Poor results fine-tunned GoogLeNet, how to improve them? - deep-learning

Question

I've trained the GooogLeNet from scratch on the MNIST dataset. It achieved very good results (top-1 accuracy of 99% on test set).
Now I want to do transfer learning in order to adapt it to the FashionMNIST dataset. For that I'm doing the following:
# Loading trained model on MNIST
googlenet = torch.load('googlenet-mnist.pth')
# Freeze the network
def freeze(net):
for param in net.parameters():
param.requires_grad = False
return net
# Override all the Linear layers and initialize them
# (including the ones that produce auxiliarity logits)
def forget_FC(net):
net.aux1.fc1 = nn.Linear(in_features=net.aux1.fc1.in_features, out_features=net.aux1.fc1.out_features, bias=True)
net.aux1.fc2 = nn.Linear(in_features=net.aux1.fc2.in_features, out_features=net.aux1.fc2.out_features, bias=True)
net.aux2.fc1 = nn.Linear(in_features=net.aux2.fc1.in_features, out_features=net.aux2.fc1.out_features, bias=True)
net.aux2.fc2 = nn.Linear(in_features=net.aux2.fc2.in_features, out_features=net.aux2.fc2.out_features, bias=True)
# Override the classification layer
net.fc = nn.Sequential(
nn.Linear(num_in_features, num_in_features),
nn.Linear(num_in_features, num_in_features),
nn.Linear(num_in_features, 10))
# Initialize weights auxiliarity logits branches
torch.nn.init.trunc_normal_(net.aux1.fc1.weight, mean=0.0, std=0.01, a=-2, b=2)
torch.nn.init.trunc_normal_(net.aux1.fc2.weight, mean=0.0, std=0.01, a=-2, b=2)
torch.nn.init.trunc_normal_(net.aux2.fc1.weight, mean=0.0, std=0.01, a=-2, b=2)
torch.nn.init.trunc_normal_(net.aux2.fc2.weight, mean=0.0, std=0.01, a=-2, b=2)
# Initialize weights each Linear module in the classification layer
for module in net.fc.modules():
if isinstance(module, nn.Linear):
torch.nn.init.trunc_normal_(module.weight, mean=0.0, std=0.01, a=-2, b=2)
return net
# The training algorithm
def train(net, train_iter, test_iter, num_epochs, lr, device, plot_title, fine_tune=False):
"""Train a model with a GPU.
"""
# def init_weights(m):
# if type(m) == nn.Linear or type(m) == nn.Conv2d:
# nn.init.xavier_uniform_(m.weight)
# net.apply(init_weights)
print('training on', device)
progress = ""
net.to(device)
if fine_tune:
optimizer = torch.optim.SGD(net.fc.parameters(), lr=lr, momentum=.9)
optimizer = torch.optim.SGD(net.parameters(), lr=lr, momentum=.9)
scheduler = ReduceLROnPlateau(optimizer, 'min', patience=1, verbose=True)
loss = nn.CrossEntropyLoss()
animator = Animator(xlabel='epoch', xlim=[1, num_epochs], title=plot_title, ylim=[0, 1], figsize=(5,5),
legend=['train loss', 'train acc', 'val acc'])
timer, num_batches = d2l.Timer(), len(train_iter)
for epoch in range(num_epochs):
# Sum of training loss, sum of training accuracy, sum of top 5 training accuracy, no. of examples
metric = d2l.Accumulator(4)
net.train()
# Training
for i, (X, y) in enumerate(train_iter):
timer.start()
optimizer.zero_grad()
X, y = X.to(device), y.to(device)
# Mini-batch inference
y_hat = net(X)
# Take into account the auxiliarity logits (see link cell above)
if isinstance(y_hat, GoogLeNetOutputs):
aux_logit1, aux_logit2, y_hat = y_hat
l1 = loss(y_hat, y)
l2 = loss(aux_logit1, y)
l3 = loss(aux_logit2, y)
l = l1 + .3 * (l2 + l3)
else:
l = loss(y_hat, y)
l.backward()
optimizer.step()
# Training accuracies
with torch.no_grad():
acc_1, acc_5 = accuracy(y_hat, y)
metric.add(l * X.shape[0], acc_1, acc_5, X.shape[0])
timer.stop()
train_l = metric[0] / metric[3]
train_acc_1 = metric[1] / metric[3]
train_acc_5 = metric[2] / metric[3]
if (i + 1) % (num_batches // 5) == 0 or i == num_batches - 1:
animator.add(epoch + (i + 1) / num_batches,
(train_l, train_acc_1, None), plot_title)
# Validation, (validation loss computed when model in eval mode, is that correct?)
val_l, test_acc_1, test_acc_5 = evaluate_accuracy_gpu(net, test_iter)
scheduler.step(val_l)
animator.add(epoch + 1, (None, None, test_acc_1), plot_title)
# Un-comment to see memory consumption, modify batch size to see effects
# print(os.popen('nvidia-smi').read())
# break
progress += f"----\nEpoch {epoch}/{num_epochs}\n\ttrain loss={train_l}[{train_acc_1}]\tval loss={val_l} [{test_acc_1}]\n----"
print(progress)
print(f'loss={train_l:.3f}, train=[1-acc {train_acc_1:.3f}, 5-acc {train_acc_5:.3f}]'
f'test=[1-acc {test_acc_1:.3f}, 5-acc {test_acc_5:.3f}]')
print(f'{metric[3] * num_epochs / timer.sum():.1f} examples/sec '
f'on {str(device)}')
print(f'total training time: {timer.sum()} seconds')
With this approach 34% training accuracy is achieved. Honestly, I was expecting more close to results obtained in MNIST. What is wrong with my current approach?