├── Readme.md ├── cv └── lm_lstm_epoch30.00_1.3904.t7 ├── data └── tinyshakespeare │ ├── data.t7 │ ├── input.txt │ └── vocab.t7 ├── inspect_checkpoint.lua ├── model ├── GRU.lua ├── LSTM.lua └── RNN.lua ├── sample.lua ├── server.py ├── templates └── main.html ├── train.lua └── util ├── CharSplitLMMinibatchLoader.lua ├── OneHot.lua ├── misc.lua └── model_utils.lua /Readme.md: -------------------------------------------------------------------------------- 1 | # char-rnn 2 | A multi-layer Recurrent Neural Network (RNN, LSTM, and GRU) for training/sampling from character-level language models. The input is a single text file and the model learns to predict the next character in the sequence. More info here and here. Created by (https://twitter.com/karpathy). 3 | 4 | # char-rnn-API 5 | An API and web frontend for char-rnn, running on python/flask. 6 | Hoping to see many public char-rnn micro-api´s with different models spring up, so we can experiment together more easily. Created by (https://twitter.com/samim). 7 | 8 |  9 | 10 | # instructions 11 | - install torch: http://torch.ch/docs/getting-started.html 12 | - install `luarocks install nngraph` and `luarocks install optim` 13 | - install flask: http://flask.pocoo.org/docs/0.10/installation/ 14 | - install flask flask cors: `pip install -U flask-cors` 15 | - `git clone https://github.com/samim23/char-rnn-api` 16 | - python server.py 17 | - goto https://thisserver.com:8080 18 | 19 | # API calls 20 | Post json request to: http://thisserver.com/api/v1.0 21 | {"primetext":"mytext", "temperature":"1", "length":"2000", "gpuid":"-1", "model":"model.t7","seed":"123", "sample":"1" } 22 | 23 | 24 | # char-rnn 25 | 26 | This code implements **multi-layer Recurrent Neural Network** (RNN, LSTM, and GRU) for training/sampling from character-level language models. The input is a single text file and the model learns to predict the next character in the sequence. 27 | 28 | The context of this code base is described in detail in my [blog post](http://karpathy.github.io/2015/05/21/rnn-effectiveness/). 29 | 30 | There is also a [project page](http://cs.stanford.edu/people/karpathy/char-rnn/) that has some pointers and datasets. 31 | 32 | This code is based on Oxford University Machine Learning class [practical 6](https://github.com/oxford-cs-ml-2015/practical6), which is in turn based on [learning to execute](https://github.com/wojciechz/learning_to_execute) code from Wojciech Zaremba. Chunks of it were also developed in collaboration with my labmate [Justin Johnson](https://github.com/jcjohnson/). 33 | 34 | ## Requirements 35 | 36 | This code is written in Lua and requires [Torch](http://torch.ch/). 37 | Additionally, you need to install the `nngraph` and `optim` packages using [LuaRocks](https://luarocks.org/) which you will be able to do after installing Torch 38 | 39 | ```bash 40 | $ luarocks install nngraph 41 | $ luarocks install optim 42 | ``` 43 | 44 | ## Usage 45 | 46 | 47 | ### Data 48 | 49 | All input data is stored inside the `data/` directory. You'll notice that there is an example dataset included in the repo (in folder `data/tinyshakespeare`) which consists of a subset of works of Shakespeare. I'm providing a few more datasets on the [project page](http://cs.stanford.edu/people/karpathy/char-rnn/). 50 | 51 | **Your own data**: If you'd like to use your own data create a single file `input.txt` and place it into a folder in `data/`. For example, `data/some_folder/input.txt`. The first time you run the training script it will write two more convenience files into `data/some_folder`. **Note**: If you change the file `input.txt` in place you currently must delete the two intermediate files manually to force the preprocessing to re-run. 52 | 53 | Note that if your data is too small (1MB is already considered very small) the RNN won't learn very effectively. Remember that it has to learn everything completely from scratch. 54 | 55 | ### Training 56 | 57 | Start training the model using `train.lua`, for example: 58 | 59 | ``` 60 | $ th train.lua -data_dir data/some_folder -gpuid -1 61 | ``` 62 | 63 | The `-data_dir` flag is most important since it specifies the dataset to use. Notice that in this example we're also setting `gpuid` to -1 which tells the code to train using CPU, otherwise it defaults to GPU 0. There are many other flags for various options. Consult `$ th train.lua -help` for comprehensive settings. Here's another example: 64 | 65 | ``` 66 | $ th train.lua -data_dir data/some_folder -rnn_size 512 -num_layers 2 -dropout 0.5 67 | ``` 68 | 69 | While the model is training it will periodically write checkpoint files to the `cv` folder. The frequency with which these checkpoints are written is controlled with number of iterations, as specified with the `eval_val_every` option (e.g. if this is 1 then a checkpoint is written every iteration). 70 | 71 | We can use these checkpoints to generate text (discussed next). 72 | 73 | ### Sampling 74 | 75 | Given a checkpoint file (such as those written to `cv`) we can generate new text. For example: 76 | 77 | ``` 78 | $ th sample.lua cv/some_checkpoint.t7 -gpuid -1 79 | ``` 80 | 81 | Make sure that if your checkpoint was trained with GPU it is also sampled from with GPU, or vice versa. Otherwise the code will (currently) complain. As with the train script, see `$ th sample.lua -help` for full options. One important one is (for example) `-length 10000` which would generate 10,000 characters (default = 2000). 82 | 83 | **Temperature**. An important parameter you may want to play with a lot is `-temparature`, which takes a number in range (0, 1] (notice 0 not included), default = 1. The temperature is dividing the predicted log probabilities before the Softmax, so lower temperature will cause the model to make more likely, but also more boring and conservative predictions. Higher temperatures cause the model to take more chances and increase diversity of results, but at a cost of more mistakes. 84 | 85 | **Priming**. It's also possible to prime the model with some starting text using `-primetext`. 86 | 87 | Happy sampling! 88 | 89 | ## Tips and Tricks 90 | 91 | ### Monitoring Validation Loss vs. Training Loss 92 | If you're somewhat new to Machine Learning or Neural Networks it can take a bit of expertise to get good models. The most important quantity to keep track of is the difference between your training loss (printed during training) and the validation loss (printed once in a while when the RNN is run on the validation data (by default every 1000 iterations)). In particular: 93 | 94 | - If your training loss is much lower than validation loss then this means the network is **overfitting**. Solutions to this are to decrease your network size, or to increase dropout. For example you could try dropout of 0.5 and so on. 95 | - If your training/validation loss are about equal then your model is **underfitting**. Increase the size of your model (either number of layers or the raw number of neurons per layer) 96 | 97 | ### Approximate number of parameters 98 | 99 | The two most important parameters that control the model are `rnn_size` and `num_layers`. I would advise that you always use `num_layers` of about 3. The `rnn_size` can be adjusted based on how much data you have. The two important quantities to keep track of here are: 100 | 101 | - The number of parameters in your model. This is printed when you start training. 102 | - The size of your dataset. 1MB file is approximately 1 million characters. 103 | 104 | These two should be about the same order of magnitude. It's a little tricky to tell. Here are some examples: 105 | 106 | - I have a 100MB dataset and I'm using the default parameter settings (which currently print 150K parameters). My data size is significantly larger (100 mil >> 0.15 mil), so I expect to heavily underfit. I am thinking I can comfortably afford to make `rnn_size` larger. 107 | - I have a 10MB dataset and running a 10 million parameter model. I'm slightly nervous and I'm carefully monitoring my validation loss. If it's larger than my training loss then I may want to increase dropout a bit. 108 | 109 | ### Best models strategy 110 | 111 | The winning strategy to obtaining very good models (if you have the compute time) is to always err on making the network larger (as large as you're willing to wait for it to compute) and then try different dropout values (between 0,1). Whatever model has the best validation performance (the loss, written in the checkpoint filename, low is good) is the one you should use in the end. 112 | 113 | It is very common in deep learning to run many different models with many different hyperparameter settings, and in the end take whatever checkpoint gave the best validation performance. 114 | 115 | By the way, the size of your training and validation splits are also parameters. Make sure you have a decent amount of data in your validation set or otherwise the validation performance will be noisy and not very informative. 116 | 117 | ## License 118 | 119 | MIT 120 | 121 | 122 | ## Datasets 123 | - text from https://cs.stanford.edu/people/karpathy/char-rnn/ 124 | 125 | Cleaner version of this page coming soon, but for now some fun datasets: 126 | 127 | - Linux Kernel (6.2MB) 128 | https://cs.stanford.edu/people/karpathy/char-rnn/linux_input.txt 129 | 130 | The above is only the kernel. The examples in my blog post were trained on the full Linux code base. That is: 131 | 132 | $ git clone https://github.com/torvalds/linux.git 133 | $ cd linux 134 | $ find . -name "*.[c|h]" | shuf | xargs cat > linux.txt 135 | 136 | (This gives a 474MB file that I plugged in) 137 | 138 | - All works of Shakespeare concatenated (4.6MB) 139 | https://cs.stanford.edu/people/karpathy/char-rnn/shakespeare_input.txt 140 | 141 | - Leo Tolstoy's War and Peace (3.3MB) 142 | https://cs.stanford.edu/people/karpathy/char-rnn/warpeace_input.txt 143 | 144 | - Free books: (in general, including War and Peace) can be found in https://www.gutenberg.org/. 145 | 146 | - Wikipedia: 147 | 100MB Wikipedia data Hutter Prize 148 | http://prize.hutter1.net/ 149 | 150 | - The Stacks Project 151 | http://stacks.math.columbia.edu/ 152 | which is where the Latex dataset on Algebraic Geometry came from. 153 | 154 | -------------------------------------------------------------------------------- /cv/lm_lstm_epoch30.00_1.3904.t7: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/samim23/char-rnn-api/a81d3894f59e2fc4dea069efc0eb418145d80d35/cv/lm_lstm_epoch30.00_1.3904.t7 -------------------------------------------------------------------------------- /data/tinyshakespeare/vocab.t7: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/samim23/char-rnn-api/a81d3894f59e2fc4dea069efc0eb418145d80d35/data/tinyshakespeare/vocab.t7 -------------------------------------------------------------------------------- /inspect_checkpoint.lua: -------------------------------------------------------------------------------- 1 | -- simple script that loads a checkpoint and prints its opts 2 | 3 | require 'torch' 4 | require 'nn' 5 | require 'nngraph' 6 | require 'cutorch' 7 | require 'cunn' 8 | 9 | require 'util.OneHot' 10 | require 'util.misc' 11 | 12 | cmd = torch.CmdLine() 13 | cmd:text() 14 | cmd:text('Load a checkpoint and print its options and validation losses.') 15 | cmd:text() 16 | cmd:text('Options') 17 | cmd:argument('-model','model to load') 18 | cmd:option('-gpuid',0,'gpu to use') 19 | cmd:text() 20 | 21 | -- parse input params 22 | opt = cmd:parse(arg) 23 | 24 | print('using CUDA on GPU ' .. opt.gpuid .. '...') 25 | require 'cutorch' 26 | require 'cunn' 27 | cutorch.setDevice(opt.gpuid + 1) 28 | 29 | local model = torch.load(opt.model) 30 | 31 | print('opt:') 32 | print(model.opt) 33 | print('val losses:') 34 | print(model.val_losses) 35 | 36 | -------------------------------------------------------------------------------- /model/GRU.lua: -------------------------------------------------------------------------------- 1 | 2 | local GRU = {} 3 | 4 | --[[ 5 | Creates one timestep of one GRU 6 | Paper reference: http://arxiv.org/pdf/1412.3555v1.pdf 7 | ]]-- 8 | function GRU.gru(input_size, rnn_size, n) 9 | 10 | -- there are n+1 inputs (hiddens on each layer and x) 11 | local inputs = {} 12 | table.insert(inputs, nn.Identity()()) -- x 13 | for L = 1,n do 14 | table.insert(inputs, nn.Identity()()) -- prev_h[L] 15 | end 16 | 17 | function new_input_sum(insize, xv, hv) 18 | local i2h = nn.Linear(insize, rnn_size)(xv) 19 | local h2h = nn.Linear(rnn_size, rnn_size)(hv) 20 | return nn.CAddTable()({i2h, h2h}) 21 | end 22 | 23 | local x, input_size_L 24 | local outputs = {} 25 | for L = 1,n do 26 | 27 | local prev_h = inputs[L+1] 28 | if L == 1 then x = inputs[1] else x = outputs[L-1] end 29 | if L == 1 then input_size_L = input_size else input_size_L = rnn_size end 30 | 31 | -- GRU tick 32 | -- forward the update and reset gates 33 | local update_gate = nn.Sigmoid()(new_input_sum(input_size_L, x, prev_h)) 34 | local reset_gate = nn.Sigmoid()(new_input_sum(input_size_L, x, prev_h)) 35 | -- compute candidate hidden state 36 | local gated_hidden = nn.CMulTable()({reset_gate, prev_h}) 37 | local p2 = nn.Linear(rnn_size, rnn_size)(gated_hidden) 38 | local p1 = nn.Linear(input_size_L, rnn_size)(x) 39 | local hidden_candidate = nn.Tanh()(nn.CAddTable()({p1,p2})) 40 | -- compute new interpolated hidden state, based on the update gate 41 | local zh = nn.CMulTable()({update_gate, hidden_candidate}) 42 | local zhm1 = nn.CMulTable()({nn.AddConstant(1,false)(nn.MulConstant(-1,false)(update_gate)), prev_h}) 43 | local next_h = nn.CAddTable()({zh, zhm1}) 44 | 45 | table.insert(outputs, next_h) 46 | end 47 | 48 | return nn.gModule(inputs, outputs) 49 | end 50 | 51 | return GRU 52 | 53 | -------------------------------------------------------------------------------- /model/LSTM.lua: -------------------------------------------------------------------------------- 1 | 2 | local LSTM = {} 3 | function LSTM.lstm(input_size, rnn_size, n, dropout) 4 | dropout = dropout or 0 5 | 6 | -- there will be 2*n+1 inputs 7 | local inputs = {} 8 | table.insert(inputs, nn.Identity()()) -- x 9 | for L = 1,n do 10 | table.insert(inputs, nn.Identity()()) -- prev_c[L] 11 | table.insert(inputs, nn.Identity()()) -- prev_h[L] 12 | end 13 | 14 | local x, input_size_L 15 | local outputs = {} 16 | for L = 1,n do 17 | -- c,h from previos timesteps 18 | local prev_h = inputs[L*2+1] 19 | local prev_c = inputs[L*2] 20 | -- the input to this layer 21 | if L == 1 then x = inputs[1] else x = outputs[(L-1)*2] end 22 | if L == 1 then input_size_L = input_size else input_size_L = rnn_size end 23 | -- evaluate the input sums at once for efficiency 24 | local i2h = nn.Linear(input_size_L, 4 * rnn_size)(x) 25 | local h2h = nn.Linear(rnn_size, 4 * rnn_size)(prev_h) 26 | local all_input_sums = nn.CAddTable()({i2h, h2h}) 27 | -- decode the gates 28 | local sigmoid_chunk = nn.Narrow(2, 1, 3 * rnn_size)(all_input_sums) 29 | sigmoid_chunk = nn.Sigmoid()(sigmoid_chunk) 30 | local in_gate = nn.Narrow(2, 1, rnn_size)(sigmoid_chunk) 31 | local forget_gate = nn.Narrow(2, rnn_size + 1, rnn_size)(sigmoid_chunk) 32 | local out_gate = nn.Narrow(2, 2 * rnn_size + 1, rnn_size)(sigmoid_chunk) 33 | -- decode the write inputs 34 | local in_transform = nn.Narrow(2, 3 * rnn_size + 1, rnn_size)(all_input_sums) 35 | in_transform = nn.Tanh()(in_transform) 36 | -- perform the LSTM update 37 | local next_c = nn.CAddTable()({ 38 | nn.CMulTable()({forget_gate, prev_c}), 39 | nn.CMulTable()({in_gate, in_transform}) 40 | }) 41 | -- gated cells form the output 42 | local next_h = nn.CMulTable()({out_gate, nn.Tanh()(next_c)}) 43 | -- add dropout to output, if desired 44 | if dropout > 0 then next_h = nn.Dropout(dropout)(next_h) end 45 | 46 | table.insert(outputs, next_c) 47 | table.insert(outputs, next_h) 48 | end 49 | 50 | return nn.gModule(inputs, outputs) 51 | end 52 | 53 | return LSTM 54 | 55 | -------------------------------------------------------------------------------- /model/RNN.lua: -------------------------------------------------------------------------------- 1 | local RNN = {} 2 | 3 | function RNN.rnn(input_size, rnn_size, n) 4 | 5 | -- there are n+1 inputs (hiddens on each layer and x) 6 | local inputs = {} 7 | table.insert(inputs, nn.Identity()()) -- x 8 | for L = 1,n do 9 | table.insert(inputs, nn.Identity()()) -- prev_h[L] 10 | end 11 | 12 | local x, input_size_L 13 | local outputs = {} 14 | for L = 1,n do 15 | 16 | local prev_h = inputs[L+1] 17 | if L == 1 then x = inputs[1] else x = outputs[L-1] end 18 | if L == 1 then input_size_L = input_size else input_size_L = rnn_size end 19 | 20 | -- RNN tick 21 | local i2h = nn.Linear(input_size_L, rnn_size)(x) 22 | local h2h = nn.Linear(rnn_size, rnn_size)(prev_h) 23 | local next_h = nn.Tanh()(nn.CAddTable(){i2h, h2h}) 24 | 25 | table.insert(outputs, next_h) 26 | end 27 | 28 | return nn.gModule(inputs, outputs) 29 | end 30 | 31 | return RNN 32 | -------------------------------------------------------------------------------- /sample.lua: -------------------------------------------------------------------------------- 1 | 2 | --[[ 3 | 4 | This file samples characters from a trained model 5 | 6 | Code is based on implementation in 7 | https://github.com/oxford-cs-ml-2015/practical6 8 | 9 | ]]-- 10 | 11 | require 'torch' 12 | require 'nn' 13 | require 'nngraph' 14 | require 'optim' 15 | require 'lfs' 16 | 17 | require 'util.OneHot' 18 | require 'util.misc' 19 | 20 | cmd = torch.CmdLine() 21 | cmd:text() 22 | cmd:text('Sample from a character-level language model') 23 | cmd:text() 24 | cmd:text('Options') 25 | -- required: 26 | cmd:argument('-model','model checkpoint to use for sampling') 27 | -- optional parameters 28 | cmd:option('-seed',123,'random number generator\'s seed') 29 | cmd:option('-sample',1,' 0 to use max at each timestep, 1 to sample at each timestep') 30 | cmd:option('-primetext'," ",'used as a prompt to "seed" the state of the LSTM using a given sequence, before we sample.') 31 | cmd:option('-length',2000,'number of characters to sample') 32 | cmd:option('-temperature',1,'temperature of sampling') 33 | cmd:option('-gpuid',0,'which gpu to use. -1 = use CPU') 34 | cmd:text() 35 | 36 | -- parse input params 37 | opt = cmd:parse(arg) 38 | 39 | if opt.gpuid >= 0 then 40 | print('using CUDA on GPU ' .. opt.gpuid .. '...') 41 | require 'cutorch' 42 | require 'cunn' 43 | cutorch.setDevice(opt.gpuid + 1) -- note +1 to make it 0 indexed! sigh lua 44 | end 45 | torch.manualSeed(opt.seed) 46 | 47 | -- load the model checkpoint 48 | if not lfs.attributes(opt.model, 'mode') then 49 | print('Error: File ' .. opt.model .. ' does not exist. Are you sure you didn\'t forget to prepend cv/ ?') 50 | end 51 | checkpoint = torch.load(opt.model) 52 | 53 | 54 | local vocab = checkpoint.vocab 55 | local ivocab = {} 56 | for c,i in pairs(vocab) do ivocab[i] = c end 57 | 58 | protos = checkpoint.protos 59 | local rnn_idx = #protos.softmax.modules - 1 60 | opt.rnn_size = protos.softmax.modules[rnn_idx].weight:size(2) 61 | 62 | -- initialize the rnn state 63 | local current_state, state_predict_index 64 | local model = checkpoint.opt.model 65 | 66 | print('creating an LSTM...') 67 | local num_layers = checkpoint.opt.num_layers or 1 -- or 1 is for backward compatibility 68 | current_state = {} 69 | for L=1,checkpoint.opt.num_layers do 70 | -- c and h for all layers 71 | local h_init = torch.zeros(1, opt.rnn_size) 72 | if opt.gpuid >= 0 then h_init = h_init:cuda() end 73 | table.insert(current_state, h_init:clone()) 74 | table.insert(current_state, h_init:clone()) 75 | end 76 | state_predict_index = #current_state -- last one is the top h 77 | local seed_text = opt.primetext 78 | local prev_char 79 | 80 | protos.rnn:evaluate() -- put in eval mode so that dropout works properly 81 | 82 | -- do a few seeded timesteps 83 | print('seeding with ' .. seed_text) 84 | for c in seed_text:gmatch'.' do 85 | prev_char = torch.Tensor{vocab[c]} 86 | if opt.gpuid >= 0 then prev_char = prev_char:cuda() end 87 | local embedding = protos.embed:forward(prev_char) 88 | current_state = protos.rnn:forward{embedding, unpack(current_state)} 89 | if type(current_state) ~= 'table' then current_state = {current_state} end 90 | end 91 | 92 | -- start sampling/argmaxing 93 | for i=1, opt.length do 94 | 95 | -- softmax from previous timestep 96 | local next_h = current_state[state_predict_index] 97 | next_h = next_h / opt.temperature 98 | local log_probs = protos.softmax:forward(next_h) 99 | 100 | if opt.sample == 0 then 101 | -- use argmax 102 | local _, prev_char_ = log_probs:max(2) 103 | prev_char = prev_char_:resize(1) 104 | else 105 | -- use sampling 106 | local probs = torch.exp(log_probs):squeeze() 107 | prev_char = torch.multinomial(probs:float(), 1):resize(1):float() 108 | end 109 | 110 | -- forward the rnn for next character 111 | local embedding = protos.embed:forward(prev_char) 112 | current_state = protos.rnn:forward{embedding, unpack(current_state)} 113 | if type(current_state) ~= 'table' then current_state = {current_state} end 114 | 115 | io.write(ivocab[prev_char[1]]) 116 | end 117 | io.write('\n') io.flush() 118 | 119 | -------------------------------------------------------------------------------- /server.py: -------------------------------------------------------------------------------- 1 | from flask import Flask 2 | from flask import jsonify,render_template,redirect,url_for,request,abort 3 | from flask.ext.cors import CORS, cross_origin 4 | import json 5 | import subprocess 6 | 7 | app = Flask(__name__) 8 | 9 | modelsDirectory = 'cv' 10 | 11 | @app.route('/') 12 | def index(): 13 | return render_template('main.html') 14 | 15 | @app.route('/api/v1.0', methods=['POST']) 16 | def api_v1(): 17 | if not request.json or not 'primetext' in request.json: 18 | abort(400) 19 | 20 | primetext = request.json['primetext'] 21 | temperature = request.json['temperature'] 22 | length = request.json['length'] 23 | model = request.json['model'] 24 | seed = request.json['seed'] 25 | sample = request.json['sample'] 26 | gpuid = request.json['gpuid'] 27 | # override for public APIs 28 | gpuid = '-1' 29 | 30 | searchstring = 'th ../char-rnn/sample.lua ../char-rnn/'+modelsDirectory+'/' + str(model) 31 | searchstring += ' -gpuid ' + str(gpuid) 32 | searchstring += ' -primetext "' + str(primetext) + '"' 33 | searchstring += ' -temperature ' + str(temperature) 34 | searchstring += ' -length ' + str(length) 35 | searchstring += ' -seed ' + str(seed) 36 | searchstring += ' -sample ' + str(sample) 37 | 38 | responds = subprocess.Popen(searchstring, shell=True, stdout=subprocess.PIPE).stdout.read() 39 | 40 | # remove console stats output 41 | responds = responds.split('\n', 1)[1].split('\n', 1)[1].split('\n', 1)[1] 42 | 43 | return jsonify({'responds': responds}), 201 44 | 45 | @app.route('/api/v1.0/model', methods=['POST']) 46 | def api_v1_model(): 47 | searchstring = '(cd ../char-rnn/cv/ && ls -t)' 48 | responds = subprocess.Popen(searchstring, shell=True, stdout=subprocess.PIPE).stdout.read() 49 | responds = responds.splitlines(); 50 | return jsonify({'models': responds}), 201 51 | 52 | if __name__ == "__main__": 53 | app.run(host='0.0.0.0', port=8080) 54 | 55 | -------------------------------------------------------------------------------- /templates/main.html: -------------------------------------------------------------------------------- 1 | 2 | 3 |
4 |