I have looked at some similar posts but couldn't find any solution to my specific problem.
I'm running this code :
class MyNet(nn.Module):
def __init__(self,input_dim):
super(MyNet, self).__init__()
self.conv1 = nn.Conv2d(input_dim, nChannel, kernel_size=3, stride=1, padding=1 )
self.bn1 = nn.BatchNorm2d(nChannel)
self.conv2 = nn.ModuleList()
self.bn2 = nn.ModuleList()
for i in range(nConv-1):
self.conv2.append( nn.Conv2d(nChannel, nChannel, kernel_size=3, stride=1, padding=1 ) )
self.bn2.append( nn.BatchNorm2d(nChannel) )
self.conv3 = nn.Conv2d(nChannel, nChannel, kernel_size=1, stride=1, padding=0 )
self.bn3 = nn.BatchNorm2d(nChannel)
def forward(self, x):
x = self.conv1(x)
x = F.relu( x )
x = self.bn1(x)
for i in range(nConv-1):
x = self.conv2[i](x)
x = F.relu( x )
x = self.bn2[i](x)
x = self.conv3(x)
x = self.bn3(x)
return x
img_list = sorted(glob.glob(input_path+'0.png'))
im = cv2.imread(img_list[0])
# train
model = MyNet( im.shape[2] )
if use_cuda:
model.cuda()
model.train()
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.9)
for batch_idx in range(maxIter):
print('Training started. '+str(datetime.datetime.now())+' '+str(batch_idx+1)+' / '+str(maxIter))
for im_file in range(int(len(img_list)/batch_size)):
for loop in tqdm.tqdm(range(maxUpdate)):
im = []
for batch_count in range(batch_size):
# load image
resized_im = cv2.imread(img_list[batch_size*im_file + batch_count])
print(resized_im.shape)
resized_im = cv2.resize(resized_im, dsize=(224, 224))
resized_im = resized_im.transpose( (2, 0, 1) ).astype('float32')/255.
print(resized_im.shape)
im.append(resized_im)
data = torch.from_numpy( np.array(im) )
print(data.shape)
if use_cuda:
data = data.cuda()
data = Variable(data)
# forwarding
optimizer.zero_grad()
output = model( data )
I have no problem while training the model here are the different shapes the prints are giving :
(200, 200, 3) for the first resized_im
(3, 224, 224) for the transposed resized_im
torch.Size([1, 3, 224, 224]) for the tensor data
But I'm not able to run it with this part of slightly modified code where reference_img is a pre-loaded image :
# Load image
print(reference_img[batch_size*img_idx + batch_count].shape)
resized_img = cv2.resize(reference_img[batch_size*img_idx + batch_count], dsize=(224, 224))
print(resized_img.shape)
resized_img = resized_img.transpose((2, 0, 1)).astype('float32')/255.
print(resized_img.shape)
img.append(resized_img)
data = torch.from_numpy(np.array(img))
print(data.shape)
I got this error :
RuntimeError: Given groups=1, weight of size [100, 200, 3, 3], expected input[1, 3, 224, 224] to have 200 channels, but got 3 channels instead
Here are the shapes in the second case :
(200, 200, 3) for the reference_img[batch_size*img_idx + batch_count]
(224, 224, 3) for the resized_img
(3, 224, 224) for the transposed resized_img
torch.Size([1, 3, 224, 224]) for the tensor data
Thank you in advance for your answers and sorry if I made some obvious mistakes
My guess is that this line model = MyNet( im.shape[2]) is causing your issue. Your 2D conv layers expect an input of size [_,200,_,_], because your input_dim for the conv layer is set by the above line. Print out the shape of im and verify it is as expected.