I am fine tuning a pre-trained fcn resnet101 segmentation model with my own data set and would like to change the input size to something like (750, 1000) instead of (224, 224).
My understanding is that this model was originally trained on images that are 224 by 224, but I would like to keep the size of my images and its mask 750 by 1000 because I would like to get a more accurate, higher level of detail from the segmentation. If I simply shrink my images to 224 by 224 the images will become too pixelated, and the segmented part will most likely lose some information that I care about such as the shape/contour of the segment.
I am following this tutorial and this is the part where I see input size is set:
def initialize_model(model_name, num_classes, feature_extract, use_pretrained=True):
# Initialize these variables which will be set in this if statement. Each of these
# variables is model specific.
model_ft = None
input_size = 0
if model_name == "resnet":
""" Resnet18
"""
model_ft = models.resnet18(pretrained=use_pretrained)
set_parameter_requires_grad(model_ft, feature_extract)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, num_classes)
input_size = 224
So if I want to make the input size 750 by 1000, would I simply change the value of the input_size variable to 750, 100 (so input_size = 750,1000)?
Would I have to change any other parameters besides input_size so I don’t get errors when fine tuning?
And finally are there any potential negative consequences of fine tuning a model with a different, bigger input size than the original pretrained model? I think a potential one might be that it would take more time and GPU power to fine tune on bigger size images because when I used the the pretrained model out of the box to evaluate images that were about 750 by 1000 it took significantly more time and GPU resources to predict the segments, but overall I got decent results on those larger images. I am more concerned about negative effects like loosing accuracy, not so much loosing speed.