lzhengning
/
JittorMirror
mirror of https://github.com/Jittor/Jittor
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
JittorMirror/python/jittor/nn.py

3047 lines
110 KiB
Python
Raw Blame History

# ***************************************************************
# Copyright (c) 2022 Jittor. All Rights Reserved.
# Maintainers:
# Guowei Yang <471184555@qq.com>
# Guoye Yang <498731903@qq.com>
# Wenyang Zhou <576825820@qq.com>
# Meng-Hao Guo <guomenghao1997@gmail.com>
# Dun Liang <randonlang@gmail.com>.
# Zheng-Ning Liu <lzhengning@gmail.com>
#
# This file is subject to the terms and conditions defined in
# file 'LICENSE.txt', which is part of this source code package.
# ***************************************************************
from abc import abstractmethod
import jittor as jt
from jittor import flatten, init, Module
import numpy as np
import collections
import math
from collections import OrderedDict
from jittor.pool import *
from jittor.optim import *
from jittor.misc import _pair, _triple
from jittor_utils import LOG
def matmul_transpose(a, b):
'''
returns a * b^T
'''
assert a.shape[-1] == b.shape[-1], (a.shape, b.shape)
if len(a.shape) != 2:
aa = a.reshape((-1, a.shape[-1]))
cc = matmul_transpose(aa, b)
return cc.reshape(a.shape[:-1]+(-1,))
assert len(a.shape) == 2 and len(b.shape) == 2
shape = list(a.shape)[:-1] + list(b.shape)
with jt.flag_scope(amp_reg = jt.flags.amp_reg | 4):
a = a.broadcast(shape, [len(shape)-2])
b = b.broadcast(shape)
return (a*b).sum(len(shape)-1)
def bmm_transpose(a, b):
'''
returns a * b^T
'''
if jt.flags.use_cuda and jt.compile_extern.cublas_ops:
return jt.compile_extern.cublas_ops.cublas_batched_matmul(a, b, 0, 1)
t = list(range(b.ndim))
t[-1], t[-2] = t[-2], t[-1]
return bmm(a, b.transpose(t))
def bmm(a, b):
''' batch matrix multiply,
shape of input a is [batch, n, m],
shape of input b is [batch, m, k],
return shape is [batch, n, k]
Example::
import jittor as jt
from jittor import nn
batch, n, m, k = 100, 5, 6, 7
a = jt.random((batch, n, m))
b = jt.random((batch, m, k))
c = nn.bmm(a, b)
'''
assert len(a.shape) > 2 and len(b.shape) > 2
return matmul(a, b)
def baddbmm(input, batch1, batch2, beta=1, alpha=1):
res = bmm(batch1, batch2)
if alpha != 1: res = res * alpha
if beta == 0: return res
return beta * input + res
def matmul(a, b):
''' matrix multiply,
Example::
a = jt.random([3])
b = jt.random([3])
c = jt.matmul(a, b)
assert c.shape == [1]
a = jt.random([3, 4])
b = jt.random([4])
c = jt.matmul(a, b)
assert c.shape == [3]
a = jt.random([10, 3, 4])
b = jt.random([4])
c = jt.matmul(a, b)
assert c.shape == [10, 3]
a = jt.random([10, 3, 4])
b = jt.random([4, 5])
c = jt.matmul(a, b)
assert c.shape == [10, 3, 5]
a = jt.random([10, 3, 4])
b = jt.random([10, 4, 5])
c = jt.matmul(a, b)
assert c.shape == [10, 3, 5]
a = jt.random([8, 1, 3, 4])
b = jt.random([10, 4, 5])
c = jt.matmul(a, b)
assert c.shape == [8, 10, 3, 5]
'''
with jt.flag_scope(amp_reg = jt.flags.amp_reg | 4):
len_a = len(a.shape)
len_b = len(b.shape)
if len_b == 1:
# a: [n, m], b:[m], c:[n]
return (a*b).sum(-1)
if len_a == 1:
# a: [n], b:[n,k], c:[k]
return (a.broadcast(b, [-1]) * b).sum(0)
if len_a>=3 and len_a==len_b:
# bmm
# a: [..., n, m], b: [..., m, k], c:[..., n, k]
if jt.flags.use_cuda and jt.compile_extern.cublas_ops:
return jt.compile_extern.cublas_ops.cublas_batched_matmul(a, b, 0, 0)
shape = []
len_c = max(len_a, len_b)
(n, m), (m_, k) = a.shape[-2:], b.shape[-2:]
assert m == m_, f"dimension not match, a.shape:{a.shape}, b.shape:{b.shape}"
# a: [..., n, m]
# b: [..., m, k]
# cc:[..., n, m, k]
# -->
# 012
if len_b == 2 and len_a>2:
# TODO:ugly implementation for tuner
aa = a.reshape((-1, m))
cc = matmul(aa, b)
# print(a.shape, b.shape, cc.shape)
return cc.reshape(a.shape[:-1] + [k])
for i in range(len_c-2):
ai = len_a-(len_c-i)
bi = len_b-(len_c-i)
an = a.shape[ai] if ai>=0 else 1
bn = b.shape[bi] if bi>=0 else 1
if an!=1 and bn!=1:
assert an == bn, f"dimension not match, a.shape:{a.shape}, b.shape:{b.shape}"
cn = max(an, bn)
shape.append(cn)
shape.extend([n, m, k])
a = a.broadcast(shape, [-1])
b = b.broadcast(shape, [-3])
return (a*b).sum(-2)
jt.Var.matmul = jt.Var.__matmul__ = matmul
jt.Var.__imatmul__ = lambda a,b: a.assign(matmul(a,b))
def get_init_var_rand(shape, dtype):
return jt.array(np.random.normal(0.0, 1.0, shape).astype(np.float32))
def relu(x):
r''' Applies the element-wise function:
.. math::
\text{ReLU6}(x) = \max(0,x)
:param x: the input var
:type x: jt.Var
Example:
>>> a = jt.randn(3)
>>> a
jt.Var([-0.38380373 1.1338731 6.128115 ], dtype=float32)
>>> nn.relu(a)
jt.Var([0. 1.1338731 6.128115 ], dtype=float32)
'''
cond = x>0.0
return jt.ternary_out_hint(cond, x, 0.0)
def leaky_relu(x, scale=0.01):
r''' Applies the element-wise function:
.. math::
\text{LeakyRELU}(x) =
\begin{cases}
x, & \text{ if } x \geq 0 \\
\text{scale} \times x, & \text{ otherwise }
\end{cases}
:param x: the input var
:type x: jt.Var
:param scale: the :math:`\scale` value for the leaky relu formulation. Default: 0.01
:param scale: float, optional
Example:
>>> a = jt.randn(3)
>>> a
jt.Var([-0.38380373 1.1338731 6.128115 ], dtype=float32)
>>> nn.leaky_relu(a)
jt.Var([-3.8380371e-03 1.1338731e+00 6.1281152e+00], dtype=float32)
'''
return jt.ternary(x>0, x, x*scale)
def relu6(x):
r''' Applies the element-wise function:
.. math::
\text{ReLU6}(x) = \min(\max(0,x), 6)
:param x: the input var
:type x: jt.Var
Example:
>>> a = jt.randn(3)
>>> a
jt.Var([-0.38380373 1.1338731 6.128115 ], dtype=float32)
>>> nn.relu6(a)
jt.Var([0. 1.1338731 6. ], dtype=float32)
'''
return jt.minimum(jt.maximum(x, 0.0), 6.0)
def elu(x: jt.Var, alpha: float = 1.0) -> jt.Var:
r''' Applies the element-wise function:
.. math::
\text{ELU}(x) = \begin{cases}
x, & \text{ if } x > 0\\
\alpha * (\exp(x) - 1), & \text{ if } x \leq 0
\end{cases}
:param x: the input var
:type x: jt.Var
:param alpha: the :math:`\alpha` value for the ELU formulation. Default: 1.0
:param alpha: float, optional
Example:
>>> a = jt.randn(3)
>>> a
jt.Var([-0.38380373 -1.1338731 2.128115 ], dtype=float32)
>>> nn.elu(a)
jt.Var([-0.31873488 -0.6782155 2.128115 ], dtype=float32)
'''
return jt.ternary(x>0,x,alpha*(x.exp()-1))
def sign(x: jt.Var) -> jt.Var:
''' returns the signs of elements of x
:param x: the input Var
:type x: jt.Var
Example:
>>> a = jt.float32([0.99, 0, -0.99])
>>> nn.sign(a)
jt.Var([ 1. 0. -1.], dtype=float32)
'''
one = jt.ones(x.shape)
x = jt.ternary(x>0, one, x)
return jt.ternary(x<0, -one, x)
def gelu(x):
r''' Applies the element-wise function:
.. math::
\text{GELU}(x) = x * \Phi(x)
where :math:`\Phi(x)` is the Cumulative Distribution Function for Gaussian Distribution.
:param x: the input var
:type x: jt.Var
Example:
>>> a = jt.randn(3)
>>> a
jt.Var([-0.38380373 -1.1338731 2.128115 ], dtype=float32)
>>> nn.gelu(a)
jt.Var([-0.134547 0.9882567 6.128115 ], dtype=float32)
'''
_sqrt2 = 1.4142135623730951
erf = jt.erf(x/_sqrt2)+1
r = erf*x*.5
return r
def silu(x):
r''' Applies the element-wise function:
.. math::
\text{SILU}(x) = x * Sigmoid(x)
:param x: the input var
:type x: jt.Var
Example:
>>> a = jt.randn(3)
>>> a
jt.Var([-0.38380373 -1.1338731 2.128115 ], dtype=float32)
>>> nn.silu(a)
jt.Var([-0.15552104 -0.27603802 1.9016962 ], dtype=float32)
'''
return x * x.sigmoid()
class ELU(Module):
r''' Applies the element-wise function:
.. math::
\text{ELU}(x) = \begin{cases}
x, & \text{ if } x > 0\\
\alpha * (\exp(x) - 1), & \text{ if } x \leq 0
\end{cases}
:param x: the input var
:type x: jt.Var
:param alpha: the :math:`\alpha` value for the ELU formulation. Default: 1.0
:param alpha: float, optional
Example:
>>> a = jt.randn(3)
>>> a
jt.Var([-0.38380373 -1.1338731 2.128115 ], dtype=float32)
>>> nn.elu(a)
jt.Var([-0.31873488 -0.6782155 2.128115 ], dtype=float32)
'''
def __init__(self,alpha=1.0):
self.alpha=alpha
def execute(self,x):
return elu(x,self.alpha)
class PReLU(Module):
r''' Applies the element-wise function:
.. math::
\text{PReLU}(x) =
\begin{cases}
x, & \text{ if } x \geq 0 \\
ax, & \text{ otherwise }
\end{cases}
:param x: the input var
:type x: jt.Var
:param num_parameters: number of :math:`a` to learn, can be either 1 or the number of channels at input. Default: 1
:type num_parameters: int, optional
:param init: the initial value of :math:`a`. Default: 0.25
:param init: float, optional
Example:
>>> a = jt.randn(3)
>>> prelu = nn.PReLU()
>>> prelu(a)
jt.Var([-0.09595093 1.1338731 6.128115 ], dtype=float32)
'''
def __init__(self, num_parameters=1, init_=0.25):
self.num_parameters = num_parameters
self.weight = init.constant((num_parameters,), "float32", init_)
def execute(self, x):
if self.num_parameters != 1:
assert self.num_parameters == x.size(1), f"num_parameters does not match input channels in PReLU"
return jt.maximum(0, x) + self.weight.broadcast(x, [0,2,3]) * jt.minimum(0, x)
else:
return jt.maximum(0, x) + self.weight * jt.minimum(0, x)
#TODO dims is 4 will cause slowly execution
def cross_entropy_loss(output, target, weight=None, ignore_index=None,reduction='mean'):
target_shape = target.shape
if len(output.shape) == 4:
c_dim = output.shape[1]
output = output.transpose((0, 2, 3, 1))
output = output.reshape((-1, c_dim))
target = target.reshape((-1, ))
target_weight = jt.ones(target.shape[0], dtype='float32')
if weight is not None:
target_weight = weight[target]
if ignore_index is not None:
target_weight = jt.ternary(
target==ignore_index,
jt.array(0).broadcast(target_weight),
target_weight
)
target = target.broadcast(output, [1])
target = target.index(1) == target
output = output - output.max([1], keepdims=True)
logsum = output.exp().sum(1).log()
loss = (logsum - (output*target).sum(1)) * target_weight
if reduction == 'sum':
return loss.sum()
elif reduction == 'mean':
return loss.mean() / target_weight.mean()
else:
return loss.reshape(target_shape)
def mse_loss(output, target, reduction="mean"):
return (output-target).sqr().reduce(reduction)
def bce_loss(output, target, weight=None, size_average=True):
loss = - (target * jt.log(jt.maximum(output, 1e-20)) + (1 - target) * jt.log(jt.maximum(1 - output, 1e-20)))
if weight is not None:
loss *= weight
if size_average:
return loss.mean()
else:
return loss.sum()
def l1_loss(output, target):
return (output-target).abs().mean()
def smooth_l1_loss(y_true, y_pred,reduction="mean"):
"""Implements Smooth-L1 loss.
y_true and y_pred are typically: [N, 4], but could be any shape.
Args:
y_true - ground truth
y_pred - predictions
reduction - the mode of cal loss which must be in ['mean','sum','none']
"""
diff = jt.abs(y_true - y_pred)
less_than_one = (diff<1.0).float32()
loss = (less_than_one * 0.5 * diff.sqr()) + (1 - less_than_one) * (diff - 0.5)
if reduction=="mean":
return loss.mean()
elif reduction=="sum":
return loss.sum()
elif reduction=="none":
return loss
else:
raise ValueError(f'not support {reduction}')
def nll_loss(output,target,weight=None,ignore_index=-100,reduction='mean'):
assert output.ndim<=2 and output.ndim>0 and target.ndim==1
n_classes = output.shape[-1]
assert weight is None or weight.numel()==n_classes
assert ignore_index<0 or ignore_index<n_classes
if weight is None:
weight = jt.ones((n_classes,))
if ignore_index>0:
weight[ignore_index]=0
if output.ndim==2:
index = jt.index((output.shape[0],),dim=0)
loss = -output[index,target]*weight[target]
else:
loss = -output[target[0]]*weight[target[0]]
if reduction=="mean":
total_weight = weight[target].sum() if output.ndim==2 else weight[target[0]].sum()
return loss.sum()/total_weight
elif reduction=="sum":
return loss.sum()
elif reduction=="none":
return loss
else:
raise ValueError(f'not support {reduction}')
class CrossEntropyLoss(Module):
def __init__(self, weight=None, ignore_index=None):
self.weight = weight
self.ignore_index = ignore_index
def execute(self, output, target):
return cross_entropy_loss(output, target, self.weight, self.ignore_index)
class MSELoss(Module):
def __init__(self, reduction='mean'):
self.reduction = reduction
def execute(self, output, target):
return mse_loss(output, target, self.reduction)
class BCELoss(Module):
def __init__(self, weight=None, size_average=True):
self.weight = weight
self.size_average = size_average
def execute(self, output, target):
return bce_loss(output, target, self.weight, self.size_average)
class L1Loss(Module):
def __init__(self):
pass
def execute(self, output, target):
return l1_loss(output, target)
def binary_cross_entropy_with_logits(output, target, weight=None, pos_weight=None, size_average=True):
max_val = jt.clamp(-output,min_v=0)
if pos_weight is not None:
log_weight = (pos_weight-1)*target + 1
loss = (1-target)*output+(log_weight*(((-max_val).exp()+(-output - max_val).exp()).log()+max_val))
else:
loss = (1-target)*output+max_val+((-max_val).exp()+(-output -max_val).exp()).log()
if weight is not None:
loss *=weight
if size_average:
return loss.mean()
else:
return loss.sum()
class BCEWithLogitsLoss(Module):
def __init__(self, weight=None, pos_weight=None, size_average=True):
self.pos_weight = pos_weight
self.weight = weight
self.size_average = size_average
def execute(self, output, target):
return binary_cross_entropy_with_logits(output,target,self.weight,self.pos_weight,self.size_average)
def softmax(x, dim=None, log=False):
import jittor.other.code_softmax as code_softmax
if code_softmax.can_softmax_v1(x, dim) and jt.compiler.is_cuda:
return code_softmax.softmax_v1(x, log)
if dim is None: dim = ()
if log:
a = x - jt.max(x, dim, keepdims=True)
return a - a.exp().sum(dim, keepdims=True).log()
x = (x - jt.max(x, dim, keepdims=True)).exp()
return x / x.sum(dim, keepdims=True)
jt.Var.softmax = softmax
def log_softmax(x,dim=None):
return softmax(x,dim=dim, log=True)
jt.Var.log_softmax = log_softmax
def log_sigmoid(x):
return jt.log(jt.sigmoid(x))
jt.Var.log_sigmoid = log_sigmoid
def logsumexp(x, dim, keepdims=False, keepdim=False):
return x.exp().sum(dim, keepdim or keepdims).log()
jt.Var.logsumexp = logsumexp
class Identity(Module):
def __init__(self, *args, **kwargs):
super(Identity, self).__init__()
def execute(self, input):
return input
def identity(input): return input
class Dropout(Module):
def __init__(self, p=0.5, is_train=False):
assert p >= 0 and p <= 1, "dropout probability has to be between 0 and 1, but got {}".format(p)
self.p = p
self.is_train = is_train
#TODO: test model.train() to change self.is_train
def execute(self, input):
output = input
if self.p > 0 and self.is_train:
if self.p == 1:
noise = jt.zeros(input.shape)
output = output * noise
else:
noise = jt.random(input.shape)
noise = (noise > self.p).int()
output = output * noise / (1.0 - self.p) # div keep prob
return output
def dropout(x,p=0.5,is_train=False):
return Dropout(p=p,is_train=is_train)(x)
class Dropout2d(Module):
def __init__(self, p=0.5, is_train=False):
'''
Randomly zero out entire channels, from "Efficient Object Localization Using Convolutional Networks"
input:
x: [N,C,H,W] or [N,C,L]
output:
y: same shape as x
'''
assert p >= 0 and p <= 1, "dropout probability has to be between 0 and 1, but got {}".format(p)
self.p = p
self.is_train = is_train
#TODO: test model.train() to change self.is_train
def execute(self, input):
output = input
shape = input.shape[:-2]
if self.p > 0 and self.is_train:
if self.p == 1:
output = jt.zeros(input.shape)
else:
noise = jt.random(shape)
noise = (noise > self.p).int()
output = output * noise.broadcast(input.shape, dims=[-2,-1]) / (1.0 - self.p) # div keep prob
return output
def dropout2d(x,p=0.5,is_train=False):
return Dropout2d(p=p,is_train=is_train)(x)
class DropPath(Module):
'''Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
'''
def __init__(self, p=0.5, is_train=False):
'''
:param p: Specifies the probability of each batch retention. Defaults to 0.5.
:type p: float dtype
:param is_train: Specify whether it is a training model. Defaults to False.
:type is_train: bool
'''
self.p = p
self.is_train = is_train
#TODO: test model.train() to change self.is_train
def execute(self, x):
if self.p == 0. or not self.is_train:
return x
keep_prob = 1 - self.p
shape = (x.shape[0], ) + (1, ) * (x.ndim - 1)
random_tensor = keep_prob + jt.rand(shape, dtype=x.dtype)
output = x.divide(keep_prob) * random_tensor.floor()
return output
def droppath(x,p=0.5,is_train=False):
return DropPath(p=p,is_train=is_train)(x)
class Linear(Module):
def __init__(self, in_features, out_features, bias=True):
self.in_features = in_features
self.out_features = out_features
self.weight = init.invariant_uniform((out_features, in_features), "float32")
bound = 1.0/math.sqrt(in_features)
self.bias = init.uniform((out_features,), "float32",-bound,bound) if bias else None
def execute(self, x):
x = matmul_transpose(x, self.weight)
if self.bias is not None:
return x + self.bias
return x
def linear(x, weight, bias=None):
''' Returns x * weight^T
'''
x = matmul_transpose(x, weight)
if bias is not None:
return x + bias
return x
class BatchNorm(Module):
def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True, is_train=True, sync=True):
self.sync = sync
self.num_features = num_features
self.is_train = is_train
self.eps = eps
self.momentum = momentum
self.affine = affine
self.weight = init.constant((num_features,), "float32", 1.0) if affine else 1.0
self.bias = init.constant((num_features,), "float32", 0.0) if affine else 0.0
self.running_mean = init.constant((num_features,), "float32", 0.0).stop_grad()
self.running_var = init.constant((num_features,), "float32", 1.0).stop_grad()
def execute(self, x):
dims = [0]+list(range(2,x.ndim))
if self.is_train:
xmean = jt.mean(x, dims=dims)
x2mean = jt.mean(x*x, dims=dims)
if self.sync and jt.in_mpi:
xmean = xmean.mpi_all_reduce("mean")
x2mean = x2mean.mpi_all_reduce("mean")
xvar = (x2mean-xmean*xmean).maximum(0.0)
w = self.weight / jt.sqrt(xvar+self.eps)
b = self.bias - xmean * w
norm_x = x * w.broadcast(x, dims) + b.broadcast(x, dims)
self.running_mean.update(self.running_mean +
(xmean.reshape((-1,)) - self.running_mean) * self.momentum)
self.running_var.update(self.running_var +
(xvar.reshape((-1,))-self.running_var)*self.momentum)
return norm_x
else:
w = self.weight / jt.sqrt(self.running_var+self.eps)
b = self.bias - self.running_mean * w
norm_x = x * w.broadcast(x, dims) + b.broadcast(x, dims)
return norm_x
BatchNorm3d = BatchNorm2d = BatchNorm1d = BatchNorm
def batch_norm(x, running_mean, running_var, weight=1, bias=0, training=False, momentum=0.1, eps=1e-05):
dims = [0]+list(range(2,x.ndim))
assert not training
w = weight / jt.sqrt(running_var+eps)
b = bias - running_mean * w
norm_x = x * w.broadcast(x, dims) + b.broadcast(x, dims)
return norm_x
class InstanceNorm(Module):
def __init__(self, num_features, eps=1e-05, momentum=0.1, affine=True, is_train=True, sync=True):
self.sync = sync
self.num_features = num_features
self.is_train = is_train
self.eps = eps
self.momentum = momentum
self.affine = affine
self.weight = init.constant((num_features,), "float32", 1.0) if affine else 1.0
self.bias = init.constant((num_features,), "float32", 0.0) if affine else 0.0
def execute(self, x):
dims = list(range(2,x.ndim))
xmean = jt.mean(x, dims=dims)
x2mean = jt.mean(x*x, dims=dims)
xvar = (x2mean-xmean*xmean).maximum(0.0)
w = self.weight / jt.sqrt(xvar+self.eps)
b = self.bias - xmean * w
return x * w.broadcast(x, dims) + b.broadcast(x, dims)
InstanceNorm3d = InstanceNorm2d = InstanceNorm1d = InstanceNorm
def fp32_guard(func):
def wrapper(*args, **kw):
if jt.flags.amp_level == 0:
return func(*args, **kw)
new_args = []
for a in args:
if isinstance(a, jt.Var) and a.dtype == "float16":
new_args.append(a.float32())
else:
new_args.append(a)
with jt.flag_scope(amp_level=0):
a = func(*new_args, **kw)
if isinstance(a, jt.Var) and a.dtype == "float32":
a = a.float16()
return a
return wrapper
def instance_norm(x,
running_mean = None,
running_var = None,
weight = 1,
bias = 0,
momentum = 0.1,
eps = 1e-5):
dims = list(range(2,x.ndim))
xmean = jt.mean(x, dims=dims)
x2mean = jt.mean(x*x, dims=dims)
xvar = (x2mean-xmean*xmean).maximum(0.0)
w = weight / jt.sqrt(xvar+eps)
b = bias - xmean * w
return x * w.broadcast(x, dims) + b.broadcast(x, dims)
class LayerNorm(Module):
def __init__(self, normalized_shape, eps: float = 1e-5, elementwise_affine: bool = True) -> None:
if isinstance(normalized_shape, int):
normalized_shape = (normalized_shape,)
self.normalized_shape = tuple(normalized_shape)
self.eps = eps
self.elementwise_affine = elementwise_affine
self.weight = init.constant(normalized_shape, "float32", 1.0) if elementwise_affine else 1.0
self.bias = init.constant(normalized_shape, "float32", 0.0) if elementwise_affine else 0.0
@fp32_guard
def execute(self, x):
dims = [-i for i in range(len(self.normalized_shape), 0, -1)]
xmean = jt.mean(x, dims=dims, keepdims=1)
x2mean = jt.mean(x*x, dims=dims, keepdims=1)
xvar = (x2mean-xmean*xmean).maximum(0.0)
w = self.weight / jt.sqrt(xvar+self.eps)
b = self.bias - xmean * w
return x * w + b
LayerNorm3d = LayerNorm2d = LayerNorm1d = LayerNorm
@fp32_guard
def layer_norm(x,
normalized_shape,
weight = 1,
bias = 0,
eps: float = 1e-5,
elementwise_affine: bool = True):
dims = [-i for i in range(len(normalized_shape), 0, -1)]
xmean = jt.mean(x, dims=dims, keepdims=1)
x2mean = jt.mean(x*x, dims=dims, keepdims=1)
xvar = (x2mean-xmean*xmean).maximum(0.0)
w = weight / jt.sqrt(xvar+eps)
b = bias - xmean * w
return x * w + b
class GroupNorm(Module):
def __init__(self, num_groups, num_channels, eps=1e-05, affine=True, is_train=True):
self.num_groups = num_groups
self.num_channels = num_channels
self.eps = eps
self.affine = affine
self.weight = init.constant((num_channels,), "float32", 1.0) if affine else 1.0
self.bias = init.constant((num_channels,), "float32", 0.0) if affine else 0.0
def execute(self, x):
N = x.shape[0]
C = self.num_channels
output_shape = (N,-1)
# TODO: 3d group norm
if x.ndim==4:
output_shape = x.shape
assert C % self.num_groups == 0
x = x.reshape((N, self.num_groups, C//self.num_groups, -1))
xmean = jt.mean(x, dims=[2,3]).reshape((N, self.num_groups, 1))
x2mean = jt.mean(x*x, dims=[2,3]).reshape((N, self.num_groups, 1))
xvar = (x2mean-xmean*xmean).maximum(0.0)
if self.affine:
w = self.weight.reshape((1, self.num_groups, -1))
b = self.bias.reshape((1, self.num_groups, -1))
else:
w = 1
b = 0
w = w / jt.sqrt(xvar+self.eps)
b = b - xmean * w
x = x * w.broadcast(x, [3]) + b.broadcast(x, [3])
return x.reshape(output_shape)
def group_norm(x,
num_groups,
weight = 1,
bias = 0,
eps=1e-05):
N = x.shape[0]
C = x.shape[1]
output_shape = (N,-1)
# TODO: 3d group norm
if x.ndim==4:
output_shape = x.shape
assert C % num_groups == 0
x = x.reshape((N, num_groups, C//num_groups, -1))
xmean = jt.mean(x, dims=[2,3]).reshape((N, num_groups, 1))
x2mean = jt.mean(x*x, dims=[2,3]).reshape((N, num_groups, 1))
xvar = (x2mean-xmean*xmean).maximum(0.0)
if isinstance(weight, jt.Var):
weight = weight.reshape((1, num_groups, -1))
if isinstance(bias, jt.Var):
bias = bias.reshape((1, num_groups, -1))
weight = weight / jt.sqrt(xvar+eps)
bias = bias - xmean * weight
x = x * weight.broadcast(x, [3]) + bias.broadcast(x, [3])
return x.reshape(output_shape)
Relu = jt.make_module(relu)
ReLU = Relu
Leaky_relu = jt.make_module(leaky_relu, 2)
LeakyReLU = Leaky_relu
ReLU6 = jt.make_module(relu6)
Softmax = jt.make_module(softmax, 2)
GELU = jt.make_module(gelu)
class Flatten(Module):
''' Flattens the contiguous range of dimensions in a Var.
:param start_dim: the first dimension to be flattened. Defaults: 1.
:type start_dim: int
:param end_dim: the last dimension to be flattened. Defaults: -1.
:type end_dim: int
'''
def __init__(self, start_dim=1, end_dim=-1):
self.start_dim = start_dim
self.end_dim = end_dim
def execute(self, x) -> jt.Var:
return x.flatten(self.start_dim, self.end_dim)
from jittor.depthwise_conv import DepthwiseConv
class Conv(Module):
''' Applies a 2D convolution over an input signal composed of several input planes.
:param in_channels: Number of channels in the input feature map
:type in_channels: int
:param out_channels: Number of channels in the output feature map
:type out_channels: int
:param kernel_size: Size of the convolving kernel
:type kernel_size: int or tuple
:param stride: Stride of the convolution. Default: 1
:type stride: int or tuple, optional
:param padding: Padding added to all four sides of the input. Default: 0
:type padding: int or tuple, optional
:param dilation: Spacing between kernel elements. Default: 1
:type dilation: int or tuple, optional
:param groups: Number of blocked connections from input channels to output channels. Default: 1
:type groups: int, optional
:param bias: If True, adds a learnable bias to the output. Default: True
:type bias: bool, optional
Example:
>>> conv = nn.Conv2d(24, 32, 3)
>>> conv = nn.Conv2d(24, 32, (3,3))
>>> conv = nn.Conv2d(24, 32, 3, stride=2, padding=1)
>>> conv = nn.Conv2d(24, 32, 3, dilation=(3, 1))
>>> input = jt.randn(4, 24, 100, 100)
>>> output = conv(input)
'''
def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True):
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size if isinstance(kernel_size, tuple) else (kernel_size, kernel_size)
self.stride = stride if isinstance(stride, tuple) else (stride, stride)
self.padding = padding if isinstance(padding, tuple) else (padding, padding)
self.dilation = dilation if isinstance(dilation, tuple) else (dilation, dilation)
self.groups = groups
self.is_depthwise_conv = self.groups == self.out_channels and self.groups == self.in_channels
if self.is_depthwise_conv and jt.flags.use_cuda and jt.compiler.is_cuda:
self.depthwise_conv = DepthwiseConv(stride, padding, dilation)
assert in_channels % groups == 0, 'in_channels must be divisible by groups'
assert out_channels % groups == 0, 'out_channels must be divisible by groups'
Kh, Kw = self.kernel_size
# self.weight = init.relu_invariant_gauss([out_channels, in_channels//groups, Kh, Kw], dtype="float", mode="fan_out")
self.weight = init.invariant_uniform([out_channels, in_channels//groups, Kh, Kw], dtype="float")
if bias:
fan=1
for i in self.weight.shape[1:]:
fan *= i
bound = 1 / math.sqrt(fan)
self.bias = init.uniform([out_channels], dtype="float", low=-bound, high=bound)
else:
self.bias = None
def execute(self, x):
if hasattr(self, 'depthwise_conv'):
y = self.depthwise_conv(x, self.weight)
if self.bias is not None:
b = self.bias.broadcast(y.shape, [0,2,3])
y = y + b
return y
elif self.groups == 1:
N,C,H,W = x.shape
Kh, Kw = self.kernel_size
assert C==self.in_channels
oh = (H+self.padding[0]*2-Kh*self.dilation[0]+self.dilation[0]-1)//self.stride[0]+1
ow = (W+self.padding[1]*2-Kw*self.dilation[1]+self.dilation[1]-1)//self.stride[1]+1
assert oh>0 and ow>0
with jt.flag_scope(amp_reg = jt.flags.amp_reg | 4):
xx = x.reindex([N,self.out_channels,C,oh,ow,Kh,Kw], [
'i0', # Nid
'i2', # Cid
f'i3*{self.stride[0]}-{self.padding[0]}+i5*{self.dilation[0]}', # Hid+Khid
f'i4*{self.stride[1]}-{self.padding[1]}+i6*{self.dilation[1]}', # Wid+KWid
])
ww = self.weight.broadcast(xx.shape, [0,3,4])
yy = xx*ww
y = yy.sum([2,5,6]) # Kc, Kh, Kw
if self.bias is not None:
b = self.bias.broadcast(y.shape, [0,2,3])
y = y + b
return y
else:
N,C,H,W = x.shape
Kh, Kw = self.kernel_size
G = self.groups
CpG = C // G # channels per group
assert C==self.in_channels
oc = self.out_channels
oh = (H+self.padding[0]*2-Kh*self.dilation[0]+self.dilation[0]-1)//self.stride[0]+1
ow = (W+self.padding[1]*2-Kw*self.dilation[1]+self.dilation[1]-1)//self.stride[1]+1
assert oh>0 and ow>0
xx = x.reindex([N,G,oc//G,CpG,oh,ow,Kh,Kw], [
'i0', # Nid
f'i1*{CpG}+i3', # Gid
f'i4*{self.stride[0]}-{self.padding[0]}+i6*{self.dilation[0]}', # Hid+Khid
f'i5*{self.stride[1]}-{self.padding[1]}+i7*{self.dilation[1]}', # Wid+KWid
])
# w: [oc, CpG, Kh, Kw]
ww = self.weight.reindex([N, G, oc//G, CpG, oh, ow, Kh, Kw], [
f'i1*{oc//G}+i2',
'i3',
'i6',
'i7'
])
ww.compile_options = xx.compile_options = {"G":G,"C":C}
yy = xx*ww
y = yy.reindex_reduce('add', [N, oc, oh, ow], [
'i0',
f'i1*{oc//G}+i2',
'i4',
'i5'
])
if self.bias is not None:
b = self.bias.broadcast(y.shape, [0,2,3])
y = y + b
return y
Conv2d = Conv
class Conv1d(Module):
''' Applies a 1D convolution over an input signal composed of several input planes.
:param in_channels: Number of channels in the input feature map
:type in_channels: int
:param out_channels: Number of channels in the output feature map
:type out_channels: int
:param kernel_size: Size of the convolving kernel
:type kernel_size: int or tuple
:param stride: Stride of the convolution. Default: 1
:type stride: int or tuple, optional
:param padding: Padding added to all four sides of the input. Default: 0
:type padding: int or tuple, optional
:param dilation: Spacing between kernel elements. Default: 1
:type dilation: int or tuple, optional
:param groups: Number of blocked connections from input channels to output channels. Default: 1
:type groups: int, optional
:param bias: If True, adds a learnable bias to the output. Default: True
:type bias: bool, optional
Example:
>>> conv = nn.Conv1d(24, 32, 3)
>>> conv = nn.Conv1d(24, 32, (3,3))
>>> conv = nn.Conv1d(24, 32, 3, stride=2, padding=1)
>>> conv = nn.Conv1d(24, 32, 3, dilation=(3, 1))
>>> input = jt.randn(4, 24, 100)
>>> output = conv(input)
'''
def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True):
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = (kernel_size, 1)
self.stride = (stride, 1)
self.padding = (padding, 0)
self.dilation = (dilation, 1)
self.groups = groups
self.bias = bias
assert in_channels % groups == 0, 'in_channels must be divisible by groups'
assert out_channels % groups == 0, 'out_channels must be divisible by groups'
# using list to escape module dfs
self._conv = [Conv(self.in_channels, self.out_channels, self.kernel_size, self.stride, self.padding, self.dilation, self.groups, self.bias)]
self.weight = self._conv[0].weight.squeeze(-1)
self.bias = self._conv[0].bias
def execute(self, x):
N,C,D = x.shape
assert C==self.in_channels
self._conv[0].weight = self.weight.unsqueeze(-1)
x = x.unsqueeze(-1)
x = self._conv[0](x)
y = x.squeeze(-1)
return y
class Conv3d(Module):
''' Applies a 3D convolution over an input signal composed of several input planes.
:param in_channels: Number of channels in the input feature map
:type in_channels: int
:param out_channels: Number of channels in the output feature map
:type out_channels: int
:param kernel_size: Size of the convolving kernel
:type kernel_size: int or tuple
:param stride: Stride of the convolution. Default: 1
:type stride: int or tuple, optional
:param padding: Padding added to all four sides of the input. Default: 0
:type padding: int or tuple, optional
:param dilation: Spacing between kernel elements. Default: 1
:type dilation: int or tuple, optional
:param groups: Number of blocked connections from input channels to output channels. Default: 1
:type groups: int, optional
:param bias: If True, adds a learnable bias to the output. Default: True
:type bias: bool, optional
Example:
>>> conv = nn.Conv3d(24, 32, 3)
>>> conv = nn.Conv3d(24, 32, (3,3))
>>> conv = nn.Conv3d(24, 32, 3, stride=2, padding=1)
>>> conv = nn.Conv3d(24, 32, 3, dilation=(3, 1))
>>> input = jt.randn(4, 24, 50, 50, 50)
>>> output = conv(input)
'''
def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True):
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size if isinstance(kernel_size, tuple) else (kernel_size, kernel_size, kernel_size)
self.stride = stride if isinstance(stride, tuple) else (stride, stride, stride)
self.padding = padding if isinstance(padding, tuple) else (padding, padding, padding)
self.dilation = dilation if isinstance(dilation, tuple) else (dilation, dilation, dilation)
self.groups = groups
assert in_channels % groups == 0, 'in_channels must be divisible by groups'
assert out_channels % groups == 0, 'out_channels must be divisible by groups'
Kh, Kw, Kd = self.kernel_size
self.groups = groups
assert in_channels % groups == 0, 'in_channels must be divisible by groups'
assert out_channels % groups == 0, 'out_channels must be divisible by groups'
self.weight = init.invariant_uniform([out_channels, in_channels//groups, Kh, Kw, Kd], dtype="float")
if bias:
fan=1
for i in self.weight.shape[1:]:
fan *= i
bound = 1 / math.sqrt(fan)
self.bias = init.uniform([out_channels], dtype="float", low=-bound, high=bound)
else:
self.bias = None
def execute(self, x):
return conv3d(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
class Conv1d_sp(Linear):
def __init__(self, inchannels, outchannels, kernel_size=1, bias=True):
super().__init__(inchannels, outchannels, bias=bias)
assert kernel_size == 1
def execute(self, x):
x = x.transpose(0, 2, 1)
x = super().execute(x)
x = x.transpose(0, 2, 1)
return x
def conv2d(x, weight, bias=None, stride=1, padding=0, dilation=1, groups=1):
''' Applies a 2D convolution over an input signal composed of several input planes.
:param x: the input image
:type x: jt.Var
:param weight: the convolution kernel
:type weight: jt.Var
:param bias: the bias after convolution
:type bias: jt,Var, optional
:param stride: Stride of the convolution. Default: 1
:type stride: int or tuple, optional
:param padding: Padding added to all four sides of the input. Default: 0
:type padding: int or tuple, optional
:param dilation: Spacing between kernel elements. Default: 1
:type dilation: int or tuple, optional
:param groups: Number of blocked connections from input channels to output channels. Default: 1
:type groups: int, optional
Example:
>>> x = jt.randn(4, 24, 100, 100)
>>> w = jt.randn(32, 24, 3, 3)
>>> y = nn.conv2d(x, w)
'''
padding = _pair(padding)
stride = _pair(stride)
dilation = _pair(dilation)
out_channels = weight.shape[0]
if groups == 1:
N,C,H,W = x.shape
Kh, Kw = weight.shape[-2:]
oh = (H+padding[0]*2-Kh*dilation[0]+dilation[0]-1)//stride[0]+1
ow = (W+padding[1]*2-Kw*dilation[1]+dilation[1]-1)//stride[1]+1
with jt.flag_scope(amp_reg = jt.flags.amp_reg | 4):
xx = x.reindex([N,out_channels,C,oh,ow,Kh,Kw], [
'i0', # Nid
'i2', # Cid
f'i3*{stride[0]}-{padding[0]}+i5*{dilation[0]}', # Hid+Khid
f'i4*{stride[1]}-{padding[1]}+i6*{dilation[1]}', # Wid+KWid
])
ww = weight.broadcast(xx.shape, [0,3,4])
yy = xx*ww
y = yy.sum([2,5,6]) # Kc, Kh, Kw
if bias is not None:
b = bias.broadcast(y.shape, [0,2,3])
y = y + b
return y
else:
N,C,H,W = x.shape
Kh, Kw = weight.shape[-2:]
G = groups
CpG = C // G # channels per group
oc = out_channels
oh = (H+padding[0]*2-Kh*dilation[0]+dilation[0]-1)//stride[0]+1
ow = (W+padding[1]*2-Kw*dilation[1]+dilation[1]-1)//stride[1]+1
xx = x.reindex([N,G,oc//G,CpG,oh,ow,Kh,Kw], [
'i0', # Nid
f'i1*{CpG}+i3', # Gid
f'i4*{stride[0]}-{padding[0]}+i6*{dilation[0]}', # Hid+Khid
f'i5*{stride[1]}-{padding[1]}+i7*{dilation[1]}', # Wid+KWid
])
xx.compile_options = {"G":G}
# w: [oc, CpG, Kh, Kw]
ww = weight.reindex([N, G, oc//G, CpG, oh, ow, Kh, Kw], [
f'i1*{oc//G}+i2',
'i3',
'i6',
'i7'
])
yy = xx*ww
y = yy.reindex_reduce('add', [N, oc, oh, ow], [
'i0',
f'i1*{oc//G}+i2',
'i4',
'i5'
])
if bias is not None:
b = bias.broadcast(y.shape, [0,2,3])
y = y + b
return y
conv = conv2d
def conv3d(x, weight, bias=None, stride=1, padding=0, dilation=1, groups=1):
''' Applies a 3D convolution over an input signal composed of several input planes.
:param x: the input volume
:type x: jt.Var
:param weight: the convolution kernel
:type weight: jt.Var
:param bias: the bias after convolution
:type bias: jt,Var, optional
:param stride: Stride of the convolution. Default: 1
:type stride: int or tuple, optional
:param padding: Padding added to all four sides of the input. Default: 0
:type padding: int or tuple, optional
:param dilation: Spacing between kernel elements. Default: 1
:type dilation: int or tuple, optional
:param groups: Number of blocked connections from input channels to output channels. Default: 1
:type groups: int, optional
Example:
>>> x = jt.randn(4, 24, 50, 50, 50)
>>> w = jt.randn(32, 24, 3, 3, 3)
>>> y = nn.conv2d(x, w)
'''
padding = _triple(padding)
stride = _triple(stride)
dilation = _triple(dilation)
out_channels = weight.shape[0]
if jt.flags.use_cuda and jt.cudnn:
y = jt.cudnn.ops.cudnn_conv3d(x, weight, *stride, *padding, *dilation, groups)
elif groups == 1:
N,C,D,H,W = x.shape
Kd, Kh, Kw = weight.shape[-3:]
od = (D+padding[0]*2-Kd*dilation[0]+dilation[0]-1)//stride[0]+1
oh = (H+padding[1]*2-Kh*dilation[1]+dilation[1]-1)//stride[1]+1
ow = (W+padding[2]*2-Kw*dilation[2]+dilation[2]-1)//stride[2]+1
xx = x.reindex([N,out_channels,C,od,oh,ow,Kd,Kh,Kw], [
'i0', # Nid
'i2', # Cid
f'i3*{stride[0]}-{padding[0]}+i6*{dilation[0]}', # Hid+Khid
f'i4*{stride[1]}-{padding[1]}+i7*{dilation[1]}', # Wid+KWid
f'i5*{stride[2]}-{padding[2]}+i8*{dilation[2]}', # Did+KDid
])
ww = weight.broadcast(xx.shape, [0,3,4,5])
yy = xx*ww
y = yy.sum([2,6,7,8]) # Kc, Kh, Kw,Kd
else:
N,C,D,H,W = x.shape
Kd, Kh, Kw = weight.shape[-3:]
G = groups
CpG = C // G # channels per group
oc = out_channels
od = (D+padding[0]*2-Kd*dilation[0]+dilation[0]-1)//stride[0]+1
oh = (H+padding[1]*2-Kh*dilation[1]+dilation[1]-1)//stride[1]+1
ow = (W+padding[2]*2-Kw*dilation[2]+dilation[2]-1)//stride[2]+1
xx = x.reindex([N,G,oc//G,CpG,od,oh,ow,Kd,Kh,Kw], [
'i0', # Nid
f'i1*{CpG}+i3', # Gid
f'i4*{stride[0]}-{padding[0]}+i7*{dilation[0]}', # Hid+Khid
f'i5*{stride[1]}-{padding[1]}+i8*{dilation[1]}', # Wid+KWid
f'i6*{stride[2]}-{padding[2]}+i9*{dilation[2]}', # Did+KDid
])
xx.compile_options = {"G":G}
# w: [oc, CpG, Kh, Kw, Kd]
ww = weight.reindex([N, G, oc//G, CpG, oh, ow, od, Kh, Kw, Kd], [
f'i1*{oc//G}+i2',
'i3',
'i7',
'i8',
'i9'
])
yy = xx*ww
y = yy.reindex_reduce('add', [N, oc, od, oh, ow], [
'i0',
f'i1*{oc//G}+i2',
'i4',
'i5',
'i6'
])
if bias is not None:
b = bias.broadcast(y.shape, [0,2,3,4])
y = y + b
return y
class ConvTranspose(Module):
def __init__(self, in_channels, out_channels, kernel_size, stride=1, \
padding=0, output_padding=0, groups=1, bias=True, dilation=1):
self.in_channels = in_channels
self.out_channels = out_channels
# added
self.dilation = dilation
self.groups = groups
self.kernel_size = kernel_size if isinstance(kernel_size, tuple) else (kernel_size, kernel_size)
self.stride = stride if isinstance(stride, tuple) else (stride, stride)
self.dilation = dilation if isinstance(dilation, tuple) else (dilation, dilation)
# added
self.padding = padding if isinstance(padding, tuple) else (padding, padding)
self.real_padding = (self.dilation[0] * (self.kernel_size[0] - 1) - self.padding[0],
self.dilation[1] * (self.kernel_size[1] - 1) - self.padding[1])
self.output_padding = output_padding if isinstance (output_padding, tuple) else (output_padding, output_padding)
assert self.output_padding[0] < max(self.stride[0], self.dilation[0]) and \
self.output_padding[1] < max(self.stride[1], self.dilation[1]), \
"output padding must be smaller than max(stride, dilation)"
assert in_channels % groups == 0, 'in_channels must be divisible by groups'
assert out_channels % groups == 0, 'out_channels must be divisible by groups'
self.weight = init.invariant_uniform((in_channels, out_channels//groups) + self.kernel_size, dtype="float")
if bias:
fan=1
for i in self.weight.shape[1:]:
fan *= i
bound = 1 / math.sqrt(fan)
self.bias = init.uniform([out_channels], dtype="float", low=-bound, high=bound)
else:
self.bias = None
def execute(self, x):
if self.groups == 1:
N,C,H,W = x.shape
i,o,h,w = self.weight.shape
assert C==i
stride_h, stride_w = self.stride
padding_h, padding_w = self.padding
dilation_h, dilation_w = self.dilation
h_out = (H-1) * stride_h + self.output_padding[0] - 2*padding_h + 1 + (h-1)*dilation_h
w_out = (W-1) * stride_w + self.output_padding[1] - 2*padding_w + 1 + (w-1)*dilation_w
out_shape = (N, o, h_out, w_out)
shape = (N, i, o, H, W, h, w)
xx = x.broadcast(shape, (2, 5, 6)) # i,h,w
ww = self.weight.broadcast(shape, (0, 3, 4)) # N,H,W
y = (ww*xx).reindex_reduce("add", out_shape, [
'i0', # N
'i2', # o
f'i3*{stride_h}-{padding_h}+i5*{dilation_h}', # Hid+Khid
f'i4*{stride_w}-{padding_w}+i6*{dilation_w}', # Wid+KWid
])
if self.bias is not None:
b = self.bias.broadcast(y.shape, [0,2,3])
y = y + b
return y
else:
N,C,H,W = x.shape
Kh, Kw = self.kernel_size
i,o,h,w = self.weight.shape
oc = self.out_channels
G = self.groups
CpG = C // G # channels per group
assert C==self.in_channels
stride_h, stride_w = self.stride
padding_h, padding_w = self.padding
dilation_h, dilation_w = self.dilation
oh = (H-1) * stride_h + self.output_padding[0] - 2*padding_h + 1 + (h-1)*dilation_h
ow = (W-1) * stride_w + self.output_padding[1] - 2*padding_w + 1 + (w-1)*dilation_w
out_shape = (N, oc, oh, ow)
shape = [N,G,oc//G,CpG,oh,ow,Kh,Kw]
xx = x.reindex(shape, [
'i0',
f'i1*{oc//G}+i2',
'i4',
'i5'
])
ww = self.weight.reindex(shape, [
f'i1*{oc//G}+i2',
'i3',
'i6',
'i7'
])
ww.compile_options = xx.compile_options = {"G":G,"C":C}
y = (ww*xx).reindex_reduce("add", out_shape, [
'i0', # Nid
f'i1*{CpG}+i3', # Gid
f'i4*{self.stride[0]}-{self.padding[0]}+i6*{self.dilation[0]}', # Hid+Khid
f'i5*{self.stride[1]}-{self.padding[1]}+i7*{self.dilation[1]}', # Wid+KWid
])
if self.bias is not None:
b = self.bias.broadcast(y.shape, [0,2,3])
y = y + b
return y
ConvTranspose2d = ConvTranspose
class ConvTranspose3d(Module):
def __init__(self, in_channels, out_channels, kernel_size, stride=1, \
padding=0, output_padding=0, groups=1, bias=True, dilation=1):
self.in_channels = in_channels
self.out_channels = out_channels
# added
self.dilation = dilation
self.group = groups
assert groups==1, "Group conv not supported yet."
self.kernel_size = kernel_size if isinstance(kernel_size, tuple) else (kernel_size, kernel_size, kernel_size)
self.stride = stride if isinstance(stride, tuple) else (stride, stride, stride)
self.dilation = dilation if isinstance(dilation, tuple) else (dilation, dilation, dilation)
# added
self.padding = padding if isinstance(padding, tuple) else (padding, padding, padding)
self.real_padding = (
self.dilation[0] * (self.kernel_size[0] - 1) - self.padding[0],
self.dilation[1] * (self.kernel_size[1] - 1) - self.padding[1],
self.dilation[2] * (self.kernel_size[2] - 1) - self.padding[2])
self.output_padding = output_padding if isinstance (output_padding, tuple) else (output_padding, output_padding, output_padding)
assert self.output_padding[0] < max(self.stride[0], self.dilation[0]) and \
self.output_padding[1] < max(self.stride[1], self.dilation[1]) and \
self.output_padding[2] < max(self.stride[2], self.dilation[2]), \
"output padding must be smaller than max(stride, dilation)"
self.weight = init.invariant_uniform((in_channels, out_channels) + self.kernel_size, dtype="float")
if bias:
fan=1
for i in self.weight.shape[1:]:
fan *= i
bound = 1 / math.sqrt(fan)
self.bias = init.uniform([out_channels], dtype="float", low=-bound, high=bound)
else:
self.bias = None
def execute(self, x):
return conv_transpose3d(x, self.weight, self.bias, self.stride, self.padding, self.output_padding, self.group, self.dilation)
def conv_transpose(input, weight, bias=None, stride=1, padding=0, output_padding=0, groups=1, dilation=1):
if groups == 1:
x = input
N,C,H,W = x.shape
i,o,h,w = weight.shape
assert C==i
stride = stride if isinstance(stride, tuple) else (stride, stride)
dilation = dilation if isinstance(dilation, tuple) else (dilation, dilation)
# added
padding = padding if isinstance(padding, tuple) else (padding, padding)
output_padding = output_padding if isinstance (output_padding, tuple) else (output_padding, output_padding)
assert output_padding[0] < max(stride[0], dilation[0]) and \
output_padding[1] < max(stride[1], dilation[1]), \
"output padding must be smaller than max(stride, dilation)"
stride_h, stride_w = stride
padding_h, padding_w = padding
dilation_h, dilation_w = dilation
h_out = (H-1) * stride_h + output_padding[0] - 2*padding_h + 1 + (h-1)*dilation_h
w_out = (W-1) * stride_w + output_padding[1] - 2*padding_w + 1 + (w-1)*dilation_w
out_shape = (N, o, h_out, w_out)
shape = (N, i, o, H, W, h, w)
xx = x.broadcast(shape, (2, 5, 6)) # i,h,w
ww = weight.broadcast(shape, (0, 3, 4)) # N,H,W
y = (ww*xx).reindex_reduce("add", out_shape, [
'i0', # N
'i2', # o
f'i3*{stride_h}-{padding_h}+i5*{dilation_h}', # Hid+Khid
f'i4*{stride_w}-{padding_w}+i6*{dilation_w}', # Wid+KWid
])
if isinstance(bias, jt.Var):
b = bias.broadcast(y.shape, [0,2,3])
y = y + b
else:
assert not bias, "Bias should be none or jittor var"
return y
else:
N,C,H,W = input.shape
i,o,h,w = weight.shape
G = groups
oc = o * G
CpG = C // G # channels per group
assert C % G == 0
assert C==i, (C, i)
stride = stride if isinstance(stride, tuple) else (stride, stride)
dilation = dilation if isinstance(dilation, tuple) else (dilation, dilation)
# added
padding = padding if isinstance(padding, tuple) else (padding, padding)
output_padding = output_padding if isinstance (output_padding, tuple) else (output_padding, output_padding)
assert output_padding[0] < max(stride[0], dilation[0]) and \
output_padding[1] < max(stride[1], dilation[1]), \
"output padding must be smaller than max(stride, dilation)"
stride_h, stride_w = stride
padding_h, padding_w = padding
dilation_h, dilation_w = dilation
oh = (H-1) * stride_h + output_padding[0] - 2*padding_h + 1 + (h-1)*dilation_h
ow = (W-1) * stride_w + output_padding[1] - 2*padding_w + 1 + (w-1)*dilation_w
out_shape = (N, oc, oh, ow)
shape = [N,G,oc//G,CpG,oh,ow,h,w]
xx = input.reindex(shape, [
'i0',
f'i1*{oc//G}+i2',
'i4',
'i5'
])
ww = weight.reindex(shape, [
f'i1*{oc//G}+i2',
'i3',
'i6',
'i7'
])
ww.compile_options = xx.compile_options = {"G":G,"C":C}
y = (ww*xx).reindex_reduce("add", out_shape, [
'i0', # Nid
f'i1*{CpG}+i3', # Gid
f'i4*{stride[0]}-{padding[0]}+i6*{dilation[0]}', # Hid+Khid
f'i5*{stride[1]}-{padding[1]}+i7*{dilation[1]}', # Wid+KWid
])
if bias is not None:
b = bias.broadcast(y.shape, [0,2,3])
y = y + b
return y
conv_transpose2d = conv_transpose
def conv_transpose3d(input, weight, bias=None, stride=1, padding=0, output_padding=0, groups=1, dilation=1):
x = input
N,C,D,H,W = x.shape
i,o,d,h,w = weight.shape
assert C==i
assert groups==1, "Group conv not supported yet."
stride = stride if isinstance(stride, tuple) else (stride, stride, stride)
dilation = dilation if isinstance(dilation, tuple) else (dilation, dilation, dilation)
# added
padding = padding if isinstance(padding, tuple) else (padding, padding, padding)
output_padding = output_padding if isinstance (output_padding, tuple) else (output_padding, output_padding, output_padding)
assert output_padding[0] < max(stride[0], dilation[0]) and \
output_padding[1] < max(stride[1], dilation[1]) and \
output_padding[2] < max(stride[2], dilation[2]), \
"output padding must be smaller than max(stride, dilation)"
stride_d, stride_h, stride_w = stride
padding_d, padding_h, padding_w = padding
dilation_d, dilation_h, dilation_w = dilation
d_out = (D-1) * stride_d + output_padding[0] - 2*padding_d + 1 + (d-1)*dilation_d
h_out = (H-1) * stride_h + output_padding[1] - 2*padding_h + 1 + (h-1)*dilation_h
w_out = (W-1) * stride_w + output_padding[2] - 2*padding_w + 1 + (w-1)*dilation_w
out_shape = (N, o, d_out, h_out, w_out)
if jt.flags.use_cuda and jt.cudnn:
return jt.cudnn.ops.cudnn_conv3d_backward_x(weight, x, *out_shape[2:], *stride, *padding, *dilation, groups)
shape = (N, i, o, D, H, W, d, h, w)
xx = x.broadcast(shape, (2, 6, 7, 8)) # i,h,w
ww = weight.broadcast(shape, (0, 3, 4, 5)) # N,H,W
y = (ww*xx).reindex_reduce("add", out_shape, [
'i0', # N
'i2', # o
f'i3*{stride_d}-{padding_d}+i6*{dilation_d}', # Did+Kdid
f'i4*{stride_h}-{padding_h}+i7*{dilation_h}', # Hid+Khid
f'i5*{stride_w}-{padding_w}+i8*{dilation_w}', # Wid+KWid
])
if isinstance(bias, jt.Var):
b = bias.broadcast(y.shape, [0,2,3,4])
y = y + b
else:
assert not bias, "Bias should be none or jittor var"
return y
conv_transpose2d = conv_transpose
def pad(x,padding, mode='constant', value=0):
assert mode in ['constant','replicate','reflect','circular'],'only support constant,replicate,reflect,circular pad'
assert len(padding)%2==0 and len(padding)//2<=x.ndim
padding = list(padding)
left = [0]*(x.ndim-len(padding)//2)+padding[::2][::-1]
right = [0]*(x.ndim-len(padding)//2)+padding[1::2][::-1]
out_dims = []
out_shape = []
for i,n,l,r in zip(range(x.ndim),x.shape,left,right):
out_shape.append(n+l+r)
if mode == 'constant':
out_dims.append(f'i{i}-{l}')
elif mode == 'replicate':
out_dims.append(f"i{i}<{l} ? 0 : i{i} > {n+l-1} ? {n-1} : i{i}-{l}")
elif mode == 'reflect':
out_dims.append(f"i{i}<{l} ? {l}-i{i} : i{i} > {n+l-1} ? {2*(n-1)+l}-i{i} : i{i}-{l}")
elif mode == 'circular':
out_dims.append(f"i{i}<{l} ? {n-l}+i{i} : i{i} > {n+l-1} ? i{i}-{n+l} : i{i}-{l}")
return x.reindex(out_shape,out_dims,overflow_value=value)
class ReflectionPad2d(Module):
def __init__(self, padding):
self.padding = padding
if isinstance(self.padding, int):
self.pl = self.padding
self.pr = self.padding
self.pt = self.padding
self.pb = self.padding
elif isinstance(self.padding, tuple):
self.pl, self.pr, self.pt, self.pb = self.padding
else:
raise TypeError(f"ReflectionPad2d padding just support int or tuple, but found {type(padding)}")
def execute(self, x):
n,c,h,w = x.shape
assert (self.pl < w and self.pr < w), f"padding_left and padding_right should be smaller than input width"
assert (self.pt < h and self.pb < h), f"padding_top and padding_bottom should be smaller than input height"
oh=h+self.pt+self.pb
ow=w+self.pl+self.pr
l = self.pl
r = self.pl + w - 1
t = self.pt
b = self.pt + h - 1
return x.reindex([n,c,oh,ow], ["i0","i1",
f"i2<{t} ? {t}-i2 : i2 > {b} ? {h-1+b}-i2 : i2-{t}",
f"i3<{l} ? {l}-i3 : i3 > {r} ? {w-1+r}-i3 : i3-{l}",
])
class ZeroPad2d(Module):
def __init__(self, padding):
self.padding = padding
if isinstance(self.padding, int):
self.pl = self.padding
self.pr = self.padding
self.pt = self.padding
self.pb = self.padding
elif isinstance(self.padding, (tuple,list)):
self.pl, self.pr, self.pt, self.pb = self.padding
else:
raise TypeError(f"ZeroPad2d padding just support int or tuple, but found {type(padding)}")
def execute(self, x):
n,c,h,w = x.shape
return x.reindex([n,c,h+self.pt+self.pb,w+self.pl+self.pr], ["i0","i1",f"i2-{self.pt}",f"i3-{self.pl}"])
class ConstantPad2d(Module):
def __init__(self, padding, value):
self.padding = padding
if isinstance(self.padding, int):
self.pl = self.padding
self.pr = self.padding
self.pt = self.padding
self.pb = self.padding
elif isinstance(self.padding, tuple):
self.pl, self.pr, self.pt, self.pb = self.padding
else:
raise TypeError(f"ConstantPad2d padding just support int or tuple, but found {type(padding)}")
self.value = value
def execute(self, x):
assert len(x.shape) >= 2
shape = x.shape
tar_shape = shape[0:-2] + [shape[-2]+self.pt+self.pb,shape[-1]+self.pl+self.pr]
tar_dims = []
for i in range(len(shape)-2):
tar_dims.append(f"i{i}")
tar_dims.append(f"i{i+1}-{self.pt}")
tar_dims.append(f"i{i+2}-{self.pl}")
return x.reindex(tar_shape, tar_dims, overflow_value=self.value)
class ReplicationPad2d(Module):
def __init__(self, padding):
self.padding = padding
if isinstance(self.padding, int):
self.pl = self.padding
self.pr = self.padding
self.pt = self.padding
self.pb = self.padding
elif isinstance(self.padding, tuple):
self.pl, self.pr, self.pt, self.pb = self.padding
else:
raise TypeError(f"ReplicationPad2d padding just support int or tuple, but found {type(padding)}")
def execute(self, x):
n,c,h,w = x.shape
oh=h+self.pt+self.pb
ow=w+self.pl+self.pr
l = self.pl
r = self.pl + w - 1
t = self.pt
b = self.pt + h - 1
return x.reindex([n,c,oh,ow], ["i0","i1",
f"i2<{t} ? 0 : i2 > {b} ? {h-1} : i2-{t}",
f"i3<{l} ? 0 : i3 > {r} ? {w-1} : i3-{l}"
])
class Embedding(Module):
''' A simple lookup table that stores embeddings of a fixed dictionary and size.
:param num: size of the dictionary of embeddings
:type num: int
:param dim: the size of each embedding vector
:type dim: int
Example:
>>> embedding = nn.Embedding(10, 3)
>>> x = jt.int32([1, 2, 3, 3])
>>> embedding(x)
jt.Var([[ 1.1128596 0.19169547 0.706642]
[ 1.2047412 1.9668795 0.9932192]
[ 0.14941819 0.57047683 -1.3217674]
[ 0.14941819 0.57047683 -1.3217674]], dtype=float32)
'''
def __init__(self, num_embeddings, embedding_dim, dtype="float32"):
self.num = num_embeddings
self.dim = embedding_dim
self.weight = jt.init.gauss([self.num, self.dim], dtype).stop_grad()
def execute(self, x):
res = self.weight[x]
return res
def embedding(input, weight):
return weight[input]
class PixelShuffle(Module):
def __init__(self, upscale_factor):
self.upscale_factor = upscale_factor
def execute(self, x):
n,c,h,w = x.shape
r = self.upscale_factor
assert c%(r*r)==0, f"input channel needs to be divided by upscale_factor's square in PixelShuffle"
return x.reindex([n,int(c/r**2),h*r,w*r], [
"i0",
f"i1*{r*r}+i2%{r}*{r}+i3%{r}",
f"i2/{r}",
f"i3/{r}"
])
class Tanh(Module):
def __init__(self):
super().__init__()
def execute(self, x) :
return x.tanh()
class Sigmoid(Module):
def __init__(self):
super().__init__()
def execute(self, x) :
return x.sigmoid()
def softplus(x,beta=1.0,threshold=20.0):
return 1 / beta * jt.log(1 + (beta * x).minimum(threshold).exp()) + \
(x - threshold/beta).maximum(0.0)
def hardtanh(x,min_val=-1,max_val=1):
return jt.clamp(x,min_v=min_val,max_v=max_val)
class Softplus(Module):
r'''
SoftPlus is a smooth approximation to the ReLU function and can be used to constrain the output of a machine to always be positive.
Args:
[in] beta (float): the beta value for the Softplus formulation. Default: 1.
[in] threshold (float): values above this revert to a linear function. Default: 20.
'''
def __init__(self, beta=1, threshold=20):
self.beta = beta
self.threshold = threshold
def execute(self, x):
return softplus(x, self.beta, self.threshold)
class Resize(Module):
def __init__(self, size, mode="nearest", align_corners=False):
super().__init__()
self.size = size
self.mode = mode
self.align_corners = align_corners
def execute(self, x):
return resize(x, self.size, self.mode, self.align_corners)
def _bicubic(x, a, func):
# normal ver
if func == 1:
return (a+2)*(jt.abs(x)**3)-(a+3)*(x**2)+1
if func == 2:
return a*(jt.abs(x)**3)-5*a*(x**2)+8*a*(jt.abs(x))-4*a
return 0
def _interpolate(img, x, y, ids, mode):
if mode == "nearest":
return img.reindex([*ids, x.floor_int(), y.floor_int()])
if mode == "bilinear":
fx, fy = x.floor_int(), y.floor_int()
cx, cy = fx + 1, fy + 1
dx, dy = x - fx, y - fy
a = img.reindex_var([*ids, fx, fy])
b = img.reindex_var([*ids, cx, fy])
c = img.reindex_var([*ids, fx, cy])
d = img.reindex_var([*ids, cx, cy])
dnx, dny = 1 - dx, 1 - dy
ab = dx * b + dnx * a
cd = dx * d + dnx * c
o = ab * dny + cd * dy
return o
if mode=="bicubic": # ugly ver.
n,c,h,w = img.shape
fx, fy = x.floor_int(), y.floor_int()
dix, diy = x - fx, y - fy
ax, ay = _bicubic(dix+1,-0.75,2), _bicubic(diy+1,-0.75,2)
bx, by = _bicubic(dix,-0.75,1), _bicubic(diy,-0.75,1)
cx, cy = _bicubic(1-dix,-0.75,1), _bicubic(1-diy,-0.75,1)
dx, dy = _bicubic(2-dix,-0.75,2), _bicubic(2-diy,-0.75,2)
afx, afy = jt.maximum(jt.minimum(fx-1,h-1),0), jt.maximum(jt.minimum(fy-1,w-1),0)
bfx, bfy = jt.maximum(jt.minimum(fx,h-1),0), jt.maximum(jt.minimum(fy,w-1),0)
cfx, cfy = jt.maximum(jt.minimum(fx+1,h-1),0), jt.maximum(jt.minimum(fy+1,w-1),0)
dfx, dfy = jt.maximum(jt.minimum(fx+2,h-1),0), jt.maximum(jt.minimum(fy+2,w-1),0)
a = ax*(img.reindex_var([*ids,afx,afy])*ay+img.reindex_var([*ids,afx,bfy])*by+img.reindex_var([*ids,afx,cfy])*cy+img.reindex_var([*ids,afx,dfy])*dy)
b = bx*(img.reindex_var([*ids,bfx,afy])*ay+img.reindex_var([*ids,bfx,bfy])*by+img.reindex_var([*ids,bfx,cfy])*cy+img.reindex_var([*ids,bfx,dfy])*dy)
c = cx*(img.reindex_var([*ids,cfx,afy])*ay+img.reindex_var([*ids,cfx,bfy])*by+img.reindex_var([*ids,cfx,cfy])*cy+img.reindex_var([*ids,cfx,dfy])*dy)
d = dx*(img.reindex_var([*ids,dfx,afy])*ay+img.reindex_var([*ids,dfx,bfy])*by+img.reindex_var([*ids,dfx,cfy])*cy+img.reindex_var([*ids,dfx,dfy])*dy)
o = a + b + c + d
return o
raise (f"Not support interpolation mode: {mode}")
# TODO: tf_mode to another function
def resize(img, size, mode="nearest", align_corners=False, tf_mode=False):
n, c, h, w = img.shape
H, W = size
nid, cid, hid, wid = jt.index((n, c, H, W))
if align_corners:
x = hid * ((h - 1) / max(1, H - 1))
y = wid * ((w - 1) / max(1, W - 1))
elif mode == "bicubic":
x = (hid + 0.5) * (h / H) - 0.5
y = (wid + 0.5) * (w / W) - 0.5
elif mode == 'nearest':
x = hid * (h / H)
y = wid * (w / W)
elif mode == "area":
'''
Area interpolation uses AdaptivePool2D to resize origin images.
'''
stride = (h // H, w // W)
assert stride[0] > 0 and stride[1] > 0
x, y = jt.meshgrid(jt.arange(0, H, 1), jt.arange(0, W, 1))
startH = jt.floor(x*h/H).int32()
endH = jt.ceil((x+1)*h/H).int32()
maxH = int(jt.max(endH - startH).data)
startW = jt.floor(y*w/W).int32()
endW = jt.ceil((y+1)*w/W).int32()
maxW = int(jt.max(endW - startW).data)
pixel_count = (endH - startH) * (endW - startW)
adaptive_output = img.reindex([img.shape[0], img.shape[1], H, W, maxH, maxW], ["i0", "i1", "@e0(i2, i3) + i4", "@e2(i2, i3) + i5"], extras=[startH, endH, startW, endW], overflow_conditions=["i4 >= @e1(i2, i3) - @e0(i2, i3)", "i5 >= @e3(i2, i3) - @e2(i2, i3)"], overflow_value=0)
adaptive_output = adaptive_output.reduce("sum", [4,5]) / pixel_count[None, None, ...]
return adaptive_output
else:
if (tf_mode):
x = hid * (h / H)
if H > h: x = x.clamp(0, h - 1)
y = wid * (w / W)
if W > w: y = y.clamp(0, w - 1)
else:
x = hid * (h / H) + (h / H * 0.5 - 0.5)
if H > h: x = x.clamp(0, h - 1)
y = wid * (w / W) + (w / W * 0.5 - 0.5)
if W > w: y = y.clamp(0, w - 1)
return _interpolate(img, x, y, (nid, cid), mode)
upsample = resize
def interpolate(X, size=None, scale_factor=None, mode='bilinear', align_corners=False, tf_mode=False):
if scale_factor is not None:
size = [int(X.shape[-2] * scale_factor), int(X.shape[-1] * scale_factor)]
if isinstance(size, int):
size = (size, size)
if scale_factor is not None and scale_factor > 1:
return upsample(X, size, mode, align_corners, tf_mode)
else:
return resize(X, size, mode, align_corners, tf_mode)
def grid_sample_v0(input, grid, mode='bilinear', padding_mode='zeros'):
r'''
Given an input and a flow-field grid, computes the output using input values and pixel locations from grid.
grid specifies the sampling pixel locations normalized by the input spatial dimensions. Therefore, it should have most values in the range of [-1, 1]. For example, values x = -1, y = -1 is the left-top pixel of input, and values x = 1, y = 1 is the right-bottom pixel of input.
Args:
[in] input (var): the source input var, whose shape is (N, C, Hi, Wi)
[in] grid (var): the pixel locations, whose shape is (N, Ho, Wo, 2)
[in] mode (string): the interpolate way, default: bilinear.
[in] padding_mode (string): the padding way, default: zeros.
[out] output (var): the output var, whose shape is (N, C, Ho, Wo)
Example:
>>> x = jt.array([[[[1,2],[3,4]]]])
>>> print(x)
[[[[1 2]
[3 4]]]]
>>> grid = jt.array([[[[0.5, 0.5]]]])
>>> print(x.shape, grid.shape)
[1,1,2,2,], [1,1,2,2,]
>>> nn.grid_sample(x, grid)
[[[[3.25]]]]
'''
assert padding_mode == 'zeros'
Ni, Ci, Hi, Wi = input.shape
No, Ho, Wo, D = grid.shape
assert D == 2
assert Ni == No
assert len(input.shape) == 4 and len(grid.shape)
nid, cid, hid, wid = jt.index((Ni, Ci, Ho, Wo))
x = ((grid[:, :, :, 1].unsqueeze(1).repeat([1, Ci, 1, 1]) + 1) / 2) * (Hi - 1)
y = ((grid[:, :, :, 0].unsqueeze(1).repeat([1, Ci, 1, 1]) + 1) / 2) * (Wi - 1)
return _interpolate(input, x, y, (nid, cid), mode)
def linspace_from_neg_one(grid,num_steps,align_corners):
if num_steps <= 1:
return jt.array([],dtype=grid.dtype)
# TODO: use jt.index
ra = np.linspace(-1,1,num_steps)
if not align_corners:
ra = ra*(num_steps-1)/num_steps
return jt.array(ra,dtype=grid.dtype)
def make_base_grid_4D(theta,N,C,H,W,align_corners):
base_grid = jt.zeros((N, H, W, 3), dtype=theta.dtype)
base_grid[...,0] = linspace_from_neg_one(theta, W, align_corners)
base_grid[...,1] = jt.unsqueeze(linspace_from_neg_one(theta, H, align_corners),-1)
base_grid[...,-1] = 1
return base_grid
def make_base_grid_5D(theta,N,C,D,H,W,align_corners):
base_grid = jt.zeros((N, D, H, W, 4), dtype=theta.dtype)
base_grid[...,0] = linspace_from_neg_one(theta, W, align_corners)
base_grid[...,1] = jt.unsqueeze(linspace_from_neg_one(theta, H, align_corners),-1)
base_grid[...,2] = jt.unsqueeze(jt.unsqueeze(linspace_from_neg_one(theta, D, align_corners),-1),-1)
base_grid[...,-1] = 1
return base_grid
def affine_grid_generator_4D(theta,N,C,H,W,align_corners):
base_grid = make_base_grid_4D(theta, N, C, H, W, align_corners)
grid = jt.nn.bmm(base_grid.reshape(N, H * W, 3),theta.transpose(0,2,1))
return grid.reshape(N, H, W, 2)
def affine_grid_generator_5D(theta,N,C,D,H,W,align_corners):
base_grid = make_base_grid_5D(theta, N, C, D, H, W, align_corners)
grid = jt.nn.bmm(base_grid.reshape(N, D * H * W, 4),theta.transpose(0,2,1))
return grid.reshape(N, D, H, W, 3)
def affine_grid(theta, size, align_corners=False):
assert str(theta.dtype) in ['float','float32','float64']
assert min(size)>0
assert len(size) in [4,5]
if len(size)== 4:
assert theta.ndim == 3 and theta.shape[-2] == 2 and theta.shape[-1] == 3
return affine_grid_generator_4D(theta, size[0], size[1], size[2], size[3], align_corners)
elif len(size)==5:
assert theta.ndim == 3 and theta.shape[-2] == 3 and theta.shape[-1] == 4
return affine_grid_generator_5D(theta, size[0], size[1], size[2], size[3], size[4], align_corners)
def grid_sampler_unnormalize(coord,size,align_corners):
if align_corners:
#unnormalize coord from [-1, 1] to [0, size - 1]
return ((coord + 1) / 2) * (size - 1)
else:
#unnormalize coord from [-1, 1] to [-0.5, size - 0.5]
return ((coord + 1) * size - 1) / 2
def clip_coordinates(x,clip_limit):
return jt.clamp(x,min_v=0,max_v=clip_limit-1)
def reflect_coordinates(x,twice_low,twice_high):
if twice_low == twice_high:
return jt.zeros_like(x)
m = twice_low / 2
span = (twice_high - twice_low) / 2
x = (x - m).abs()
#`fmod` returns same sign as `in`, which is positive after the `fabs` above.
extra = x.mod(span)
flips = (x / span).floor_int()
result1 = extra+m
result2 = span-extra+m
con = flips%2==0
not_con = flips%2!=0
result1[not_con]=0.0
result2[con]=0.0
return result1+result2
def grid_sampler_compute_source_index(coord,size,padding_mode,align_corners):
coord = grid_sampler_unnormalize(coord, size, align_corners)
if padding_mode == 'border':
#clip coordinates to image borders
coord = clip_coordinates(coord, size)
elif padding_mode == 'reflection':
#reflect coordinates by image borders
if align_corners:
coord = reflect_coordinates(coord, 0, 2*(size - 1))
else:
coord = reflect_coordinates(coord, -1, 2*size - 1)
#clip coordinates to image borders
coord = clip_coordinates(coord, size)
return coord
def grid_sampler_3d(X,grid,mode,padding_mode,align_corners):
N = X.shape[0]
C = X.shape[1]
inp_D = X.shape[2]
inp_H = X.shape[3]
inp_W = X.shape[4]
D = grid.shape[1]
H = grid.shape[2]
W = grid.shape[3]
x = grid[:,:,:,:,0]
y = grid[:,:,:,:,1]
z = grid[:,:,:,:,2]
shape = [N,C,D,H,W]
cid = jt.index(shape, dim=1)
nid = jt.index(shape, dim=0)
x = grid_sampler_compute_source_index(x,inp_W,padding_mode,align_corners)
y = grid_sampler_compute_source_index(y,inp_H,padding_mode,align_corners)
z = grid_sampler_compute_source_index(z,inp_D,padding_mode,align_corners)
xid = x.reindex(shape,['i0','i2','i3','i4'])
yid = y.reindex(shape,['i0','i2','i3','i4'])
zid = z.reindex(shape,['i0','i2','i3','i4'])
if mode=='nearest':
return X.reindex([nid,cid,zid.round_int(),yid.round_int(),xid.round_int()])
elif mode=='bilinear':
fx,fy,fz = xid.floor_int(),yid.floor_int(),zid.floor_int()
cx,cy,cz = fx+1,fy+1,fz+1
dx,dy,dz = xid-fx,yid-fy,zid-fz
dnx,dny,dnz = cx-xid,cy-yid,cz-zid
a = X.reindex([nid,cid,fz,fy,fx])
b = X.reindex([nid,cid,cz,fy,fx])
c = X.reindex([nid,cid,fz,cy,fx])
d = X.reindex([nid,cid,fz,fy,cx])
e = X.reindex([nid,cid,fz,cy,cx])
f = X.reindex([nid,cid,cz,fy,cx])
g = X.reindex([nid,cid,cz,cy,fx])
h = X.reindex([nid,cid,cz,cy,cx])
o = a*dnx*dny*dnz+b*dnx*dny*dz+c*dnx*dy*dnz+d*dx*dny*dnz+e*dx*dy*dnz+f*dx*dny*dz+g*dnx*dy*dz+h*dx*dy*dz
return o
def grid_sampler_2d(X,grid,mode,padding_mode,align_corners):
N = X.shape[0]
C = X.shape[1]
inp_H = X.shape[2]
inp_W = X.shape[3]
H = grid.shape[1]
W = grid.shape[2]
x = grid[:,:,:,0]
y = grid[:,:,:,1]
shape = [N,C,H,W]
cid = jt.index(shape, dim=1)
nid = jt.index(shape, dim=0)
x = grid_sampler_compute_source_index(x,inp_W,padding_mode,align_corners)
y = grid_sampler_compute_source_index(y,inp_H,padding_mode,align_corners)
xid = x.reindex(shape,['i0','i2','i3'])
yid = y.reindex(shape,['i0','i2','i3'])
if mode=='nearest':
return X.reindex([nid,cid,yid.round_int(),xid.round_int()])
elif mode=='bilinear':
#xid,yid = (xid+0.00001),(yid+0.00001)
fx,fy = (xid).floor_int(),(yid).floor_int()
cx,cy = fx+1,fy+1
dx,dy = xid-fx,yid-fy
dnx,dny = cx-xid,cy-yid
a = X.reindex([nid,cid,fy,fx],overflow_value=0.0)
b = X.reindex([nid,cid,cy,fx],overflow_value=0.0)
c = X.reindex([nid,cid,fy,cx],overflow_value=0.0)
d = X.reindex([nid,cid,cy,cx],overflow_value=0.0)
o = a*dnx*dny+b*dnx*dy+c*dx*dny+d*dx*dy
return o
def grid_sampler(X, grid, mode, padding_mode, align_corners):
assert X.dtype==grid.dtype
assert ((X.ndim==4 or X.ndim==5) and X.ndim==grid.ndim)
assert X.shape[0]==grid.shape[0] and grid.shape[-1]==X.ndim-2
assert X.numel()>0
if X.ndim == 4:
return grid_sampler_2d(X, grid, mode, padding_mode, align_corners)
else:
return grid_sampler_3d(X, grid, mode, padding_mode, align_corners)
def grid_sample(input, grid, mode='bilinear', padding_mode='zeros', align_corners=False):
assert mode in ['bilinear','nearest']
assert padding_mode in ['zeros','border','reflection']
return grid_sampler(input, grid, mode, padding_mode, align_corners)
class Upsample(Module):
def __init__(self, scale_factor=None, mode='nearest'):
self.scale_factor = scale_factor if isinstance(scale_factor, tuple) else (scale_factor, scale_factor)
self.mode = mode
def execute(self, x):
return upsample(x,
size=(
int(x.shape[2]*self.scale_factor[0]),
int(x.shape[3]*self.scale_factor[1])),
mode=self.mode)
class UpsamplingBilinear2d(Upsample):
def __init__(self, scale_factor=None):
Upsample.__init__(self, scale_factor, 'bilinear')
class UpsamplingNearest2d(Upsample):
def __init__(self, scale_factor=None):
Upsample.__init__(self, scale_factor, 'nearest')
class Sequential(Module):
def __init__(self, *args):
self.layers = collections.OrderedDict()
for mod in args:
if isinstance(mod, collections.OrderedDict):
for k, m in mod.items():
self.add_module(k, m)
elif isinstance(mod,list):
for m in mod:
self.append(m)
else:
self.append(mod)
def __getitem__(self, idx):
if isinstance(idx, slice) or idx not in self.layers:
return list(self.layers.values())[idx]
return self.layers[idx]
def __iter__(self):
return self.layers.values().__iter__()
def keys(self):
return self.layers.keys()
def values(self):
return self.layers.values()
def items(self):
return self.layers.items()
def execute(self, x):
for k, layer in self.layers.items():
x = layer(x)
return x
def dfs(self, parents, k, callback, callback_leave, recurse=True):
n_children = len(self.layers)
ret = callback(parents, k, self, n_children)
if ret == False:
return
parents.append(self)
if recurse:
for k,v in self.layers.items():
if isinstance(v, Module):
v.dfs(parents, k, callback, callback_leave)
parents.pop()
if callback_leave:
callback_leave(parents, k, self, n_children)
def append(self, mod):
assert callable(mod), f"Module <{type(mod)}> is not callable"
assert not isinstance(mod, type), f"Module is not a type"
self.layers[len(self.layers)]=mod
def add_module(self, name, mod):
assert callable(mod), f"Module <{type(mod)}> is not callable"
assert not isinstance(mod, type), f"Module is not a type"
self.layers[name]=mod
def __len__(self):
return len(self.layers)
class ParameterList(Module):
def __init__(self, *args):
self.params = collections.OrderedDict()
for var in args:
if isinstance(var, (collections.OrderedDict, dict)):
for k, v in var.items():
self.add_param(k, v)
elif isinstance(var, list):
for v in var:
self.append(v)
else:
self.append(var)
def __getitem__(self, idx):
if idx not in self.params:
return list(self.params.values())[idx]
return self.params[idx]
def __iter__(self):
return self.params.values().__iter__()
def keys(self):
return self.params.keys()
def values(self):
return self.params.values()
def items(self):
return self.params.items()
def execute(self, x):
raise NotImplementedError("Parameters is not executable")
def append(self, var):
assert isinstance(var, jt.Var), f"argument <{type(var)}> is not jittor var"
self.params[len(self.params)] = var
def add_param(self, name, var):
assert isinstance(var, jt.Var), f"argument <{type(var)}> is not jittor var"
self.params[name]=var
def __setitem__(self, name, var):
self.add_param(name, var)
def __len__(self):
return len(self.params)
ParameterDict = ParameterList
def Parameter(data, requires_grad=True):
''' The `Parameter` interface isn't needed in Jittor, this interface
does nothings and it is just used for compatible.
A Jittor Var is a Parameter
when it is a member of Module, if you don't want a Jittor
Var menber is treated as a Parameter, just name it startswith
underscore `_`.
'''
LOG.w(Parameter.__doc__)
data = data.clone()
data.requires_grad = requires_grad
return data
def backward(v, *args, **kw):
''' The `backward` variable interface doesn't exist in Jittor.
please use `optimizer.backward(loss)` or
`optimizer.step(loss)` instead.
For example, if your code looks like this::
optimizer.zero_grad()
loss.backward()
optimizer.step()
It can be changed to this::
optimizer.zero_grad()
optimizer.backward(loss)
optimizer.step()
Or more concise::
optimizer.step(loss)
The step function will automatically zero grad and backward.
'''
LOG.f(backward.__doc__)
jt.Var.backward = backward
def unfold(X, kernel_size, dilation=1, padding=0, stride=1):
assert X.ndim == 4
if not isinstance(kernel_size, tuple):
kernel_size = (kernel_size, kernel_size)
if not isinstance(dilation, tuple):
dilation = (dilation, dilation)
if not isinstance(padding, tuple):
padding = (padding, padding)
if not isinstance(stride, tuple):
stride = (stride, stride)
n, c, h, w = X.shape
shape = X.shape
area = kernel_size[0] * kernel_size[1]
block_nums = []
for i in range(2, 4):
block_nums.append(
(shape[i] + 2 * padding[i - 2] - dilation[i - 2] * (kernel_size[i - 2] - 1) - 1) // stride[i - 2] + 1)
if padding[0] != 0 or padding[1] != 0:
X = X.reindex([n, c, h + padding[0] * 2, w + padding[1] * 2],
["i0", "i1", f"i2-{padding[0]}", f"i3-{padding[1]}"])
output = X.reindex([n, c * area, block_nums[0] * block_nums[1]], ["i0", f"i1/{area}",
f"i2/{block_nums[1]}*{stride[0]}+(i1%{area})/{kernel_size[1]}*{dilation[0]}",
f"i2%{block_nums[1]}*{stride[1]}+(i1%{area})%{kernel_size[1]}*{dilation[1]}"])
return output
def fold(X,output_size,kernel_size,dilation=1,padding=0,stride=1):
assert X.ndim==3
if not isinstance(kernel_size,tuple):
kernel_size = (kernel_size,kernel_size)
if not isinstance(dilation,tuple):
dilation = (dilation,dilation)
if not isinstance(padding,tuple):
padding = (padding,padding)
if not isinstance(stride,tuple):
stride = (stride,stride)
n,cl,num = X.shape
area = kernel_size[0] * kernel_size[1]
block_nums = []
for i in range(2,4):
block_nums.append((output_size[i-2]+2*padding[i-2]-dilation[i-2]*(kernel_size[i-2]-1)-1) // stride[i-2]+1)
output = X.reindex_reduce("add",[n,cl // area,output_size[0]+2*padding[0],output_size[1]+2*padding[1]],["i0",f"i1/{area}",f"i2/{block_nums[1]}*{stride[0]}+(i1%{area})/{kernel_size[1]}*{dilation[0]}",f"i2%{block_nums[1]}*{stride[1]}+(i1%{area})%{kernel_size[1]}*{dilation[1]}"])
return output[:,:,padding[0]:padding[0]+output_size[0],padding[1]:padding[1]+output_size[1]]
ModuleList = Sequential
class LSTMCell(jt.Module):
''' A long short-term memory (LSTM) cell.
:param input_size: The number of expected features in the input
:type input_size: int
:param hidden_size: The number of features in the hidden state
:type hidden_size: int
:param bias: If False, then the layer does not use bias weights b_ih and b_hh. Default: True.
:type bias: bool, optional
Example:
>>> rnn = nn.LSTMCell(10, 20) # (input_size, hidden_size)
>>> input = jt.randn(2, 3, 10) # (time_steps, batch, input_size)
>>> hx = jt.randn(3, 20) # (batch, hidden_size)
>>> cx = jt.randn(3, 20)
>>> output = []
>>> for i in range(input.shape[0]):
hx, cx = rnn(input[i], (hx, cx))
output.append(hx)
>>> output = jt.stack(output, dim=0)
'''
def __init__(self, input_size, hidden_size, bias=True):
super().__init__()
self.hidden_size = hidden_size
self.bias = bias
k = math.sqrt(1 / hidden_size)
self.weight_ih = init.uniform((4 * hidden_size, input_size), 'float32', -k, k)
self.weight_hh = init.uniform((4 * hidden_size, hidden_size), 'float32', -k, k)
if bias:
self.bias_ih = init.uniform((4 * hidden_size,), 'float32', -k, k)
self.bias_hh = init.uniform((4 * hidden_size,), 'float32', -k, k)
def execute(self, input, hx = None):
if hx is None:
zeros = jt.zeros((input.shape[0], self.hidden_size), dtype=input.dtype)
h, c = zeros, zeros
else:
h, c = hx
y = matmul_transpose(input, self.weight_ih) + matmul_transpose(h, self.weight_hh)
if self.bias:
y = y + self.bias_ih + self.bias_hh
i = y[:, :self.hidden_size].sigmoid()
f = y[:, self.hidden_size : 2 * self.hidden_size].sigmoid()
g = y[:, 2 * self.hidden_size : 3 * self.hidden_size].tanh()
o = y[:, 3 * self.hidden_size:].sigmoid()
c = f * c + i * g
h = o * c.tanh()
return h, c
class RNNCell(jt.Module):
''' An Elman RNN cell with tanh or ReLU non-linearity.
:param input_size: The number of expected features in the input
:type input_size: int
:param hidden_size: The number of features in the hidden state
:type hidden_size: int
:param bias: If False, then the layer does not use bias weights b_ih and b_hh. Default: True.
:type bias: bool, optional
:param nonlinearity: The non-linearity to use. Can be either 'tanh' or 'relu'. Default: 'tanh'.
:type nonlinearity: str, optional
Example:
>>> rnn = nn.RNNCell(10, 20)
>>> input = jt.randn((6, 3, 10))
>>> hx = jt.randn((3, 20))
>>> output = []
>>> for i in range(6):
hx = rnn(input[i], hx)
output.append(hx)
'''
def __init__(self, input_size, hidden_size, bias=True, nonlinearity = "tanh"):
super().__init__()
self.hidden_size = hidden_size
self.bias = bias
self.nonlinearity = nonlinearity
k = math.sqrt(1 / hidden_size)
self.weight_ih = init.uniform((hidden_size, input_size), 'float32', -k, k)
self.weight_hh = init.uniform((hidden_size, hidden_size), 'float32', -k, k)
if bias:
self.bias_ih = init.uniform((hidden_size,), 'float32', -k, k)
self.bias_hh = init.uniform((hidden_size,), 'float32', -k, k)
def execute(self, input, hx = None):
if hx is None:
hx = jt.zeros((input.shape[0], self.hidden_size), dtype=input.dtype)
y = matmul_transpose(input, self.weight_ih)+matmul_transpose(hx, self.weight_hh)
if self.bias:
y= y + self.bias_ih + self.bias_hh
if self.nonlinearity == 'tanh':
y = y.tanh()
elif self.nonlinearity == 'relu':
y = relu(y)
else:
raise RuntimeError("Unknown nonlinearity: {}".format(self.nonlinearity))
return y
class GRUCell(jt.Module):
''' A gated recurrent unit (GRU) cell.
:param input_size: The number of expected features in the input
:type input_size: int
:param hidden_size: The number of features in the hidden state
:type hidden_size: int
:param bias: If False, then the layer does not use bias weights b_ih and b_hh. Default: True.
:type bias: bool, optional
Example:
>>> rnn = nn.GRUCell(10, 20)
>>> input = jt.randn((6, 3, 10))
>>> hx = jt.randn((3, 20))
>>> output = []
>>> for i in range(6):
hx = rnn(input[i], hx)
output.append(hx)
'''
def __init__(self, input_size, hidden_size, bias=True):
super().__init__()
self.hidden_size = hidden_size
self.bias = bias
k = math.sqrt(1 / hidden_size)
self.weight_ih = init.uniform((3*hidden_size, input_size), 'float32', -k, k)
self.weight_hh = init.uniform((3*hidden_size, hidden_size), 'float32', -k, k)
if bias:
self.bias_ih = init.uniform((3*hidden_size,), 'float32', -k, k)
self.bias_hh = init.uniform((3*hidden_size,), 'float32', -k, k)
def execute(self, input, hx = None):
if hx is None:
hx = jt.zeros((input.shape[0], self.hidden_size), dtype=input.dtype)
gi = matmul_transpose(input, self.weight_ih)
gh = matmul_transpose(hx, self.weight_hh)
if self.bias:
gi += self.bias_ih
gh += self.bias_hh
i_r, i_i, i_n = gi.chunk(3, 1)
h_r, h_i, h_n = gh.chunk(3, 1)
resetgate = jt.sigmoid(i_r + h_r)
inputgate = jt.sigmoid(i_i + h_i)
newgate = jt.tanh(i_n + resetgate * h_n)
hy = newgate + inputgate * (hx - newgate)
return hy
class RNNBase(Module):
def __init__(self, mode: str, input_size: int, hidden_size: int,
num_layers: int = 1, bias: bool = True, batch_first: bool = False,
dropout: float = 0, bidirectional: bool = False,
proj_size: int = 0, nonlinearity: str = None) -> None:
super().__init__()
self.mode = mode
self.input_size = input_size
self.hidden_size = hidden_size
self.num_layers = num_layers
self.bias = bias
self.batch_first = batch_first
self.dropout = dropout
self.bidirectional = bidirectional
self.proj_size = proj_size
self.nonlinearity = nonlinearity
if mode == 'LSTM':
gate_size = 4 * hidden_size
elif mode == 'GRU':
gate_size = 3 * hidden_size
elif mode == 'RNN':
gate_size = hidden_size
else:
raise ValueError("Unrecognized RNN mode: " + mode)
num_directions = 1 + bidirectional
k = math.sqrt(1 / hidden_size)
def build_unit(name, in_channels, out_channels=None):
if out_channels is not None:
shape = (in_channels, out_channels)
else:
shape = (in_channels,)
setattr(self, name, init.uniform(shape, 'float32', -k, k))
if self.bidirectional:
setattr(self, name + '_reverse', init.uniform(shape, 'float32', -k, k))
for layer in range(num_layers):
if layer == 0:
build_unit(f'weight_ih_l{layer}', gate_size, input_size)
else:
if proj_size > 0:
build_unit(f'weight_ih_l{layer}', gate_size, num_directions * proj_size)
else:
build_unit(f'weight_ih_l{layer}', gate_size, num_directions * hidden_size)
if proj_size > 0:
build_unit(f'weight_hh_l{layer}', gate_size, proj_size)
build_unit(f'weight_hr_l{layer}', proj_size, hidden_size)
else:
build_unit(f'weight_hh_l{layer}', gate_size, hidden_size)
if bias:
build_unit(f'bias_ih_l{layer}', gate_size)
build_unit(f'bias_hh_l{layer}', gate_size)
def _cudnn_flatten_weights(self, cudnn_mode):
def copy_to_flatten_weight(param_name, offset_idx, num_gates):
def copy_to(param_name, offset_idx, idx):
cur_offset = self._cudnn_weight_offset[offset_idx]
param = getattr(self, param_name)
param = param[self.hidden_size * idx: self.hidden_size * (idx + 1)]
ft_weight[cur_offset:cur_offset + param.numel()] = param.flatten()
if self.bias:
for idx in range(num_gates):
copy_to('weight' + param_name, offset_idx + idx * 2, idx)
copy_to('bias' + param_name, offset_idx + idx * 2 + 1, idx)
return num_gates * 2
else:
for idx in range(num_gates):
copy_to('weight' + param_name, offset_idx + idx, idx)
return num_gates
if jt.flags.use_cuda and jt.cudnn and jt.compiler.is_cuda:
if getattr(self, '_cudnn_weight_size', None) is None:
offset_array = jt.cudnn.cudnn_rnn_weight_offset(
cudnn_mode,
self.input_size,
self.hidden_size,
self.num_layers,
self.proj_size,
self.bias,
self.bidirectional
)
self._cudnn_weight_size = offset_array[0]
self._cudnn_weight_offset = offset_array[1:]
num_gates = {
"RNN": 1, "LSTM": 4, "GRU": 3
}[self.mode]
ft_weight = jt.zeros(self._cudnn_weight_size, dtype=jt.float32)
cnt = 0
for layer in range(self.num_layers):
suffix = ''
cnt += copy_to_flatten_weight(f'_ih_l{layer}' + suffix, cnt, num_gates)
cnt += copy_to_flatten_weight(f'_hh_l{layer}' + suffix, cnt, num_gates)
if self.bidirectional:
suffix = '_reverse'
cnt += copy_to_flatten_weight(f'_ih_l{layer}' + suffix, cnt, num_gates)
cnt += copy_to_flatten_weight(f'_hh_l{layer}' + suffix, cnt, num_gates)
return ft_weight
else:
raise RuntimeError("Not Cudnn found")
@abstractmethod
def call_rnn_cell(self, input, hidden, suffix):
pass
def call_rnn_sequence(self, input, hidden, suffix):
if 'reverse' in suffix:
input = input[::-1]
output = []
for s in range(input.shape[0]):
out, hidden = self.call_rnn_cell(input[s], hidden, suffix)
output.append(out)
if 'reverse' in suffix:
output = output[::-1]
output = jt.stack(output, dim=0)
return output, hidden
def _execute_cudnn_rnn(self, input, hx):
cudnn_mode = {
('RNN', 'tanh'): 'tanh',
('RNN', 'relu'): 'relu',
('LSTM', None): 'lstm',
('GRU', None): 'gru'
}[(self.mode, self.nonlinearity)]
ft_weight = self._cudnn_flatten_weights(cudnn_mode)
if self.mode == 'LSTM':
ret = jt.cudnn.ops.cudnn_rnn(input, hx[0], hx[1], ft_weight,
cudnn_mode, self.input_size, self.hidden_size, self.num_layers, 0,
self.dropout, self.bias, self.bidirectional, self.is_training()
)
return ret[0], (ret[1], ret[2])
else:
ret = jt.cudnn.ops.cudnn_rnn(input, hx, ft_weight,
cudnn_mode, self.input_size, self.hidden_size, self.num_layers, 0,
self.dropout, self.bias, self.bidirectional, self.is_training()
)
return ret[0], ret[1]
def execute(self, input, hx=None):
if self.batch_first:
input = input.permute(1, 0, 2)
num_directions = 2 if self.bidirectional else 1
if hx is None:
if self.mode in ['RNN', 'GRU']:
hx = jt.zeros((num_directions * self.num_layers, input.shape[1], self.hidden_size), dtype=input.dtype)
elif self.mode == 'LSTM':
hx = (jt.zeros((num_directions * self.num_layers, input.shape[1], self.hidden_size), dtype=input.dtype),
jt.zeros((num_directions * self.num_layers, input.shape[1], self.hidden_size), dtype=input.dtype))
if jt.flags.use_cuda and jt.cudnn and self.proj_size == 0 and jt.compiler.is_cuda:
return self._execute_cudnn_rnn(input, hx)
else:
hidden_n = []
for l in range(self.num_layers):
output = []
if isinstance(hx, tuple):
hidden = [h[l * num_directions] for h in hx]
else:
hidden = hx[l * num_directions]
output, _hidden = self.call_rnn_sequence(input, hidden, f'l{l}')
hidden_n.append(_hidden)
if self.bidirectional:
if isinstance(hx, tuple):
hidden = [h[l * num_directions + 1] for h in hx]
else:
hidden = hx[l * num_directions + 1]
output_b, _hidden = self.call_rnn_sequence(input, hidden, f'l{l}_reverse')
output = jt.concat([output, output_b], dim=-1)
hidden_n.append(_hidden)
if self.dropout > 0:
input = dropout(output, p=self.dropout)
else:
input = output
if isinstance(hx, tuple):
hidden_n = tuple(jt.stack(hn, dim=0) for hn in zip(*hidden_n))
else:
hidden_n = jt.stack(hidden_n, dim=0)
return output, hidden_n
class RNN(RNNBase):
''' Applies a multi-layer Elman RNN with tanh ReLU non-linearity to an input sequence.
:param input_size: The number of expected features in the input.
:type input_size: int
:param hidden_size: The number of features in the hidden state.
:type hidden_size: int
:param num_layers: Number of recurrent layers. E.g., setting num_layers=2 would mean stacking two RNNs together to form a stacked RNN, with the second RNN taking in outputs of the first RNN and computing the final results. Default: 1
:type num_layers: int, optinal
:param nonlinearity: The non-linearity to use. Can be either 'tanh' or 'relu'. Default: 'tanh'
:type nonlinearity: str, optional
:param bias: If False, then the layer does not use bias weights b_ih and b_hh. Default: True.
:type bias: bool, optional
:param batch_first: If True, then the input and output tensors are provided as (batch, seq, feature). Default: False
:type bias: bool, optional
:param dropout: If non-zero, introduces a Dropout layer on the outputs of each RNN layer except the last layer, with dropout probability equal to dropout. Default: 0
:type dropout: float, optional
:param bidirectional: If True, becomes a bidirectional RNN. Default: False
:type bidirectional: bool, optional
Example:
>>> rnn = nn.RNN(10, 20, 2)
>>> input = jt.randn(5, 3, 10)
>>> h0 = jt.randn(2, 3, 20)
>>> output, hn = rnn(input, h0)
'''
def __init__(self, input_size: int, hidden_size: int, num_layers: int = 1,
nonlinearity: str = 'tanh', bias: bool = True, batch_first: bool = False,
dropout: float = 0, bidirectional: bool = False) -> None:
super().__init__('RNN', input_size, hidden_size, num_layers=num_layers,
bias=bias, batch_first=batch_first, dropout=dropout,
bidirectional=bidirectional)
if not nonlinearity in ['tanh', 'relu']:
raise ValueError('Unrecognized nonlinearity: ' + nonlinearity)
self.nonlinearity = nonlinearity
def call_rnn_cell(self, input, hidden, suffix):
y = matmul_transpose(input, getattr(self, f'weight_ih_{suffix}'))
y = y + matmul_transpose(hidden, getattr(self, f'weight_hh_{suffix}'))
if self.bias:
y = y + getattr(self, f'bias_ih_{suffix}') + getattr(self, f'bias_hh_{suffix}')
if self.nonlinearity == 'tanh':
h = jt.tanh(y)
else:
h = jt.nn.relu(y)
return h, h
class LSTM(RNNBase):
''' Applies a multi-layer long short-term memory (LSTM) RNN to an input sequence.
:param input_size: The number of expected features in the input.
:type input_size: int
:param hidden_size: The number of features in the hidden state.
:type hidden_size: int
:param num_layers: Number of recurrent layers. E.g., setting num_layers=2 would mean stacking two LSTMs together to form a stacked LSTM, with the second LSTM taking in outputs of the first LSTM and computing the final results. Default: 1
:type num_layers: int, optinal
:param bias: If False, then the layer does not use bias weights b_ih and b_hh. Default: True.
:type bias: bool, optional
:param batch_first: If True, then the input and output tensors are provided as (batch, seq, feature). Default: False
:type bias: bool, optional
:param dropout: If non-zero, introduces a Dropout layer on the outputs of each LSTM layer except the last layer, with dropout probability equal to dropout. Default: 0
:type dropout: float, optional
:param bidirectional: If True, becomes a bidirectional LSTM. Default: False
:type bidirectional: bool, optional
:param proj_size: If > 0, will use LSTM with projections of corresponding size. Default: 0
:type proj_size: int, optional
Example:
>>> rnn = nn.LSTM(10, 20, 2)
>>> input = jt.randn(5, 3, 10)
>>> h0 = jt.randn(2, 3, 20)
>>> c0 = jt.randn(2, 3, 20)
>>> output, (hn, cn) = rnn(input, (h0, c0))
'''
def __init__(self, input_size, hidden_size, num_layers=1, bias=True,
batch_first=False, dropout=0, bidirectional=False, proj_size=0):
super().__init__('LSTM', input_size, hidden_size, num_layers=num_layers,
bias=bias, batch_first=batch_first, dropout=dropout,
bidirectional=bidirectional, proj_size=proj_size)
def call_rnn_cell(self, input, hidden, suffix):
h, c = hidden
y = matmul_transpose(input, getattr(self, f'weight_ih_{suffix}'))
y = y + matmul_transpose(h, getattr(self, f'weight_hh_{suffix}'))
if self.bias:
y = y + getattr(self, f'bias_ih_{suffix}') + getattr(self, f'bias_hh_{suffix}')
i = y[:, :self.hidden_size].sigmoid()
f = y[:, self.hidden_size : 2 * self.hidden_size].sigmoid()
g = y[:, 2 * self.hidden_size : 3 * self.hidden_size].tanh()
o = y[:, 3 * self.hidden_size:].sigmoid()
c = f * c + i * g
h = o * c.tanh()
if self.proj_size > 0:
h = matmul_transpose(h, getattr(self, f'weight_hr_{suffix}'))
return h, (h, c)
class GRU(RNNBase):
''' Applies a multi-layer gated recurrent unit (GRU) RNN to an input sequence.
:param input_size: The number of expected features in the input.
:type input_size: int
:param hidden_size: The number of features in the hidden state.
:type hidden_size: int
:param num_layers: Number of recurrent layers. E.g., setting num_layers=2 would mean stacking two GRUs together to form a stacked GRU, with the second GRU taking in outputs of the first GRU and computing the final results. Default: 1
:type num_layers: int, optinal
:param bias: If False, then the layer does not use bias weights b_ih and b_hh. Default: True.
:type bias: bool, optional
:param batch_first: If True, then the input and output tensors are provided as (batch, seq, feature). Default: False
:type bias: bool, optional
:param dropout: If non-zero, introduces a Dropout layer on the outputs of each GRU layer except the last layer, with dropout probability equal to dropout. Default: 0
:type dropout: float, optional
:param bidirectional: If True, becomes a bidirectional GRU. Default: False
:type bidirectional: bool, optional
Example:
>>> rnn = nn.GRU(10, 20, 2)
>>> input = jt.randn(5, 3, 10)
>>> h0 = jt.randn(2, 3, 20)
>>> output, hn = rnn(input, h0)
'''
def __init__(self, input_size: int, hidden_size: int, num_layers: int = 1,
bias: bool = True, batch_first: bool = False, dropout: float = 0,
bidirectional: bool = False) -> None:
super().__init__('GRU', input_size, hidden_size, num_layers=num_layers,
bias=bias, batch_first=batch_first, dropout=dropout,
bidirectional=bidirectional)
def call_rnn_cell(self, input, hidden, suffix):
ih = matmul_transpose(input, getattr(self, f'weight_ih_{suffix}'))
hh = matmul_transpose(hidden, getattr(self, f'weight_hh_{suffix}'))
if self.bias:
ih = ih + getattr(self, f'bias_ih_{suffix}')
hh = hh + getattr(self, f'bias_hh_{suffix}')
hs = self.hidden_size
r = (ih[:, :hs] + hh[:, :hs]).sigmoid()
z = (ih[:, hs: 2 * hs] + hh[:, hs: 2 * hs]).sigmoid()
n = (ih[:, 2 * hs:] + r * hh[:, 2 * hs:]).tanh()
h = (1 - z) * n + z * hidden
return h, h
def bilinear(in1, in2, weight, bias):
w = weight.transpose((1,0,2))
w = w.reshape((w.shape[0], -1))
x = jt.matmul(in1, w)
x = x.reshape(x.shape[:-1]+[weight.shape[0], weight.shape[2]])
y = in2.broadcast(x, (-2,))
z = (x*y).sum(-1)
if bias is not None:
z += bias
return z
class Bilinear(Module):
''' bilinear transformation $out = in1^T W in2 + bias$, Example::
m = nn.Bilinear(20, 30, 40)
input1 = jt.randn(128, 20)
input2 = jt.randn(128, 30)
output = m(input1, input2)
print(output.shape)
# [128, 40]
'''
def __init__(self, in1_features, in2_features, out_features, bias=True, dtype="float32"):
bound = 1 / math.sqrt(in1_features)
self.weight = jt.init.uniform([out_features, in1_features, in2_features], dtype, -bound, bound)
self.bias = bias
if bias:
self.bias = jt.init.uniform([out_features], dtype, -bound, bound)
def execute(self, in1, in2):
return bilinear(in1, in2, self.weight, self.bias)
#TODO: support FFT2D only now.
def _fft2(x, inverse=False):
assert(jt.flags.use_cuda==1)
assert(len(x.shape) == 4)
assert(x.shape[3] == 2)
y = jt.compile_extern.cufft_ops.cufft_fft(x, inverse)
if inverse:
y /= x.shape[1] * x.shape[2]
return y
def one_hot(x: jt.Var, num_classes: int=-1) -> jt.Var:
''' Returns the one_hot encoding of inputs.
:param x: class values of any shape
:type x: jt.Var with bool or integer dtype
:param num_classes: Total number of classes. If set to -1, the number of classes will be inferred as one greater than the largest class value in the input tensor.
:type num_classes: int, optional
:return: a Var with one more dimension with 1 values at the index
of last dimension indicated by the input, and 0 everywhere else.
:rtype: jt.Var
.. note::
if the values in x are greater than num_class or less than 0,
the returned one_hot will be all zeros.
Example:
>>> jt.nn.one_hot(jt.arange(5) % 3)
jt.Var([[1 0 0]
[0 1 0]
[0 0 1]
[1 0 0]
[0 1 0]], dtype=int32)
>>> jt.nn.one_hot(jt.arange(5) % 3, num_classes=5)
jt.Var([[1 0 0 0 0]
[0 1 0 0 0]
[0 0 1 0 0]
[1 0 0 0 0]
[0 1 0 0 0]], dtype=int32)
>>> jt.nn.one_hot(jt.arange(6).reshape(3,2) % 3)
jt.Var([[[1 0 0]
[0 1 0]]
[[0 0 1]
[1 0 0]]
[[0 1 0]
[0 0 1]]], dtype=int32)
'''
assert x.dtype in [jt.bool, jt.int8, jt.int16, jt.int32, jt.int64, jt.uint8, jt.uint16, jt.uint32, jt.uint64]
if num_classes == -1:
num_classes = x.max().item() + 1
N = len(x.shape)
indices = ["i"+str(i) for i in range(N)]
y = jt.ones_like(x).reindex(
x.shape + [num_classes],
indices,
extras=[x],
overflow_conditions=[f"i{N} != @e0({','.join(indices)})"],
overflow_value=0)
return y
class KLDivLoss(Module):
''' Computes the Kullback-Leibler divergence loss.
'''
def __init__(self, reduction: str = 'mean', log_target: bool = False):
'''
:param reduction: Specifies the reduction to apply to the output. Can be 'mean', 'sum', 'batchmean', or 'none'. Defaults to 'mean'.
:type reduction: str, optional
:param log_target: Specifies whether target is the log space. Defaults to False.
:type log_target: bool, optional
'''
self.reduction = reduction
self.log_target = log_target
def execute(self, input: jt.Var, target: jt.Var) -> jt.Var:
if not self.log_target:
loss_pointwise = target * (target.log() - input)
else:
loss_pointwise = target.exp() * (target - input)
if self.reduction == "mean":
loss = loss_pointwise.mean()
elif self.reduction == "batchmean":
loss = loss_pointwise.sum() / input.size(0)
elif self.reduction == "sum":
loss = loss_pointwise.sum()
else:
loss = loss_pointwise
return loss
class Mish(Module):
def __init__(self, inplace=False):
'''
Applies the Mish function, element-wise.
reference: Mish - A Self Regularized Non-Monotonic Neural Activation Function.
'''
pass
def execute(self, x):
return x * jt.tanh(jt.softplus(x))
def mish(x, inplace=False):
return x * jt.tanh(jt.softplus(x))
def skip_init(module_cls, *args, **kw):
return module_cls(*args, **kw)
Reference in New Issue View Git Blame Copy Permalink