Tensorflow implementation of handwritten sequense of small letter recognition.

The handwritten dataset used is IAM.

In order to make TFRecord file use the function from util.py

In order to run the training:

python trainer/run.py --input_path data/handwritten-test.tfrecords --input_path_test data/handwritten-test.tfrecords --model_dir models --board_path TFboard --filenameNr 1 --save_step 500  --batch_size 10 --max_steps 1000 --display_step 100

In order to get a sample from trained model:

python run.py --sample --shuffle_batch --batch_size 1 --input_path data/handwritten-test.tfrecords --input_path_test data/handwritten-test.tfrecords --model_dir models --board_path TFboard --filenameNr 1

In order to see statistics in tensorboard:

tensorboard --logdir=gs://my-first-bucket-mosnoi/handwritten/m2/TFboard2 --port=8080

params:

• --rnn_cell [LSTM,GRU,LSTMGRID2,GRUGRID2,BasicLSTM,LSTMGRID]
• --optimizer [ADAM,RMSP]
• --initializer [graves]
• --sample {make a sample}
• --shuffle_batch {shuffle the batch}
• --insertLastState {insert last state at next training step}
• --ctc_decoder [greedy,beam_search]

Google cloud ML:

rm -rf gs://my-first-bucket-mosnoi/handwritten3x200GRUGRID2

gcloud beta ml jobs submit training handwrittenRMSP3x200LSTM1 \
  --package-path=trainer \
  --module-name=trainer.run \
  --staging-bucket=gs://my-first-bucket-mosnoi/ \
  --region=us-central1 \
  --scale-tier=BASIC_GPU \
  -- \
  --input_path gs://my-first-bucket-mosnoi/handwritten/m2/tf-data/handwritten-test-{}.tfrecords \
  --input_path_test gs://my-first-bucket-mosnoi/handwritten/m2/tf-data/handwritten-test-55.tfrecords \
  --board_path gs://my-first-bucket-mosnoi/handwritten/m2/TFboard2_handwrittenRMSP3x200LSTM1 \
  --model_dir gs://my-first-bucket-mosnoi/handwritten/m2/models2 \
  --filenameNr 50 \
  --save_step 5000 \
  --display_step 100 \
  --max_steps 10000 \
  --batch_size 50 \
  --learning_rate 0.001 \
  --keep_prob 0.8 \
  --layers 3 \
  --hidden 250 \
  --rnn_cell LSTM \
  --optimizer RMSP \
  --initializer  graves \
  --bias -0.1 \
  --shuffle_batch \
  --gpu

  //--optimizer RMSP --momentum 0.9 --decay 0.95
  // python run.py --layers 1 --hidden 20 --rnn_cell GRUGRID2 --optimizer RMSP --insertLastState
  // python run.py --shuffle_batch  --layers 3   --sample --batch_size 1 --hidden 200 --insertLastState