A sincere apology about kaggle-dstl HOT 9 CLOSED

Hi, sorry for closing that issue - I didn't treat "excuse me" as contempt, and english is also not my native language. Could you please share the problem that you had? I will try to help, but I wrote this code long time ago and it is not production quality, so please excuse any errors that you get.

from kaggle-dstl.

hi,
1. I'm using raw data.Run according to the example you gave (below).
Making a submission
Train a CNN (choose number of epochs and other hyper-params running without --all):

$ ./train.py checkpoint-folder --all --hps dice_loss=10,n_epochs=70

2. There was an error(below).
Traceback (most recent call last):
File "/usr/lib/python3.5/concurrent/futures/thread.py", line 55, in run
result = self.fn(*self.args, **self.kwargs)
File "./train.py", line 306, in gen_batch
torch.from_numpy(np.array(dist_outputs)))
RuntimeError: the given numpy array has zero-sized dimensions. Zero-sized dimensions are not supported in PyTorch

3. Error location is located at train.py(Upside down count second lines are 306 lines)

def train_on_images(self, train_images: List[Image],
subsample: int=1,
square_validation: bool=False,
no_mp: bool=False):
self.net.train()
b = self.hps.patch_border
s = self.hps.patch_inner
# Extra margin for rotation
m = int(np.ceil((np.sqrt(2) - 1) * (b + s / 2)))
mb = m + b # full margin
mean_area = np.mean(
[im.size[0] * im.size[1] for im in train_images])
n_batches = int(
mean_area / (s + b) / self.hps.batch_size / subsample / 2)

def gen_batch(_):
    inputs, outputs, dist_outputs = [], [], []
    for _ in range(self.hps.batch_size):
        im, (x, y) = self.sample_im_xy(train_images, square_validation)
        if random.random() < self.hps.oversample:
            for _ in range(1000):
                if im.mask[x: x + s, y: y + s].sum():
                    break
                im, (x, y) = self.sample_im_xy(
                    train_images, square_validation)
        patch = im.data[:, x - mb: x + s + mb, y - mb: y + s + mb]
        mask = im.mask[:, x - m: x + s + m, y - m: y + s + m]
        if self.hps.needs_dist:
            dist_mask = im.dist_mask[:, x - m: x + s + m, y - m: y + s + m]

        if self.hps.augment_flips:
            if random.random() < 0.5:
                patch = np.flip(patch, 1)
                mask = np.flip(mask, 1)
                if self.hps.needs_dist:
                    dist_mask = np.flip(dist_mask, 1)
            if random.random() < 0.5:
                patch = np.flip(patch, 2)
                mask = np.flip(mask, 2)
                if self.hps.needs_dist:
                    dist_mask = np.flip(dist_mask, 2)

        if self.hps.augment_rotations:
            assert self.hps.augment_rotations != 1  # old format
            angle = (2 * random.random() - 1.) * self.hps.augment_rotations
            patch = utils.rotated(patch, angle)
            mask = utils.rotated(mask, angle)
            if self.hps.needs_dist:
                dist_mask = utils.rotated(dist_mask, angle)

        if self.hps.augment_channels:
            ch_shift = np.random.normal(
                1, self.hps.augment_channels, patch.shape[0])
            patch = patch * ch_shift[:, None, None]

        inputs.append(patch[:, m: -m, m: -m].astype(np.float32))
        outputs.append(mask[:, m: -m, m: -m].astype(np.float32))
        if self.hps.needs_dist:
            dist_outputs.append(
                dist_mask[:, m: -m, m: -m].astype(np.float32))

    return (torch.from_numpy(np.array(inputs)),
            torch.from_numpy(np.array(outputs)),
            torch.from_numpy(np.array(dist_outputs)))

self._train_on_feeds(gen_batch, n_batches, no_mp=no_mp)

from kaggle-dstl.

I see - this is indeed the same error as #2 (I noticed your message in this issue only here). Unfortunately, I don't have any other ideas besides missing input data.

Do you mind if I close this issue as a duplicate of #2?

from kaggle-dstl.

Is it missing data, or lack of data ?

from kaggle-dstl.

Training script needs training data which was provided for this competition. This repo does not contain this data.

from kaggle-dstl.

however i already download this data.
url： [https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection/data]
grid_sizes.csv.zip
sample_submission.csv.zip
sixteen_band.zip
three_band.zip
train_geojson_v3.zip
train_wkt_v4.csv.zip

from kaggle-dstl.

This data also needs to be unpacked, for an idea of the paths see here https://github.com/lopuhin/kaggle-dstl/blob/master/utils.py#L47-L53

In general, this repo is not supported, this is a code dump after the competition so you can check the general approach, but getting this to work will require either running this against your own data, or debugging issues such as the one you met.

from kaggle-dstl.

You're right, but I've unzipped it, and I hope you can “try catch"before open file or path. But your source code doesn't work,I'll have to debug. Thank you very much.

from kaggle-dstl.

set args "dist_loss=10,n_epochs=70" , the project will run.

from kaggle-dstl.