How inference is done after training model?

[sujit420]

There is no such example or sample code given for inference.
Supposing, I have trained my model, how do I use it for prediction?
@ZhuFengdaaa can u help me
Thanks in advance

[kevgeo]

Even I am trying to do the same thing. My doubt is whether there is a simpler way to load the layers of the model? Since the model is saved using torch.save(model.state_dict(), model_path), only the trained parameters are saved; so we will have to load the layers first and then load the trained parameters.

But I am confused as to how to load the layers of the model for simple testing on an image. Below is permalink where the model where the model layers are loaded.

bottom-up-attention-vqa/main.py

Lines 37 to 39 in f51446b

	constructor = 'build_%s' % args.model
	model = getattr(base_model, constructor)(train_dset, args.num_hid).cuda()
	model.w_emb.init_embedding('data/glove6b_init_300d.npy')

Is it necessary to use train_dset in line 2 to construct the model? I am not able to understand why dataset is needed? Is there a simpler way to load the model layers for testing on a single image?

@hengyuan-hu @ZhuFengdaaa Hope you guys could help. Thanks in advance.

[ZhuFengdaaa]

like this line, this implementation uses number of tokens to construct the embedding layer.