RuntimeError: The expanded size of the tensor (64) must match the existing size (256) at non-singleton dimension 0. Target sizes: [64, 2]. Tensor sizes: [256, 2]

from future import print_function

import os
import argparse

import torch
import numpy as np
from torch.optim import Adam, lr_scheduler, SGD
from torch.nn import functional as F

import data
import utils
import metrics
import config_bayesian as cfg
from models.BayesianModels.Bayesian3Conv3FC import BBB3Conv3FC
from models.BayesianModels.BayesianAlexNet import BBBAlexNet
from models.BayesianModels.BayesianLeNet import BBBLeNet
from models.BayesianModels.BayesianVGG import BBBVGGNet

CUDA settings

device = torch.device(“cuda:0” if torch.cuda.is_available() else “cpu”)

device = torch.device(“cpu”)

def getModel(net_type, inputs, outputs, priors, layer_type, activation_type):
if net_type == ‘lenet’:
return BBBLeNet(outputs, inputs, priors, layer_type, activation_type)
elif net_type == ‘alexnet’:
return BBBAlexNet(outputs, inputs, priors, layer_type, activation_type)
elif net_type == ‘3conv3fc’:
return BBB3Conv3FC(outputs, inputs, priors, layer_type, activation_type)
elif net_type == ‘VGG’:
return BBBVGGNet(outputs, inputs, priors, layer_type, activation_type)
else:
raise ValueError(‘Network should be either [LeNet / AlexNet / 3Conv3FC’)

def train_model(net, optimizer, criterion, trainloader, num_ens=1, beta_type=0.1, epoch=None, num_epochs=None):
net.train()
training_loss = 0.0
accs = []
kl_list = []
for i, (inputs, labels) in enumerate(trainloader, 1):

    optimizer.zero_grad()

    inputs, labels = inputs.to(device), labels.to(device)
    print(inputs.shape[0])
    outputs = torch.zeros(inputs.shape[0], net.num_classes, num_ens).to(device)

    kl = 0.0
    for j in range(num_ens):
        net_out, _kl = net(inputs)
        kl += _kl
        outputs[:, :, j] = F.log_softmax(net_out, dim=1)

    kl = kl / num_ens
    kl_list.append(kl.item())
    log_outputs = utils.logmeanexp(outputs, dim=2)

    beta = metrics.get_beta(i - 1, len(trainloader), beta_type, epoch, num_epochs)
    loss = criterion(log_outputs, labels, kl, beta)
    loss.backward()
    optimizer.step()

    accs.append(metrics.acc(log_outputs.data, labels))
    training_loss += loss.cpu().data.numpy()
return training_loss / len(trainloader), np.mean(accs), np.mean(kl_list)

def validate_model(net, criterion, validloader, num_ens=1, beta_type=0.1, epoch=None, num_epochs=None):
“”“Calculate ensemble accuracy and NLL Loss”""
net.train()
valid_loss = 0.0
accs = []

for i, (inputs, labels) in enumerate(validloader):
    inputs, labels = inputs.to(device), labels.to(device)
    outputs = torch.zeros(inputs.shape[0], net.num_classes, num_ens).to(device)
    kl = 0.0
    for j in range(num_ens):
        net_out, _kl = net(inputs)
        kl += _kl
        outputs[:, :, j] = F.log_softmax(net_out, dim=1).data

    log_outputs = utils.logmeanexp(outputs, dim=2)

    beta = metrics.get_beta(i - 1, len(validloader), beta_type, epoch, num_epochs)
    valid_loss += criterion(log_outputs, labels, kl, beta).item()
    accs.append(metrics.acc(log_outputs, labels))

return valid_loss / len(validloader), np.mean(accs)

def run(dataset, net_type):
# Hyper Parameter settings
layer_type = cfg.layer_type
activation_type = cfg.activation_type
priors = cfg.priors

train_ens = cfg.train_ens
valid_ens = cfg.valid_ens
n_epochs = cfg.n_epochs
lr_start = cfg.lr_start
num_workers = cfg.num_workers
valid_size = cfg.valid_size
batch_size = cfg.batch_size
beta_type = cfg.beta_type

trainset, testset, inputs, outputs = data.getDataset(dataset)
train_loader, valid_loader, test_loader = data.getDataloader(
    trainset, testset, valid_size, batch_size, num_workers)
net = getModel(net_type, inputs, outputs, priors, layer_type, activation_type).to(device)
# net.load_state_dict(torch.load(weight_path))

ckpt_dir = f'checkpoints/{dataset}/bayesian'
ckpt_name = f'checkpoints/{dataset}/bayesian/model_{net_type}_{layer_type}_{activation_type}.pt'

if not os.path.exists(ckpt_dir):
    os.makedirs(ckpt_dir, exist_ok=True)

criterion = metrics.ELBO(len(trainset)).to(device)
optimizer = Adam(net.parameters(), lr=lr_start)
lr_sched = lr_scheduler.ReduceLROnPlateau(optimizer, patience=6, verbose=True)
valid_loss_max = np.Inf
for epoch in range(n_epochs):  # loop over the dataset multiple times

    train_loss, train_acc, train_kl = train_model(net, optimizer, criterion, train_loader, num_ens=train_ens,
                                                  beta_type=beta_type, epoch=epoch, num_epochs=n_epochs)
    valid_loss, valid_acc = validate_model(net, criterion, valid_loader, num_ens=valid_ens, beta_type=beta_type,
                                           epoch=epoch, num_epochs=n_epochs)
    lr_sched.step(valid_loss)

    print(
        'Epoch: {} \tTraining Loss: {:.4f} \tTraining Accuracy: {:.4f} \tValidation Loss: {:.4f} \tValidation Accuracy: {:.4f} \ttrain_kl_div: {:.4f}'.format(
            epoch, train_loss, train_acc, valid_loss, valid_acc, train_kl))

    # save model if validation accuracy has increased
    if valid_loss <= valid_loss_max:
        print('Validation loss decreased ({:.6f} --> {:.6f}).  Saving model ...'.format(
            valid_loss_max, valid_loss))
        torch.save(net.state_dict(), ckpt_name)
        valid_loss_max = valid_loss

if name == ‘main’:
parser = argparse.ArgumentParser(description=“PyTorch Bayesian Model Training”)
parser.add_argument(’–net_type’, default=‘VGG’, type=str, help=‘model’)
parser.add_argument(’–dataset’, default=‘crack’, type=str, help=‘dataset = ‘dataset’)
# parser.add_argument(’–weights_path’, default=‘checkpoints/crack/bayesian/model_VGG_lrt_softplus.pt’, type=str,
# help=‘weights for model’)
args = parser.parse_args()

run(args.dataset, args.net_type)

When i try to run the code above for other than 64*64 image size …I get the following error:

File “/home/sharif/PyTorch-BayesianCNN-master/main_bayesian.py”, line 156, in
run(args.dataset, args.net_type)
File “/home/sharif/PyTorch-BayesianCNN-master/main_bayesian.py”, line 130, in run
train_loss, train_acc, train_kl = train_model(net, optimizer, criterion, train_loader, num_ens=train_ens,
File “/home/sharif/PyTorch-BayesianCNN-master/main_bayesian.py”, line 57, in train_model
outputs[:, :, j] = F.log_softmax(net_out, dim=1)
RuntimeError: The expanded size of the tensor (64) must match the existing size (256) at non-singleton dimension 0. Target sizes: [64, 2]. Tensor sizes: [256, 2]

My best guess is that you are using a view or reshape operation inside the model’s forward, which might be using hard-coded and wrong shapes.
Based on the error message it seems that you are changing the batch size, so I guess in your code a similar operation such as:

x = x.view(-1, 2)

might be used while:

x = x.view(x.size(0), -1)

would be the recommended way.

In any case, this change might just yield another shape mismatch if you haven’t written the model in a way it can accept variable spatial input shapes. Could you describe a bit more how the model is handling these different spatial shapes (e.g. via adaptive pooling layers)?