Acheiving Comparable performance using Resnet

[varghesealex90]

Hi Zoozog;

This is not an issue, rather an update.
Using your code as the base, I trained a ResNet-18 ( PRETRAINED on imagenet) for the CXR task. I see that the performance is comparable to a large extent. I assume that ResNET18 would have fewer parameters than densenet121.

AUROC mean 0.847266414478356
Atelectasis 0.8323852409157189
Cardiomegaly 0.9118541996631041
Effusion 0.8866060621559915
Infiltration 0.7162746001795058
Mass 0.8627326595643292
Nodule 0.7897284404542668
Pneumonia 0.7753048198007456
Pneumothorax 0.8821548678140269
Consolidation 0.8131018114945643
Edema 0.8982024375815074
Emphysema 0.9355988306093312
Fibrosis 0.8532190562301323
Pleural_Thickening 0.7942477917778453
Hernia 0.9103189844559144

I am also experimenting with training the network from scratch ( sans pre-trained weights). Shall update you once that is done.

[zoogzog]

Nice results! Thanks for sharing!

[Ahsanr312]

Hi varghesealex90;
Can you share your code for ResNet-18 ?

[varghesealex90]

Thanks for the reply. However, I had surrender the model, code to my previous institute.

…

On Tue, 9 Jul 2019 at 11:36, Ahsan Raza ***@***.***> wrote: Hi varghesealex90; Can you share your code for ResNet-18 ? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ACNI7TSFBHGGOTKTR3CBO7TP6QTGBA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZPGGJI#issuecomment-509502245>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACNI7TSCNNCR5FB2JI55ZZTP6QTGBANCNFSM4ERUD5LA> .

[Ahsanr312]

Thankyou for replying. But can't you share the idea ? I have been searching alot but haven't find anything accurate. DenseNet architecture requires alot of memory. And i get cuda memory error. So it would be so kind of you if you can share it with me. On 14 Jul 2019 11:08 a.m., varghese <[email protected]> wrote: Thanks for the reply. However, I had surrender the model, code to my previous institute.

On Tue, 9 Jul 2019 at 11:36, Ahsan Raza ***@***.***> wrote: Hi varghesealex90; Can you share your code for ResNet-18 ? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ACNI7TSFBHGGOTKTR3CBO7TP6QTGBA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZPGGJI#issuecomment-509502245>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACNI7TSCNNCR5FB2JI55ZZTP6QTGBANCNFSM4ERUD5LA> .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub<#4?email_source=notifications&email_token=AJJZIU7E4C6CMIF5RM7U2ATP7K7GLA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ364VA#issuecomment-511176276>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AJJZIU4TDID6LEKDL73R4ODP7K7GLANCNFSM4ERUD5LA>.

[varghesealex90]

Surely i can help you with you with the idea. I initialized with pretrained weights, reduced the batch size ( i dont remember what the batch size was). All other hyperparameters were kept to same as what zoozog provided

…

On Sun, 14 Jul, 2019, 1:46 PM Ahsan Raza, ***@***.***> wrote: Thankyou for replying. But can't you share the idea ? I have been searching alot but haven't find anything accurate. DenseNet architecture requires alot of memory. And i get cuda memory error. So it would be so kind of you if you can share it with me. On 14 Jul 2019 11:08 a.m., varghese ***@***.***> wrote: Thanks for the reply. However, I had surrender the model, code to my previous institute. On Tue, 9 Jul 2019 at 11:36, Ahsan Raza ***@***.***> wrote: > Hi varghesealex90; > Can you share your code for ResNet-18 ? > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > < #4?email_source=notifications&email_token=ACNI7TSFBHGGOTKTR3CBO7TP6QTGBA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZPGGJI#issuecomment-509502245>, > or mute the thread > < https://github.com/notifications/unsubscribe-auth/ACNI7TSCNNCR5FB2JI55ZZTP6QTGBANCNFSM4ERUD5LA> > . > — You are receiving this because you commented. Reply to this email directly, view it on GitHub< #4?email_source=notifications&email_token=AJJZIU7E4C6CMIF5RM7U2ATP7K7GLA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ364VA#issuecomment-511176276>, or mute the thread< https://github.com/notifications/unsubscribe-auth/AJJZIU4TDID6LEKDL73R4ODP7K7GLANCNFSM4ERUD5LA>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ACNI7TVBWXU7XGB44P25HQTP7LOGFA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ4ARLQ#issuecomment-511183022>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACNI7TXSS6WOAEATEIALPBTP7LOGFANCNFSM4ERUD5LA> .

[Ahsanr312]

You used pretrained model of resnet-18 ? And through transfer learning achieved approximately similar results as DenseNet? On 14 Jul 2019 2:07 p.m., varghese <[email protected]> wrote: Surely i can help you with you with the idea. I initialized with pretrained weights, reduced the batch size ( i dont remember what the batch size was). All other hyperparameters were kept to same as what zoozog provided

On Sun, 14 Jul, 2019, 1:46 PM Ahsan Raza, ***@***.***> wrote: Thankyou for replying. But can't you share the idea ? I have been searching alot but haven't find anything accurate. DenseNet architecture requires alot of memory. And i get cuda memory error. So it would be so kind of you if you can share it with me. On 14 Jul 2019 11:08 a.m., varghese ***@***.***> wrote: Thanks for the reply. However, I had surrender the model, code to my previous institute. On Tue, 9 Jul 2019 at 11:36, Ahsan Raza ***@***.***> wrote: > Hi varghesealex90; > Can you share your code for ResNet-18 ? > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > < #4?email_source=notifications&email_token=ACNI7TSFBHGGOTKTR3CBO7TP6QTGBA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZPGGJI#issuecomment-509502245>, > or mute the thread > < https://github.com/notifications/unsubscribe-auth/ACNI7TSCNNCR5FB2JI55ZZTP6QTGBANCNFSM4ERUD5LA> > . > — You are receiving this because you commented. Reply to this email directly, view it on GitHub< #4?email_source=notifications&email_token=AJJZIU7E4C6CMIF5RM7U2ATP7K7GLA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ364VA#issuecomment-511176276>, or mute the thread< https://github.com/notifications/unsubscribe-auth/AJJZIU4TDID6LEKDL73R4ODP7K7GLANCNFSM4ERUD5LA>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ACNI7TVBWXU7XGB44P25HQTP7LOGFA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ4ARLQ#issuecomment-511183022>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACNI7TXSS6WOAEATEIALPBTP7LOGFANCNFSM4ERUD5LA> .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub<#4?email_source=notifications&email_token=AJJZIU7PMZYZJH43PR23LOLP7LUGJA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ4BJSI#issuecomment-511186121>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AJJZIUZZ2RQNBOPT3OR5NUDP7LUGJANCNFSM4ERUD5LA>.

[Ahsanr312]

You used pretrained ResNet-18 model? And using transfer learning trained it on CheXNet dataset? On 14 Jul 2019 1:08 p.m., Ahsan Raza <[email protected]> wrote: Thankyou for replying. But can't you share the idea ? I have been searching alot but haven't find anything accurate. DenseNet architecture requires alot of memory. And i get cuda memory error. So it would be so kind of you if you can share it with me. On 14 Jul 2019 11:08 a.m., varghese <[email protected]> wrote: Thanks for the reply. However, I had surrender the model, code to my previous institute.

On Tue, 9 Jul 2019 at 11:36, Ahsan Raza ***@***.***> wrote: Hi varghesealex90; Can you share your code for ResNet-18 ? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ACNI7TSFBHGGOTKTR3CBO7TP6QTGBA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZPGGJI#issuecomment-509502245>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACNI7TSCNNCR5FB2JI55ZZTP6QTGBANCNFSM4ERUD5LA> .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub<#4?email_source=notifications&email_token=AJJZIU7E4C6CMIF5RM7U2ATP7K7GLA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ364VA#issuecomment-511176276>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AJJZIU4TDID6LEKDL73R4ODP7K7GLANCNFSM4ERUD5LA>.

[varghesealex90]

Yes. I used resnet pretrained on imagenet for my experiments

…

On Wed, 17 Jul, 2019, 11:05 PM Ahsan Raza, ***@***.***> wrote: You used pretrained model of resnet-18 ? And through transfer learning achieved approximately similar results as DenseNet? On 14 Jul 2019 2:07 p.m., varghese ***@***.***> wrote: Surely i can help you with you with the idea. I initialized with pretrained weights, reduced the batch size ( i dont remember what the batch size was). All other hyperparameters were kept to same as what zoozog provided On Sun, 14 Jul, 2019, 1:46 PM Ahsan Raza, ***@***.***> wrote: > Thankyou for replying. > But can't you share the idea ? I have been searching alot but haven't find > anything accurate. DenseNet architecture requires alot of memory. And i get > cuda memory error. > So it would be so kind of you if you can share it with me. > > On 14 Jul 2019 11:08 a.m., varghese ***@***.***> wrote: > Thanks for the reply. However, I had surrender the model, code to my > previous institute. > > On Tue, 9 Jul 2019 at 11:36, Ahsan Raza ***@***.***> wrote: > > > Hi varghesealex90; > > Can you share your code for ResNet-18 ? > > > > — > > You are receiving this because you authored the thread. > > Reply to this email directly, view it on GitHub > > < > #4?email_source=notifications&email_token=ACNI7TSFBHGGOTKTR3CBO7TP6QTGBA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZPGGJI#issuecomment-509502245>, > > > or mute the thread > > < > https://github.com/notifications/unsubscribe-auth/ACNI7TSCNNCR5FB2JI55ZZTP6QTGBANCNFSM4ERUD5LA> > > > . > > > > — > You are receiving this because you commented. > Reply to this email directly, view it on GitHub< > #4?email_source=notifications&email_token=AJJZIU7E4C6CMIF5RM7U2ATP7K7GLA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ364VA#issuecomment-511176276>, > or mute the thread< > https://github.com/notifications/unsubscribe-auth/AJJZIU4TDID6LEKDL73R4ODP7K7GLANCNFSM4ERUD5LA>. > > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > < #4?email_source=notifications&email_token=ACNI7TVBWXU7XGB44P25HQTP7LOGFA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ4ARLQ#issuecomment-511183022>, > or mute the thread > < https://github.com/notifications/unsubscribe-auth/ACNI7TXSS6WOAEATEIALPBTP7LOGFANCNFSM4ERUD5LA> > . > — You are receiving this because you commented. Reply to this email directly, view it on GitHub< #4?email_source=notifications&email_token=AJJZIU7PMZYZJH43PR23LOLP7LUGJA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ4BJSI#issuecomment-511186121>, or mute the thread< https://github.com/notifications/unsubscribe-auth/AJJZIUZZ2RQNBOPT3OR5NUDP7LUGJANCNFSM4ERUD5LA>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ACNI7TTSQESBO5CC75NXZ63P75J6NA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2FYFLQ#issuecomment-512459438>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACNI7TWHPO3BFUGPLVDADRTP75J6NANCNFSM4ERUD5LA> .

[Ahsanr312]

Thankyou for your help. On 18 Jul 2019 9:20 a.m., varghese <[email protected]> wrote: Yes. I used resnet pretrained on imagenet for my experiments

On Wed, 17 Jul, 2019, 11:05 PM Ahsan Raza, ***@***.***> wrote: You used pretrained model of resnet-18 ? And through transfer learning achieved approximately similar results as DenseNet? On 14 Jul 2019 2:07 p.m., varghese ***@***.***> wrote: Surely i can help you with you with the idea. I initialized with pretrained weights, reduced the batch size ( i dont remember what the batch size was). All other hyperparameters were kept to same as what zoozog provided On Sun, 14 Jul, 2019, 1:46 PM Ahsan Raza, ***@***.***> wrote: > Thankyou for replying. > But can't you share the idea ? I have been searching alot but haven't find > anything accurate. DenseNet architecture requires alot of memory. And i get > cuda memory error. > So it would be so kind of you if you can share it with me. > > On 14 Jul 2019 11:08 a.m., varghese ***@***.***> wrote: > Thanks for the reply. However, I had surrender the model, code to my > previous institute. > > On Tue, 9 Jul 2019 at 11:36, Ahsan Raza ***@***.***> wrote: > > > Hi varghesealex90; > > Can you share your code for ResNet-18 ? > > > > — > > You are receiving this because you authored the thread. > > Reply to this email directly, view it on GitHub > > < > #4?email_source=notifications&email_token=ACNI7TSFBHGGOTKTR3CBO7TP6QTGBA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZPGGJI#issuecomment-509502245>, > > > or mute the thread > > < > https://github.com/notifications/unsubscribe-auth/ACNI7TSCNNCR5FB2JI55ZZTP6QTGBANCNFSM4ERUD5LA> > > > . > > > > — > You are receiving this because you commented. > Reply to this email directly, view it on GitHub< > #4?email_source=notifications&email_token=AJJZIU7E4C6CMIF5RM7U2ATP7K7GLA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ364VA#issuecomment-511176276>, > or mute the thread< > https://github.com/notifications/unsubscribe-auth/AJJZIU4TDID6LEKDL73R4ODP7K7GLANCNFSM4ERUD5LA>. > > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > < #4?email_source=notifications&email_token=ACNI7TVBWXU7XGB44P25HQTP7LOGFA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ4ARLQ#issuecomment-511183022>, > or mute the thread > < https://github.com/notifications/unsubscribe-auth/ACNI7TXSS6WOAEATEIALPBTP7LOGFANCNFSM4ERUD5LA> > . > — You are receiving this because you commented. Reply to this email directly, view it on GitHub< #4?email_source=notifications&email_token=AJJZIU7PMZYZJH43PR23LOLP7LUGJA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ4BJSI#issuecomment-511186121>, or mute the thread< https://github.com/notifications/unsubscribe-auth/AJJZIUZZ2RQNBOPT3OR5NUDP7LUGJANCNFSM4ERUD5LA>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ACNI7TTSQESBO5CC75NXZ63P75J6NA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2FYFLQ#issuecomment-512459438>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACNI7TWHPO3BFUGPLVDADRTP75J6NANCNFSM4ERUD5LA> .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub<#4?email_source=notifications&email_token=AJJZIU6ZCIAOU66CNL4YNTTP77VRRA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2HI4MA#issuecomment-512658992>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AJJZIU4TNZHUBC6PIHOMT33P77VRRANCNFSM4ERUD5LA>.

[Ahsanr312]

I am facing the same cuda memory error while using ResNet 18. I am confused what to do. Can you help me out? On 18 Jul 2019 10:22 p.m., Ahsan Raza <[email protected]> wrote: Thankyou for your help. On 18 Jul 2019 9:20 a.m., varghese <[email protected]> wrote: Yes. I used resnet pretrained on imagenet for my experiments

On Wed, 17 Jul, 2019, 11:05 PM Ahsan Raza, ***@***.***> wrote: You used pretrained model of resnet-18 ? And through transfer learning achieved approximately similar results as DenseNet? On 14 Jul 2019 2:07 p.m., varghese ***@***.***> wrote: Surely i can help you with you with the idea. I initialized with pretrained weights, reduced the batch size ( i dont remember what the batch size was). All other hyperparameters were kept to same as what zoozog provided On Sun, 14 Jul, 2019, 1:46 PM Ahsan Raza, ***@***.***> wrote: > Thankyou for replying. > But can't you share the idea ? I have been searching alot but haven't find > anything accurate. DenseNet architecture requires alot of memory. And i get > cuda memory error. > So it would be so kind of you if you can share it with me. > > On 14 Jul 2019 11:08 a.m., varghese ***@***.***> wrote: > Thanks for the reply. However, I had surrender the model, code to my > previous institute. > > On Tue, 9 Jul 2019 at 11:36, Ahsan Raza ***@***.***> wrote: > > > Hi varghesealex90; > > Can you share your code for ResNet-18 ? > > > > — > > You are receiving this because you authored the thread. > > Reply to this email directly, view it on GitHub > > < > #4?email_source=notifications&email_token=ACNI7TSFBHGGOTKTR3CBO7TP6QTGBA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZPGGJI#issuecomment-509502245>, > > > or mute the thread > > < > https://github.com/notifications/unsubscribe-auth/ACNI7TSCNNCR5FB2JI55ZZTP6QTGBANCNFSM4ERUD5LA> > > > . > > > > — > You are receiving this because you commented. > Reply to this email directly, view it on GitHub< > #4?email_source=notifications&email_token=AJJZIU7E4C6CMIF5RM7U2ATP7K7GLA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ364VA#issuecomment-511176276>, > or mute the thread< > https://github.com/notifications/unsubscribe-auth/AJJZIU4TDID6LEKDL73R4ODP7K7GLANCNFSM4ERUD5LA>. > > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > < #4?email_source=notifications&email_token=ACNI7TVBWXU7XGB44P25HQTP7LOGFA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ4ARLQ#issuecomment-511183022>, > or mute the thread > < https://github.com/notifications/unsubscribe-auth/ACNI7TXSS6WOAEATEIALPBTP7LOGFANCNFSM4ERUD5LA> > . > — You are receiving this because you commented. Reply to this email directly, view it on GitHub< #4?email_source=notifications&email_token=AJJZIU7PMZYZJH43PR23LOLP7LUGJA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZ4BJSI#issuecomment-511186121>, or mute the thread< https://github.com/notifications/unsubscribe-auth/AJJZIUZZ2RQNBOPT3OR5NUDP7LUGJANCNFSM4ERUD5LA>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ACNI7TTSQESBO5CC75NXZ63P75J6NA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2FYFLQ#issuecomment-512459438>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACNI7TWHPO3BFUGPLVDADRTP75J6NANCNFSM4ERUD5LA> .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub<#4?email_source=notifications&email_token=AJJZIU6ZCIAOU66CNL4YNTTP77VRRA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2HI4MA#issuecomment-512658992>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AJJZIU4TNZHUBC6PIHOMT33P77VRRANCNFSM4ERUD5LA>.

[prasanth-rgb]

Hi,
My question may be stupid. Anyway, i am also in a struggle for building a mode for Chest XRay classification.

I have read that for transfer learning we have to find a model which had already trained with similar data set (medical image of chest xray). But here it is using Imagenet right?

Can we use Imagenet trained model for retraining with chexpert data set?

[varghesealex90]

Yes

…

On Tue, 3 Dec, 2019, 5:58 PM prasanth-rgb, ***@***.***> wrote: Hi, My question may be stupid. Anyway, i am also in a struggle for building a mode for Chest XRay classification. I have read that for transfer learning we have to find a model which had already trained with similar data set (medical image of chest xray). But here it is using Imagenet right? Can we use Imagenet trained model for retraining with chexpert data set? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ACNI7TRPWWG7OYY2LJXF5XTQWZGGFA5CNFSM4ERUD5LKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFZGPDA#issuecomment-561145740>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACNI7TWG5MWHFYKWO37QAKTQWZGGFANCNFSM4ERUD5LA> .

[Aliktk]

@varghesealex90 it's too late but as a researcher, I found this study and going to replicate it. Now is it possible now to help in replication?
If yes then I am sharing what I have done.
I want to retrain the model by changing the runTrain() and putting data into their folder as required.
I use the debugger to run the code as many errors coming during. The model loaded for training but before starting training it gives error which I mentioned after your response.
What was your system specs, what packages version you installed.
I am using Window IBM inctance with 64GB Ram and 16GB GPU.
Need you help thanks

[varghesealex90]

Hi @Aliktk can you please paste your errors. I ran the code on Ubuntu Machine and probably on a GPU with 12 Gb

[Aliktk]

@varghesealex90 Thank you so much for your quick response:
I have installed
python 3.8.0
CUDA.11
torch1.8.1
torchaudio 0.8.1
tensorflow-gpu 2.4.1
opencv-python 4.5.1.48
If I installed the wrong version please correct me. Also, tell me a simple method to replicate with same results
Following are the few problems that I have to suffer till.
Error-1: When I run the script it says you have an old version of CUDA please update the CUDA by this this....(RuntimeError: The NVIDIA driver on your system is too old (found version 10020). Please update your GPU driver by downloading and installing a new version from)
Error-2: AttributeError: Can't pickle local object 'ChexnetTrainer.test..'
Error-3: self = reduction.pickle.load(from_parent) EOFError: Ran out of input
Please tell me the exact version which I installed to relicate

[varghesealex90]

As mentioned previously, I did this activity 3 years back. I am guessing pytorch version should be 0.4.0