|
a |
|
b/experiments/reconstruction_test.py |
|
|
1 |
import matplotlib.pyplot as plt |
|
|
2 |
|
|
|
3 |
from models.segmentation_models import * |
|
|
4 |
from data.hyperkvasir import KvasirSegmentationDataset |
|
|
5 |
from torch.utils.data import DataLoader |
|
|
6 |
import torch |
|
|
7 |
import torch.nn as nn |
|
|
8 |
import copy |
|
|
9 |
|
|
|
10 |
|
|
|
11 |
class SplicedReconstructor(nn.Module): |
|
|
12 |
def __init__(self): |
|
|
13 |
super(SplicedReconstructor, self).__init__() |
|
|
14 |
inductivenet = InductiveNet() |
|
|
15 |
inductivenet.load_state_dict(torch.load("Predictors/Augmented/InductiveNet/consistency_1")) |
|
|
16 |
self.decoder = copy.deepcopy(inductivenet.reconstruction_decoder) |
|
|
17 |
self.head = copy.deepcopy(inductivenet.reconstruction_head) |
|
|
18 |
del inductivenet |
|
|
19 |
deeplab = DeepLab() |
|
|
20 |
deeplab.load_state_dict(torch.load("Predictors/Augmented/DeepLab/consistency_1")) |
|
|
21 |
self.encoder = copy.deepcopy(deeplab.encoder) |
|
|
22 |
del deeplab |
|
|
23 |
|
|
|
24 |
def predict(self, x): |
|
|
25 |
features = self.encoder(x) |
|
|
26 |
reconstructor_output = self.decoder(*features) |
|
|
27 |
reconstructed = self.head(reconstructor_output) |
|
|
28 |
return reconstructed |
|
|
29 |
|
|
|
30 |
|
|
|
31 |
if __name__ == '__main__': |
|
|
32 |
model = SplicedReconstructor().to("cuda").eval() |
|
|
33 |
|
|
|
34 |
for x, y, _ in DataLoader(KvasirSegmentationDataset("Datasets/HyperKvasir/", "test")): |
|
|
35 |
with torch.no_grad(): |
|
|
36 |
reconstruction = model.predict(x.to("cuda")).cpu() |
|
|
37 |
fig, ax = plt.subplots(ncols=1, nrows=2, sharey=True, sharex=True, figsize=(2, 1), dpi=1000) |
|
|
38 |
fig.subplots_adjust(wspace=0, hspace=0) |
|
|
39 |
ax[0].imshow(reconstruction[0].T) |
|
|
40 |
ax[1].imshow(x[0].T) |
|
|
41 |
plt.show() |
|
|
42 |
print("Showing...") |