Open Dawson-huang opened 7 years ago
@Dawson-huang Hey I actually plan to add a LSTM layers after the CNN feature extraction. After that, the results should be much better.
Yeah, maybe add a LSTM layer to better handle the contextual semantic issues of CNN features, let's work it together!
Sure! It will be great to have someone work with together! Feel free to submit any pull request.
sir the prediction of the text is not related to my test image what i have to do now did you add LSTM layer if you have done then please explain how to add it
Predicted Vectors are: [-0.06664543 0.03540059 -0.00320938 ..., 0.03019159 -0.08383787 0.09057866] 2017-10-16 09:44:21.008353: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GT 1030, pci bus id: 0000:01:00.0) ['dun', 'eagle', 'and', 'Altiplano', 'divers', 'eagle', 'Altiplano', 'pedestrian', 'reflecting', 'mushrooms', 'crackers', 'toilet', 'linesmen', 'vandalised', 'image', 'causing', 'text', 'tiredly', 'starfish', 'arm', 'engines', 'nuts', 'slats', 'nice', 'Futuristic', 'Lego', 'cups', 'blurred', 'bombs', 'swim']
I trained 10,000 times the model, but the prediction is not a true answer to the test image. For example, the above prediction is not a description of the image (./iaprtc12/images/00/25.jpg). Can tell you me how to actually reproduce your project, thank you!