np.ndArray in tensor

Part 1 (2018) Beginner (2018)
1

Hi everyone, I am trying to run Resnet18 on STL_10, but get the following error message:

RuntimeError: can’t convert a given np.ndarray to a tensor - it has an invalid type. The only supported types are: double, float, int64, int32, and uint8.

On the face of it, my arrays look ok shape-wise

image
image.png739×167 7.13 KB

I pass them through, all fine
arch = resnet18 #start off with this, and see how we go
data = ImageClassifierData.from_arrays(PATH, train, valid, tfms=tfms)
learn = ConvLearner.pretrained(arch, data, precompute=True)
learn.lr_find()

but then on learn.lr_find():

~/fastai/courses/dl1/fastai/dataloader.py in get_tensor(batch, pin)
 25 def get_tensor(batch, pin):
 26     if isinstance(batch, (np.ndarray, np.generic)):
---> 27         batch = torch.from_numpy(batch).contiguous()
 28         return batch.pin_memory() if pin else batch
 29     elif isinstance(batch, string_classes): return batch

RuntimeError: can't convert a given np.ndarray to a tensor - it has an invalid type. The only supported types are: double, float, int64, int32, and uint8.

Has anyone else seen this or know what I need to do? How can I tell which is ending up as an ndarray, and how can I change it an array? Thanks!

2

Type train_X.dtype, it will show you what the datatype of the numpy array. There are some numpy array types that are not compatible with PyTorch. You will need to say train_X.astype(np.float32) or something similar

1 Like
3

I casted it to int64 and all is good! The colours of the images went funny when casted it to float. Thanks for your help :slight_smile:

4

So I’m now getting this:

error                                     Traceback (most recent call last)
<ipython-input-95-0ffabd5bbedd> in <module>()
      2 data = ImageClassifierData.from_arrays(PATH, train, valid, tfms=tfms)
      3 learn = ConvLearner.pretrained(arch, data, precompute=False)
----> 4 learn.lr_find()

~/fastai/courses/dl1/fastai/transforms.py in scale_min(im, targ)
     13     ratio = targ/min(r,c)
     14     sz = (scale_to(c, ratio, targ), scale_to(r, ratio, targ))
---> 15     return cv2.resize(im, sz)
     16 
     17 def zoom_cv(x,z):

error: /io/opencv/modules/imgproc/src/imgwarp.cpp:3361: error: (-215) func != 0 in function resize

Things I’ve tried -

  1. I’ve checked for any null values, negative values, or very large values in my data array, but there aren’t any
  2. Catching the exception and returning an empty array ( this just moves the problem down the line)
  3. Changing to array to different dtypes, which hasn’t seemed to have any effect

Anybody have any ideas?

5

Can you try changing it to np.uint8. If these are images, the values range from 0-255 and uint8 or int16 would be enough.

6

will do

7

I’m now getting this when I run fit! It looks like it’s not expecting the ByteTensor, just working out what that corresponds to…
TypeError: CudaClassNLLCriterion_updateOutput received an invalid combination of arguments - got (int, torch.cuda.FloatTensor, torch.cuda.ByteTensor, torch.cuda.FloatTensor, bool, NoneType, torch.cuda.FloatTensor, int), but expected (int state, torch.cuda.FloatTensor input, torch.cuda.LongTensor target, torch.cuda.FloatTensor output, bool sizeA`verage, [torch.cuda.FloatTensor weights or None], torch.cuda.FloatTensor total_weight, int ignore_index)

8

use np.float32. When you need to plot it typecast to np.uint8. I have seen that pattern followed in many places.

1 Like
9

I gave that a go:
train_x = np.float32(train_x)
test_x = np.float32(test_x)
train_y = np.uint32(train_y)
test_y = np.uint32(test_y)
arch = resnet18 #start off with this, and see how we go
data = ImageClassifierData.from_arrays(PATH, train, valid, tfms=tfms)
learn = ConvLearner.pretrained(arch, data, precompute=False)
learn.lr_find()

~/src/anaconda3/envs/fastai/lib/python3.6/site-packages/torch/nn/_functions/thnn/auto.py in forward(ctx, input, target, *args)
45 output = input.new(1)
46 getattr(ctx._backend, update_output.name)(ctx._backend.library_state, input, target,
—> 47 output, *ctx.additional_args)
48 return output
49

now it’s just complaining about a Float instead of a long. I tried turning the y’s to uints, but that didn’t seem to do anything
TypeError: CudaClassNLLCriterion_updateOutput received an invalid combination of arguments - got (int, torch.cuda.FloatTensor, torch.cuda.FloatTensor, torch.cuda.FloatTensor, bool, NoneType, torch.cuda.FloatTensor, int), but expected (int state, torch.cuda.FloatTensor input, torch.cuda.LongTensor target, torch.cuda.FloatTensor output, bool sizeAverage, [torch.cuda.FloatTensor weights or None], torch.cuda.FloatTensor total_weight, int ignore_index)

1 Like
10

It might be complaining about your train_y dtype. Change that to np.int32 instead. Also, if you take a look at Lesson_1 notebook, it has X as uint8 and y as np.int64. If this doesn’t work, can you put your notebook in gist.github.com to take a look?

11

Ahh good point. I’ll give that a crack first thing tomorrow. Thanks again for your help :smiley:

12

It removes that, but creates
error: /io/opencv/modules/imgproc/src/imgwarp.cpp:3361: error: (-215) func != 0 in function resize
I think I must have an aloof image

Here’s the gist

13

@sjdlloyd - Figured it out.
TLDR: Basically the input images needs to be uint8 and Target needs to be int64.

Both train_x, and train_y was loaded as np.uint8. So you had to change just the train_y to int64 (that maps to torch.LongTensor). Also, the y was starting at 1, you will have to start the index at 0.

Details in the updated gist here

1 Like
14

Ah that’s done it! Amazing, thank you so much :smile: ; I’ve learnt a lot about Torch Tensors and array manipulation from this