torchfile.py

"""
Copyright (c) 2016, Brendan Shillingford
All rights reserved.

Redistribution and use in source and binary forms, with or without modification, are permitted provided that the 
following conditions are met: 

1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following 
disclaimer. 

2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the 
following disclaimer in the documentation and/or other materials provided with the distribution. 

3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote 
products derived from this software without specific prior written permission. 

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, 
INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 
WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 
----------------------------------------------------------------------------------------------------------------------
The file was taken from https://github.com/bshillingford/python-torchfile and slightly modified
----------------------------------------------------------------------------------------------------------------------

Mostly direct port of the Lua and C serialization implementation to 
Python, depending only on `struct`, `array`, and numpy.

Supported types:
 * `nil` to Python `None`
 * numbers to Python floats, or by default a heuristic changes them to ints or
   longs if they are integral
 * booleans
 * strings: read as byte strings (Python 3) or normal strings (Python 2), like
   lua strings which don't support unicode, and that can contain null chars
 * tables converted to a special dict (*); if they are list-like (i.e. have
   numeric keys from 1 through n) they become a python list by default
 * Torch classes: supports Tensors and Storages, and most classes such as 
   modules. Trivially extensible much like the Torch serialization code.
   Trivial torch classes like most `nn.Module` subclasses become 
   `TorchObject`s. The `torch_readers` dict contains the mapping from class
   names to reading functions.
 * functions: loaded into the `LuaFunction` `namedtuple`,
   which simply wraps the raw serialized data, i.e. upvalues and code.
   These are mostly useless, but exist so you can deserialize anything.

(*) Since Lua allows you to index a table with a table but Python does not, we 
    replace dicts with a subclass that is hashable, and change its
    equality comparison behaviour to compare by reference.
    See `hashable_uniq_dict`.

Currently, the implementation assumes the system-dependent binary Torch 
format, but minor refactoring can give support for the ascii format as well.
"""

TYPE_NIL = 0
TYPE_NUMBER = 1
TYPE_STRING = 2
TYPE_TABLE = 3
TYPE_TORCH = 4
TYPE_BOOLEAN = 5
TYPE_FUNCTION = 6
TYPE_RECUR_FUNCTION = 8
LEGACY_TYPE_RECUR_FUNCTION = 7

import struct
from array import array
import numpy as np
import sys
from collections import namedtuple

LuaFunction = namedtuple('LuaFunction',
                         ['size', 'dumped', 'upvalues'])


class hashable_uniq_dict(dict):
    """
    Subclass of dict with equality and hashing semantics changed:
    equality and hashing is purely by reference/instance, to match
    the behaviour of lua tables.

    Supports lua-style dot indexing.

    This way, dicts can be keys of other dicts.
    """

    def __hash__(self):
        return id(self)

    def __getattr__(self, key):
        return self.get(key)

    def __eq__(self, other):
        return id(self) == id(other)
    # TODO: dict's __lt__ etc. still exist

torch_readers = {}


def add_tensor_reader(typename, dtype):
    def read_tensor_generic(reader, version):
        # source:
        # https://github.com/torch/torch7/blob/master/generic/Tensor.c#L1243
        ndim = reader.read_int()

        # read size:
        size = reader.read_long_array(ndim)
        # read stride:
        stride = reader.read_long_array(ndim)
        # storage offset:
        storage_offset = reader.read_long() - 1
        # read storage:
        storage = reader.read_obj()

        if storage is None or ndim == 0 or len(size) == 0 or len(stride) == 0:
            # empty torch tensor
            return np.empty((0), dtype=dtype)

        # convert stride to numpy style (i.e. in bytes)
        stride = [storage.dtype.itemsize * x for x in stride]

        # create numpy array that indexes into the storage:
        return np.lib.stride_tricks.as_strided(
            storage[storage_offset:],
            shape=size,
            strides=stride)
    torch_readers[typename] = read_tensor_generic
add_tensor_reader(b'torch.ByteTensor', dtype=np.uint8)
add_tensor_reader(b'torch.CharTensor', dtype=np.int8)
add_tensor_reader(b'torch.ShortTensor', dtype=np.int16)
add_tensor_reader(b'torch.IntTensor', dtype=np.int32)
add_tensor_reader(b'torch.LongTensor', dtype=np.int64)
add_tensor_reader(b'torch.FloatTensor', dtype=np.float32)
add_tensor_reader(b'torch.DoubleTensor', dtype=np.float64)
add_tensor_reader(b'torch.CudaTensor', np.float32)  # float
add_tensor_reader(b'torch.CudaByteTensor', dtype=np.uint8)
add_tensor_reader(b'torch.CudaCharTensor', dtype=np.int8)
add_tensor_reader(b'torch.CudaShortTensor', dtype=np.int16)
add_tensor_reader(b'torch.CudaIntTensor', dtype=np.int32)
add_tensor_reader(b'torch.CudaDoubleTensor', dtype=np.float64)


def add_storage_reader(typename, dtype):
    def read_storage(reader, version):
        # source:
        # https://github.com/torch/torch7/blob/master/generic/Storage.c#L244
        size = reader.read_long()
        return np.fromfile(reader.f, dtype=dtype, count=size)
    torch_readers[typename] = read_storage
add_storage_reader(b'torch.ByteStorage', dtype=np.uint8)
add_storage_reader(b'torch.CharStorage', dtype=np.int8)
add_storage_reader(b'torch.ShortStorage', dtype=np.int16)
add_storage_reader(b'torch.IntStorage', dtype=np.int32)
add_storage_reader(b'torch.LongStorage', dtype=np.int64)
add_storage_reader(b'torch.FloatStorage', dtype=np.float32)
add_storage_reader(b'torch.DoubleStorage', dtype=np.float64)
add_storage_reader(b'torch.CudaStorage', dtype=np.float32)  # float
add_storage_reader(b'torch.CudaByteStorage', dtype=np.uint8)
add_storage_reader(b'torch.CudaCharStorage', dtype=np.int8)
add_storage_reader(b'torch.CudaShortStorage', dtype=np.int16)
add_storage_reader(b'torch.CudaIntStorage', dtype=np.int32)
add_storage_reader(b'torch.CudaDoubleStorage', dtype=np.float64)


class TorchObject(object):
    """
    Simple torch object, used by `add_trivial_class_reader`.
    Supports both forms of lua-style indexing, i.e. getattr and getitem.
    Use the `torch_typename` method to get the object's torch class name.

    Equality is by reference, as usual for lua (and the default for Python
    objects).
    """

    def __init__(self, typename, obj):
        self._typename = typename
        self._obj = obj

    def __getattr__(self, k):
        return self._obj.get(k)

    def __getitem__(self, k):
        return self._obj.get(k)

    def torch_typename(self):
        return self._typename

    def __repr__(self):
        return "TorchObject(%s, %s)" % (self._typename, repr(self._obj))

    def __str__(self):
        return repr(self)

    def __dir__(self):
        keys = list(self._obj.keys())
        keys.append('torch_typename')
        return keys


def add_trivial_class_reader(typename):
    def reader(reader, version):
        obj = reader.read_obj()
        return TorchObject(typename, obj)
    torch_readers[typename] = reader
for mod in [b"nn.ConcatTable", b"nn.SpatialAveragePooling",
            b"nn.TemporalConvolutionFB", b"nn.BCECriterion", b"nn.Reshape", b"nn.gModule",
            b"nn.SparseLinear", b"nn.WeightedLookupTable", b"nn.CAddTable",
            b"nn.TemporalConvolution", b"nn.PairwiseDistance", b"nn.WeightedMSECriterion",
            b"nn.SmoothL1Criterion", b"nn.TemporalSubSampling", b"nn.TanhShrink",
            b"nn.MixtureTable", b"nn.Mul", b"nn.LogSoftMax", b"nn.Min", b"nn.Exp", b"nn.Add",
            b"nn.BatchNormalization", b"nn.AbsCriterion", b"nn.MultiCriterion",
            b"nn.LookupTableGPU", b"nn.Max", b"nn.MulConstant", b"nn.NarrowTable", b"nn.View",
            b"nn.ClassNLLCriterionWithUNK", b"nn.VolumetricConvolution",
            b"nn.SpatialSubSampling", b"nn.HardTanh", b"nn.DistKLDivCriterion",
            b"nn.SplitTable", b"nn.DotProduct", b"nn.HingeEmbeddingCriterion",
            b"nn.SpatialBatchNormalization", b"nn.DepthConcat", b"nn.Sigmoid",
            b"nn.SpatialAdaptiveMaxPooling", b"nn.Parallel", b"nn.SoftShrink",
            b"nn.SpatialSubtractiveNormalization", b"nn.TrueNLLCriterion", b"nn.Log",
            b"nn.SpatialDropout", b"nn.LeakyReLU", b"nn.VolumetricMaxPooling",
            b"nn.KMaxPooling", b"nn.Linear", b"nn.Euclidean", b"nn.CriterionTable",
            b"nn.SpatialMaxPooling", b"nn.TemporalKMaxPooling", b"nn.MultiMarginCriterion",
            b"nn.ELU", b"nn.CSubTable", b"nn.MultiLabelMarginCriterion", b"nn.Copy",
            b"nn.CuBLASWrapper", b"nn.L1HingeEmbeddingCriterion",
            b"nn.VolumetricAveragePooling", b"nn.StochasticGradient",
            b"nn.SpatialContrastiveNormalization", b"nn.CosineEmbeddingCriterion",
            b"nn.CachingLookupTable", b"nn.FeatureLPPooling", b"nn.Padding", b"nn.Container",
            b"nn.MarginRankingCriterion", b"nn.Module", b"nn.ParallelCriterion",
            b"nn.DataParallelTable", b"nn.Concat", b"nn.CrossEntropyCriterion",
            b"nn.LookupTable", b"nn.SpatialSoftMax", b"nn.HardShrink", b"nn.Abs", b"nn.SoftMin",
            b"nn.WeightedEuclidean", b"nn.Replicate", b"nn.DataParallel",
            b"nn.OneBitQuantization", b"nn.OneBitDataParallel", b"nn.AddConstant", b"nn.L1Cost",
            b"nn.HSM", b"nn.PReLU", b"nn.JoinTable", b"nn.ClassNLLCriterion", b"nn.CMul",
            b"nn.CosineDistance", b"nn.Index", b"nn.Mean", b"nn.FFTWrapper", b"nn.Dropout",
            b"nn.SpatialConvolutionCuFFT", b"nn.SoftPlus", b"nn.AbstractParallel",
            b"nn.SequentialCriterion", b"nn.LocallyConnected",
            b"nn.SpatialDivisiveNormalization", b"nn.L1Penalty", b"nn.Threshold", b"nn.Power",
            b"nn.Sqrt", b"nn.MM", b"nn.GroupKMaxPooling", b"nn.CrossMapNormalization",
            b"nn.ReLU", b"nn.ClassHierarchicalNLLCriterion", b"nn.Optim", b"nn.SoftMax",
            b"nn.SpatialConvolutionMM", b"nn.Cosine", b"nn.Clamp", b"nn.CMulTable",
            b"nn.LogSigmoid", b"nn.LinearNB", b"nn.TemporalMaxPooling", b"nn.MSECriterion",
            b"nn.Sum", b"nn.SoftSign", b"nn.Normalize", b"nn.ParallelTable", b"nn.FlattenTable",
            b"nn.CDivTable", b"nn.Tanh", b"nn.ModuleFromCriterion", b"nn.Square", b"nn.Select",
            b"nn.GradientReversal", b"nn.SpatialFullConvolutionMap", b"nn.SpatialConvolution",
            b"nn.Criterion", b"nn.SpatialConvolutionMap", b"nn.SpatialLPPooling",
            b"nn.Sequential", b"nn.Transpose", b"nn.SpatialUpSamplingNearest",
            b"nn.SpatialFullConvolution", b"nn.ModelParallel", b"nn.RReLU",
            b"nn.SpatialZeroPadding", b"nn.Identity", b"nn.Narrow", b"nn.MarginCriterion",
            b"nn.SelectTable", b"nn.VolumetricFullConvolution",
            b"nn.SpatialFractionalMaxPooling", b"fbnn.ProjectiveGradientNormalization",
            b"fbnn.Probe", b"fbnn.SparseLinear", b"cudnn._Pooling3D",
            b"cudnn.VolumetricMaxPooling", b"cudnn.SpatialCrossEntropyCriterion",
            b"cudnn.VolumetricConvolution", b"cudnn.SpatialAveragePooling", b"cudnn.Tanh",
            b"cudnn.LogSoftMax", b"cudnn.SpatialConvolution", b"cudnn._Pooling",
            b"cudnn.SpatialMaxPooling", b"cudnn.ReLU", b"cudnn.SpatialCrossMapLRN",
            b"cudnn.SoftMax", b"cudnn._Pointwise", b"cudnn.SpatialSoftMax", b"cudnn.Sigmoid",
            b"cudnn.SpatialLogSoftMax", b"cudnn.VolumetricAveragePooling", b"nngraph.Node",
            b"nngraph.JustTable", b"graph.Edge", b"graph.Node", b"graph.Graph"]:
	
    add_trivial_class_reader(mod)


class T7ReaderException(Exception):
    pass


class T7Reader:

    def __init__(self,
                 fileobj,
                 use_list_heuristic=True,
                 use_int_heuristic=True,
                 force_deserialize_classes=True,
                 force_8bytes_long=True):
        """
        Params:
        * `fileobj` file object to read from, must be actual file object
                    as it must support array, struct, and numpy
        * `use_list_heuristic`: automatically turn tables with only consecutive
                                positive integral indices into lists
                                (default True)
        * `use_int_heuristic`: cast all whole floats into ints (default True)
        * `force_deserialize_classes`: deserialize all classes, not just the
                                       whitelisted ones (default True)
        """
        self.f = fileobj
        self.objects = {}  # read objects so far

        self.use_list_heuristic = use_list_heuristic
        self.use_int_heuristic = use_int_heuristic
        self.force_deserialize_classes = force_deserialize_classes
        self.force_8bytes_long = force_8bytes_long

    def _read(self, fmt):
        sz = struct.calcsize(fmt)
        b = self.f.read(sz)
        if b == b'':
            # print('x')
            s = (0,)
        else:
            s = struct.unpack(fmt, b)

        # print(s)
        return s

    def read_boolean(self):
        return self.read_int() == 1

    def read_int(self):
        return self._read('i')[0]

    def read_long(self):
        if self.force_8bytes_long:
            return self._read('q')[0]
        else:
            return self._read('l')[0]

    def read_long_array(self, n):
        if self.force_8bytes_long:
            lst = []
            for i in range(n):
                lst.append(self.read_long())
            return lst
        else:
            arr = array('l')
            arr.fromfile(self.f, n)
            return arr.tolist()       
    
    def read_float(self):
        return self._read('f')[0]

    def read_double(self):
        return self._read('d')[0]

    def read_string(self):
        size = self.read_int()
        return self.f.read(size)

    def read_obj(self):
        typeidx = self.read_int()
        if typeidx == TYPE_NIL:
            return None
        elif typeidx == TYPE_NUMBER:
            x = self.read_double()
            # Extra checking for integral numbers:
            if self.use_int_heuristic and x.is_integer():
                return int(x)
            return x
        elif typeidx == TYPE_BOOLEAN:
            return self.read_boolean()
        elif typeidx == TYPE_STRING:
            return self.read_string()
        elif (typeidx == TYPE_TABLE or typeidx == TYPE_TORCH
                or typeidx == TYPE_FUNCTION or typeidx == TYPE_RECUR_FUNCTION
                or typeidx == LEGACY_TYPE_RECUR_FUNCTION):
            # read the index
            index = self.read_int()

            # check it is loaded already
            if index in self.objects:
                return self.objects[index]

            # otherwise read it
            if (typeidx == TYPE_FUNCTION or typeidx == TYPE_RECUR_FUNCTION
                    or typeidx == LEGACY_TYPE_RECUR_FUNCTION):
                size = self.read_int()
                dumped = self.f.read(size)
                upvalues = self.read_obj()
                obj = LuaFunction(size, dumped, upvalues)
                self.objects[index] = obj
                return obj
            elif typeidx == TYPE_TORCH:
                version = self.read_string()
                if version.startswith(b'V '):
                    versionNumber = int(version.partition(b' ')[2])
                    className = self.read_string()
                else:
                    className = version
                    versionNumber = 0  # created before existence of versioning
                # print(className)
                if className not in torch_readers:
                    if not self.force_deserialize_classes:
                        raise T7ReaderException(
                            'unsupported torch class: <%s>' % className)
                    obj = TorchObject(className, self.read_obj())
                else:
                    obj = torch_readers[className](self, version)
                self.objects[index] = obj
                return obj
            else:  # it is a table: returns a custom dict or a list
                size = self.read_int()
                obj = hashable_uniq_dict()  # custom hashable dict, can be a key
                key_sum = 0                # for checking if keys are consecutive
                keys_natural = True        # and also natural numbers 1..n.
                # If so, returns a list with indices converted to 0-indices.
                for i in range(size):
                    k = self.read_obj()
                    v = self.read_obj()
                    obj[k] = v

                    if self.use_list_heuristic:
                        if not isinstance(k, int) or k <= 0:
                            keys_natural = False
                        elif isinstance(k, int):
                            key_sum += k
                if self.use_list_heuristic:
                    # n(n+1)/2 = sum <=> consecutive and natural numbers
                    n = len(obj)
                    if keys_natural and n * (n + 1) == 2 * key_sum:
                        lst = []
                        for i in range(len(obj)):
                            lst.append(obj[i + 1])
                        obj = lst
                self.objects[index] = obj
                return obj
        else:
            raise T7ReaderException("unknown object")


def load(filename, **kwargs):
    """
    Loads the given t7 file using default settings; kwargs are forwarded
    to `T7Reader`.
    """
    with open(filename, 'rb') as f:
        reader = T7Reader(f, **kwargs)
        return reader.read_obj()