I am implementing a binary classification task with a self distillation based model which has been known to increase accuracy. I have four blocks, I compute a loss between the final predictions and labels and loss between intermediate predictions and final prediction. I am implementing everything except the loss from hints as shown in the below diagram.
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.model_zoo as model_zoo
class ResNet(nn.Module):
def __init__(self, block, layers, num_classes, KaimingInit=False):
self.inplanes = 16
super(ResNet, self).__init__()
self.conv1 = nn.Conv2d(1, 16, kernel_size=7, stride=2, padding=3, bias=False)
self.bn1 = nn.BatchNorm2d(16)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 16, layers[0])
self.layer2 = self._make_layer(block, 32, layers[1], stride=2)
self.layer3 = self._make_layer(block, 64, layers[2], stride=2)
self.layer4 = self._make_layer(block, 128, layers[3], stride=2)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.middle_fc1 = nn.Linear(16* block.expansion, 2)
self.middle_fc2=nn.Linear(32* block.expansion, 2)
self.middle_fc3=nn.Linear(64* block.expansion, 2)
self.classifier = nn.Linear(128 * block.expansion, 2)
if KaimingInit == True:
print('Using Kaiming Initialization.')
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
for m in self.modules():
if isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def _make_layer(self, block, planes, blocks, stride=1):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(self.inplanes, planes * block.expansion,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(planes * block.expansion),
)
layers = []
layers.append(block(self.inplanes, planes, stride, downsample))
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(block(self.inplanes, planes))
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
x = self.layer1(x)
# middle_output1 = torch.flatten(middle_output1, 1)
middle_output1=self.avgpool(x).view(x.size()[0], -1)
middle_output1 = self.middle_fc1(middle_output1)
x = self.layer2(x)
middle_output2=self.avgpool(x).view(x.size()[0], -1)
middle_output2 = self.middle_fc2(middle_output2)
x = self.layer3(x)
middle_output3=self.avgpool(x).view(x.size()[0], -1)
middle_output3 = self.middle_fc3(middle_output3)
x = self.layer4(x)
x = self.avgpool(x).view(x.size()[0], -1)
out = self.classifier(x)
return middle_output1,middle_output2,middle_output3,out,x
I tried working with CrossEntropyLoss with two output logits. I also tried working with BCEWithLogitsLoss with one output logit. Both are giving me accuracy for each epoch around 50 %.
import torch
import torchaudio
from torch import nn
from torch.utils.data import DataLoader
import numpy as np
BATCH_SIZE = 128
EPOCHS = 20
LEARNING_RATE = 0.001
softmax = nn.Softmax(dim=1)
def create_data_loader(train_data, batch_size):
train_dataloader = DataLoader(train_data, batch_size=batch_size)
return train_dataloader
def kd_loss_function(output, target_output):
temperature=3
output = output / temperature
output_log_softmax = torch.log_softmax(output, dim=1)
loss_kd = -torch.mean(torch.sum(output_log_softmax * target_output, dim=1))
return loss_kd
def train_single_epoch(model, data_loader, loss_fn, optimiser, device):
temperature=3
alpha=0.1
correct=0
total=0
for input, target,domain in data_loader:
input = input.to(device)
# label = target.clone().to(device)
# target = torch.tensor(target, dtype=torch.float).to(device)
target=target.to(device)
doamin=domain.to(device)
# calculate loss
middle_output1,middle_output2,middle_output3,prediction,features = model(input)
temp4 = prediction / temperature
temp4 = torch.softmax(temp4, dim=1)
probabilities = softmax(prediction)
predicted_classes = torch.argmax(probabilities, dim=1)
correct += (predicted_classes == target).sum().item()
total += target.size(0)
loss1by4 = kd_loss_function(middle_output1, temp4.detach())* (temperature**2)
loss2by4 = kd_loss_function(middle_output2, temp4.detach()) * (temperature**2)
loss3by4 = kd_loss_function(middle_output3, temp4.detach()) * (temperature**2)
loss = loss_fn(prediction, target)
total_loss = (1 - alpha) * (loss1by4 + loss2by4 + loss3by4) + alpha *loss
optimiser.zero_grad()
total_loss.backward()
optimiser.step()
print(f"loss: {loss.item()}")
# accuracy = 100 * correct / 30000
accuracy = correct / total
print(" Accuracy = {}".format(accuracy))
def train(model, data_loader, loss_fn, optimiser, device, epochs):
model.train()
for i in range(epochs):
print(f"Epoch {i+1}")
features,domains,labels=train_single_epoch(model, data_loader, loss_fn, optimiser, device)
print("---------------------------")
print("Finished training")
model = eca_resnet18().to(device)
loss_fn = torch.nn.CrossEntropyLoss()
optimiser = torch.optim.Adam(model.parameters(),
lr=LEARNING_RATE)
train(model, train_dataloader, loss_fn, optimiser, device, EPOCHS)