Skip to content
Snippets Groups Projects
Commit 76877ba9 authored by umlauf's avatar umlauf
Browse files

CEL + mean

parent 67e76a5c
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
Code/mean_deviation_results.py
+ 37
− 13
View file @ 76877ba9
......@@ -185,33 +185,57 @@ high_reloc_f1 = [0.927651139742319, 0.9285014691478942, 0.9384164222873901, 0.93
high_reloc_prec = [0.9454545454545454, 0.9349112426035503, 0.9430255402750491, 0.9482071713147411, 0.9375]
high_reloc_rec = [0.9105058365758755, 0.9221789883268483, 0.933852140077821, 0.9260700389105059, 0.933852140077821]
#MixUp:
mixup_sem_acc = [0.8821585903083701, 0.8733480176211453, 0.8777533039647577, 0.8821585903083701, 0.8788546255506607]
mixup_sem_f1 = [0.6805970149253731, 0.6587537091988132, 0.7008086253369273, 0.6898550724637681, 0.6944444444444444]
mixup_sem_prec = [0.7702702702702703, 0.74, 0.7065217391304348, 0.7531645569620253, 0.7225433526011561]
mixup_sem_rec = [ 0.6096256684491979, 0.5935828877005348, 0.6951871657754011, 0.6363636363636364, 0.6684491978609626]
#Train
mixup_sem_acc_train = [1.0, 1.0, 0.9987980769230769, 0.9987980769230769, 1.0]
mixup_sem_f1_train = [1.0, 1.0, 0.9970326409495549, 0.9970326409495549, 1.0]
mixup_sem_prec_train = [1.0, 1.0, 1.0, 1.0, 1.0]
mixup_sem_rec_train = [1.0, 1.0, 0.9940828402366864, 0.9940828402366864, 1.0]
mixup_comp_acc = [0.8087885985748219, 0.7957244655581948, 0.7790973871733967, 0.7933491686460807, 0.7957244655581948]
mixup_comp_f1 = [0.7294117647058824, 0.7034482758620688, 0.6804123711340208, 0.6958041958041957, 0.7114093959731543]
mixup_comp_prec = [0.7948717948717948, 0.7906976744186046, 0.7615384615384615, 0.796, 0.7737226277372263]
mixup_comp_rec = [0.6739130434782609, 0.6335403726708074, 0.6149068322981367, 0.6180124223602484, 0.6583850931677019]
mixup_rel_acc = [0.9478435305917753, 0.9518555667001003, 0.9408224674022067, 0.9438314944834504, 0.9358074222668004]
mixup_rel_f1 = [0.9488188976377954, 0.9525691699604744, 0.9407035175879396,0.9444444444444444, 0.9357429718875502]
mixup_rel_prec = [0.9601593625498008, 0.9678714859437751, 0.972972972972973, 0.9635627530364372, 0.966804979253112]
mixup_rel_rec = [0.9377431906614786, 0.9377431906614786, 0.9105058365758755, 0.9260700389105059, 0.9066147859922179]
# Original array
array1 = high_comp_acc
array1 = mixup_comp_acc
#array2= roberta_context_split_comp_acc
array2 = high_comp_f1
array3 = high_comp_prec
array4 = high_comp_rec
array2 = mixup_comp_f1
array3 = mixup_comp_prec
array4 = mixup_comp_rec
#print(array)
r1 = np.mean(array1)
f1 = np.mean(array2)
prec = np.mean(array3)
rec=np.mean(array4)
print("\nMean: ", r1)
print("\nMean: ", f1)
print("\nMean: ", prec)
print("\nMean: ", rec)
rec = np.mean(array4)
print("\nMean Accuracy: ", r1)
print("\nMean F1: ", f1)
print("\nMean Precision: ", prec)
print("\nMean Recall: ", rec)
r2 = np.std(array1)
f1 = np.std(array2)
prec = np.std(array3)
rec = np.std(array4)
print("\nstd: ", r2)
print("\nstd: ", f1)
print("\nstd: ", prec)
print("\nstd: ", rec)
print("\nstd Accuracy: ", r2)
print("\nstd F1: ", f1)
print("\nstd Precision: ", prec)
print("\nstd Recall: ", rec)
# r3 = np.var(array1)
# f1 = np.var(array2)
......
Code/test_loss.py
+ 14
− 22
View file @ 76877ba9
......@@ -25,10 +25,10 @@ target = torch.tensor([0.6000, 0.4000, 0.4000, 0.6000])
l = 0.4
logits = torch.tensor([[ 0.9391, -1.2892], [-0.1607, 0.1771], [-0.1707, 0.17871], [-0.1807, 0.1971]], dtype=torch.float32)
#one_hot = torch.zeros(target.shape[0], len(logits[1]), dtype=torch.float32)
#one_hot[torch.arange(target.shape[0]), target] = 1
print()
results = torch.tensor([])
for i in range(logits.shape[0]):
lg = logits[i:i+1,:]
......@@ -37,31 +37,23 @@ for i in range(logits.shape[0]):
value_r = round(value, 1)
print("Logits: ", lg)
print("Value_r", value_r)
#integer flag ?? Für abfrag ist value int? Aber wie wenn auch 1.0 = Float?
if value == 1 or value == 0:
one_hot = torch.tensor([1-value,value]) #creating one-hot vector e.g. [0. ,1.]
loss_clear_labels = -((one_hot[0] * logprobs[0][0]) + (one_hot[1] * logprobs[0][1]))
print("Clear Labels Loss: ", loss_clear_labels)
#print("Clear Labels Loss: ", loss_clear_labels)
results = torch.cat((loss_mixed_labels.view(1), results), dim=0)
print("Results Clear Labels: ", results)
#print("Results Clear Labels: ", results)
else:
if value_r == l:
mixed_vec = torch.tensor([l, 1-l])
print("Mixed Vec: ", mixed_vec)
logprobs = torch.nn.functional.log_softmax(lg, dim=1)
print("Log:", logprobs)
#loss_mixed_labels = -torch.mul(mixed_vec, logprobs).sum()
loss_mixed_labels = -((mixed_vec[0] * logprobs[0][0]) + (mixed_vec[1] * logprobs[0][1]))
print("Loss Mixed Lables l: ", loss_mixed_labels)
results = torch.cat((loss_mixed_labels.view(1), results), dim=0)
print("Results Mixed 1: ", results)
else:
mixed_vec = torch.tensor([1-l, l])
logprobs = torch.nn.functional.log_softmax(lg, dim=1)
#loss_mixed_labels = -torch.mul(mixed_vec, logprobs).sum()
loss_mixed_labels = -((mixed_vec[0] * logprobs[0][0]) + (mixed_vec[1] * logprobs[0][1]))
print("Mixed Labels 1-l: ", loss_mixed_labels)
results = torch.cat((loss_mixed_labels.view(1), results), dim=0)
print("Results Mixed 2: ", results)
mixed_vec = torch.tensor([value_r, 1-value_r])
print("Mixed Vec: ", mixed_vec)
logprobs = torch.nn.functional.log_softmax(lg, dim=1)
print("Log:", logprobs)
#loss_mixed_labels = -torch.mul(mixed_vec, logprobs).sum()
loss_mixed_labels = -((mixed_vec[0] * logprobs[0][0]) + (mixed_vec[1] * logprobs[0][1]))
print("Loss Mixed Lables l: ", loss_mixed_labels)
results = torch.cat((loss_mixed_labels.view(1), results), dim=0)
print("Results Mixed 1: ", results)
print("Results all", results)
average_loss = results.mean() #compute the average
print("Avg", average_loss)
......
Code/train.py
+ 34
− 16
View file @ 76877ba9
......@@ -139,42 +139,60 @@ def train(model, name, seed,gradient_accumulation_steps,mixup, threshold, lambda
return evaluation_test, evaluation_train
#log base e
#Fkt vom Meeting
#BINARY
def cross_entropy(logits, target, l):
results = torch.tensor([], device='cuda')
for i in range (logits.shape[0]):
lg = logits[i:i+1,:]
lg = logits[i:i+1,:] #comment
t = target[i]
#makes the logits in log (base e) probabilities
logprobs = torch.nn.functional.log_softmax(lg, dim=1)
value = t.item() #gets Item (0. or 1.)
if value == 1 or value == 0:
one_hot = torch.tensor([1-value,value], device='cuda:0') #creating one-hot vector e.g. [0. ,1.]
#class 1 and 2 mixed
loss_clear_labels = -((one_hot[0] * logprobs[0][0]) + (one_hot[1] * logprobs[0][1]))
#print("Clear Labels: ", loss_clear_labels)
results = torch.cat((loss_clear_labels.view(1), results), dim=0)
#print("Results Clear Label: ", results)
else:
value_r = round(value, 1) #to make it equal to lambda_value e.g. 0.4
if value_r == l:
#create vector: e.g. [l, 1-l]
mixed_vec = torch.tensor([l, 1-l], device='cuda')
loss_mixed_labels_1 = -((mixed_vec[0] * logprobs[0][0]) + (mixed_vec[1] * logprobs[0][1]))
#print("Mixed Labels1: ", loss_mixed_labels_1)
results = torch.cat((loss_mixed_labels_1.view(1), results), dim=0)
#print("Results Mixed Label1: ", results)
else:
mixed_vec = torch.tensor([1-l, l], device='cuda')
loss_mixed_labels_2 = -((mixed_vec[0] * logprobs[0][0]) + (mixed_vec[1] * logprobs[0][1]))
#print("Mixed Labels2: ", loss_mixed_labels_2)
results = torch.cat((loss_mixed_labels_2.view(1), results), dim=0)
#print("Results Mixed Label2: ", results)
#Wert mit Flag
mixed_vec = torch.tensor([value_r, 1-value_r])
print("Mixed Vec: ", mixed_vec)
logprobs = torch.nn.functional.log_softmax(lg, dim=1)
print("Log:", logprobs)
#loss_mixed_labels = -torch.mul(mixed_vec, logprobs).sum()
loss_mixed_labels = -((mixed_vec[0] * logprobs[0][0]) + (mixed_vec[1] * logprobs[0][1]))
print("Loss Mixed Lables l: ", loss_mixed_labels)
results = torch.cat((loss_mixed_labels.view(1), results), dim=0)
print("Results Mixed 1: ", results)
# if value_r == l:
# #create vector: e.g. [l, 1-l]
# mixed_vec = torch.tensor([l, 1-l], device='cuda')
# #class left class, right class?
# loss_mixed_labels_1 = -((mixed_vec[0] * logprobs[0][0]) + (mixed_vec[1] * logprobs[0][1]))
# #
# #print("Mixed Labels1: ", loss_mixed_labels_1)
# results = torch.cat((loss_mixed_labels_1.view(1), results), dim=0)
# #print("Results Mixed Label1: ", results)
# else:
# mixed_vec = torch.tensor([1-l, l], device='cuda')
# loss_mixed_labels_2 = -((mixed_vec[0] * logprobs[0][0]) + (mixed_vec[1] * logprobs[0][1]))
# #print("Mixed Labels2: ", loss_mixed_labels_2)
# results = torch.cat((loss_mixed_labels_2.view(1), results), dim=0)
# #print("Results Mixed Label2: ", results)
print("ALL BATCH Results: ", results)
batch_loss = results.mean() #compute average
#print("Batch Loss: ", batch_loss)
return batch_loss
#noch in eine if schleife (beide mixed labels)
#evt. nicht immer torch.cat aufrufen
#Matrix -> 2 dim
#pro Zeile Label wenn gleich -> Standard instanz
# wenn ungleich mixed
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment