def checkGradShape(self, shape):
if len(shape) != 5:
raise ModuleError("Grad must be 5d tensor")
_, outmaps, outd, outh, outw = shape
_, _, fd, fh, fw = self.W.shape
dpad, hpad, wpad = self.pad
ddilation, hdilation, wdilation = self.dilation
if outmaps != self.W.shape[1] * self.groups:
raise ModuleError("Grad has %d maps (expected: %d)" % (outmaps, self.W.shape[1] * self.groups))
if outd + 2 * dpad < ddilation * (fd - 1) + 1:
raise ModuleError(
"Grad maps depth is too small (got %d, expected at least %d)" %
(outd + 2 * dpad, ddilation * (fd - 1) + 1)
)
if outh + 2 * hpad < hdilation * (fh - 1) + 1:
raise ModuleError(
"Grad maps height is too small (got %d, expected at least %d)" %
(outh + 2 * hpad, hdilation * (fh - 1) + 1)
)
if outw + 2 * wpad < wdilation * (fw - 1) + 1:
raise ModuleError(
"Grad maps width is too small (got %d, expected at least %d)" %
(outw + 2 * wpad, wdilation * (fw - 1) + 1)
)
def checkDataShape(self, shape):
if len(shape) != 5:
raise ModuleError("Data must be 5d tensor")
_, inmaps, _, _, _ = shape
if inmaps != self.W.shape[0]:
raise ModuleError("Data has %d maps (expected: %d)" % (inmaps, self.W.shape[0]))
def checkDataShape(self, shape):
if len(shape) != 5:
raise ModuleError("Data must be 5d tensor")
_, inmaps, ind, inh, inw = shape
_, _, fd, fh, fw = self.W.shape
dpad, hpad, wpad = self.pad
ddilation, hdilation, wdilation = self.dilation
if inmaps != self.W.shape[1] * self.groups:
raise ModuleError("Data has %d maps (expected: %d)" %
(inmaps, self.W.shape[1] * self.groups))
if ind + 2 * dpad < ddilation * (fd - 1) + 1:
raise ModuleError(
"Data maps depth is too small (got %d, expected >= %d)" %
(ind + 2 * dpad, ddilation * (fd - 1) + 1))
if inh + 2 * hpad < hdilation * (fh - 1) + 1:
raise ModuleError(
"Data maps height is too small (got %d, expected >= %d)" %
(inh + 2 * hpad, hdilation * (fh - 1) + 1))
if inw + 2 * wpad < wdilation * (fw - 1) + 1:
raise ModuleError(
"Data maps width is too small (got %d, expected >= %d)" %
(inw + 2 * wpad, wdilation * (fw - 1) + 1))
def checkGradShape(self, shape):
if len(shape) != 4:
raise ModuleError("Grad must be 4d tensor")
_, _, h, w = shape
if h % self.scale != 0 or w % self.scale != 0:
raise ModuleError("Grad map size is not divisible by scale %s" %
self.scale)
def checkGradShape(self, shape):
if len(shape) != 5:
raise ModuleError("Grad must be 5d tensor")
_, outmaps, _, _, _ = shape
if outmaps != self.W.shape[0]:
raise ModuleError("Grad has %d maps (expected: %d)" %
(outmaps, self.W.shape[0]))
def checkDataShape(self, shape):
if len(shape) != 3:
raise ModuleError("Data must be 3d tensor")
_, _, insize = shape
if insize + 2 * self.pad[1] < self.size[1]:
raise ModuleError("Data maps size is too small (got %d, expected at least %d)" %
(insize + 2 * self.pad[1], self.size[1]))
def checkDataShape(self, shape):
if len(shape) != 4:
raise ModuleError("Data must be 4d tensor")
_, maps, _, _ = shape
if maps != self.maps:
raise ModuleError("Data has %d maps (expected: %d)" %
(maps, self.maps))
def checkDataShape(self, shape):
if len(shape) != 2:
raise ModuleError("Data must be 2d matrix")
if shape[1] != self.sentlength:
raise ModuleError(
"Expected %d data sentence length, %d was given" %
(self.sentlength, shape[1]))
def checkGradShape(self, shape):
if len(shape) != 5:
raise ModuleError("Grad must be 5d tensor")
_, maps, _, _, _ = shape
if maps != self.maps:
raise ModuleError("Grad has %d maps (expected: %d)" %
(maps, self.maps))
def checkGradShape(self, shape):
if len(shape) != 5:
raise ModuleError("Grad must be 5d tensor")
_, _, d, h, w = shape
if d % self.scale != 0 or h % self.scale != 0 or w % self.scale != 0:
raise ModuleError(
"Grad map size is not divisible by scale %s (got %s, %s, %s)" %
(self.scale, d, h, w))
def calcMode(self, T):
if Config.backend == Config.Backend.cuda:
if T not in {np.float16, np.float32}:
raise ModuleError("Unsupported dtype %s" % T)
elif T != np.float32:
raise ModuleError("Unsupported dtype %s" % T)
self.calctype = T
def checkDataShape(self, shapes):
if not isinstance(shapes, list):
raise ModuleError("Data must be list of tensors")
for shape in shapes:
if len(shape) != 4:
raise ModuleError("Data must consist of 4d tensors")
if shape[0] != shapes[0][0]:
raise ModuleError("Inconsistency in batch size")
def checkGradShape(self, shape):
if self.axis >= len(shape):
raise ModuleError(
"Grad dimension needs to be at least %d, (grad has %d)" %
(self.axis + 1, len(shape)))
if shape[self.axis] != self.topk:
raise ModuleError(
"Grad topk axis is wrong (got %d, expected exactly %d)" %
(shape[self.axis], self.topk))
def checkDataShape(self, shape):
if self.axis >= len(shape):
raise ModuleError(
"Data dimension needs to be at least %d, (data has %d)" %
(self.axis + 1, len(shape)))
if shape[self.axis] < self.topk:
raise ModuleError(
"Data topk axis is too small (got %d, expected at least %d)" %
(shape[self.axis], self.topk))
def checkGradShape(self, shape):
if len(shape) != 3:
raise ModuleError("Grad must be 3d tensor")
lpad, rpad = self.pad
size = shape[2]
if size < lpad + rpad + 1:
raise ModuleError(
"Grad maps size is too small (got %d, expected >= %d)" %
(size, lpad + rpad + 1))
def verifyData(self, data):
mn, mx = gpuarray.minimum(data).get(), gpuarray.maximum(data).get()
if mn < -1:
raise ModuleError(
"Embedder data verification failed, found index %s (< -1)" %
mn)
if mx >= self.W.shape[0]:
raise ModuleError(
"Embedder data verification failed, found index %s (vocabulary size is %s)"
% (mx, self.W.shape[0]))
def checkDataShape(self, shape):
if len(shape) != 3:
raise ModuleError("Data must be 3d tensor")
if self.hintBatchSize is not None and shape[1] != self.hintBatchSize:
raise ModuleError("Data batch size must be = %s (was given %s)" %
(self.hintBatchSize, shape[1]))
if shape[2] != self.insize:
raise ModuleError("Data must have data size = %s (was given %s)" %
(self.insize, shape[2]))
def checkGradShape(self, shape):
if self.axis >= len(shape):
raise ModuleError(
"Not enough dims in grad (%d were given, need at least %d)" %
(len(shape), self.axis))
if self.useWeights:
if shape[:self.axis] != self.v.shape[:self.axis]:
raise ModuleError(
"Inconsistency in grad and weights shapes (%s with %s)" %
(shape, self.v.shape))
def checkVarOnLoad(self, paramName, dataset):
if paramName == "W":
if dataset.shape[1] != self.embsize:
raise ModuleError("Expected embedding size %s, was given %s" %
(self.embsize, dataset.shape[1]))
self.setVar("W", Variable(gpuarray.to_gpu(dataset)))
else:
raise ModuleError("Unknown parameter name '%s' for embedder" %
paramName)
def checkDataShape(self, shape):
if not self.transpose:
if shape[-1] != self.W.shape[0]:
raise ModuleError(
"Expected %d data dimensions, %d were given" %
(self.W.shape[0], shape[1]))
else:
if shape[-1] != self.W.shape[1]:
raise ModuleError(
"Expected %d data dimensions, %d were given" %
(self.W.shape[1], shape[1]))
def checkGradShape(self, shape):
if len(shape) != 3:
raise ModuleError("Grad must be 3d tensor")
batchsize, sentlen, embsize = shape
if embsize != self.embsize:
raise ModuleError("Expected %d grad embedding size, %d was given" %
(self.embsize, embsize))
if batchsize != self.inData.shape[0]:
raise ModuleError("Expected %d grad batch size, %d was given" %
(self.inData.shape[0], batchsize))
def checkDataShape(self, shape):
if len(shape) != 3:
raise ModuleError("Data must be 3d tensor")
lpad, rpad = self.pad
size = shape[2]
pad = max(lpad, rpad)
if size < pad + 1:
raise ModuleError(
"Data maps size is too small (got %d, expected >= %d)" %
(size, pad + 1))
def checkDataShape(self, shapes):
dshape, tshape = shapes
if len(tshape) != 3 or tshape[1:] != (2, 3):
raise ModuleError("Bad transform shape (%s was given)" % tshape)
if len(dshape) != 4:
raise ModuleError("Data must be 4d tensor")
if tshape[0] != dshape[0]:
raise ModuleError(
"Inconsistency in transform and data batch size (%d in transform vs %d in data)"
% (tshape[0], dshape[0]))
def checkGradShape(self, shape):
if len(shape) != 4:
raise ModuleError("Grad must be 4d tensor")
if self.shape is not None:
if self.shape != shape[1:]:
raise ModuleError(
"Bad grad shape (was given %s, expected %s)" %
(shape[1:], self.shape))
else:
if self.inData[0].shape != shape:
raise ModuleError(
"Bad grad shape (was given %s, expected %s)" %
(shape, self.inData[0].shape))
def checkGradShape(self, shape):
if len(shape) != 2:
raise ModuleError("Grad must be 2d matrix")
if not self.transpose:
if shape[1] != self.W.shape[1]:
raise ModuleError(
"Expected %d grad dimensions, %d were given" %
(self.W.shape[1], shape[1]))
else:
if shape[1] != self.W.shape[0]:
raise ModuleError(
"Expected %d grad dimensions, %d were given" %
(self.W.shape[0], shape[1]))
def checkDataShape(self, shape):
if len(shape) != 4:
raise ModuleError("Data must be 4d tensor")
_, _, inh, inw = shape
if inh + 2 * self.pad[0] < self.size[0]:
raise ModuleError(
"Data maps height is too small (got %d, expected at least %d)"
% (inh + 2 * self.pad[0], self.size[0]))
if inw + 2 * self.pad[1] < self.size[1]:
raise ModuleError(
"Data maps width is too small (got %d, expected at least %d)" %
(inw + 2 * self.pad[1], self.size[1]))
def checkDataShape(self, shape):
if len(shape) < self.axis:
raise ModuleError(
"Not enough dims in data (%d were given, need at least %d)" %
(len(shape), self.axis))
concatDim = 0
for sec in self.sections:
concatDim += sec
if concatDim != shape[self.axis]:
raise ModuleError(
"Data shape %s is inconsistent with given sections %s"
"(expected size %d on axis %d, %d was given)" %
(shape, self.sections, concatDim, self.axis, shape[self.axis]))
def checkGradShape(self, shape):
if len(shape) != 4:
raise ModuleError("Grad must be 4d tensor")
depth, h, w = 0, 0, 0
for data in self.inData:
sh = data.shape
depth += sh[1]
h, w = max(h, sh[2]), max(w, sh[3])
gradshape = (self.inData[0].shape[0], depth, h, w)
if shape != gradshape:
raise ModuleError("Bad grad shape (%s given, %s expected)" %
(shape, gradshape))
def checkAttrOnLoad(self, attrName, dataset):
if attrName == "vocab":
self.setAttr("vocab", dataset)
else:
raise ModuleError("Unknown attribute name '%s' for embedder" %
attrName)
def checkGradShape(self, shape):
ln = max(self.src, self.dst)
if len(shape) - 1 < ln:
raise ModuleError(
"Grad dimension needs to be at least %d, (grad has %d)" %
(ln + 1, len(shape)))