**Model trained On the AM0.1-FLUKE dataset**
Setup environment¶
Imports¶
from cellsegment.core import *
from cellsegment.inference_utils import *
from cellsegment.set_directories import *
import pandas
from fastai import *
from fastai.vision import *
Define directories¶
local_datapath = '../testdata_2/'
# local_datapath = '../../data/FEC/03-M100-Fluke-2019-11/'
dirs = Dirs('data') if IN_COLAB else Dirs(local_datapath)
dirs.model = dirs.model if IN_COLAB else '/home/john/github/data/FEC/03-M100-Fluke-2019-11/models'
print(dirs)
Load the Training Images¶
(if in colab)
%%bash
[[ ! -e /tools/google-cloud-sdk ]] && exit # if in colab
export fileid=1SEW0Kf1CI4e4-up4TGsDqwDwVk_QZEUf
export filename=Fluke-Train-2019-12-01.zip
## CURL ##
curl -L -c cookies.txt 'https://docs.google.com/uc?export=download&id='$fileid \
| sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1/p' > confirm.txt
curl -L -b cookies.txt -o $filename \
'https://docs.google.com/uc?export=download&id='$fileid'&confirm='$(<confirm.txt)
rm -f confirm.txt cookies.txt
unzip -u -q $filename -d data
Load exported Model¶
%%bash
[[ ! -e /tools/google-cloud-sdk ]] && exit # if in colab
switch=true
if $switch; then
export fileid=11cZWhg23QDag_3b7jcd02U8Pzq3W6U5Y
export filename=export-fluke-2019-12-01.pkl
## CURL ##
curl -L -c cookies.txt 'https://docs.google.com/uc?export=download&id='$fileid \
| sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1/p' > confirm.txt
curl -L -b cookies.txt -o $filename \
'https://docs.google.com/uc?export=download&id='$fileid'&confirm='$(<confirm.txt)
rm -f confirm.txt cookies.txt
fi
Run model inference on all of the test tiles¶
path = Path(dirs.crop)
fnames = get_image_files(path/'Test')
print (f'Number of test tiles {len(fnames)}')
Create Learner from exported model¶
# defaults.device = 'cpu'
defaults.device = 'cuda'
if torch.cuda.is_available():
def acc_metric1(input, target):
target = target.squeeze(1)
return (input.argmax(dim=1)==target).float().mean()
def acc_metric2(input, target):
target = target.squeeze(1)
return (input.argmax(dim=1)[target>0]==target[target>0]).float().mean()
fn_model = f'{dirs.model}/export-fluke-2019-12-01.pkl'
learn = load_learner('', fn_model)
learn.model.float()
# summary(learn.model, (3, 32, 32))
print("Learner loaded")
else:
print("Learner not loaded as torch.cuda.is_available() = ", torch.cuda.is_available())
Inferences¶
Running an inference produces 3 tensors with the same xy dimensions Each xy position contains the predicted class, label and probabilities for pixel at that location. We concentrate here on the raw predictions as this gis use more information as the first two are binary thresholded results. Plotting each of the layers of raw_preds shows the probability of the pixel at that location being in one of the n classes. Here we have 4 classes,
Background,Fluke-Liver,Fluke-Rumen&other
def run_and_plot_raw_pred(fn):
fig, axes = plt.subplots(1, 5, figsize=(15, 3))
titles = ['Background','Fluke-Liver','Fluke-Rumen','other']
img = open_image(fn)
pc,pi,raw_pred = learn.predict(open_image(fn))
ax = axes.flat[0]
im = ax.imshow(img.data.permute(1,2,0).numpy())
ax.set_axis_off()
ax.set_title(fn.name)
for i, ax in enumerate(axes.flat[1:]):
arr = (raw_pred[i,:,:]).numpy()
im = ax.imshow(arr, cmap='inferno', vmin=0, vmax=1)
ax.set_axis_off()
ax.set_title(titles[i])
df = pd.read_csv(path/"test_df.csv")
df = df[['Name', 'Label', 'Op']]
df
Raw_preds for a Liver Fluke¶
run_and_plot_raw_pred(path/'Test'/df.iloc[0].Name)
print('Label is ', df.iloc[0].Label)
Raw_preds for a Rumen Fluke¶
run_and_plot_raw_pred(path/'Test'/df.iloc[2].Name)
print('Label is ', df.iloc[2].Label)
Probability Calculation¶
The background layer is used the generate a binary mask, this . For each of the n class layers ( here n = 4 ) the average of all the pixels within the mask region is calculated and returned. These should sum to 1
Run inferences on all the test tiles¶
path_img = path/'Train'
path_lbl = path/'Label'
path_tst = path/'Test'
fnames = sorted(get_image_files(path/'Test'))
print (f'Number of test tiles to run {len(fnames)}')
preds= run_inferences(learn, fnames, number_files='all')
print (f'Number of predictions generated {len(preds)}')
Dataframe of inference probabilities¶
print(f"Populate dataframe in path {path/'Test'}")
df = pd.read_csv(path/"test_df.csv")
df = df[['Name', 'Label', 'Op']]
col_heads = ['Background', 'Fluke_Liver', 'Fluke_Rumen', 'Other']
print(f'Adding inference probs to the dataframe')
preds_to_df(preds, col_heads, df)
df.to_csv(dirs.crop+'/results_df.csv', index=False)
df.sort_values(by=['Name'], inplace=True)
print(f'Finding maximums')
df["max_prob"] = df[["Background", "Fluke_Liver", "Fluke_Rumen"]].max(axis=1).round(2)
df.loc[df.Fluke_Liver == df.max_prob, "max_label"] = 'Fluke-Liver'
df.loc[df.Fluke_Rumen == df.max_prob, "max_label"] = 'Fluke-Rumen'
df.to_csv(dirs.crop+'/results_df.csv', index=False)
#print(df[['Name', 'Label', 'Pstr', 'max_prob']].tail(10))
summary_df = df[['Name', 'Label', 'max_label', 'max_prob']]
summary_df
Explanation of results¶
Plot of all the test tiles¶
plot_inferences(preds, df, src_path=path_tst, label='max_label', start=0, cols=4, rows=4)
# add the actual label todo
print("make image columns number even")
fnames = sorted(get_image_files(dirs.train))
img = open_image(fnames[1])
if img.shape[2] % 2 == 1:img.px = img.px[:,:,1:]
pc,pi,raw_pred = learn.predict(img)
find_prediction_blobs(img, CONF=0.5, min_area=500,offset=0)
import torch.nn.functional as F
source = torch.ones((3, 5, 5))
# now we expand to size (7, 11) by appending a row of 0s at pos 0 and pos 6,
# and a column of 0s at pos 10
result = F.pad(input=source, pad=(1, 1,1,1), mode='constant', value=(0)
print(source.shape)
print(result.shape)
result
result.shape
old - todo¶
from skimage.measure import label, regionprops
def _get_props(tens, layer, min_conf):
min_area = 100
arr = o[layer,:,:].numpy()
if layer == 0:
label_arr = label(arr < 1-min_conf)
else:
label_arr = label(arr > min_conf)
region_props = []
for region in regionprops(label_arr, arr, cache=True):
if region.area > min_area:
cx = int(region.centroid[1])
cy = int(region.centroid[0])
region_props.append({
"class_layer": layer,
"centroid": region.centroid,
"mean_intensity": region.mean_intensity,
"area": region.area,
"coords": region.coords,
})
return region_props
def _get_bool_mask(o):
min_conf = 0.5
arr = o[0,:,:].numpy()
mask = np.zeros((arr.shape), dtype=bool)
bgnd = _get_props(o, 0, min_conf)
AAcoords = bgnd[0]["coords"]
mask[AAcoords.T.tolist()] = True
return mask
def _calc_probs(o):
mask = _get_bool_mask(o)
p = [None]*o.shape[0]
for i in range(o.shape[0]):
p[i] = o[i,:,:].numpy()[mask].mean()
return p
old_ Calculate probability of marked up region¶
# # get boolean mask of a region
# mask = get_bool_mask(o)
# show_img(mask)
p = calc_probs(o)
print(f'{p[0]:3.2f} {p[1]:3.2f} {p[2]:3.2f}')
p
# print(f'{p:3.2f}')p
# print(f'{p0.mean():3.2f},{p1.mean():3.2f},{p2.mean():3.2f},{p3.mean():3.2f}')
# print(f'{(p0+p1+p2).mean():3.2f}')
min_conf = 0.2
bgnd = __get_props(o, 0, min_conf)
lyr1 = __get_props(o, 1, min_conf)
lyr1
lyr2 = __get_props(o, 2, min_conf)
lyr2
print(bgnd[0]["area"] )
print(lyr1[0]["area"], lyr1[0]["mean_intensity"], lyr1[0]["mean_intensity"] * lyr1[0]["area"] / bgnd[0]["area"])
print(lyr2[0]["area"], lyr2[0]["mean_intensity"], lyr2[0]["mean_intensity"] * lyr2[0]["area"] / bgnd[0]["area"])
# region.mean_intensity * region.area / 314
bgnd[0]["centroid"]
print(o[0,90,90], o[1,90,90], o[2,90,90], o[3,90,90])
AAcoords = bgnd[0]["coords"]
AAcoords
arr = o[0,:,:].numpy()
bin = label(arr < 0.5) > 0
bin.max()
AA0 = o[0,:,:].numpy()[bin]
AA0.mean()
AA1 = o[1,:,:].numpy()[bin]
AA1.mean()
AA2 = o[2,:,:].numpy()[bin]
AA2.mean()
(AA0+AA1+AA2).mean()
def find_prediction_blobs(img, min_conf=0.1, min_area=100, plot=False):
RESIZE = 2.0
img = cv2.resize(img,None,fx=1.0/RESIZE, fy=1.0/RESIZE, interpolation = cv2.INTER_AREA)
min_area = int(min_area/RESIZE)
# region props seems to have region.max_intensity errors if no data not np.int
SCALE = 100
min_conf *= SCALE
img = (img * (SCALE/img.max())).astype(np.int)
img[img[:,:,0]<min_conf,0] = 0
img[img[:,:,1]<min_conf,1] = 0
predictions = []
label_image0 = label(img[:,:,0] > min_conf)
label_image1 = label(img[:,:,1] > min_conf)
for region in regionprops(label_image0, img[:,:,0], cache=True):
if region.area > min_area:
cx = int(region.centroid[1]*RESIZE)
cy = int(region.centroid[0]*RESIZE)
predictions.append({
"label": 'Strongyle',
"point": [cx,cy],
"probability": region.max_intensity.round(2),
"area": region.area
})
fill = (0,255,0)
for region in regionprops(label_image1, img[:,:,1]):
if region.area > min_area:
cx = int(region.centroid[1]*RESIZE)
cy = int(region.centroid[0]*RESIZE)
predictions.append({
"label": 'Nematodirus',
"point": [cx,cy],
"probability": region.max_intensity.round(2),
"area": region.area
})
if plot:
# show_img(imglab, figsize = (15,15))
plt.figure(figsize=(15, 15))
plt.subplot(121)
plt.imshow(label_image0, cmap='nipy_spectral')
# plt.imshow(img[:,:,0] > min_conf, cmap='nipy_spectral')
plt.axis('off')
plt.subplot(122)
plt.imshow(label_image1, cmap='nipy_spectral')
plt.axis('off')
plt.tight_layout()
plt.show()
return predictions, img