Hi everyone,
I'm trying to train my own OCR Service based on my own set of images words, as the sample servicces words_mnist or the multiword_ocr.
But the predictions service return are either meansless or blank.
So I decided to startfrom scrach, and train sample service words_mnist with the set of works provided.
But the result is the same, predictions are either meansless or blank.
Here are the requests:
Service creation
curl -X PUT "http://localhost:8080/services/word_mnist" -d '
{
"mllib": "caffe",
"description": "word_mnist",
"type": "supervised",
"parameters": {
"input": {
"connector": "image",
"width": 128,
"height": 80,
"bw": false,
"db": true,
"ctc": true
},
"mllib": {
"template": "crnn",
"nclasses": 100,
"rotate": false,
"mirror": false,
"scale": 1.0,
"layers": [],
"db": true,
"noise": {
"all_effects": true,
"prob": 0.001
},
"distort": {
"all_effects": true,
"prob": 0.001
},
"gpu": true,
"gpuid": 1
},
"output": {}
},
"model": {
"templates": "../templates/caffe/",
"repository": "/opt/models/words_mnist"
}
}
'
Service prediction
curl -X POST "http://localhost:8080/train" '
{
"service": "word_mnist",
"async": true,
"parameters": {
"input": {
"test_split": 0.0,
"shuffle": true,
"db": true
},
"mllib": {
"gpu": true,
"net": {
"batch_size": 1,
"test_batch_size": 1
},
"solver": {
"test_initialization": false,
"iterations": 10000,
"test_interval": 1000,
"snapshot": 1000,
"base_lr": 0.0001,
"solver_type": "AMSGRAD",
"lookahead": true,
"rectified": true,
"iter_size": 1
},
"timesteps": 4
},
"output": {
"measure": [
"acc"
],
"target_repository": ""
}
},
"data": [
"/opt/samples/word_mnist/train.txt",
"/opt/samples/word_mnist/test.txt"
]
}
'
Someone now what I am doing wrong?
I forgot to paste the "-d" on the chat, but my problem is other.
I got the trainnning running successful , but if i set the batch_size to 128 and test_batch_size 32 the dd server stop working.
The log server reports this problem (Segmentation fault) :
[words_mnist] [info] detected network type is ctc
Segmentation fault
My test enviroment is a docker dd over Windows ( 6 CPUs, 18 GB RAM )
I'm guessing that I'm runnnig out of resources, problably memory.
I have lower it as much as I can, but only seems to complete all the iterantions if I set the batch size to one or maximum two.
Even setting the batch size to 1, the result of the predictions is meanless, the respond is either some strange characters or non at all.
As far as I understand, is this example, the batch size whould be the number of images to process every iteration. Therefore lowering the batch size would make the training process slower but the result should be the same.
That is the issue that i dont undenstant, lowering the batch size seems to work and completes the training, but the result of the predictions is meanless.
Best
YaYaB.
