Learner: predict_with_targs vs. predict - what's the difference?

Part 2 & Alumni (2018)
1

I see the learner has 2 predict functions. predict_with_targs & predict. What is the difference between the two?

Follow up question while I was trying to understand that myself opening the learner.py: Is this function recursively calling itself? I know it should be calling the model.py’s predict function at the end.

def predict(self, is_test=False, use_swa=False):
    dl = self.data.test_dl if is_test else self.data.val_dl
    m = self.swa_model if use_swa else self.model
    return predict(m, dl)
1 Like
2

It’s not recursively calling itself, because in Python to refer to the method you’d need to use self.predict. (I know in other languages like Java that’s not the case, which is probably the source of the confusion). Instead it’s using the function predict that came from the imports.

4 Likes
3

predict gives you only the predictions. predict_with_targs gives you both predictions and true Labels. That you can then use to measure Accuracy or feed to confusion matrix. Otherwise the code is the same -

def predict_with_targs(m, dl):
    preda,targa = predict_with_targs_(m, dl)
    return to_np(torch.cat(preda)), to_np(torch.cat(targa))


def predict(m, dl):
    preda,_ = predict_with_targs_(m, dl)
    return to_np(torch.cat(preda))
11 Likes
4

That’s very helpful, Thank you William!

5

That helps, Thank you Ramesh!

6

Weirdly, I am receiving two different length np’s from predict_with_targs

I am using predict_with_targs on the IMDB sentiment classification from lesson 10:
log_preds, y = learn.predict_with_targs()

log_preds.shape is (39780, 2)
y.shape is (2699424,)

What am I doing wrong?

Also including my model summary below:

<bound method Learner.summary of SequentialRNN(
  (0): MultiBatchRNN(
(encoder): Embedding(60002, 400, padding_idx=1)
(encoder_with_dropout): EmbeddingDropout(
  (embed): Embedding(60002, 400, padding_idx=1)
)
(rnns): ModuleList(
  (0): WeightDrop(
    (module): LSTM(400, 1150)
  )
  (1): WeightDrop(
    (module): LSTM(1150, 1150)
  )
  (2): WeightDrop(
    (module): LSTM(1150, 400)
  )
)
(dropouti): LockedDropout(
)
(dropouths): ModuleList(
  (0): LockedDropout(
  )
  (1): LockedDropout(
  )
  (2): LockedDropout(
  )
)
  )
  (1): PoolingLinearClassifier(
(layers): ModuleList(
  (0): LinearBlock(
    (lin): Linear(in_features=1200, out_features=50, bias=True)
    (drop): Dropout(p=0.2)
    (bn): BatchNorm1d(1200, eps=1e-05, momentum=0.1, affine=True)
  )
  (1): LinearBlock(
    (lin): Linear(in_features=50, out_features=2, bias=True)
    (drop): Dropout(p=0.1)
    (bn): BatchNorm1d(50, eps=1e-05, momentum=0.1, affine=True)
  )
)
  )
)>
7

Sounds like you might have language model data, instead of classification data.

2 Likes
8

Thanks, that was it! A bug in my code

57%20PM
Screen Shot 2018-05-30 at 4.28.57 PM.png750×718 45.8 KB

9

Hi, sorry to bother you but could you share the code you used to create the confusion matrix for the ULMFiT results? Thanks!

10

@cayman I don’t have the code handy with me now, but if you look at lesson 2 of fast.ai or the notebook that goes with it you will find it:

cm = confusion_matrix(val_classes, preds)
plot_confusion_matrix(cm, val_batches.class_indices)