This code generates CrossFit WODs (workout of the day) with recurrent neural networks (Andrej Karpathy's char-rnn model). The network is trained on 5k WOD samples collected from various crossfit boxes. Despite of small dataset, the results are both fun and satisfying.
Some collection of the good results can be seen from here.
Of course, not all the results are great. Here are some funny ones:
Zero WOD
0 rounds for time 15 minutes
800 mt run
10 sumo deadlift high pull
9 front squat
1 muscle-up
Rest WOD
5 rounds for time
20 minutes rest
Nearly imposibble WOD
5 rounds for time 20 minutes
100 double under
50 wall-ball
In this WOD, there is no row but the network still wants burpees over the row
7 rounds for time
18 dumbbell squat clean
15 burpees over the row
250 mt run
Some imaginary exercises
brusseane push-up
t-up
touster
lean & jerk
publ-up
dumbell burpee over the bar
hanging ring in
louble under
ode-hand dip
roundstand walk
tempo kim back extension
muscle sprint
pistol squat snatch
over clean
elite push-up
inverted barbell
rest clean
pill-up
All the WODs are collected from:
- crossfit.com
- findawod.com
- crossfitbalabanlevent.blogspot.com
After collecting, all of WODs are syntactically normalized by hand. That's why the dataset is that small, I wish to feed a lot more WODs but all the sources uses different format, some uses abbreviation some don't. To get best results, whole dataset should be in the same format.
The dataset can be found in data/wods.txt, or better they are in wod table in sqlite database located db.sqlite.
Feel free to use it.
Requirements:
- Python 2.7 (not tested in 3)
- Tensorflow > 1.0
> cd char-rnn
> python train.py --help
usage: train.py [-h] [--input_file INPUT_FILE] [--rnn_size RNN_SIZE]
[--num_layers NUM_LAYERS] [--model MODEL]
[--batch_size BATCH_SIZE] [--seq_length SEQ_LENGTH]
[--num_epochs NUM_EPOCHS] [--log_step LOG_STEP]
[--grad_clip GRAD_CLIP] [--learning_rate LEARNING_RATE]
[--decay_rate DECAY_RATE]
[--input_dropout_keep_prob INPUT_DROPOUT_KEEP_PROB]
[--output_dropout_keep_prob OUTPUT_DROPOUT_KEEP_PROB]
[--train_root_dir TRAIN_ROOT_DIR] [--vocab_size VOCAB_SIZE]
optional arguments:
-h, --help show this help message and exit
--input_file INPUT_FILE
Input text file to train on
--rnn_size RNN_SIZE The size of RNN hidden state
--num_layers NUM_LAYERS
The number of layers in the RNN
--model MODEL RNN model: rnn, gru, lstm, or nas
--batch_size BATCH_SIZE
Batch size
--seq_length SEQ_LENGTH
RNN sequence length
--num_epochs NUM_EPOCHS
Number of epochs for training
--log_step LOG_STEP Logging period in terms of iteration
--grad_clip GRAD_CLIP
Clip gradients value
--learning_rate LEARNING_RATE
Learning rate for adam optimizer
--decay_rate DECAY_RATE
Learning rate for adam optimizer
--input_dropout_keep_prob INPUT_DROPOUT_KEEP_PROB
Input dropout keep probability
--output_dropout_keep_prob OUTPUT_DROPOUT_KEEP_PROB
Output dropout keep probability
--train_root_dir TRAIN_ROOT_DIR
Root directory to put the training data
> cd char-rnn
> python sample.py --help
usage: sample.py [-h] [--data_dir DATA_DIR] [--seperator_char SEPERATOR_CHAR]
[--num_sample NUM_SAMPLE] [--save_to_db [SAVE_TO_DB]]
[--nosave_to_db]
optional arguments:
-h, --help show this help message and exit
--data_dir DATA_DIR Training data directory. If empty, latest folder in
training/ folder will be used
--seperator_char SEPERATOR_CHAR
WOD item seperator character, default `|`
--num_sample NUM_SAMPLE
The number of WODs to be sampled, default 1
--save_to_db [SAVE_TO_DB]
Should save into sqlite, default false
--nosave_to_db
