Predict issue about im2latex-tensorflow

[ritheshkumar95]

Hi Robin,

So the reason I dont use it anywhere particularly is that I generally train models on a detachable (screen) terminal with my script running in ipython. I generally just call my predict() function with no particular arguments once the training process gets over.

The predict function should work by default if you have all the data prepared as I have instructed. Let me know if you have any issues.

PS: Follow all my preprocessing instructions, (each and every line). Predict function requires files like test_buckets.npy and val_buckets.npy (which are basically images bucketed based on their size).

Predict function just randomly samples a mini-batch from those buckets and makes a free-run prediction using just the images.

[ritheshkumar95]

Hi Robin,

This architecture would certainly work well, for all problems related to OCR (In a way, im2latex itself is a type of OCR task). Handwriting is a harder version of OCR owing to the sheer variance in writing styles. But I must say it should work decently well atleast.

PS: Considering this architecture, it would be best to preprocess the handwriting data in a similar way, in the sense that you'd have to restrict it to a few bucket sizes not exceed by far in length. The bucket sizes (of images) used in this paper are: (120,50), (160,40), (200,40),(200,50), (240,40), (240,50), (280,40), (280,50), (320,40),(320,50), (360,40), (360,50), (360,60), (360, 100), (400,50),(400,160), (500,100).

Hope this helps! Good luck!

Best,
Rithesh

[RobinXL]

Hi Rithesh,
Thanks for being patient with. I have run a few test and the perplexity is pretty low. However, when I ipython to call the predict function as you mentioned, the predict error pops up:

FailedPreconditionError (see above for traceback): Attempting to use uninitialized value conv3_1//beta
[[Node: conv3_1//beta/read = IdentityT=DT_FLOAT, _class=["loc:@conv3_1//beta"], _device="/job:localhost/replica:0/task:0/gpu:0"]]
[[Node: scan_1/while/Identity/_964 = _Recvclient_terminated=false, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0/task:0/gpu:0", send_device_incarnation=1, tensor_name="edge_249_scan_1/while/Identity", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/cpu:0"]]

Do you have any idea about this error?

[ritheshkumar95]

Hey Robin,

I fixed this issue (seemed to be pretty new when I ran on my laptop as well). Check the latest commit.

Also, I added a link to my best weight files (from the latest checkpoint). Good luck!

Best,
Rithesh

[RobinXL]

Hey Rithesh,
Thank you for fixing the issue :)