🧾 View as a summary

spicy = randoms[:12].clone()

spicy[0] *= 10000
spicy[1] /= 10000
spicy[3] = float('inf')
spicy[4] = float('-inf')
spicy[5] = float('nan')
spicy = spicy.reshape((2,6))

---

lovely

def lovely(
    t:Tensor, # Tensor of interest
    verbose:bool=False, # Whether to show the full tensor
    plain:bool=False, # Just print if exactly as before
    depth:int=0, # Show stats in depth
    color:NoneType=None, # Force color (True/False) or auto.
    show_histogram:NoneType=None, # Show the histogram: '▁▂▃▃▆█▆▃▁▁'
):

Examples

print(lovely(randoms[0]))
print(lovely(randoms[:2]))
print(lovely(randoms[:6].view(2, 3))) # More than 2 elements -> show statistics
print(lovely(randoms[:11]))           # More than 10 -> suppress data output

tensor 1.927
tensor[2] μ=1.707 σ=0.311 [1.927, 1.487]
tensor[2, 3] n=6 x∈[-2.106, 1.927] μ=0.276 σ=1.594 [[1.927, 1.487, 0.901], [-2.106, 0.678, -1.235]]
tensor[11] x∈[-2.106, 1.927] μ=0.046 σ=1.384

Weird types

print(lovely(torch.tensor([1., 2., 3.], dtype=torch.float8_e4m3fn)))
print(lovely(torch.tensor([1., 2., 3.], dtype=torch.float8_e4m3fnuz)))
print(lovely(torch.tensor([1., 2., 3.], dtype=torch.float8_e5m2)))
print(lovely(torch.tensor([1., 2., 3.], dtype=torch.float8_e5m2fnuz)))

# Note: This one does positive powers of 2 only, since it has exponent only without mantissa or sign.
print(lovely(torch.tensor([1., 2., 3., -4., 5., 6.], dtype=torch.float8_e8m0fnu)))

tensor[3] float8_e4m3fn x∈[1.000, 3.000] μ=2.000 σ=1.000 [1.000, 2.000, 3.000]
tensor[3] float8_e4m3fnuz x∈[1.000, 3.000] μ=2.000 σ=1.000 [1.000, 2.000, 3.000]
tensor[3] float8_e5m2 x∈[1.000, 3.000] μ=2.000 σ=1.000 [1.000, 2.000, 3.000]
tensor[3] float8_e5m2fnuz x∈[1.000, 3.000] μ=2.000 σ=1.000 [1.000, 2.000, 3.000]
tensor[6] float8_e8m0fnu x∈[1.000, 8.000] μ=3.833 σ=2.401 [1.000, 2.000, 4.000, 4.000, 4.000, 8.000]

The gradient

torch.manual_seed(1)


grad = torch.randn((10, 10), requires_grad=True, dtype=torch.float64)
grad_plus_one = grad+1

print(f"Before .backward:\n{lovely(grad)}\n")

# We can't access .grad of non-leaf tensors
print(f"Before .backward, non-leaf node:\n{lovely(grad_plus_one)}\n")

grad_plus_one.prod().backward()

grad.grad[0,0] = float("-inf") # type: ignore

print(f"After .backward():\n{lovely(grad)}\n")

grad.grad.zero_() # type: ignore

print(f"After .zero_() on .grad:\n{lovely(grad)}")

Before .backward:
tensor[10, 10] f64 n=100 x∈[-3.705 |▁ ▂▂▇█▇█▃▂| 2.537] μ=0.105 σ=1.066 grad=None

Before .backward, non-leaf node:
tensor[10, 10] f64 n=100 x∈[-2.705 |▁ ▂▂▇█▇█▃▂| 3.537] μ=1.105 σ=1.066 grad (non-leaf) AddBackward0

After .backward():
tensor[10, 10] f64 n=100 x∈[-3.705 |▁ ▂▂▇█▇█▃▂| 2.537] μ=0.105 σ=1.066 grad={ x∈[-1.351e-05 |▁▁▁▁  ▁█▁▁| 4.236e-06] μ=-1.145e-07 σ=2.284e-06 -Inf! }

After .zero_() on .grad:
tensor[10, 10] f64 n=100 x∈[-3.705 |▁ ▂▂▇█▇█▃▂| 2.537] μ=0.105 σ=1.066 grad={ all_zeros }

if torch.cuda.is_available():
    print(lovely(torch.tensor(1., device=torch.device("cuda:0"))))
    test_eq(str(lovely(torch.tensor(1., device=torch.device("cuda:0")))), "tensor cuda:0 1.000")

tensor cuda:0 1.000

Do we have any floating point nasties? Is the tensor all zeros?

# Statistics and range are calculated on good values only, if there are at lest 3 of them.
lovely(spicy)

tensor[2, 6] n=12 x∈[-1.605, 1.927e+04] μ=2.141e+03 σ=6.423e+03 +Inf! -Inf! NaN!

lovely(spicy, color=False)

tensor[2, 6] n=12 x∈[-1.605, 1.927e+04] μ=2.141e+03 σ=6.423e+03 +Inf! -Inf! NaN!

lovely(torch.tensor([float("nan")]*11))

tensor[11] NaN!

lovely(torch.zeros(12))

tensor[12] all_zeros

lovely(torch.randn([0,0,0], dtype=torch.float16))

tensor[0, 0, 0] f16 empty

lovely(torch.tensor([1,2,3], dtype=torch.int32))

tensor[3] i32 x∈[1, 3] μ=2.000 σ=1.000 [1, 2, 3]

torch.set_printoptions(linewidth=120)
lovely(spicy, verbose=True)

tensor[2, 6] n=12 x∈[-1.605, 1.927e+04] μ=2.141e+03 σ=6.423e+03 +Inf! -Inf! NaN!
tensor([[ 1.9269e+04,  1.4873e-04,  9.0072e-01,         inf,        -inf,         nan],
        [-4.3067e-02, -1.6047e+00, -7.5214e-01,  1.6487e+00, -3.9248e-01, -1.4036e+00]])

lovely(spicy, plain=True)

tensor([[ 1.9269e+04,  1.4873e-04,  9.0072e-01,         inf,        -inf,         nan],
        [-4.3067e-02, -1.6047e+00, -7.5214e-01,  1.6487e+00, -3.9248e-01, -1.4036e+00]])

image = torch.load("mysteryman.pt")
image[1,2,3] = float('nan')

lovely(image, depth=2) # Limited by set_config(deeper_lines=N)

tensor[3, 196, 196] n=115248 (0.4Mb) x∈[-2.118 |▂▅█▇▂▁▁▂▂▁| 2.640] μ=-0.388 σ=1.073 NaN!
  tensor[196, 196] n=38416 x∈[-2.118 |▂▃▅█▂▁▁▂▃▂| 2.249] μ=-0.324 σ=1.036
    tensor[196] x∈[-1.912 |▃▁▂█▃    ▁| 2.249] μ=-0.673 σ=0.522
    tensor[196] x∈[-1.861 |▂▂██▁    ▁| 2.163] μ=-0.738 σ=0.418
    tensor[196] x∈[-1.758 |▂▂█▂▁    ▁| 2.198] μ=-0.806 σ=0.397
    tensor[196] x∈[-1.656 |▂▆█▁▁    ▁| 2.249] μ=-0.849 σ=0.369
    tensor[196] x∈[-1.673 |▁▆█▁▁▁   ▁| 2.198] μ=-0.857 σ=0.357
    tensor[196] x∈[-1.656 |▂██▂▁▁   ▁| 2.146] μ=-0.848 σ=0.372
    tensor[196] x∈[-1.433 |▃█▅▁▁▁▁  ▁| 2.215] μ=-0.784 σ=0.397
    tensor[196] x∈[-1.279 |▆█▄▁▁▁▁▁ ▁| 2.249] μ=-0.695 σ=0.486
    tensor[196] x∈[-1.364 |▆██▂▁▁▁▁▁▁| 2.249] μ=-0.637 σ=0.539
    ...
  tensor[196, 196] n=38416 x∈[-1.966 |▁▂▅█▁▁▁▁▂▁| 2.429] μ=-0.274 σ=0.973 NaN!
    tensor[196] x∈[-1.861 |▂▁▁█▁    ▁| 2.411] μ=-0.529 σ=0.556
    tensor[196] x∈[-1.826 |▁▁▁█     ▁| 2.359] μ=-0.562 σ=0.473
    tensor[196] x∈[-1.756 |▂▁▃█▁    ▁| 2.376] μ=-0.622 σ=0.459 NaN!
    tensor[196] x∈[-1.633 |▂▂█▄     ▁| 2.429] μ=-0.664 σ=0.430
    tensor[196] x∈[-1.651 |▂▂█▄     ▁| 2.376] μ=-0.669 σ=0.399
    tensor[196] x∈[-1.633 |▂▃█▃▁    ▁| 2.376] μ=-0.701 σ=0.391
    tensor[196] x∈[-1.563 |▂▃█▂▁    ▁| 2.429] μ=-0.670 σ=0.380
    tensor[196] x∈[-1.475 |▂▄█▁▁    ▁| 2.429] μ=-0.616 σ=0.386
    tensor[196] x∈[-1.511 |▁▄█▁▁▁   ▁| 2.429] μ=-0.593 σ=0.399
    ...
  tensor[196, 196] n=38416 x∈[-1.804 |▃█▂▁▁▁▁▁▂▁| 2.640] μ=-0.567 σ=1.178
    tensor[196] x∈[-1.717 |▃█▆      ▁| 2.396] μ=-0.982 σ=0.350
    tensor[196] x∈[-1.752 |▂█▄      ▁| 2.326] μ=-1.034 σ=0.314
    tensor[196] x∈[-1.648 |▂█▁▁     ▁| 2.379] μ=-1.086 σ=0.314
    tensor[196] x∈[-1.630 |▃█▁      ▁| 2.466] μ=-1.121 σ=0.305
    tensor[196] x∈[-1.717 |▂█▁      ▁| 2.448] μ=-1.120 σ=0.302
    tensor[196] x∈[-1.717 |▃█▁▁     ▁| 2.431] μ=-1.166 σ=0.314
    tensor[196] x∈[-1.560 |█▆▁▁     ▁| 2.448] μ=-1.124 σ=0.326
    tensor[196] x∈[-1.421 |█▂▁▁ ▁   ▁| 2.431] μ=-1.064 σ=0.383
    tensor[196] x∈[-1.526 |██▁▁▁ ▁  ▁| 2.396] μ=-1.047 σ=0.417
    ...

t = torch.zeros(2, 3, 4, names=('N', 'C', None))
test_eq(str(lovely(t)), "tensor[N=2, C=3, 4] n=24 \x1b[38;2;127;127;127mall_zeros\x1b[0m")

lovely(t)

/tmp/ipykernel_1044556/3561422158.py:1: UserWarning: Named tensors and all their associated APIs are an experimental feature and subject to change. Please do not use them for anything important until they are released as stable. (Triggered internally at /pytorch/c10/core/TensorImpl.h:1971.)
  t = torch.zeros(2, 3, 4, names=('N', 'C', None))

tensor[N=2, C=3, 4] n=24 all_zeros

Meta device

t = torch.empty(3,3, device="meta")
lovely(t)

tensor[3, 3] n=9 meta meta

CUDA memory is not leaked

def memstats():
    allocated = int(torch.cuda.memory_allocated() // (1024*1024))
    max_allocated = int(torch.cuda.max_memory_allocated() // (1024*1024))
    return f"Allocated: {allocated} MB, Max: {max_allocated} Mb"

if torch.cuda.is_available():
    cudamem = torch.cuda.memory_allocated()
    print(f"before allocation: {memstats()}")
    numbers = torch.randn((3, 1024, 1024), device="cuda") # 12Mb image
    torch.cuda.synchronize()

    print(f"after allocation: {memstats()}")
    # Note, the return value of lovely() is not a string, but a
    # StrProxy that holds reference to 'numbers'. You have to del
    # the references to it, but once it's gone, the reference to
    # the tensor is gone too.
    display(lovely(numbers) )
    print(f"after repr: {memstats()}")

    del numbers
    # torch.cuda.memory.empty_cache()

    print(f"after cleanup: {memstats()}")
    test_eq(cudamem >= torch.cuda.memory_allocated(), True)

before allocation: Allocated: 0 MB, Max: 0 Mb
after allocation: Allocated: 12 MB, Max: 12 Mb

tensor[3, 1024, 1024] n=3145728 (12Mb) x∈[-5.109 | ▁▁▄██▄▁▁▁| 5.141] μ=5.667e-05 σ=1.000 cuda:0

after repr: Allocated: 12 MB, Max: 99 Mb
after cleanup: Allocated: 0 MB, Max: 99 Mb

# We don't really supposed complex numbers yet
c = torch.randn(5, dtype=torch.complex64)
lovely(c)

tensor([ 0.6125-0.2495j,  0.2462+0.8040j, -0.2361-1.0412j,  0.5159-0.0928j, -0.4503+0.7375j])