Concatenating a tuple of tensors using Python eval()

1

Suppose a module returns a tuple of classes and offsets at L feature layers (something like an SSD) for an object detection problem:

pred0_raw = SomeModule(params, 0)(x)
pred1_raw = SomeModule(params, 1)(x)  
pred2_raw = SomeModule(params, 2)(x)

Here, len(pred0_raw) is 2, corresponding to a tuple of (class, offset) predictions at layer L=0. class and offset are tensors of some shape [B, N, C]. I would like to concatenate them so I do this:

L=3
classes = [eval(f"pred{i}_raw[0]") for i in range(L)]  # class preds at predL_raw[0]
classes = torch.cat(classes, -2)  # classes.shape = B, N*L, C
# similarly offsets.shape = B, N*L, 4

When I execute these statements in a Jupyter cell for experimenting, it seems to work but I get a NameError: name 'pred0_raw' is not defined when I put all of that in a Module. What am I doing wrong here? Any alternative suggestions?
Thanks!

2

Alright! so this was simply due to the fact that the list comprehension puts eval in a different scope altogether! It was not a pytorch specific issue: python - eval fails in list comprehension - Stack Overflow!