Tabular Learner and Embeddings Sizes

(Andrea de Luca) #1

Hi. While using the tabular learner, we can let fastai decide the appropriate sizes or use emb_sz to force a desired dimension.

These were my dimensions as established by fastai as defaults:

learn.data.get_emb_szs()
>[(3, 3), (553, 55), (119, 23), (4, 3), (3, 3)]

Now, I deliberately forced the second (553x55) to 10:

learn = tabular_learner(data, layers=[1024,192,15], metrics=accuracy, 
                        emb_szs={'Ticket': 10}, 
                        emb_drop=0.3, ps=0.75, callback_fns=ShowGraph)

But nothing has changed:

learn.data.get_emb_szs()
>[(3, 3), (553, 55), (119, 23), (4, 3), (3, 3)]

Could someone check if I’m doing mistakes or it’s the library which actually ignores emb_szs?

Thanks.

0 Likes