Hi Kranthi Kiran, Please find my replies below:
Let me know if you have more questions. Thanks, TG *--* *Thamme Gowda* TG | @thammegowda <https://twitter.com/thammegowda> ~Sent via somebody's Webmail server! On Tue, Mar 21, 2017 at 12:21 PM, Kranthi Kiran G V < [email protected]> wrote: > Hello Thamme Gowda, > > Thank you for letting me know of the developer mailing list. I have > created an issue [1] and I would be working on it. > The change is not straightforward since Inception V3 pre-trained model has > a graph while the Inception V3 pre-trained model is packaged in the form of > a check-point (ckpt) [2]. > Okay, I see Inception-V3 has a graph, V4 has a checkpoint. I assume there should be a way to restore model from checkpoint? Please refer https://www.tensorflow.org/programmers_guide/variables#checkpoint_files > > What do you think of using Keras to implement the Inception V4 model? It > would make the job of scaling it on CPU clusters easier if we can use > deeplearning4j's model import. > > Should I proceed in that direction? > > Regarding GSoC, what kind of computation resources are we given access to? > We would have to train the show and tell network. It takes a lot of > computation resources. > > If GPUs are not used, we would have to use a CPU cluster. So, the code has > to be re-written (from the Google implementation of Inception V4). > > Training IncpetionV4 from scratch requires too much effort, time, and resources. We are not aiming for such things, atleast not as part of Tika and GSoC. The suggestion i mentioned earlier was to upgrade IncpetionV3 model with Inception V4 pretrained model/checkpoint since that will be more benificial to Tika users community :-) > > [1] https://issues.apache.org/jira/browse/TIKA-2306 > [2] https://github.com/tensorflow/models/tree/master/ > slim#pre-trained-models > > > > > > On Mon, Mar 20, 2017 at 3:17 AM, Thamme Gowda <[email protected]> > wrote: > >> Hi Kranthi Kiran, >> >> Welcome to Tika Community. we are glad you are interested in working on >> the issue. >> Please remember to CC dev@tika mailing list for future discussions >> related to tika. >> >> *Should the model be trainable by the user?* >> The basic minimum requirement is to provide a pre-trained model and make >> the parser work out of the box without Training (expect no GPUs; expect >> a JVM and nothing else). >> Of course, the parser configuration should have options to change the >> models by changing the path. >> >> As part of this GSoC project, integration isn't enough work. If you go >> through the links provided in the Jira page you will notice that there >> models for image recognition but no ready-made models for captioning. We >> will have to train the im2text network from the dataset and make it >> available. Thus we will have to open source the training utilities, >> documentation or any supplementary tools we build along the way. We will >> have to document all these in Tika wiki for the advanced users! >> >> This is a GSoC issue and thus we expect to work on it during the summer. >> >> For now, if you want a small task to familiarise yourself with Tika, I >> have a suggestion: >> Currently, Tika uses InceptionV3 model from Google for image recognition. >> The InceptionV4 model is out recently which proved to be more accurate >> than V3. >> >> How about upgrading tika to use newer Inception model? >> >> Let me know if you have more questions. >> >> Cheers, >> TG >> >> *--* >> *Thamme Gowda* >> TG | @thammegowda <https://twitter.com/thammegowda> >> ~Sent via somebody's Webmail server! >> >> On Sun, Mar 19, 2017 at 11:56 AM, Kranthi Kiran G V < >> [email protected]> wrote: >> >>> Hello, >>> I'm Kranthi, a 3rd computer science undergrad at NIT, Warangal and a >>> member of Deep Learning research group at out college. I'm interested to >>> take up the issue. I believe it would be a great contribution to the Apache >>> Tika community. >>> >>> This is what I have done until now: >>> >>> 1) Build Tika from source using maven and explore it. >>> 2) Tried the object recognition module from the command line. (I should >>> probably start using the docker version to speed up my progress.) >>> >>> I am yet to import a keras model in dl4j. I have some doubts regarding >>> the requirements since I'm new to this community. *Should the model be >>> trainable by the user?* This is important because the Inception v3 >>> model without re-training has performed poorly for me (I'm currently >>> training it with less number of steps due to limited computational >>> resources I have -- GTX 1070). >>> >>> TODO (Before submitting the proposal): >>> >>> 1) Create a test REST API for Tika >>> 2) Import a few models in dl4j. >>> 3) Train im2txt on my computer. >>> >>> Thank you, >>> Kranthi Kiran >>> >> >> >
