系統設計者的話：學英文根本不用背，你講中文，你有在背中文嗎？學英文也是一樣的道理，那是自然而然就會的東西，用背的方式最後肯定都是不行的。 更多...

最好的學習就是有英文的父母，不斷的跟你說話、耐心的教導你，時間長了，你也就慢慢就融入其中了。

但是我們沒有英文父母怎麼辦？看影片學英文就學的起來嗎？學不起來。

我們從小到大，也看過了不少英文電影，至少也看過1、200部也跑不掉，但是為什麼看了這麼多的英文電影，英文還是很爛呢？

主要的原因是我們沒有從基礎性的強化學習。舉個例子，西方的幼兒出生時，他至少聽了各種大人講英文至少聽了2年，才慢慢開始說英文、與大人互動。

而且都是一字一句慢慢的說、慢慢的教、漸進式的、慢慢的、越來越習慣、越來越聽的懂、越來越會說的過程。

那麼我們看的影片、電影過程中，都是大人成熟型的講話速度、方式、語調，我們怎麼可能跟的上？

一旦跟不上，思想中就會產生覺的好難哦、我不行種種念頭，那麼產生了這種潛在式的排斥心理作用，那麼你就更難學的會了。


所以我們必須打好基礎、深厚的基礎，萬丈高樓平地起，唯有打下深厚的基礎，後面你走的路才會越來越順。

程式界裏有句話，程式債欠的越多，最後就只能放棄。學英文也是一樣，英文單字債欠的越多，你就直接上來文法、句型、語調，最後也只能放棄。

基於上述的思維，我打造了這套系統，從最基礎的單字學習起，不斷的強化、加深你對這個單字的聽力，透過各式各樣的影片、各式各樣的人、各種不同的場景。

就只是為了訓練一個單字的聽力，無數的人跟你講同一個單字，講個20遍、50篇、100遍(且不重覆)，我就不信學不起來。


這套系統做出來後，我自己在使用的過程中，確實真的能聽出來他們講的這個單字，而且他們講的速度再快也能聽的出來。

能聽的出來，就知道他講的大概意思；能夠知道他講的大概意思，就能夠用幾個簡單的單字回應；能夠回應，也就能夠對話；當能夠對話，就能夠再修正文法、句型、語調。


我的目地就是希望自己至少能聽的懂人家在說什麼，用英文跟人家對話，大家共同努力、學習。若有其它想法或對這套系統設計的建議，也可以與我聯絡：ichich2013@gmail.com

Sequence Models and the RNN API (TensorFlow Dev Summit 2017)
觀看次數：768
basically
it's
two
recurrent
neural
networks
one
on
the
left
is
the
encoder
and
encoder
reads
in
one
word
or
word
it
creates
some
and
it
creates
some
decoder
which
is
a
second
recurrent
neural
network
or
rnn
it
receives
a
that
again
here
there's
an
encoder
arnon
right
we
have
a
decoder
ironin
and
we're
so
here
we're
going
to
be
focusing
on
Google
protobufs
are
because
you've
sequences
so
some
of
the
pros
it
sequences
for
example
it
supports
a
efficient
parser
that
reads
in
a
serialized
proto
string
and
emits
tensors
or
and
or
sparse
tensors
tensile
batching
mechanism
the
function
pads
all
of
the
other
sequences
up
to
that
maxim
maxim
m1
so
this
is
a
little
because
you're
only
padding
up
to
the
larger
and
larger
64
32
32
64
128
you're
there's
a
little
bit
of
padding
but
this
spillover
from
the
very
long
sequences
or
variational
auto-encoders
for
example
truncated
back
propagation
to
time
so
any
sequences
or
segments
which
are
not
for
that
it's
called
bad
sequences
of
and
that's
it
let's
talk
about
arnage
calculation
flows
you
get
inputs
from
intermediate
calculation
emits
a
new
LS
tm
or
long
short-term
memory
it's
a
same
python
as
a
for
loop
so
let's
let's
amidst
new
outputs
and
state
you
also
tensor
it's
not
just
one
for
a
gru
which
functions
to
objects
i'll
talk
about
iterations
we'll
talk
about
that
next
compatibility
between
the
two
but
let's
each
tensor
has
num
units
columns
so
tensor
and
as
num
units
columns
finally
impotency
and
a
state
tuple
in
this
case
it
decomposes
the
state
into
two
tensors
performs
intermediate
calculations
for
library
of
different
implementations
of
research
oriented
architectures
one
by
a
reinforcement
learning
and
pop
on
as
of
this
week
on
templates
will
github
radically
different
aren't
an
dynamic
calculation
in
tensorflow
tensors
and
also
supports
backdrop
okay
accepts
three
arguments
the
cond
lambda
let's
start
with
a
tuple
of
tuples
ijk
jk
increment
I
by
1
and
updates
jk
with
tuples
of
tensors
so
the
final
value
is
which
is
a
tuple
of
tensors
which
calculation
there
10
times
okay
so
while
unpack
or
unstack
the
matrix
into
that
calculation
between
these
two
primitives
variants
that's
rnn
and
also
we
have
a
eight
stacked
LS
tm
layers
and
each
one
residual
connections
so
residual
0
gp
1
and
so
on
cell
it
takes
a
tensor
of
many
batches
of
inputs
it
takes
a
vector
with
the
the
trade-offs
between
memory
flexible
API
for
customizing
implementing
your
own
RN
and
architectures
dynamic
decoding
and
they're
also
handwritten
c++
kernels
implementations
of
say
LS
pm
Pro
time
designed
fully
fused
aren't
implementations
that
fuse
the
entire
trade-offs
that
you're
making
so
you're
these
implementations
don't
work
well
in
called
a
compiled
wrapper
which
we
performs
that
builds
the
jit
compilation
multi
RN
n
cell
with
three
sub
wheaties
separately
fused
ost
M
layers
running
slightly
different
architectures
of
the
takeaways
first
of
all
if
you're
there's
currently
not
much
benefit
so
looking
into
a
header
or
you're
for
an
embedded
device
say
a
android
binary
size
and
again
while
I
didn't
benchmark
these
aspects
of
it
this
is
benchmarks
from
excellet
we
have
some
handwritten
c++
kernels
for
the
lcm
and
gru
cells
devices
okay
so
try
them
out
if
you're
interested
in
fully
fused
calculations
working
you
have
nvidia
gpus
you're
using
to
dnn
we
have
cout
dnn
wrappers
you're
if
you're
working
these
devices
and
it's
fairly
flexible
supports
bi-directional
and
ends
stacks
are
n
architectures
and
as
of
about
a
month
one
of
the
other
lsdm
implementations
fact
decoding
tends
to
be
the
most
object-oriented
API
it's
under
active
decoder
is
to
make
a
plug-and-play
and
an
in
graphene
search
and
then
decode
so
method
so
during
decoding
when
you're
you
feed
it
that's
ten
serve
decoder
decoders
so
use
another
helper
in
i'll
use
that
as
lookup
in
an
embedding
outputs
of
the
rnn
treat
them
as
logits
encoder
and
then
you
build
your
decoder
attentional
decoder
where
you
provide
an
your
decoder
just
decode
call
dynamic
tensor
and
the
final
decoder
state
so
the
output
for
a
simple
decoder
they're
IDs
so
if
you
chose
sampling
decoder
calculations
within
temperature
flow
and
activations
to
the
cpu
reducing
the
decoder
library
that
we
hope
you'll

相關影片

加Line好友 每日8AM 推播1單字

加入好友

其他人正在學習的單字

1. next
2024-04-29 07:30
2. altogether
2024-04-29 07:30
3. sensible
2024-04-29 07:30
4. plug
2024-04-29 07:30
5. marsh
2024-04-29 07:30
6. creatures
2024-04-29 07:30
7. require
2024-04-29 07:30
8. sense
2024-04-29 07:30
9. able
2024-04-29 07:30
10. mere
2024-04-29 07:30
11. jerry
2024-04-29 07:30
12. bully
2024-04-29 07:30
13. compelling
2024-04-29 07:30
14. homing
2024-04-29 07:29
15. form
2024-04-29 07:29

本月觀看影片單字排行榜

本月查閱單字排行榜

本月聆聽單字排行榜