SPRINGLab/asr-task-data · Datasets at Hugging Face

audio audioduration (s) 0.62 28.4	text stringlengths 2 287
	understanding so one of the major a point is that if you look into most part of this
	code then they are actually something which is quite modularly written down so if i want
	over there so if i say i want to add down a few more linear number of channels
	so i can just keep on adding down over there so say like i i just want to add down another
	layer i can write it as n n dot linear so i have six number of outputs from there from
	that six number of outputs i can again think of going down to say another six number of
	outputs over there ok thats also pretty much possible then i can do a n n dot linear from
	is done over here till this point i just had with the self dot l two and what i can add
	defines these architectures and you will just have to make a shorter change so by now its
	behavior and it was much slower in actually falling down so the error did decrease to
	so i just need to execute only this part of it keep one thing in mind if you are making
	some changes to a previous one then you need to execute all the subsequent blocks so that
	python code for you and it will appear in the same modular way within your python environment
	and you can just execute that py file within whatever is your choice of environment most
	most likely most of us would be making use of anaconda python in order to do that and
	linearly falling down and much jagged and the error was also much higher so i can make
	a change over there and say make this as fifty and see so these changes which i am making
	over there on the learning rate thats the factor which is going to update my in
	in my learning rule and that does make a very important statement in terms of how fast its
	also pretty much true ah and and thats what we do observe so its not necessary that always
	getting down a slower is going to make you a better convergent or or even getting
	the future lectures we would be ah getting more of a clearer understanding into how this
	with more of these lectures and do ah keep on coding as we continue with the theoretical
	order to ah hierarchically keep on building so there were two different particular kinds
	to what was present in that encoder so if an encoder you were reducing the size over
	there in decoder you are just going to increase the sizes and ah then one way is where you
	decoder over there
	so we we are actually going to ah replicate both of these ah methods over here within
	same codes as you had done ah in the earlier ones and as with how it goes down is that
	and that creates
	and which will be just looking into how many number of ah parallel loaders you can work
	it out and as such its its equal to number four and then you also set down your batch
	size and which is equal to the number of images it fetches found in one particular batch so
	you
	you used gpu flag is not set down then none of your models or data gets converted on to
	layer definition and initialization is complete then the next part is that you write down
	your forward function and your forward function what it does is that it will do a forward
	pass over the encoder which consists of a linear layer and a tan hyperbolic ah as a
	ok so for the purpose of training ah its kept down as small so we just have ten iterations
	perfect way and the best way of doing it is to take down an l two ah norm as a loss function
	run it on a gpu the next part as it goes down is that ah you know to zero down all the gradient
	residuals within a particular so yeah so basically your gradients which were computed in the
	earlier epoch should not play any role in the current epoch and thats the whole purpose
	of this ah zeroing down on the gradients then the next part is just a forward pass ah you
	of an image then you have a batch then you give a batch of images
	output and the ah input image batches which are formed and from there you have your loss
	and then you do a derivative of the loss using your loss ah dot backward compute once that
	it comes down to point one zero you see clearly see that there is a decrease in error and
	then it keeps on ah steadily decreasing now one thing to keep in mind is that since its
	a fully connected layer ah with lot of neurons so it would take a bit of time and thats where
	ah it it takes down a bit of time for us as well and ah after some time you see that ah
	to come down to an optimal
	so the relative change the relative difference between these ah errors which comes down between
	two epochs is also going down quite slower now you can actually play around over there
	so ah what you can do is you can change down your learning rates and and play and see if
	but then this is not ah always necessary that this would be the only combination which is
	part of the ah classes so here you can just make a change around with your learning rates
	as well and then you would be seeing that this changes now here this was the first approach
	so what i would do is i have my network which is strain which was net ok now i take my encoder
	part over there and then start a ah i add another new module to it ok i call this as
	seven hundred eighty four neurons going down to four hundred neurons and then again reconstructed
	to seven eighty four neurons here what i do is seven eighty four neurons going down to
	four hundred neurons going down to two hundred and fifty six neurons from there coming up
	to four hundred neurons from there going to seven hundred and eighty four neurons so this
	becomes my first part of it now similarly in the decoder part where i have two hundred
	and fifty six neurons to be connected down to four hundred neurons this comes down
	over there after ah the output of this four hundred neurons comes out and then ah on the
	before to the decoder i also have another tan hyperbolic function now if you clearly
	of these four hundred neurons so that means that any of the values which comes down in
	these four hundred neurons over there are in the range of minus one to plus one
	neurons and for that reason what we have is that ah four hundred neurons ah so whenever
	hyperbolic transfer function and not any other transfer function so you need to keep this
	parts are just added over there and then i can ah print my network so lets just see what
	four neurons to four hundred neurons then a tan hyperbolic the output of this one goes
	into another sequential connection from four hundred to two fifty six and then tan hyperbolic
	now input to the next decoder layer is ah two hundred and fifty six to four hundred
	so thats up to you how you define your ah hierarchical strategies but till its a linear
	network it does not ah make much of a difference coming down
	pass and get down your outputs and then you have your ah loss function defined and derivative
	in the earlier case ah though you have sort of ah better train model over there
	so if you look down at the mean square error which comes down around point ah seven six
	initialized and that does mean that whatever is coming down on the learnt representations
	and getting forwarded to the decoder layer subsequently is now pretty much random that
	and thats why the starting error is a bit more than what was the ending error over that
	be a case that you might not be able to reach down the exact ah error rates which you had
	achieved by using just one hidden layer and there is nothing to panic around it the only
	ah we are trying to solve down ah the actual problem of classification over here using
	ah auto encoder for representation learning and then we are using a stacking policy and
	one particular task whereas the network is supposed to be used for another task and thats
	classification so given all of these things in mind what you need to have like really
	errors however clever stacking and change of error rates can actually bring it down
	so at this point i would leave it to you to actually go up and as ah update this learning
	now play around with this learning rate and you can definitely see a change coming ah
	was to create a neural network where i have two hidden layers and then a final output
	that this is just a ten digits classification problem zero to nine written in hundred and
	and you have to classify it out now over there i would need to modify it down some part of
	the network because i dont need the decoder as such anywhere now so now what i can do
	this layer ok these two layers are something which i can do away with and the best way
	over here now once that is done the next part is that i would need to add down so now what
	to two hundred and fifty six neurons from two hundred and fifty six neurons i will have

understanding so one of the major a point is that if you look into most part of this

code then they are actually something which is quite modularly written down so if i want

over there so if i say i want to add down a few more linear number of channels

so i can just keep on adding down over there so say like i i just want to add down another

layer i can write it as n n dot linear so i have six number of outputs from there from

that six number of outputs i can again think of going down to say another six number of

outputs over there ok thats also pretty much possible then i can do a n n dot linear from

is done over here till this point i just had with the self dot l two and what i can add

defines these architectures and you will just have to make a shorter change so by now its

behavior and it was much slower in actually falling down so the error did decrease to

so i just need to execute only this part of it keep one thing in mind if you are making

some changes to a previous one then you need to execute all the subsequent blocks so that

python code for you and it will appear in the same modular way within your python environment

and you can just execute that py file within whatever is your choice of environment most

most likely most of us would be making use of anaconda python in order to do that and

linearly falling down and much jagged and the error was also much higher so i can make

a change over there and say make this as fifty and see so these changes which i am making

over there on the learning rate thats the factor which is going to update my in

in my learning rule and that does make a very important statement in terms of how fast its

also pretty much true ah and and thats what we do observe so its not necessary that always

getting down a slower is going to make you a better convergent or or even getting

the future lectures we would be ah getting more of a clearer understanding into how this

with more of these lectures and do ah keep on coding as we continue with the theoretical

order to ah hierarchically keep on building so there were two different particular kinds

to what was present in that encoder so if an encoder you were reducing the size over

there in decoder you are just going to increase the sizes and ah then one way is where you

decoder over there

so we we are actually going to ah replicate both of these ah methods over here within

same codes as you had done ah in the earlier ones and as with how it goes down is that

and that creates

and which will be just looking into how many number of ah parallel loaders you can work

it out and as such its its equal to number four and then you also set down your batch

size and which is equal to the number of images it fetches found in one particular batch so

you

you used gpu flag is not set down then none of your models or data gets converted on to

layer definition and initialization is complete then the next part is that you write down

your forward function and your forward function what it does is that it will do a forward

pass over the encoder which consists of a linear layer and a tan hyperbolic ah as a

ok so for the purpose of training ah its kept down as small so we just have ten iterations

perfect way and the best way of doing it is to take down an l two ah norm as a loss function

run it on a gpu the next part as it goes down is that ah you know to zero down all the gradient

residuals within a particular so yeah so basically your gradients which were computed in the

earlier epoch should not play any role in the current epoch and thats the whole purpose

of this ah zeroing down on the gradients then the next part is just a forward pass ah you

of an image then you have a batch then you give a batch of images

output and the ah input image batches which are formed and from there you have your loss

and then you do a derivative of the loss using your loss ah dot backward compute once that

it comes down to point one zero you see clearly see that there is a decrease in error and

then it keeps on ah steadily decreasing now one thing to keep in mind is that since its

a fully connected layer ah with lot of neurons so it would take a bit of time and thats where

ah it it takes down a bit of time for us as well and ah after some time you see that ah

to come down to an optimal

so the relative change the relative difference between these ah errors which comes down between

two epochs is also going down quite slower now you can actually play around over there

so ah what you can do is you can change down your learning rates and and play and see if

but then this is not ah always necessary that this would be the only combination which is

part of the ah classes so here you can just make a change around with your learning rates

as well and then you would be seeing that this changes now here this was the first approach

so what i would do is i have my network which is strain which was net ok now i take my encoder

part over there and then start a ah i add another new module to it ok i call this as

seven hundred eighty four neurons going down to four hundred neurons and then again reconstructed

to seven eighty four neurons here what i do is seven eighty four neurons going down to

four hundred neurons going down to two hundred and fifty six neurons from there coming up

to four hundred neurons from there going to seven hundred and eighty four neurons so this

becomes my first part of it now similarly in the decoder part where i have two hundred

and fifty six neurons to be connected down to four hundred neurons this comes down

over there after ah the output of this four hundred neurons comes out and then ah on the

before to the decoder i also have another tan hyperbolic function now if you clearly

of these four hundred neurons so that means that any of the values which comes down in

these four hundred neurons over there are in the range of minus one to plus one

neurons and for that reason what we have is that ah four hundred neurons ah so whenever

hyperbolic transfer function and not any other transfer function so you need to keep this

parts are just added over there and then i can ah print my network so lets just see what

four neurons to four hundred neurons then a tan hyperbolic the output of this one goes

into another sequential connection from four hundred to two fifty six and then tan hyperbolic

now input to the next decoder layer is ah two hundred and fifty six to four hundred

so thats up to you how you define your ah hierarchical strategies but till its a linear

network it does not ah make much of a difference coming down

pass and get down your outputs and then you have your ah loss function defined and derivative

in the earlier case ah though you have sort of ah better train model over there

so if you look down at the mean square error which comes down around point ah seven six

initialized and that does mean that whatever is coming down on the learnt representations

and getting forwarded to the decoder layer subsequently is now pretty much random that

and thats why the starting error is a bit more than what was the ending error over that

be a case that you might not be able to reach down the exact ah error rates which you had

achieved by using just one hidden layer and there is nothing to panic around it the only

ah we are trying to solve down ah the actual problem of classification over here using

ah auto encoder for representation learning and then we are using a stacking policy and

one particular task whereas the network is supposed to be used for another task and thats

classification so given all of these things in mind what you need to have like really

errors however clever stacking and change of error rates can actually bring it down

so at this point i would leave it to you to actually go up and as ah update this learning

now play around with this learning rate and you can definitely see a change coming ah

was to create a neural network where i have two hidden layers and then a final output

that this is just a ten digits classification problem zero to nine written in hundred and

and you have to classify it out now over there i would need to modify it down some part of

the network because i dont need the decoder as such anywhere now so now what i can do

this layer ok these two layers are something which i can do away with and the best way

over here now once that is done the next part is that i would need to add down so now what

to two hundred and fifty six neurons from two hundred and fifty six neurons i will have