Retrain the saved model

[arpittandon]

I made a multi class image classification model which does the identification of images into three catgegories. It worked great.
Now I have more similar data, but has 4 labels. I was wondering if I could use the above model( its weights) to retrain it for 4 labels. Does it sound good? If yes, how would I do that. I can certainly load the model like, but after that I am clueless.

sz=224
arch=resnet50
data = ImageClassifierData.from_paths(PATH, tfms=tfms_from_model(arch, sz))
learn = ConvLearner.pretrained(arch, data, precompute=False)
learn.load(“three_class_aug_no_16th_Slice_besT_model”)

[arpittandon]

I think I found it. I used following for reference:

Transfer learning... twice?

So, in the end, this is how my code looks like: First, transfer learning (follows notebook from lesson3), Load Resnet50 with the weights from imagenet: from fastai.vision import * from fastai import * path = Config.data_path()/‘MYPATH’ path.mkdir(parents=True, exist_ok=True) df = pd.read_csv(path/‘train_v2.csv’) tfms = get_transforms(flip_vert=True, max_lighting=0.1, max_zoom=1.05, max_warp=0.) np.random.seed(42) src = (ImageList.from_csv(path, ‘train_v2.csv’, folder=‘train-jpg’, suf…

Though, instead of
learn.model[-1][-1]=nn.Linear(in_features=512,out_features=5, bias=True)

I did learn.model[-2] = nn.Linear(in_features=512,out_features=5, bias=True), as for me the last layer was softmax

[arpittandon]

Though, I dont see any improvement. In fact I see either same level of performace or slightly worse. Is it due to less number of training data? The images are very similar though.