How to convert a torch tensor into a byte string?
Question:
I’m trying to serialize a torch tensor using protobuf and it seems using BytesIO along with torch.save() doesn’t work. I have tried:
import torch
import io
x = torch.randn(size=(1,20))
buff = io.BytesIO()
torch.save(x, buff)
print(f'buffer: {buff.read()}')
to no avail as it results in b'' in the output! How should I be going about this?
Answers:
You need to seek to the beginning of the buffer before reading:
import torch
import io
x = torch.randn(size=(1,20))
buff = io.BytesIO()
torch.save(x, buff)
buff.seek(0) # <-- this is what you were missing
print(f'buffer: {buff.read()}')
gives you this magnificent output:
buffer: b'PKx03x04x00x00x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x10x00x12x00archive/data.pklFBx0ex00ZZZZZZZZZZZZZZx80x02ctorch._utilsn_rebuild_tensor_v2nqx00((Xx07x00x00x00storageqx01ctorchnFloatStoragenqx02Xx0fx00x00x00140417054790352qx03Xx03x00x00x00cpuqx04Kx14tqx05QKx00Kx01Kx14x86qx06Kx14Kx01x86qx07x89ccollectionsnOrderedDictnqx08)RqttqnRqx0b.PKx07x08xf3x08ux13xa8x00x00x00xa8x00x00x00PKx03x04x00x00x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x1cx00x0ex00archive/data/140417054790352FBnx00ZZZZZZZZZZxbaxf3x?xb5xe2xc4=)Rx89xbfMx08x19xbfo%Yxbfx05xc0_xbfx03N4xbexdd_ xc0&xc4xb5?xa7xfdxc4?fxf1$?Llxa6?xeex8ex80xbfx88Uq?.<xd8?{x08xb2?xb3xa3xba>qxcdxbc?xbaxe3hxbdxcanx11xc0PKx07x08Axf3xdc>Px00x00x00Px00x00x00PKx03x04x00x00x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x0fx003x00archive/versionFB/x00ZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZ3nPKx07x08xd1x9egUx02x00x00x00x02x00x00x00PKx01x02x00x00x00x00x08x08x00x00x00x00x00x00xf3x08ux13xa8x00x00x00xa8x00x00x00x10x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00archive/data.pklPKx01x02x00x00x00x00x08x08x00x00x00x00x00x00Axf3xdc>Px00x00x00Px00x00x00x1cx00x00x00x00x00x00x00x00x00x00x00x00x00xf8x00x00x00archive/data/140417054790352PKx01x02x00x00x00x00x08x08x00x00x00x00x00x00xd1x9egUx02x00x00x00x02x00x00x00x0fx00x00x00x00x00x00x00x00x00x00x00x00x00xa0x01x00x00archive/versionPKx06x06,x00x00x00x00x00x00x00x1ex03-x00x00x00x00x00x00x00x00x00x03x00x00x00x00x00x00x00x03x00x00x00x00x00x00x00xc5x00x00x00x00x00x00x00x12x02x00x00x00x00x00x00PKx06x07x00x00x00x00xd7x02x00x00x00x00x00x00x01x00x00x00PKx05x06x00x00x00x00x03x00x03x00xc5x00x00x00x12x02x00x00x00x00'
Use BytesIO.getvalue method.
Apart from seek-ing and read-ing, you can also use the getvalue method of the io.BytesIO object. It does the seek–read internally and returns the stored bytes:
In [1121]: x = torch.randn(size=(1,20))
buff = io.BytesIO()
torch.save(x, buff)
print(f'buffer: {buff.getvalue()}')
buffer: b'PKx03x04x00x00x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x10x00x12x00archive/data.pklFBx0e\
x00ZZZZZZZZZZZZZZx80x02ctorch._utilsn_rebuild_tensor_v2nqx00((Xx07x00x00x00storageqx01ctorchnFloatStoragenqx02Xx01x00x00x000qx03Xx03x00x0
0x00cpuqx04Kx14tqx05QKx00Kx01Kx14x86qx06Kx14Kx01x86qx07x89ccollectionsnOrderedDictnqx08)RqttqnRqx0b.PKx07x08x949f)x9ax00x00x00x9a\
x00x00x00PKx03x04x00x00x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x0ex00*x00archive/data/0FB&x00ZZZZZZZZZZZZZZZ
ZZZZZZZZZZZZZZZZZZZZZZZxff*x1fxbfMxaax16?x9fBxbd?x14xeexb4xbexbcx83^>l.xba>x8dxc0x1fxbfZx06x03xbexe0(Bxbe^[xf8xbeEx83x9fxbfUoxc0\
xbdxbaXxb7?x83MHxbfxc0x0cxbbxbfxa4sxc9?x84x8bxd9xbfxa1x91xa1xbfxc6,x0c?kxW?PKx07x08/nx02&Px00x00x00Px00x00x00PKx03x04x00x00\
x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x0fx003x00archive/versionFB/x00ZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZ
ZZZ3nPKx07x08xd1x9egUx02x00x00x00x02x00x00x00PKx01x02x00x00x00x00x08x08x00x00x00x00x00x00x949f)x9ax00x00x00x9ax00x00x00x1
0x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00archive/data.pklPKx01x02x00x00x00x00x08x08x00x00x00x00x00x00/nx02&Px00x
00x00Px00x00x00x0ex00x00x00x00x00x00x00x00x00x00x00x00x00xeax00x00x00archive/data/0PKx01x02x00x00x00x00x08x08x00x00x00x00x0
0x00xd1x9egUx02x00x00x00x02x00x00x00x0fx00x00x00x00x00x00x00x00x00x00x00x00x00xa0x01x00x00archive/versionPKx06x06,x00x00x00\
x00x00x00x00x1ex03-x00x00x00x00x00x00x00x00x00x03x00x00x00x00x00x00x00x03x00x00x00x00x00x00x00xb7x00x00x00x00x00x00x00x
12x02x00x00x00x00x00x00PKx06x07x00x00x00x00xc9x02x00x00x00x00x00x00x01x00x00x00PKx05x06x00x00x00x00x03x00x03x00xb7x00x00
x00x12x02x00x00x00x00'
getvalue also works the same for any io.StringIO object but instead of bytes, it returns the stored string, as expected.
Instead of using ByteIO directly you could use pickle.loads/dumps.
According to this discuss thread and the linked PR discussion pytorch’s custom pickling handler does ultimately use torch.save anyways, but needs to serialize less objects, resulting in a 469 length vs 811 length bytes string. Not sure which is faster though.
import pickle, torch
x = torch.randn(size=(1,20))
pickled = pickle.dumps(x)
print(f"length of {len(pickled)} vs 811 for the BytesIO approaches")
# => length of 469 vs 811 for the BytesIO approaches
print(pickled)
# => b'x80x04x95xcax01x00x00x00x00x00x00x8cx0ctorch._utilsx94x8cx12_rebuild_tensor_v2x94x93x94(x8crtorch.storagex94x8cx10_load_from_bytesx94x93x94B?x01x00x00x80x02x8anlxfcx9cFxf9 jxa8Px19.x80x02Mxe9x03.x80x02}qx00(Xx10x00x00x00protocol_versionqx01Mxe9x03Xrx00x00x00little_endianqx02x88Xnx00x00x00type_sizesqx03}qx04(Xx05x00x00x00shortqx05Kx02Xx03x00x00x00intqx06Kx04Xx04x00x00x00longqx07Kx04uu.x80x02(Xx07x00x00x00storageqx00ctorchnFloatStoragenqx01Xx08x00x00x0097668208qx02Xx03x00x00x00cpuqx03Kx14Ntqx04Q.x80x02]qx00Xx08x00x00x0097668208qx01a.x14x00x00x00x00x00x00x00jx0bxeb?4xccx1b?xfaxa9$xbex86xc69>xabxd0#?x14}H>xd16#xc0x04x8ex0b?xadgx1a?xe2xb5g??xf0x83xbfx02Ex89xbf0xf8xe0>x04xc8L?x89xfcx06?xa0xcf)xbfxb9xa7x1b@^xd3xa0xbfJxdaxbdxbfEx95x99xbex94x85x94Rx94Kx00Kx01Kx14x86x94Kx14Kx01x86x94x89x8cx0bcollectionsx94x8cx0bOrderedDictx94x93x94)Rx94tx94Rx94.'
print(pickle.loads(pickled))
# => tensor([[ 0.7662, 0.7422, 0.1888, ..., -0.2035, 1.0845, -0.9637]])
I’m trying to serialize a torch tensor using protobuf and it seems using BytesIO along with torch.save() doesn’t work. I have tried:
import torch
import io
x = torch.randn(size=(1,20))
buff = io.BytesIO()
torch.save(x, buff)
print(f'buffer: {buff.read()}')
to no avail as it results in b'' in the output! How should I be going about this?
You need to seek to the beginning of the buffer before reading:
import torch
import io
x = torch.randn(size=(1,20))
buff = io.BytesIO()
torch.save(x, buff)
buff.seek(0) # <-- this is what you were missing
print(f'buffer: {buff.read()}')
gives you this magnificent output:
buffer: b'PKx03x04x00x00x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x10x00x12x00archive/data.pklFBx0ex00ZZZZZZZZZZZZZZx80x02ctorch._utilsn_rebuild_tensor_v2nqx00((Xx07x00x00x00storageqx01ctorchnFloatStoragenqx02Xx0fx00x00x00140417054790352qx03Xx03x00x00x00cpuqx04Kx14tqx05QKx00Kx01Kx14x86qx06Kx14Kx01x86qx07x89ccollectionsnOrderedDictnqx08)RqttqnRqx0b.PKx07x08xf3x08ux13xa8x00x00x00xa8x00x00x00PKx03x04x00x00x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x1cx00x0ex00archive/data/140417054790352FBnx00ZZZZZZZZZZxbaxf3x?xb5xe2xc4=)Rx89xbfMx08x19xbfo%Yxbfx05xc0_xbfx03N4xbexdd_ xc0&xc4xb5?xa7xfdxc4?fxf1$?Llxa6?xeex8ex80xbfx88Uq?.<xd8?{x08xb2?xb3xa3xba>qxcdxbc?xbaxe3hxbdxcanx11xc0PKx07x08Axf3xdc>Px00x00x00Px00x00x00PKx03x04x00x00x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x0fx003x00archive/versionFB/x00ZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZ3nPKx07x08xd1x9egUx02x00x00x00x02x00x00x00PKx01x02x00x00x00x00x08x08x00x00x00x00x00x00xf3x08ux13xa8x00x00x00xa8x00x00x00x10x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00archive/data.pklPKx01x02x00x00x00x00x08x08x00x00x00x00x00x00Axf3xdc>Px00x00x00Px00x00x00x1cx00x00x00x00x00x00x00x00x00x00x00x00x00xf8x00x00x00archive/data/140417054790352PKx01x02x00x00x00x00x08x08x00x00x00x00x00x00xd1x9egUx02x00x00x00x02x00x00x00x0fx00x00x00x00x00x00x00x00x00x00x00x00x00xa0x01x00x00archive/versionPKx06x06,x00x00x00x00x00x00x00x1ex03-x00x00x00x00x00x00x00x00x00x03x00x00x00x00x00x00x00x03x00x00x00x00x00x00x00xc5x00x00x00x00x00x00x00x12x02x00x00x00x00x00x00PKx06x07x00x00x00x00xd7x02x00x00x00x00x00x00x01x00x00x00PKx05x06x00x00x00x00x03x00x03x00xc5x00x00x00x12x02x00x00x00x00'
Use BytesIO.getvalue method.
Apart from seek-ing and read-ing, you can also use the getvalue method of the io.BytesIO object. It does the seek–read internally and returns the stored bytes:
In [1121]: x = torch.randn(size=(1,20))
buff = io.BytesIO()
torch.save(x, buff)
print(f'buffer: {buff.getvalue()}')
buffer: b'PKx03x04x00x00x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x10x00x12x00archive/data.pklFBx0e\
x00ZZZZZZZZZZZZZZx80x02ctorch._utilsn_rebuild_tensor_v2nqx00((Xx07x00x00x00storageqx01ctorchnFloatStoragenqx02Xx01x00x00x000qx03Xx03x00x0
0x00cpuqx04Kx14tqx05QKx00Kx01Kx14x86qx06Kx14Kx01x86qx07x89ccollectionsnOrderedDictnqx08)RqttqnRqx0b.PKx07x08x949f)x9ax00x00x00x9a\
x00x00x00PKx03x04x00x00x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x0ex00*x00archive/data/0FB&x00ZZZZZZZZZZZZZZZ
ZZZZZZZZZZZZZZZZZZZZZZZxff*x1fxbfMxaax16?x9fBxbd?x14xeexb4xbexbcx83^>l.xba>x8dxc0x1fxbfZx06x03xbexe0(Bxbe^[xf8xbeEx83x9fxbfUoxc0\
xbdxbaXxb7?x83MHxbfxc0x0cxbbxbfxa4sxc9?x84x8bxd9xbfxa1x91xa1xbfxc6,x0c?kxW?PKx07x08/nx02&Px00x00x00Px00x00x00PKx03x04x00x00\
x08x08x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x0fx003x00archive/versionFB/x00ZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZ
ZZZ3nPKx07x08xd1x9egUx02x00x00x00x02x00x00x00PKx01x02x00x00x00x00x08x08x00x00x00x00x00x00x949f)x9ax00x00x00x9ax00x00x00x1
0x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00x00archive/data.pklPKx01x02x00x00x00x00x08x08x00x00x00x00x00x00/nx02&Px00x
00x00Px00x00x00x0ex00x00x00x00x00x00x00x00x00x00x00x00x00xeax00x00x00archive/data/0PKx01x02x00x00x00x00x08x08x00x00x00x00x0
0x00xd1x9egUx02x00x00x00x02x00x00x00x0fx00x00x00x00x00x00x00x00x00x00x00x00x00xa0x01x00x00archive/versionPKx06x06,x00x00x00\
x00x00x00x00x1ex03-x00x00x00x00x00x00x00x00x00x03x00x00x00x00x00x00x00x03x00x00x00x00x00x00x00xb7x00x00x00x00x00x00x00x
12x02x00x00x00x00x00x00PKx06x07x00x00x00x00xc9x02x00x00x00x00x00x00x01x00x00x00PKx05x06x00x00x00x00x03x00x03x00xb7x00x00
x00x12x02x00x00x00x00'
getvalue also works the same for any io.StringIO object but instead of bytes, it returns the stored string, as expected.
Instead of using ByteIO directly you could use pickle.loads/dumps.
According to this discuss thread and the linked PR discussion pytorch’s custom pickling handler does ultimately use torch.save anyways, but needs to serialize less objects, resulting in a 469 length vs 811 length bytes string. Not sure which is faster though.
import pickle, torch
x = torch.randn(size=(1,20))
pickled = pickle.dumps(x)
print(f"length of {len(pickled)} vs 811 for the BytesIO approaches")
# => length of 469 vs 811 for the BytesIO approaches
print(pickled)
# => b'x80x04x95xcax01x00x00x00x00x00x00x8cx0ctorch._utilsx94x8cx12_rebuild_tensor_v2x94x93x94(x8crtorch.storagex94x8cx10_load_from_bytesx94x93x94B?x01x00x00x80x02x8anlxfcx9cFxf9 jxa8Px19.x80x02Mxe9x03.x80x02}qx00(Xx10x00x00x00protocol_versionqx01Mxe9x03Xrx00x00x00little_endianqx02x88Xnx00x00x00type_sizesqx03}qx04(Xx05x00x00x00shortqx05Kx02Xx03x00x00x00intqx06Kx04Xx04x00x00x00longqx07Kx04uu.x80x02(Xx07x00x00x00storageqx00ctorchnFloatStoragenqx01Xx08x00x00x0097668208qx02Xx03x00x00x00cpuqx03Kx14Ntqx04Q.x80x02]qx00Xx08x00x00x0097668208qx01a.x14x00x00x00x00x00x00x00jx0bxeb?4xccx1b?xfaxa9$xbex86xc69>xabxd0#?x14}H>xd16#xc0x04x8ex0b?xadgx1a?xe2xb5g??xf0x83xbfx02Ex89xbf0xf8xe0>x04xc8L?x89xfcx06?xa0xcf)xbfxb9xa7x1b@^xd3xa0xbfJxdaxbdxbfEx95x99xbex94x85x94Rx94Kx00Kx01Kx14x86x94Kx14Kx01x86x94x89x8cx0bcollectionsx94x8cx0bOrderedDictx94x93x94)Rx94tx94Rx94.'
print(pickle.loads(pickled))
# => tensor([[ 0.7662, 0.7422, 0.1888, ..., -0.2035, 1.0845, -0.9637]])