Data Loader seems to be pulling the same image

Hi. It seems to me that my data loader is pulling the same image. It’s transforming it differently but it’s still the same image.

I would expect it to pull 10 different images with each batch of 10.

Below is the data loader:

class Inspection_Dataset(Dataset):
        df: Dataframe containing all categorical, numerical and image columns
        numerical columns: list of numerical columns
        cat_columns: list of categorical columns
        image: column containing image file name
        root_dir: column containing root directory
        def __init__(self, df, numerical_columns = None,
                     cat_columns = None,
                     image = None, 
                     root_dir = None, 
                     label = None, 
                     transform = None):
            self.df = df
            self.transform = transform
            self.image_column = image
            self.root_dir = root_dir
            self.n = df.shape[0]
            #output column
            self.label = np.array(self.df.loc[:, label])
            #cat columns
            self.cat_columns = cat_columns if cat_columns else []
            self.numerical_columns = [col for col in df[numerical_columns]]
            if self.cat_columns:
                for column in self.cat_columns:
                    df[column] = df.loc[:, column].astype('category')
                    df[column] = df[column]
                self.cat_columns = np.array(df[cat_columns]) 
                self.cat_columns = np.zeros((self.n, 1))  
            #numerical columns
            if self.numerical_columns:
                self.numerical_columns = df[self.numerical_columns].astype(np.float32).values
                self.numerical_columns = np.zeros((self.n, 1))   
        def __len__(self):
            return self.n

        def __getitem__(self, idx):
            idx = list(self.df.index)
            image =[idx, self.root_dir].values[0],
                                              self.df.loc[idx, self.image_column].values[0]))
            image = self.transform(image)

            return self.label[idx], self.numerical_columns[idx], self.cat_columns[idx], image

Below is the code I used to test the data loader/print the images.

train_data = Inspection_Dataset(train_sample,
                                numerical_columns = numerical_columns,
                                cat_columns = non_loca_cat_columns,
                                image = 'file',
                                root_dir = 'root',
                                label = 'target',
                                transform = train_transform)

train_loader = DataLoader(train_data, batch_size = 10, shuffle = True)

count = 50
for i in range(count):
    for b , (label, numericals, cats, image) in enumerate(train_loader):
print('Label:', label.numpy())

im = make_grid(image, nrow=5)  # the default nrow is 8

# Inverse normalize the images
inv_normalize = transforms.Normalize(
    mean=[-0.485/0.229, -0.456/0.224, -0.406/0.225],
    std=[1/0.229, 1/0.224, 1/0.225]
im_inv = inv_normalize(im)

# Print the images
plt.imshow(np.transpose(im_inv.numpy(), (1, 2, 0)));

What am I missing here?

Could you check the path you are using to open the image:

os.path.join(self.df.loc[idx, self.root_dir].values[0], self.df.loc[idx, self.image_column].values[0])

Could it be that .values[0] is indexing the first image only?

The image path is legit. When I take the [0] off of the values, I get the following error:

  File "<ipython-input-120-d05164b59f9d>", line 79, in <module>
    for label, numericals, cats, image in train_loader:

  File "C:\Users\JORDAN.HOWELL.GITDIR\AppData\Local\Continuum\anaconda3\envs\torch_env\lib\site-packages\torch\utils\data\", line 345, in __next__
    data = self._next_data()

  File "C:\Users\JORDAN.HOWELL.GITDIR\AppData\Local\Continuum\anaconda3\envs\torch_env\lib\site-packages\torch\utils\data\", line 385, in _next_data
    data = self._dataset_fetcher.fetch(index)  # may raise StopIteration

  File "C:\Users\JORDAN.HOWELL.GITDIR\AppData\Local\Continuum\anaconda3\envs\torch_env\lib\site-packages\torch\utils\data\_utils\", line 44, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]

  File "C:\Users\JORDAN.HOWELL.GITDIR\AppData\Local\Continuum\anaconda3\envs\torch_env\lib\site-packages\torch\utils\data\_utils\", line 44, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]

  File "<ipython-input-120-d05164b59f9d>", line 57, in __getitem__
    self.df.loc[idx, self.image_column].values))

  File "C:\Users\JORDAN.HOWELL.GITDIR\AppData\Local\Continuum\anaconda3\envs\torch_env\lib\", line 76, in join
    path = os.fspath(path)

TypeError: expected str, bytes or os.PathLike object, not numpy.ndarray

I get the same if I take the whole values[0] off.

When I put value[i] I get different pictures but the same picture per batch like the below:

I think it’s something to do with batches. When I change to batch of 1, I get different pictures each time. I’m not sure how batches works with the custom datasets and how to get them to pull correctly.

The Dataset.__getitem__ will get an index as the argument from the DataLoader to create a batch.
These indices will be in the range [0, len(dataset)-1], so you would only have to make sure to load each corresponding sample using the passed index.

I would still recommend to check the path and make sure that the idx is used to load the right path.