img)}. "\
"gcam_img and input_img should have the same shape!"
gcam_img = cm.jet_r(gcam_img)[..., :3] * 255.0
input_img = cm.gray(input_img)[..., :3] * 255.0
gcam_overlay = gcam_alpha*gcam_img.astype(np.float) + \
(1-gcam_alpha)*input_img.astype(np.float)
cv2.imwrite(filename, gcam_overlay.astype(np.uint8))
def refresh(self):
if not self.IsShown():
return
self.colPlotPan.draw()
self.lFluorSpecies.DeleteAllItems()
for key in self.pipeline.colour_mapper.species_ratios.keys():
ind = self.lFluorSpecies.InsertStringItem(UI_MAXSIZE, key)
ratio = self.pipeline.colour_mapper.species_ratios[key]
self.lFluorSpecies.SetStringItem(ind, 1, '%3.3f' % ratio)
self.lFluorSpecies.SetItemTextColour(
ind, wx.Colour(*((128 * numpy.array(cm.jet_r(ratio)))[:3])))
num_dyes = sum(self.pipeline.colourFilter._index(key))
self.lFluorSpecies.SetStringItem(ind, 2, '%d' % num_dyes)
def __init__(self, res, vc, f, ap):
mi = 'minimum'
ma = 'maksimum'
sr = 'srednje'
arr = 'array'
nplt = 500
vplt = {mi: vc.min(), ma: vc.max(), sr: vc.mean()}
vplt[arr] = np.linspace(vplt[mi], vplt[ma], nplt)
fplt = {mi: f.min(), ma: f.max(), sr: f.mean()}
fplt[arr] = np.linspace(fplt[mi], fplt[ma], nplt)
applt = {mi: ap.min(), ma: ap.max(), sr: ap.mean()}
applt[arr] = np.linspace(applt[mi], applt[ma], nplt)
t = PostojanostAlata()
self.d3 = plt.figure(figsize=(22, 11), dpi=300)
ax = self.d3.add_subplot(111, projection='3d')
ax.set_title('Postojanost alata pri razlicitim dubinama rezanja',
fntdict)
ax.set_xlabel(r'Brzina rezanja $[\frac{m}{min}]$', fntdict)
ax.set_ylabel(r'Posmak $[\frac{mm}{okr}]$', fntdict)
ax.set_zlabel(r'Postojanost alata $[min]$', fntdict)
V, F = np.meshgrid(vplt[arr], fplt[arr])
Ap = np.empty(V.shape)
Postmax = np.array([])
for i in applt[mi], applt[sr], applt[ma]:
Ap[:] = i
Post = t(res.x, V, F, Ap)
Postmax = np.append(Postmax, Post.max())
N = Post / Postmax.max()
ax.plot_surface(V,
F,
Post,
linewidth=0,
facecolors=cm.jet_r(N),
antialiased=False,
shade=False)
def preview(self, dataset):
kwargs = {"nrow": 4, "padding": 40}
for i, (_, images, labels) in enumerate(dataset):
if i == 0:
image = make_grid(images, pad_value=-1, **kwargs).numpy()
image = np.transpose(image, (1, 2, 0))
mask = np.zeros(image.shape[:2])
mask[(image != -1)[..., 0]] = 255
image = np.dstack((image, mask)).astype(np.uint8)
labels = labels[:, np.newaxis, ...]
label = make_grid(labels, pad_value=255, **kwargs).numpy()
label_ = np.transpose(label, (1, 2, 0))[..., 0].astype(np.float32)
label = cm.jet_r(label_ / 3.0) * 255
label[..., 3][(label_ == 255)] = 0
label = label.astype(np.uint8)
tiled_images = np.hstack((image, label))
# cv2.imwrite("./docs/datasets/voc12.png", tiled_images)
plt.figure(figsize=(40, 20))
plt.imshow(np.dstack((tiled_images[..., 2::-1], tiled_images[..., 3])), aspect='auto')
plt.show()
return
loader = data.DataLoader(dataset, batch_size=batch_size)
for i, (image_ids, images,
labels) in tqdm(enumerate(loader),
total=np.ceil(len(dataset) / batch_size),
leave=False):
if i == 0:
mean = torch.tensor((104.008, 116.669, 122.675))[None, :, None,
None]
images += mean.expand_as(images)
image = make_grid(images, pad_value=-1, **kwargs).numpy()
image = np.transpose(image, (1, 2, 0))
mask = np.zeros(image.shape[:2])
mask[(image != -1)[..., 0]] = 255
image = np.dstack((image, mask)).astype(np.uint8)
labels = labels[:, np.newaxis, ...]
label = make_grid(labels, pad_value=255, **kwargs).numpy()
label_ = np.transpose(label, (1, 2, 0))[..., 0].astype(np.float32)
label = cm.jet_r(label_ / 21.0) * 255
mask = np.zeros(label.shape[:2])
label[..., 3][(label_ == 255)] = 0
label = label.astype(np.uint8)
tiled_images = np.hstack((image, label))
# cv2.imwrite("./docs/datasets/voc12.png", tiled_images)
plt.imshow(
np.dstack((tiled_images[..., 2::-1], tiled_images[..., 3])))
plt.show()
break
cb.set_label('Battery Percentage')
cb.ax.tick_params(labelsize=8)
plt.savefig(os.path.join(dbdir, 'batyear.png'), dpi=300)
# yearview
#################################################
from matplotlib.colors import ListedColormap as LC
fig = plt.figure(figsize=(10,5), dpi=300)
ax = fig.add_subplot(111)
#c=ax.scatter(DOY, TOD, c=numpy.array(ssids), edgecolor='none', marker='s',
# s=1, cmap=cm.rainbow)
mp = cm.jet_r(numpy.linspace(0, 1, nSSID))
ssidcols = LC(mp, 'ssidcols')
cm.register_cmap(cmap = ssidcols)
ssidtick = AWAKE.ssids.unique()
ssidtick.sort()
ssidtick = ssidtick * (ssidtick.max() / float(nSSID))
ssidtick = ssidtick + ssidtick[1]/2.
c = ax.imshow(ssidgrid, interpolation='none', aspect='auto', cmap=ssidcols)
ax.set_xlabel('Day of Year')
ax.set_ylabel('Time of Day')
#ax.set_xlim((0, 366))
#ax.set_ylim((-288, 0))
def get_normed_colormap(inarray):
norm = colors.Normalize(inarray.min(), inarray.max())
jj = jet_r(norm(inarray))
cl = [colors.to_hex(c) for c in jj]
return cl
import matplotlib as mpl
if __name__ == '__main__':
files = list(glob.glob('evaluate-depths/predicted-*.png'))
files.extend(list(glob.glob('evaluate-depths/orig-depth-*.png')))
files.extend(list(glob.glob('evaluate-depths/gt-depth-*.png')))
for file in files:
im = np.array(Image.open(file))
colored_im = np.copy(im)
# if 'orig-' in file:
# im[im == 0] = 255
if 'gt-' in file:
colored_im[colored_im == 5] = 255
elif 'predicted-' in file:
colored_im[colored_im < 4] = 255
colored_im = Image.fromarray(np.uint8(cm.jet_r(colored_im) * 255.0))
colored_im.save('evaluate-depths/colored-' + os.path.basename(file) +
'.png')
fig = plt.figure(figsize=(1, 3))
ax1 = fig.add_axes([0, 0.05, 0.2, 0.9])
cmap = mpl.cm.jet_r
norm = mpl.colors.Normalize(vmin=50, vmax=0)
cb1 = mpl.colorbar.ColorbarBase(ax1,
cmap=cmap,
norm=norm,
orientation='vertical')
cb1.set_label('distance [m]')
plt.savefig('evaluate-depths/colorbar-cropped.png')
fig = plt.figure(figsize=(1, 3))
def colorize(labelmap):
# Assign a unique color to each label
labelmap = labelmap.astype(np.float32) / CONFIG.DATASET.N_CLASSES
colormap = cm.jet_r(labelmap)[..., :-1] * 255.0
return np.uint8(colormap)
def colors(number=None):
if (number == None):
return ["r", "coral", "gold", "limegreen", "seagreen", "aqua", \
"royalblue", "b", "navy"]
if (number != None):
return [cm.jet_r((1.0 * i) / (1.0 * number)) for i in range(number)]
fig, ax = plt.subplots()
# Plot HIV
for ih in xrange(len(S_bins) - 1):
ax.plot(binsc, hists[ih] / hists[ih, 0] * hists[-1, 0], lw=2, c='k', marker='o',
label='HIV, $S_{type M} \in ['+str(S_bins[ih])+', '+str(S_bins[ih + 1])+']$',
color=cm.jet(1.0 * ih / hists.shape[0]))
# Plot theory
al = hists[1, 0]
if add_bsc:
for (N, alpha) in sfs_bc:
ax.plot(sfs_bc[(N, alpha)], sfs_bsc[(N, alpha)],
lw=2, ls = '-',
color=cm.jet_r(1.0 * (alpha - 1)),
label = 'Beta coalescent, $\\alpha = '+str(alpha)+'$')
else:
ax.plot(binsc[:8], al*binsc[0]**2/binsc[:8]**2, lw=2, c='r')
ax.plot(binsc, al*binsc[0]/binsc, lw=2, c='b', label = 'Neutral, $\\alpha = 2$')
ax.set_xlabel('derived allele frequency')
ax.set_ylabel('SFS [density = counts / sum / binsize]')
ax.set_xlim(10**(-3.1), 1.0 - 10**(-3.1))
ax.set_ylim([3e1,3e7])
ax.set_xscale('logit')
ax.set_yscale('log')
if len(fragments) == 6:
ax.set_title('All patients, all fragments')
else:
ax.set_title('All patients, fragments '+str(fragments))
def test(target_layer_name):
model = Res101().to(device)
model.eval()
model.load_state_dict(
torch.load('res101.pt', map_location=torch.device(device)))
xs, ts, paths = data_load('../Dataset/test/images/')
target_layer = None
for name, module in model.named_modules():
if target_layer_name == name:
print('target:', name)
target_layer = module
if target_layer is None:
for name, module in model.named_modules():
print(name)
raise Exception('invalid target layer name >>', target_layer_name)
if type(target_layer) is torch.nn.Sequential:
target_layer = target_layer[-1]
print(target_layer)
fmap_pool = OrderedDict()
grad_pool = OrderedDict()
def forward_hook(key):
def forward_hook_(module, input, output):
# Save featuremaps
fmap_pool[key] = output.detach()
return forward_hook_
def backward_hook(key):
def backward_hook_(module, grad_in, grad_out):
# Save the gradients correspond to the featuremaps
grad_pool[key] = grad_out[0].detach()
return backward_hook_
# If any candidates are not specified, the hook is registered to all the layers.
for name, module in model.named_modules():
module.register_forward_hook(forward_hook(name))
module.register_backward_hook(backward_hook(name))
for i in range(len(paths)):
_x = xs[i]
t = ts[i]
path = paths[i]
x = np.expand_dims(_x, axis=0)
x = torch.tensor(x, dtype=torch.float).to(device)
# forward network
logit = model(x)
pred = F.softmax(logit, dim=1).detach().cpu().numpy()
raw_image = (_x).transpose(1, 2, 0)
plt.subplot(1, Class_N + 1, 1)
plt.imshow(raw_image)
plt.title('input')
for i, class_label in enumerate(Class_label):
# set one-hot class activity
class_index = torch.zeros(pred.shape).to(device)
_index = Class_label.index(class_label)
class_index[:, _index] = 1
logit.backward(gradient=class_index, retain_graph=True)
#target_layer_output = target_layer.forward(x)
fmaps = fmap_pool[target_layer_name]
grads = grad_pool[target_layer_name]
weights = F.adaptive_avg_pool2d(grads, 1)
gcam = torch.mul(fmaps, weights).sum(dim=1, keepdim=True)
gcam = F.relu(gcam)
gcam = F.interpolate(gcam, [img_height, img_width],
mode="bilinear",
align_corners=False)
B, C, H, W = gcam.shape
gcam = gcam.view(B, -1)
gcam -= gcam.min(dim=1, keepdim=True)[0]
gcam /= gcam.max(dim=1, keepdim=True)[0]
gcam = gcam.view(B, C, H, W)
gcam = gcam.cpu().numpy()[0, 0]
cmap = cm.jet_r(gcam)[..., :3]
gcam = (cmap.astype(np.float) + raw_image.astype(np.float)) / 2
plt.subplot(1, Class_N + 1, i + 2)
plt.imshow(gcam)
plt.title('{} :{:.2f}'.format(class_label, pred[0, i]))
plt.show()
print("in {}, predicted probabilities >> {}".format(path, pred))