Training a Inception v1 for tenwsorflow 1.x with a previous version of fastai

fastai
1

Hi,

I’ve to train an Inception-V1 and save one checkpoint in tensorflow 1 format.

I’ve seen that it has 3 softmax outputs.

I’m going to to use a gloabl_loss= w0loss0(softmax0)+ w1loss1(softmax1)+ w2*loss2(softmax2)
Where w0, w1 and w2 are non learnable weightes, like w0=0.2, w1=0.3 and w2=0.5

If it didn’t work, I think I’d try to start training until softmax-0, and freeze the network above this point (not affecting this output) And then unfreeze to train until softmax-1. And then softmax-2

I had thought in using a previous version of fastai, as far as I remember, it was initially made for tf.

Could you recommend me a fastai version, and an exeample to train my Inception-V1?

Cheers,

2

I trained a pytorch model in fastai v1.x and then created a onnx binary that I can directly use in Tensorflow.

1 Like
3

thanks a lot @kushaj , I think I could do the same training in Keras and exporting to tf