fastai 2.7.16 py_0 fastai
I have looked at other posts (Post 1 & Post 2) with people having the same problem as I have, but their solution doesn’t seem to work in my case :c
I am using a pre-trained resnet50 for classifying images into 3 categories.
My program checks the models directory to see if there is an existing model. If there isn’t it will train a new one, if there is, it will use that.
When training a new model and immediately using that after, it runs just fine, and sorts the images into their respective categories. But when loading a model, which has already been trained (and previously worked just fine), it only places the images into the first category. Only on very rare occasions will it pop out a couple images into another category.
def create_new_model(trainset_path, train=True):
# Create DataLoaders from the folder structure
#print(trainset_path)
dls = make_dls(trainset_path) # This function returns a dataloader
# Create and train the model
learn = vision_learner(dls, resnet50, metrics=accuracy, pretrained=train)
if train: learn.fine_tune(3) # I don't want to train the model, if I can load one much faster.
return learn
# This function outputs a learner which is used in another script like this: "learn.predict(img)"
def get_model(trainset_path : pathlib.PosixPath):
# Path to the file
model_path_pth = trainset_path / "models/model.pth"
# If the path exists
if model_path_pth.exists():
print(f"Loading model from {model_path_pth}.")
learn = create_new_model(trainset_path, train=False)
#learn = load_learner(model_path_pth)
learn.load("model")
learn.model.eval() # <--- Here I am trying the solution from the two other threads having the same issue
else:
print(f"Creating new model at {model_path_pth}.")
learn = create_new_model(trainset_path)
learn.save("model")
return learn
I have also done learn.summary() and compared the trained version and the loaded version using meld, and the only difference in them is the location of the optimizer:
