Dog Breed Identification challenge

Part 1 (2018)
325

This was in notebook example shared by Jeremy. Because I was using nasnet and resnext for Dog breeds, I noticed the difference. Simple as that.

326

.
:cowboy_hat_face:

327

Did you get a chance to try nasnet on dog breed?

328

No, I didnā€™t have the time to fix the code for NasNet and focused on starting the Seedlings competition instead, Iā€™d like to test a few things out before the DHS-TSA competition first round ends.

329

facing this issue ,does anybody know how to resolve it

image
image.png1366Ɨ391 19.5 KB

330

The definition of TTA() has changed - the notebooks have been updated to show the new usage.

331

iā€™m sorry,please spare me i was not able to follow you on your comment,
but i tried this approach this time
log_preds,y=learn.TTA(n_aug=4, is_test=False)
probs=np.mean(np.exp(log_preds),0)
accuracy(log_preds,y), metrics.log_loss(y,probs)
but results were same

332

Double check this lineā€¦

333

sorry, i didnā€™t get you

334

could anyone exactly point out which part of my code is wrong and the immediate step i should take to correct my code
please iā€™m newbie in this domain.

335

Hi @naveenmanwani,

Just wanted to tell you something which I learnt beacuse of THE amazingly awesome people in this forum,

Itā€™s high time to give up your fear at looking the code which just works out of the box and understanding it in pieces and then connecting the dotsā€¦
I had got rid of that fear because of this Forum

So if I can, anyone canā€¦

1 Like
336 
def TTA(self, n_aug=4, is_test=False):
        Args:
            n_aug: a number of augmentation images to use per original image
            is_test: indicate to use test images; otherwise use validation images
        Returns:
            (tuple): a tuple containing:
                log predictions (numpy.ndarray): log predictions (i.e. `np.exp(log_preds)` will return probabilities)
                targs (numpy.ndarray): target values when `is_test==False`; zeros otherwise.

So log_preds variable contains the log prediction.
y variable contains the target values in your case.

Accuracy Code

def accuracy(preds, targs):
    preds = np.argmax(preds, axis=1)
    return (preds==targs).mean()

And this one is from sklearn


def log_loss(y_true, y_pred, eps=1e-15, normalize=True, sample_weight=None,
             labels=None):
    Parameters
    ----------
    y_true : array-like or label indicator matrix
        Ground truth (correct) labels for n_samples samples.
    y_pred : array-like of float, shape = (n_samples, n_classes) or (n_samples,)
        Predicted probabilities, as returned by a classifier
    `Returns
    -------
    loss : float

Hope it helps nowā€¦

337

Hey guys,
How is there suddenly a lot of guys in top 20 of dog breed identification challenge ? Some people even jumped 300+ places. Any ideas how in last 2-3 days so many people are jumping so high ?

338

well,i try my level best in doing or rather implementing what you have advice me to do.and i relly appreciate your effort in putting these scripts for me
so ,if i understood you correctly may be iā€™m wrong in line 3 for putting accuracy(log_preds,y), metrics.log_loss(y,probs)
because iā€™m already doing mean of the log_preds .
so ,i should remove accuracy from the last line

339

Well letā€™s reveal the answer as per as I know,

accuracy(probs,y)
is the correct version to call
Can someone just confirm?

@jamesrequa(sorry for @) but is it correct now?

1 Like
340

sorry my bad because i was overlooking something i converted this small problem into a big one
jeremy clearly mentioned the definition of TTA() has been updated .so change things accordingly
in the first notebook accuracy was(log_preds,y)
but in the updated notebook it is accuracy(probs, y)

1 Like
342

Yes!

@naveenmanwani Sorry I didnā€™t tell you directly the answer only told you where to look, I think you can learn much better once you ā€œdiscoverā€ the solution on your own! Like @ecdrid suggested the more you look at the code and write code the more it will all start to make sense :slight_smile: I know this because I started as a complete beginner (to DL and programming) just 1 year ago myselfā€¦

6 Likes
344

After downloading the data into aws instance from kaggle how to group the images into according to their classes.

345
  • to use from_csv method - it assumes you have all data in a single folder
  • if you still need to move your stuff around bash scripting is your friend. take a look at these commands and if you got stuck let us know:
    • python - os.system
    • bash - mv path/where/your/file/is.jpg path/where/you/want/your/file/to/be.jpg
2 Likes
346

finally after troubling you all with silly question .i did my submission though i got 80 rank which i know way low according to the standard which is set by you all.but iā€™m happy and now after my first submission i have become more greedy.so iā€™ll to make it better
image

5 Likes