def load_prob_cnn(paths, X_test, Y_test, num=100):
softmax = torch.nn.Softmax(dim=0)
model1 = VGG16()
model1.load_state_dict(torch.load(paths))
print("loaded model!")
new_size = np.array([])
test_size = np.array([])
models_prob = np.zeros((2, num * 77))
for i in range(len(paths)):
new_size = np.append(new_size, Y_test)
test_size = np.append(test_size, np.shape(new_size)[0])
cur_model = model1
cur_model.eval()
print("\n now evaluating")
for j in range(num):
A = cur_model(X_test[j * 77:(j + 1) * 77, :, :, :].float())
B = A.detach().numpy()
models_prob[:, j * 77:(j + 1) * 77] = B.T
mp = torch.from_numpy(models_prob)
out = softmax(mp)
output = out.numpy()
return test_size, np.around(output[1, :], decimals=2)
w2 = torch.cat((ww1, ww2))
l1 = labels4[:n4] + labels3[:n3]
l2 = labels1[n1:] + labels2[n2:]
wid = window_idx_full1[n1:] + window_idx_full2[n2:]
train_windows, train_labels = shuffle(w1, l1)
test_windows, test_labels, test_id = shuffle(w2, l2, wid)
return train_windows.detach(), test_windows.detach(
), train_labels, test_labels, test_id
C = preprossingPipeline(BC_datapath=r"/work3/s173934/Fagprojekt/dataEEG")
path_s = r'/work3/s173934/Fagprojekt/spectograms_rgb'
criterion = nn.CrossEntropyLoss()
X_train, X_valid, Y_train, Y_valid, windows_id = split_dataset_balanced(
C, path_s, N=120, train_split=80, max_windows=20, num_channels=14)
modelA = VGG16()
freeze_parameters(modelA, feature_extracting=True)
list2 = np.array(list_of_features(modelA))
modelB = VGG16()
freeze_parameters(modelB, feature_extracting=True)
for i in range(2):
PATH = '/zhome/87/9/127623/FagprojektBALANCEDTESTS'
if i == 0:
activation_list = np.array([28, 29, 30, 31])
grad_parameters(modelA, list(list2[activation_list]))
optimizer = optim.Adam(modelA.parameters(), lr=0.0005)
train_acc, train_loss, val_acc, val_loss, wrong_guesses, wrong_predictions, all_pred, modelA = test_CNN(
modelA,
X_train,
Y_train,
X_valid,
train_windows, train_labels = shuffle(w1, l1)
test_windows, test_labels, test_id = shuffle(w2, l2, wid)
return train_windows.detach(), test_windows.detach(
), train_labels, test_labels, test_id
C = preprossingPipeline(BC_datapath=r"/work3/s173934/Fagprojekt/dataEEG")
path_s = r'/work3/s173934/Fagprojekt/spectograms_rgb'
criterion = nn.CrossEntropyLoss()
X_train, X_valid, Y_train, Y_valid, windows_id = split_dataset(C,
path_s,
N=120,
train_split=80,
max_windows=20,
num_channels=14)
modelB = VGG16()
list2 = np.array(list_of_features(modelB))
freeze_parameters(modelB, feature_extracting=True)
PATH = '/zhome/87/9/127623/FagprojektBALANCEDTESTS'
activation_list = np.array([20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31])
grad_parameters(modelB, list(list2[activation_list]))
optimizer = optim.Adam(modelB.parameters(), lr=0.0001)
train_acc, train_loss, val_acc, val_loss, wrong_guesses, wrong_predictions, all_pred, modelB = test_CNN(
modelB,
X_train,
Y_train,
X_valid,
Y_valid,
windows_id,
batch_size=256,
import pickle
import sys
from CNN.modifyCNN import VGG16
def new_forward(model, x):
x = model.features(x)
x = model.avgpool(x)
x = torch.flatten(x, 1)
x = model.classifier[0:5](x)
return x
test_set = torch.load(
r'C:\Users\johan\iCloudDrive\DTU\KID\4. semester\Fagprojekt\Resultater\CNN_ub\test_set_ub.pt'
)
test_labels = np.load(
r'C:\Users\johan\iCloudDrive\DTU\KID\4. semester\Fagprojekt\Resultater\CNN_ub\test_labels_ub.npy'
)
test_windows = np.load(
r'C:\Users\johan\iCloudDrive\DTU\KID\4. semester\Fagprojekt\Resultater\CNN_ub\test_windows_ub.npy'
)
correct_pred = []
model1_b = VGG16()
model1_b.load_state_dict(torch.load(PATH))
feature_vectors = np.array([1234, 4096])
for i in range(np.shape(correct_pred)[0]):
feature_vectors[i, :] = new_forward(model1_b, test_set[correct_pred])
from torch.autograd import Variable
from flashtorch.activmax import GradientAscent
import numpy as np
test_set = torch.load(r'C:\Users\johan\iCloudDrive\DTU\KID\4. semester\Fagprojekt\Resultater\CNN_b\test_set_b.pt')
test_labels = np.load(r'C:\Users\johan\iCloudDrive\DTU\KID\4. semester\Fagprojekt\Resultater\CNN_b\test_labels.npy')
test_windows = np.load(r'C:\Users\johan\iCloudDrive\DTU\KID\4. semester\Fagprojekt\Resultater\CNN_b\test_windows.npy')
cpi = np.load(r'C:\Users\johan\iCloudDrive\DTU\KID\4. semester\Fagprojekt\Kode\Gruppen\Fagprojekt2020\CNN_HPC\cpi.npy')
marks = np.load(r'C:\Users\johan\iCloudDrive\DTU\KID\4. semester\Fagprojekt\Kode\Gruppen\Fagprojekt2020\GMM\marked_windows.npy')
windows = test_windows[cpi]
labels = test_labels[cpi]
data_set = test_set[cpi,:,:,:]
images = data_set[marks,:,:,:]
w = windows[marks]
y = labels[marks]
model = VGG16()
model.load_state_dict(torch.load(r'C:\Users\johan\iCloudDrive\DTU\KID\4. semester\Fagprojekt\Resultater\CNN_b\model1_l5_b.pt'))
model.eval()
def get_image(image,num):
img1 = image[num, :, :, :].unsqueeze(0).requires_grad_(requires_grad=True)
return img1
img1 = get_image(images,0)
img2 = get_image(images,1)
img3 = get_image(images,2)
img4 = get_image(images,3)
img5 = get_image(images,4)
img6 = get_image(images,5)
img7 = get_image(images,6)
img8 = get_image(images,7)