├── dcnmt.png ├── visual ├── README.md ├── plot_curve.py └── embedding.py ├── preprocess ├── shuffle_data.py └── create_vocab.py ├── README.md ├── testing.py ├── configurations.py ├── train.sh ├── configurations_template.py ├── checkpoint.py ├── training.py ├── stream.py ├── sampling.py ├── model.py └── LICENSE /dcnmt.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SwordYork/DCNMT/HEAD/dcnmt.png -------------------------------------------------------------------------------- /visual/README.md: -------------------------------------------------------------------------------- 1 | `plot_curve.py` will plot the learning curve, you should copy `plot_curve.py` to `dcnmt_*` folder which contains the `cost_curve.npz` file. 2 | 3 | `embedding.py` will visualize the embedding of words by t-SNE, you should copy this file to parent folder, and create a `wordlist` file (one word per line). 4 | -------------------------------------------------------------------------------- /visual/plot_curve.py: -------------------------------------------------------------------------------- 1 | import numpy 2 | import matplotlib.pyplot as plt 3 | 4 | def smooth(x,window_len=11,window='hanning'): 5 | 6 | if x.ndim != 1: 7 | raise ValueError("smooth only accepts 1 dimension arrays.") 8 | 9 | if x.size < window_len: 10 | raise ValueError("Input vector needs to be bigger than window size.") 11 | 12 | if window_len<3: 13 | return x 14 | 15 | if not window in ['flat', 'hanning', 'hamming', 'bartlett', 'blackman']: 16 | raise ValueError("Window is on of 'flat', 'hanning', 'hamming', 'bartlett', 'blackman'") 17 | 18 | s=numpy.r_[x[window_len-1:0:-1],x,x[-1:-window_len:-1]] 19 | if window == 'flat': #moving average 20 | w=numpy.ones(window_len,'d') 21 | else: 22 | w=eval('numpy.'+window+'(window_len)') 23 | 24 | y=numpy.convolve(w/w.sum(),s,mode='valid') 25 | return y 26 | 27 | 28 | if __name__ == '__main__': 29 | a = numpy.load('cost_curve.npz') 30 | c_dict = a['cost_curves'].tolist() 31 | vals = numpy.array([list(v.values())[0] for v in c_dict]) 32 | 33 | stride = 500 34 | 35 | sv = smooth(vals, stride, 'hamming')[:-stride+1] 36 | plt.plot(vals, 'g', linewidth=0.4) 37 | plt.plot(sv, 'k', linewidth=4) 38 | plt.show() 39 | -------------------------------------------------------------------------------- /preprocess/shuffle_data.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/python 2 | 3 | import logging 4 | import os 5 | import subprocess 6 | import uuid 7 | import sys 8 | 9 | from picklable_itertools.extras import equizip 10 | 11 | 12 | def merge_parallel(src_filename, trg_filename, merged_filename): 13 | total = 0 14 | with open(src_filename, 'r', encoding='utf-8') as left: 15 | with open(trg_filename, 'r', encoding='utf-8') as right: 16 | with open(merged_filename, 'w') as final: 17 | for lline, rline in equizip(left, right): 18 | if (lline != '\n') and (rline != '\n'): 19 | total += 1 20 | final.write(lline[:-1] + ' ||| ' + rline) 21 | 22 | 23 | def split_parallel(merged_filename, src_filename, trg_filename): 24 | total = 0 25 | with open(merged_filename) as combined: 26 | with open(src_filename, 'w') as left: 27 | with open(trg_filename, 'w') as right: 28 | for line in combined: 29 | total += 1 30 | line = line.split('|||') 31 | left.write(line[0].strip() + '\n') 32 | right.write(line[1].strip() + '\n') 33 | 34 | 35 | def shuffle_parallel(src_filename, trg_filename): 36 | logger.info("Shuffling jointly [{}] and [{}]".format(src_filename, 37 | trg_filename)) 38 | out_src = src_filename + '.shuf' 39 | out_trg = trg_filename + '.shuf' 40 | merged_filename = str(uuid.uuid4()) 41 | shuffled_filename = str(uuid.uuid4()) 42 | if not os.path.exists(out_src) or not os.path.exists(out_trg): 43 | try: 44 | merge_parallel(src_filename, trg_filename, merged_filename) 45 | subprocess.check_call( 46 | " shuf {} > {} ".format(merged_filename, shuffled_filename), 47 | shell=True) 48 | split_parallel(shuffled_filename, out_src, out_trg) 49 | logger.info( 50 | "...files shuffled [{}] and [{}]".format(out_src, out_trg)) 51 | except Exception as e: 52 | logger.error("{}".format(str(e))) 53 | else: 54 | logger.info("...files exist [{}] and [{}]".format(out_src, out_trg)) 55 | if os.path.exists(merged_filename): 56 | os.remove(merged_filename) 57 | if os.path.exists(shuffled_filename): 58 | os.remove(shuffled_filename) 59 | 60 | 61 | if __name__ == "__main__": 62 | logging.basicConfig(level=logging.INFO) 63 | logger = logging.getLogger('shuffle_data') 64 | 65 | if len(sys.argv) != 3: 66 | sys.exit(-1) 67 | # Shuffle datasets 68 | src_file_name = sys.argv[1] 69 | trg_file_name = sys.argv[2] 70 | shuffle_parallel(src_file_name, trg_file_name) 71 | -------------------------------------------------------------------------------- /preprocess/create_vocab.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | 3 | import pickle 4 | import logging 5 | import os 6 | import sys 7 | 8 | from collections import Counter 9 | 10 | 11 | def safe_pickle(obj, filename): 12 | if os.path.isfile(filename): 13 | logger.info("Overwriting %s." % filename) 14 | else: 15 | logger.info("Saving to %s." % filename) 16 | with open(filename, 'wb') as f: 17 | pickle.dump(obj, f, protocol=pickle.HIGHEST_PROTOCOL) 18 | 19 | 20 | def create_dictionary(input_file, dictionary_file, vocab_size): 21 | input_filename = os.path.basename(input_file.name) 22 | logger.info("Counting words in %s" % input_filename) 23 | counter = Counter() 24 | sentence_count = 0 25 | for line in input_file: 26 | words = list(line.strip()) 27 | counter.update(words) 28 | sentence_count += 1 29 | logger.info("%d unique words in %d sentences with a total of %d words." 30 | % (len(counter), sentence_count, sum(counter.values()))) 31 | 32 | if vocab_size is not None: 33 | if vocab_size <= 3: 34 | logger.info('Building a dictionary with all unique words') 35 | vocab_size = len(counter) + 3 36 | vocab_count = counter.most_common(vocab_size - 3) 37 | logger.info("Creating dictionary of %s most common words, covering " 38 | "%2.1f%% of the text." 39 | % (vocab_size, 40 | 100.0 * sum([count for word, count in vocab_count]) / 41 | sum(counter.values()))) 42 | else: 43 | logger.info("Creating dictionary of all words") 44 | vocab_count = counter.most_common() 45 | 46 | vocab = {'UNK': 1, '': 0, '': vocab_size - 1} 47 | for i, (word, count) in enumerate(vocab_count): 48 | vocab[word] = i + 2 49 | 50 | print(counter, vocab_count) 51 | safe_pickle(vocab, dictionary_file) 52 | 53 | 54 | def create_vocabularies(src_file, trg_file, config): 55 | src_vocab_name = 'vocab.{}-{}.{}.pkl'.format( 56 | config['source'], config['target'], config['source']) 57 | trg_vocab_name = 'vocab.{}-{}.{}.pkl'.format( 58 | config['source'], config['target'], config['target']) 59 | 60 | logger.info("Creating source vocabulary [{}]".format(src_vocab_name)) 61 | if not os.path.exists(src_vocab_name): 62 | create_dictionary(open(src_file, 'r', encoding='utf-8'), src_vocab_name, config['src_vocab_size']) 63 | else: 64 | logger.info("...file exists [{}]".format(src_vocab_name)) 65 | 66 | logger.info("Creating target vocabulary [{}]".format(trg_vocab_name)) 67 | if not os.path.exists(trg_vocab_name): 68 | create_dictionary(open(trg_file, 'r', encoding='utf-8'), trg_vocab_name, config['trg_vocab_size']) 69 | else: 70 | logger.info("...file exists [{}]".format(trg_vocab_name)) 71 | 72 | 73 | if __name__ == "__main__": 74 | logging.basicConfig(level=logging.INFO) 75 | logger = logging.getLogger('create_vocab') 76 | if len(sys.argv) != 7: 77 | sys.exit(-1) 78 | 79 | configs = {'source': sys.argv[1], 'target': sys.argv[2], 'src_vocab_size': int(sys.argv[3]), 'trg_vocab_size': int(sys.argv[4])} 80 | src_file_name = sys.argv[5] 81 | trg_file_name = sys.argv[6] 82 | create_vocabularies(src_file_name, trg_file_name, configs) 83 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | Deep Character-Level Neural Machine Translation 2 | ============ 3 | We implement a [**Deep Character-Level Neural Machine Translation**](https://arxiv.org/abs/1608.04738) based on [Theano](https://github.com/Theano/Theano) and [Blocks](https://github.com/mila-udem/blocks). Please intall relative packages according to [Blocks](http://blocks.readthedocs.io/en/latest/setup.html) before testing our program. Note that, please use Python 3 instead of Python 2. There will be some problems with Python 2. 4 | 5 | The architecture of DCNMT is shown in the following figure which is a single, large neural network. 6 | ![DCNMT](/dcnmt.png?raw=true "The architecture of DCNMT") 7 | 8 | 9 | 10 | 11 | Training 12 | ----------------------- 13 | If you want to train your own model, please prepare a parallel linguistics corpus, like corpus in [WMT](http://www.statmt.org/wmt15/translation-task.html). A GPU with 12GB memory will be helpful. You could run `bash train.sh` or follow these steps. 14 | 1. Download the relative scripts (tokenizer.perl, multi-bleu.perl) and nonbreaking_prefix from [mose_git](https://raw.githubusercontent.com/moses-smt/mosesdecoder/master/scripts). 15 | 2. Download the datasets, then tokenize and shuffle the cropus. 16 | 3. Create the character list for both language using `create_vocab.py` in `preprocess` folder. Don't forget to pass the language setting, vocabulary size and file name to this script. 17 | 4. Create a `data` folder, and put the `vocab.*.*.pkl` and `*.shuf` in the `data` folder. 18 | 5. Prepare the tokenized validation and test set, and put them in `data` folder. 19 | 6. Edit the `configurations.py`, and run `python training.py`. It will take 1 to 2 weeks to train a good model. 20 | 21 | 22 | Testing 23 | ----------------------- 24 | We have trained several models which listed in the following table. However, because of the limitation of available GPU and long training time (two weeks or more), we don't have enough time and resource to train on more language pairs. Would you like to help us to train on more language pairs? If you run into any trouble, please open an issue or email me directly at `echo c3dvcmQueW9ya0BnbWFpbC5jb20K | base64 -d`. Thanks! 25 | 26 | 27 | | language pair | dataset | encoder_layers | transition_layers | BLEU | 28 | |:--------:|:--------:|:--------:|:--------:|:--------:| 29 | | en-fr | same as [RNNsearch](https://arxiv.org/abs/1409.0473) | 1 | 1 | 30.46 | 30 | | en-fr | same as [RNNsearch](https://arxiv.org/abs/1409.0473) | 2 | 1 | 31.98 | 31 | | en-fr | same as [RNNsearch](https://arxiv.org/abs/1409.0473) | 2 | 2 | 32.12 | 32 | | en-cs | [wmt15](http://www.statmt.org/wmt15/translation-task.html) | 1 | 1 | 16.43 | 33 | 34 | 35 | These models are all trained for about 5 epochs, and evaluate on `newstest2014` using the best validation model on `newstest2013`. You can download these models from [dropbox](https://www.dropbox.com/sh/eiaexn8q2sf277s/AADQ4RKWEsCIGkeKUUyMHh2aa?dl=0), then put them (dcnmt_\*, data, configurations.py) in this directory. To perform testing, just run `python testing.py`. It takes about an hour to do translation on 3000 sentences if you have a moderate GPU. 36 | 37 | 38 | Embedding 39 | ----------------------- 40 | Please prepare a wordlist to calculate embedding, then just run `python embedding.py` to view the results. 41 | 42 | Spelling Correction 43 | ----------------------- 44 | It is the special feature of DCNMT model. For example, 45 | 46 | > *Source:* Unlike in Canada, the American States are **responisble** for the **orgainisation** of federal elections in the United States. 47 | 48 | > *Ref:* Contrairement au Canada, les États américains sont **responsables** de **l’organisation** des élections fédérales aux États-Unis. 49 | 50 | > *Google:* Contrairement au Canada, les États américains sont **responisble** pour le **orgainisation** des élections fédérales aux États-Unis. 51 | 52 | > *DCNMT:* Contrairement au Canada, les États américains sont **responsables** de **l’organisation** des élections fédérales aux États-Unis. 53 | 54 | The performance of misspelling correction would be analyzed later. 55 | 56 | 57 | 58 | This program have been tested under the latest Theano and Blocks, it may fail to run because of different version. If you failed to run these scripts, please make sure that you can run the examples of [Blocks](https://github.com/mila-udem/blocks-examples). 59 | 60 | 61 | References: 62 | ---------------------- 63 | 1. [An Efficient Character-Level Neural Machine Translation](https://arxiv.org/abs/1608.04738) 64 | 2. [Neural Machine Translation by Jointly Learning to Align and Translate](https://arxiv.org/abs/1409.0473) 65 | -------------------------------------------------------------------------------- /testing.py: -------------------------------------------------------------------------------- 1 | from theano import tensor 2 | 3 | from blocks.filter import VariableFilter 4 | from blocks.graph import ComputationGraph 5 | from blocks.main_loop import MainLoop 6 | from blocks.model import Model 7 | 8 | from checkpoint import LoadNMT 9 | from model import BidirectionalEncoder, Decoder 10 | from sampling import BleuTester 11 | import logging 12 | import pprint 13 | 14 | import configurations 15 | from stream import get_test_stream 16 | 17 | logger = logging.getLogger(__name__) 18 | 19 | 20 | def main(config, test_stream): 21 | # Create Theano variables 22 | logger.info('Creating theano variables') 23 | source_char_seq = tensor.lmatrix('source_char_seq') 24 | source_sample_matrix = tensor.tensor3('source_sample_matrix') 25 | source_char_aux = tensor.matrix('source_char_aux') 26 | source_word_mask = tensor.matrix('source_word_mask') 27 | target_char_seq = tensor.lmatrix('target_char_seq') 28 | target_char_aux = tensor.matrix('target_char_aux') 29 | target_char_mask = tensor.matrix('target_char_mask') 30 | target_sample_matrix = tensor.tensor3('target_sample_matrix') 31 | target_word_mask = tensor.matrix('target_word_mask') 32 | target_resample_matrix = tensor.tensor3('target_resample_matrix') 33 | target_prev_char_seq = tensor.lmatrix('target_prev_char_seq') 34 | target_prev_char_aux = tensor.matrix('target_prev_char_aux') 35 | 36 | target_bos_idx = test_stream.trg_bos 37 | target_space_idx = test_stream.space_idx['target'] 38 | 39 | # Construct model 40 | logger.info('Building RNN encoder-decoder') 41 | 42 | encoder = BidirectionalEncoder(config['src_vocab_size'], config['enc_embed'], config['src_dgru_nhids'], 43 | config['enc_nhids'], config['src_dgru_depth'], config['bidir_encoder_depth']) 44 | 45 | decoder = Decoder(config['trg_vocab_size'], config['dec_embed'], config['trg_dgru_nhids'], config['trg_igru_nhids'], 46 | config['dec_nhids'], config['enc_nhids'] * 2, config['transition_depth'], config['trg_igru_depth'], 47 | config['trg_dgru_depth'], target_space_idx, target_bos_idx) 48 | 49 | 50 | representation = encoder.apply(source_char_seq, source_sample_matrix, source_char_aux, 51 | source_word_mask) 52 | cost = decoder.cost(representation, source_word_mask, target_char_seq, target_sample_matrix, 53 | target_resample_matrix, target_char_aux, target_char_mask, 54 | target_word_mask, target_prev_char_seq, target_prev_char_aux) 55 | 56 | # Set up training model 57 | logger.info("Building model") 58 | training_model = Model(cost) 59 | 60 | # Set extensions 61 | logger.info("Initializing extensions") 62 | # Extensions 63 | extensions = [] 64 | # Reload model if necessary 65 | if config['reload']: 66 | extensions.append(LoadNMT(config['saveto'])) 67 | 68 | # Set up beam search and sampling computation graphs if necessary 69 | if config['bleu_script'] is not None: 70 | logger.info("Building sampling model") 71 | generated = decoder.generate(representation, source_word_mask) 72 | search_model = Model(generated) 73 | _, samples = VariableFilter( 74 | bricks=[decoder.sequence_generator], name="outputs")( 75 | ComputationGraph(generated[config['transition_depth']])) # generated[config['transition_depth']] is next_outputs 76 | 77 | logger.info("Building bleu tester") 78 | extensions.append( 79 | BleuTester(source_char_seq, source_sample_matrix, source_char_aux, 80 | source_word_mask, samples=samples, config=config, 81 | model=search_model, data_stream=test_stream, 82 | normalize=config['normalized_bleu'])) 83 | 84 | # Initialize main loop 85 | logger.info("Initializing main loop") 86 | main_loop = MainLoop( 87 | model=training_model, 88 | algorithm=None, 89 | data_stream=None, 90 | extensions=extensions 91 | ) 92 | 93 | for extension in main_loop.extensions: 94 | extension.main_loop = main_loop 95 | main_loop._run_extensions('before_training') 96 | 97 | 98 | logger = logging.getLogger(__name__) 99 | logging.basicConfig(level=logging.INFO) 100 | 101 | if __name__ == '__main__': 102 | # Get configurations for model 103 | configuration = configurations.get_config() 104 | logger.info("Model options:\n{}".format(pprint.pformat(configuration))) 105 | # Get data streams and call main 106 | main(configuration, get_test_stream(**configuration)) 107 | 108 | 109 | -------------------------------------------------------------------------------- /configurations.py: -------------------------------------------------------------------------------- 1 | def get_config(): 2 | config = {} 3 | # Where to save model, this corresponds to 'prefix' in groundhog 4 | config['saveto'] = 'dcnmt_en2fr' 5 | 6 | # prepare data 7 | config['source_language'] = 'en' 8 | config['target_language'] = 'fr' 9 | 10 | # Model related ----------------------------------------------------------- 11 | 12 | # Sequences longer than this will be discarded 13 | config['max_src_seq_char_len'] = 300 14 | config['max_src_seq_word_len'] = 50 15 | config['max_trg_seq_char_len'] = 300 16 | config['max_trg_seq_word_len'] = 50 17 | 18 | # Number of hidden units in encoder/decoder GRU 19 | config['src_dgru_nhids'] = 512 20 | config['enc_nhids'] = 1024 21 | config['dec_nhids'] = 1024 22 | config['trg_dgru_nhids'] = 512 23 | config['trg_igru_nhids'] = 1024 24 | 25 | 26 | # Dimension of the word embedding matrix in encoder/decoder 27 | config['enc_embed'] = 64 28 | config['dec_embed'] = 64 29 | config['src_dgru_depth'] = 2 30 | config['bidir_encoder_depth'] = 2 31 | config['transition_depth'] = 1 32 | config['trg_dgru_depth'] = 1 33 | config['trg_igru_depth'] = 1 34 | 35 | 36 | # Optimization related ---------------------------------------------------- 37 | 38 | # Batch size 39 | config['batch_size'] = 80 40 | 41 | # This many batches will be read ahead and sorted 42 | config['sort_k_batches'] = 12 43 | 44 | # Optimization step rule 45 | config['step_rule'] = 'AdaDelta' 46 | 47 | # Gradient clipping threshold 48 | config['step_clipping'] = 1. 49 | 50 | # Std of weight initialization 51 | config['weight_scale'] = 0.01 52 | 53 | # Vocabulary/dataset related ---------------------------------------------- 54 | 55 | # Root directory for dataset 56 | datadir = './data/' 57 | 58 | # Module name of the stream that will be used 59 | config['stream'] = 'stream' 60 | 61 | # Source and target vocabularies 62 | config['src_vocab'] = datadir + 'vocab.{}-{}.{}.pkl'.format(config['source_language'], config['target_language'], 63 | config['source_language']) 64 | config['trg_vocab'] = datadir + 'vocab.{}-{}.{}.pkl'.format(config['source_language'], config['target_language'], 65 | config['target_language']) 66 | 67 | # Source and target datasets 68 | config['src_data'] = datadir + 'all.en-fr.en.tok.shuf' 69 | config['trg_data'] = datadir + 'all.en-fr.fr.tok.shuf' 70 | 71 | # Source and target vocabulary sizes, should include bos, eos, unk tokens 72 | config['src_vocab_size'] = 120 73 | config['trg_vocab_size'] = 120 74 | 75 | # Special tokens and indexes 76 | config['unk_id'] = 1 77 | config['bos_token'] = '' 78 | config['eos_token'] = '' 79 | config['unk_token'] = '' 80 | 81 | # Early stopping based on val related ------------------------------------ 82 | 83 | # Normalize cost according to sequence length after beam-search 84 | config['normalized_val'] = True 85 | 86 | # Normalize cost according to sequence length after beam-search 87 | config['normalized_bleu'] = True 88 | 89 | # Bleu script that will be used (moses multi-perl in this case) 90 | config['bleu_script'] = datadir + 'multi-bleu.perl' 91 | 92 | # Validation set source file 93 | config['val_set'] = datadir + 'newstest2013.en.tok' 94 | 95 | # Validation set gold file 96 | config['val_set_grndtruth'] = datadir + 'newstest2013.fr.tok' 97 | 98 | # Test set source file 99 | config['test_set'] = datadir + 'newstest2014.en.tok' 100 | 101 | # Test set gold file 102 | config['test_set_grndtruth'] = datadir + 'newstest2014.fr.tok' 103 | 104 | config['validate'] = True 105 | 106 | # Print validation output to file 107 | config['output_val_set'] = True 108 | 109 | # Validation output file 110 | config['val_set_out'] = config['saveto'] + '/validation_out.txt' 111 | 112 | # Validation output file 113 | config['test_set_out'] = config['saveto'] + '/test_out.txt' 114 | 115 | # Beam-size 116 | config['beam_size'] = 12 117 | 118 | # Timing/monitoring related ----------------------------------------------- 119 | 120 | # Maximum number of updates 121 | config['finish_after'] = 432000 122 | 123 | # Reload model from files if exist 124 | config['reload'] = True 125 | 126 | # Save model after this many updates 127 | config['save_freq'] = 500 128 | 129 | # Print training status after this many updates 130 | config['print_freq'] = 10 131 | 132 | # Show samples from model after this many updates 133 | config['sampling_freq'] = 30 134 | 135 | # Show this many samples at each sampling 136 | config['hook_samples'] = 2 137 | 138 | config['bleu_val_freq'] = 18000 139 | # Start validation after this many updates 140 | config['val_burn_in'] = 70000 141 | 142 | return config 143 | -------------------------------------------------------------------------------- /train.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env sh 2 | 3 | # The repository of moses, we need to download some scripts from moses. 4 | # https://github.com/moses-smt/mosesdecoder/tree/master/scripts 5 | mose_git=https://raw.githubusercontent.com/moses-smt/mosesdecoder/master/scripts 6 | 7 | echo 'select the source language: en, cs, fi, fr, ru, de' 8 | read -p '==> ' source_language 9 | echo "the selected source language is $source_language" 10 | 11 | 12 | echo 'select the target language: en, cs, fi, fr, ru, de' 13 | read -p '==> ' target_language 14 | echo "the selected target language is $target_language" 15 | 16 | if [ $target_language = $source_language ]; then 17 | echo 'languages should be different' 18 | exit -1 19 | fi 20 | 21 | if [ ! -d share/nonbreaking_prefixes ]; then 22 | mkdir -p share/nonbreaking_prefixes 23 | fi 24 | 25 | echo "downloading nonbreaking_prefix $source_language ..." 26 | curl -s $mose_git/share/nonbreaking_prefixes/nonbreaking_prefix.$source_language > share/nonbreaking_prefixes/nonbreaking_prefix.$source_language 27 | echo "downloading nonbreaking_prefix $target_language ..." 28 | curl -s $mose_git/share/nonbreaking_prefixes/nonbreaking_prefix.$target_language > share/nonbreaking_prefixes/nonbreaking_prefix.$target_language 29 | 30 | if [ ! -d data ]; then 31 | echo 'creating data directory...' 32 | mkdir data 33 | fi 34 | 35 | 36 | 37 | cp preprocess/create_vocab.py preprocess/shuffle_data.py data/ 38 | 39 | echo 'cd to data directory' 40 | cd data 41 | if [ ! -f tokenizer.perl ]; then 42 | echo 'downloading tokenizer.perl' 43 | curl -s $mose_git/tokenizer/tokenizer.perl > tokenizer.perl 44 | fi 45 | if [ ! -f multi-bleu.perl ]; then 46 | echo 'downloading multi-bleu.perl' 47 | curl -s $mose_git/generic/multi-bleu.perl > multi-bleu.perl 48 | fi 49 | 50 | echo 'please download corresponding datasets from WMT15 manually' 51 | echo 'and put the parallel corpus in the data directory' 52 | echo 'then enter the file name of source language dataset' 53 | read -p '==> ' source_data 54 | echo 'the file name of target language dataset' 55 | read -p '==> ' target_data 56 | 57 | if [ ! -f $source_data ]; then 58 | echo 'no such source dataset file' 59 | exit -1 60 | fi 61 | 62 | if [ ! -f $target_data ]; then 63 | echo 'no such target dataset file' 64 | exit -1 65 | fi 66 | 67 | tok_source_file=all.$source_language-$target_language.$source_language.tok 68 | tok_target_file=all.$source_language-$target_language.$target_language.tok 69 | 70 | perl tokenizer.perl -l $source_language -threads 4 -no-escape < $source_data > $tok_source_file 71 | perl tokenizer.perl -l $target_language -threads 4 -no-escape < $target_data > $tok_target_file 72 | 73 | echo 'please put the validation and test dataset in the data directory' 74 | echo 'the file name of source validation set:' 75 | read -p '==>' src_val 76 | echo 'the file name of target validation set:' 77 | read -p '==>' trg_val 78 | echo 'the file name of source test set:' 79 | read -p '==>' src_test 80 | echo 'the file name of targe test set:' 81 | read -p '==>' trg_test 82 | 83 | if [ ! -f $src_val ]; then 84 | echo 'no such source validation file' 85 | exit -1 86 | fi 87 | 88 | if [ ! -f $trg_val ]; then 89 | echo 'no such target validation file' 90 | exit -1 91 | fi 92 | 93 | if [ ! -f $src_test ]; then 94 | echo 'no such source test file' 95 | exit -1 96 | fi 97 | 98 | if [ ! -f $trg_test ]; then 99 | echo 'no such target test file' 100 | exit -1 101 | fi 102 | 103 | perl tokenizer.perl -l $source_language -threads 4 -no-escape < $src_val > $src_val.tok 104 | perl tokenizer.perl -l $source_language -threads 4 -no-escape < $src_test > $src_test.tok 105 | perl tokenizer.perl -l $target_language -threads 4 -no-escape < $trg_val > $trg_val.tok 106 | perl tokenizer.perl -l $target_language -threads 4 -no-escape < $trg_test > $trg_test.tok 107 | 108 | 109 | echo 'please ensure there is enough disk space' 110 | python shuffle_data.py $tok_source_file $tok_target_file 111 | 112 | echo 'please enter the size of source language vocabulary (120 is enough):' 113 | read -p '==>' src_vocab_size 114 | echo 'please enter the size of target language vocabulary (120 is enough):' 115 | read -p '==>' trg_vocab_size 116 | python create_vocab.py $source_language $target_language $src_vocab_size $trg_vocab_size $tok_source_file.shuf $tok_target_file.shuf 117 | 118 | cd .. 119 | if [ -f configurations.py ]; then 120 | cp configurations.py configurations_backup.py 121 | fi 122 | 123 | cp configurations_template.py configurations.py 124 | 125 | sed -i "s/--src_lang--/$source_language/" configurations.py 126 | sed -i "s/--trg_lang--/$target_language/" configurations.py 127 | sed -i "s/--src_vocab_size--/$src_vocab_size/" configurations.py 128 | sed -i "s/--trg_vocab_size--/$trg_vocab_size/" configurations.py 129 | sed -i "s/--src_val--/$src_val/" configurations.py 130 | sed -i "s/--trg_val--/$trg_val/" configurations.py 131 | sed -i "s/--src_test--/$src_test/" configurations.py 132 | sed -i "s/--trg_test--/$trg_test/" configurations.py 133 | 134 | echo "ok! just run 'python training.py'" 135 | 136 | 137 | 138 | 139 | -------------------------------------------------------------------------------- /configurations_template.py: -------------------------------------------------------------------------------- 1 | def get_config(): 2 | config = {} 3 | 4 | # prepare data 5 | config['source_language'] = '--src_lang--' 6 | config['target_language'] = '--trg_lang--' 7 | 8 | 9 | # Where to save model, this corresponds to 'prefix' in groundhog 10 | config['saveto'] = 'dcnmt_{}2{}'.format(config['source_language'], config['target_language']) 11 | 12 | # Model related ----------------------------------------------------------- 13 | 14 | # Sequences longer than this will be discarded 15 | config['max_src_seq_char_len'] = 300 16 | config['max_src_seq_word_len'] = 50 17 | config['max_trg_seq_char_len'] = 300 18 | config['max_trg_seq_word_len'] = 50 19 | 20 | # Number of hidden units in encoder/decoder GRU 21 | config['src_dgru_nhids'] = 512 22 | config['enc_nhids'] = 1024 23 | config['dec_nhids'] = 1024 24 | config['trg_dgru_nhids'] = 512 25 | config['trg_igru_nhids'] = 1024 26 | 27 | 28 | # Dimension of the word embedding matrix in encoder/decoder 29 | config['enc_embed'] = 64 30 | config['dec_embed'] = 64 31 | config['src_dgru_depth'] = 2 32 | config['bidir_encoder_depth'] = 2 33 | config['transition_depth'] = 1 34 | config['trg_dgru_depth'] = 1 35 | config['trg_igru_depth'] = 1 36 | 37 | # Optimization related ---------------------------------------------------- 38 | 39 | # Batch size 40 | config['batch_size'] = 80 41 | 42 | # This many batches will be read ahead and sorted 43 | config['sort_k_batches'] = 12 44 | 45 | # Optimization step rule 46 | config['step_rule'] = 'AdaDelta' 47 | 48 | # Gradient clipping threshold 49 | config['step_clipping'] = 1. 50 | 51 | # Std of weight initialization 52 | config['weight_scale'] = 0.01 53 | 54 | 55 | # Vocabulary/dataset related ---------------------------------------------- 56 | 57 | # Root directory for dataset 58 | datadir = './data/' 59 | 60 | # Module name of the stream that will be used 61 | config['stream'] = 'stream' 62 | 63 | # Source and target vocabularies 64 | config['src_vocab'] = datadir + 'vocab.{}-{}.{}.pkl'.format(config['source_language'], config['target_language'], 65 | config['source_language']) 66 | config['trg_vocab'] = datadir + 'vocab.{}-{}.{}.pkl'.format(config['source_language'], config['target_language'], 67 | config['target_language']) 68 | 69 | # Source and target datasets 70 | config['src_data'] = datadir + 'all.{}-{}.{}.tok.shuf'.format(config['source_language'], config['target_language'], 71 | config['source_language']) 72 | config['trg_data'] = datadir + 'all.{}-{}.{}.tok.shuf'.format(config['source_language'], config['target_language'], 73 | config['target_language']) 74 | 75 | # Source and target vocabulary sizes, should include bos, eos, unk tokens 76 | config['src_vocab_size'] = --src_vocab_size-- 77 | config['trg_vocab_size'] = --trg_vocab_size-- 78 | 79 | # Special tokens and indexes 80 | config['unk_id'] = 1 81 | config['bos_token'] = '' 82 | config['eos_token'] = '' 83 | config['unk_token'] = '' 84 | 85 | # Early stopping based on val related ------------------------------------ 86 | 87 | # Normalize cost according to sequence length after beam-search 88 | config['normalized_val'] = True 89 | 90 | # Normalize cost according to sequence length after beam-search 91 | config['normalized_bleu'] = True 92 | 93 | # Bleu script that will be used (moses multi-perl in this case) 94 | config['bleu_script'] = datadir + 'multi-bleu.perl' 95 | 96 | # Validation set source file 97 | config['val_set'] = datadir + '--src_val--.tok' 98 | 99 | # Validation set gold file 100 | config['val_set_grndtruth'] = datadir + '--trg_val--.tok' 101 | 102 | # Test set source file 103 | config['test_set'] = datadir + '--src_test--.tok' 104 | 105 | # Test set gold file 106 | config['test_set_grndtruth'] = datadir + '--trg_test--.tok' 107 | 108 | config['validate'] = True 109 | 110 | # Print validation output to file 111 | config['output_val_set'] = True 112 | 113 | # Validation output file 114 | config['val_set_out'] = config['saveto'] + '/validation_out.txt' 115 | 116 | # Validation output file 117 | config['test_set_out'] = config['saveto'] + '/test_out.txt' 118 | 119 | # Beam-size 120 | config['beam_size'] = 12 121 | 122 | # Timing/monitoring related ----------------------------------------------- 123 | 124 | # Maximum number of updates 125 | config['finish_after'] = 800000 126 | 127 | # Reload model from files if exist 128 | config['reload'] = True 129 | 130 | # Save model after this many updates 131 | config['save_freq'] = 500 132 | 133 | # Print training status after this many updates 134 | config['print_freq'] = 10 135 | 136 | # Show samples from model after this many updates 137 | config['sampling_freq'] = 30 138 | 139 | # Show this many samples at each sampling 140 | config['hook_samples'] = 2 141 | 142 | config['bleu_val_freq'] = 18000 143 | # Start validation after this many updates 144 | config['val_burn_in'] = 70000 145 | 146 | return config 147 | -------------------------------------------------------------------------------- /checkpoint.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import numpy 3 | import os 4 | import time 5 | 6 | from contextlib import closing 7 | from six.moves import cPickle 8 | 9 | from blocks.extensions.saveload import SAVED_TO, LOADED_FROM 10 | from blocks.extensions import TrainingExtension, SimpleExtension 11 | from blocks.serialization import secure_dump, load, BRICK_DELIMITER 12 | from blocks.utils import reraise_as 13 | 14 | logger = logging.getLogger(__name__) 15 | 16 | 17 | class SaveLoadUtils(object): 18 | """Utility class for checkpointing.""" 19 | 20 | @property 21 | def path_to_folder(self): 22 | return self.folder 23 | 24 | @property 25 | def path_to_parameters(self): 26 | return os.path.join(self.folder, 'params.npz') 27 | 28 | @property 29 | def path_to_iteration_state(self): 30 | return os.path.join(self.folder, 'iterations_state.pkl') 31 | 32 | @property 33 | def path_to_log(self): 34 | return os.path.join(self.folder, 'log') 35 | 36 | def load_parameter_values(self, path): 37 | with closing(numpy.load(path)) as source: 38 | param_values = {} 39 | for name, value in source.items(): 40 | if name != 'pkl': 41 | name_ = name.replace(BRICK_DELIMITER, '/') 42 | if not name_.startswith('/'): 43 | name_ = '/' + name_ 44 | param_values[name_] = value 45 | return param_values 46 | 47 | def save_parameter_values(self, param_values, path): 48 | param_values = {name.replace("/", BRICK_DELIMITER): param 49 | for name, param in param_values.items()} 50 | numpy.savez(path, **param_values) 51 | 52 | 53 | class CheckpointNMT(SimpleExtension, SaveLoadUtils): 54 | """Redefines checkpointing for NMT. 55 | 56 | Saves only parameters (npz), iteration state (pickle) and log (pickle). 57 | 58 | """ 59 | 60 | def __init__(self, saveto, **kwargs): 61 | self.folder = saveto 62 | kwargs.setdefault("after_training", True) 63 | super(CheckpointNMT, self).__init__(**kwargs) 64 | 65 | def dump_parameters(self, main_loop): 66 | params_to_save = main_loop.model.get_parameter_values() 67 | self.save_parameter_values(params_to_save, 68 | self.path_to_parameters) 69 | 70 | def dump_iteration_state(self, main_loop): 71 | secure_dump(main_loop.iteration_state, self.path_to_iteration_state) 72 | 73 | def dump_log(self, main_loop): 74 | secure_dump(main_loop.log, self.path_to_log, cPickle.dump) 75 | 76 | def dump(self, main_loop): 77 | if not os.path.exists(self.path_to_folder): 78 | os.mkdir(self.path_to_folder) 79 | print("") 80 | logger.info(" Saving model") 81 | start = time.time() 82 | logger.info(" ...saving parameters") 83 | self.dump_parameters(main_loop) 84 | logger.info(" ...saving iteration state") 85 | self.dump_iteration_state(main_loop) 86 | logger.info(" ...saving log") 87 | self.dump_log(main_loop) 88 | logger.info(" Model saved, took {} seconds.".format(time.time()-start)) 89 | 90 | def do(self, callback_name, *args): 91 | try: 92 | self.dump(self.main_loop) 93 | except Exception: 94 | raise 95 | finally: 96 | already_saved_to = self.main_loop.log.current_row.get(SAVED_TO, ()) 97 | self.main_loop.log.current_row[SAVED_TO] = (already_saved_to + 98 | (self.path_to_folder + 99 | 'params.npz',)) 100 | 101 | 102 | class LoadNMT(TrainingExtension, SaveLoadUtils): 103 | """Loads parameters log and iterations state.""" 104 | 105 | def __init__(self, saveto, **kwargs): 106 | self.folder = saveto 107 | super(LoadNMT, self).__init__(saveto, **kwargs) 108 | 109 | def before_training(self): 110 | if not os.path.exists(self.path_to_folder): 111 | logger.info("No dump found") 112 | return 113 | logger.info("Loading the state from {} into the main loop" 114 | .format(self.path_to_folder)) 115 | try: 116 | self.load_to(self.main_loop) 117 | self.main_loop.log.current_row[LOADED_FROM] = self.path_to_folder 118 | except Exception: 119 | reraise_as("Failed to load the state") 120 | 121 | def load_parameters(self): 122 | return self.load_parameter_values(self.path_to_parameters) 123 | 124 | def load_iteration_state(self): 125 | with open(self.path_to_iteration_state, "rb") as source: 126 | return load(source, use_cpickle=True) 127 | 128 | def load_log(self): 129 | with open(self.path_to_log, "rb") as source: 130 | return cPickle.load(source) 131 | 132 | def load_to(self, main_loop): 133 | """Loads the dump from the root folder into the main loop.""" 134 | logger.info(" Reloading model") 135 | try: 136 | logger.info(" ...loading model parameters") 137 | params_all = self.load_parameters() 138 | params_this = main_loop.model.get_parameter_dict() 139 | missing = set(params_this.keys()) - set(params_all.keys()) 140 | for pname in params_this.keys(): 141 | if pname in params_all: 142 | val = params_all[pname] 143 | if params_this[pname].get_value().shape != val.shape: 144 | logger.warning( 145 | " Dimension mismatch {}-{} for {}" 146 | .format(params_this[pname].get_value().shape, 147 | val.shape, pname)) 148 | 149 | params_this[pname].set_value(val) 150 | logger.info(" Loaded to CG {:15}: {}" 151 | .format(str(val.shape), pname)) 152 | else: 153 | logger.warning( 154 | " Parameter does not exist: {}".format(pname)) 155 | logger.info( 156 | " Number of parameters loaded for computation graph: {}" 157 | .format(len(params_this) - len(missing))) 158 | except Exception as e: 159 | logger.error(" Error {0}".format(str(e))) 160 | 161 | try: 162 | logger.info(" Loading iteration state...") 163 | main_loop.iteration_state = self.load_iteration_state() 164 | except Exception as e: 165 | logger.error(" Error {0}".format(str(e))) 166 | 167 | try: 168 | logger.info(" Loading log...") 169 | main_loop.log = self.load_log() 170 | except Exception as e: 171 | logger.error(" Error {0}".format(str(e))) 172 | -------------------------------------------------------------------------------- /visual/embedding.py: -------------------------------------------------------------------------------- 1 | import pickle 2 | from theano import tensor 3 | from blocks.main_loop import MainLoop 4 | from blocks.model import Model 5 | 6 | from checkpoint import LoadNMT 7 | from model import BidirectionalEncoder, Decoder 8 | from stream import get_tr_stream 9 | import numpy 10 | 11 | import argparse 12 | import logging 13 | import pprint 14 | 15 | import configurations 16 | import matplotlib.pyplot as plt 17 | from sklearn import manifold 18 | 19 | logger = logging.getLogger(__name__) 20 | logging.basicConfig(level=logging.INFO) 21 | 22 | 23 | def build_input_dict(input_, src_vocab): 24 | input_length = len(input_) 25 | input_ = numpy.array([src_vocab[i] for i in input_]) 26 | 27 | total_word = list(input_).count(src_vocab[' ']) 28 | 29 | source_sample_matrix = numpy.zeros((total_word, input_length), dtype='int8') 30 | source_sample_matrix[range(total_word), numpy.nonzero(input_ == src_vocab[' '])[0] - 1] = 1 31 | 32 | source_char_aux = numpy.ones(input_length, dtype='int8') 33 | source_char_aux[input_ == src_vocab[' ']] = 0 34 | 35 | input_dict = {'source_sample_matrix': source_sample_matrix[None, :], 36 | 'source_char_aux': source_char_aux[None, :], 37 | 'source_char_seq': input_[None, :]} 38 | return input_length, input_dict 39 | 40 | 41 | # Scale and visualize the embedding vectors 42 | def plot_embedding(X, Y, title=None): 43 | x_min, x_max = numpy.min(X, 0), numpy.max(X, 0) 44 | X = (X - x_min) / (x_max - x_min) 45 | 46 | plt.figure() 47 | for i in range(X.shape[0]): 48 | if str(Y[i]) in ['Sunday ', 'March ', 'June ', 'January ', 'February ', 49 | 'April ', 'May ', 'July ', 'August ', 'September ', 50 | 'October ', 'November ', 'December ']: 51 | plt.text(X[i, 0], X[i, 1], str(Y[i]), 52 | fontdict={'weight': 'bold', 'size': 18, 'color': 'blue'}) 53 | elif str(Y[i]) != 'exercise ' and str(Y[i]) != 'exrecise ': 54 | plt.text(X[i, 0], X[i, 1], str(Y[i]), 55 | fontdict={'size': 18}) 56 | else: 57 | plt.text(X[i, 0], X[i, 1], str(Y[i]), 58 | fontdict={'weight': 'bold', 'size': 18, 'color': 'red'}) 59 | 60 | plt.xticks([]), plt.yticks([]) 61 | if title is not None: 62 | plt.title(title) 63 | 64 | 65 | def embedding(embedding_model, src_vocab): 66 | sampling_fn = embedding_model.get_theano_function() 67 | try: 68 | f = open('wordlist') 69 | except FileNotFoundError: 70 | print('Please create a file named wordlist, and one word per line in this file') 71 | exit(0) 72 | s = f.read() 73 | core_list = s.strip().split('\n') 74 | f.close() 75 | 76 | X = [] 77 | Y = [] 78 | for word in core_list: 79 | word += ' ' 80 | _, input_dict = build_input_dict(word, src_vocab) 81 | w_v = sampling_fn(**input_dict)[0][0][0] 82 | X.append(w_v) 83 | Y.append(word) 84 | X = numpy.array(X) 85 | print(X.shape) 86 | 87 | # t-SNE embedding of the digits dataset 88 | print("Computing t-SNE embedding") 89 | tsne = manifold.TSNE(n_components=2, init='pca', random_state=0) 90 | X_tsne = tsne.fit_transform(X) 91 | 92 | plot_embedding(X_tsne, Y, "t-SNE embedding of the words") 93 | plt.show() 94 | 95 | 96 | def main(config, tr_stream): 97 | # Create Theano variables 98 | logger.info('Creating theano variables') 99 | source_char_seq = tensor.lmatrix('source_char_seq') 100 | source_sample_matrix = tensor.btensor3('source_sample_matrix') 101 | source_char_aux = tensor.bmatrix('source_char_aux') 102 | source_word_mask = tensor.bmatrix('source_word_mask') 103 | target_char_seq = tensor.lmatrix('target_char_seq') 104 | target_char_aux = tensor.bmatrix('target_char_aux') 105 | target_char_mask = tensor.bmatrix('target_char_mask') 106 | target_sample_matrix = tensor.btensor3('target_sample_matrix') 107 | target_word_mask = tensor.bmatrix('target_word_mask') 108 | target_resample_matrix = tensor.btensor3('target_resample_matrix') 109 | target_prev_char_seq = tensor.lmatrix('target_prev_char_seq') 110 | target_prev_char_aux = tensor.bmatrix('target_prev_char_aux') 111 | target_bos_idx = tr_stream.trg_bos 112 | target_space_idx = tr_stream.space_idx['target'] 113 | src_vocab = pickle.load(open(config['src_vocab'], 'rb')) 114 | 115 | logger.info('Building RNN encoder-decoder') 116 | encoder = BidirectionalEncoder(config['src_vocab_size'], config['enc_embed'], config['src_dgru_nhids'], 117 | config['enc_nhids'], config['src_dgru_depth'], config['bidir_encoder_depth']) 118 | 119 | decoder = Decoder(config['trg_vocab_size'], config['dec_embed'], config['trg_dgru_nhids'], config['trg_igru_nhids'], 120 | config['dec_nhids'], config['enc_nhids'] * 2, config['transition_depth'], config['trg_igru_depth'], 121 | config['trg_dgru_depth'], target_space_idx, target_bos_idx) 122 | 123 | representation = encoder.apply(source_char_seq, source_sample_matrix, source_char_aux, 124 | source_word_mask) 125 | cost = decoder.cost(representation, source_word_mask, target_char_seq, target_sample_matrix, 126 | target_resample_matrix, target_char_aux, target_char_mask, 127 | target_word_mask, target_prev_char_seq, target_prev_char_aux) 128 | 129 | # Set up model 130 | logger.info("Building model") 131 | training_model = Model(cost) 132 | 133 | # Set extensions 134 | logger.info("Initializing extensions") 135 | # Reload model if necessary 136 | extensions = [LoadNMT(config['saveto'])] 137 | 138 | # Initialize main loop 139 | logger.info("Initializing main loop") 140 | main_loop = MainLoop( 141 | model=training_model, 142 | algorithm=None, 143 | data_stream=None, 144 | extensions=extensions 145 | ) 146 | 147 | for extension in main_loop.extensions: 148 | extension.main_loop = main_loop 149 | main_loop._run_extensions('before_training') 150 | 151 | char_embedding = encoder.decimator.apply(source_char_seq.T, source_sample_matrix, source_char_aux.T) 152 | embedding(Model(char_embedding), src_vocab) 153 | 154 | 155 | logger = logging.getLogger(__name__) 156 | logging.basicConfig(level=logging.INFO) 157 | 158 | # Get the arguments 159 | parser = argparse.ArgumentParser() 160 | parser.add_argument("--proto", default="get_config_en2fr", 161 | help="Prototype config to use for config") 162 | args = parser.parse_args() 163 | 164 | if __name__ == '__main__': 165 | # Get configurations for model 166 | configuration = getattr(configurations, args.proto)() 167 | logger.info("Model options:\n{}".format(pprint.pformat(configuration))) 168 | # Get data streams and call main 169 | main(configuration, get_tr_stream(**configuration)) 170 | -------------------------------------------------------------------------------- /training.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import sys 3 | 4 | from collections import Counter 5 | 6 | from theano import tensor 7 | from toolz import merge 8 | 9 | from blocks.algorithms import (GradientDescent, StepClipping, AdaDelta, CompositeRule) 10 | from blocks.extensions import FinishAfter, Printing, Timing 11 | from blocks.filter import VariableFilter 12 | from blocks.graph import ComputationGraph, apply_noise 13 | from blocks.initialization import IsotropicGaussian, Orthogonal, Constant 14 | from blocks.main_loop import MainLoop 15 | from blocks.model import Model 16 | from blocks.select import Selector 17 | 18 | from blocks.monitoring import aggregation 19 | from checkpoint import CheckpointNMT, LoadNMT 20 | from model import BidirectionalEncoder, Decoder 21 | from sampling import BleuValidator, Sampler, CostCurve 22 | 23 | import pprint 24 | 25 | import configurations 26 | from stream import get_tr_stream, get_dev_stream 27 | 28 | logger = logging.getLogger(__name__) 29 | logging.basicConfig(level=logging.INFO) 30 | 31 | 32 | def main(config, tr_stream, dev_stream): 33 | # Create Theano variables 34 | logger.info('Creating theano variables') 35 | source_char_seq = tensor.lmatrix('source_char_seq') 36 | source_sample_matrix = tensor.btensor3('source_sample_matrix') 37 | source_char_aux = tensor.bmatrix('source_char_aux') 38 | source_word_mask = tensor.bmatrix('source_word_mask') 39 | target_char_seq = tensor.lmatrix('target_char_seq') 40 | target_char_aux = tensor.bmatrix('target_char_aux') 41 | target_char_mask = tensor.bmatrix('target_char_mask') 42 | target_sample_matrix = tensor.btensor3('target_sample_matrix') 43 | target_word_mask = tensor.bmatrix('target_word_mask') 44 | target_resample_matrix = tensor.btensor3('target_resample_matrix') 45 | target_prev_char_seq = tensor.lmatrix('target_prev_char_seq') 46 | target_prev_char_aux = tensor.bmatrix('target_prev_char_aux') 47 | target_bos_idx = tr_stream.trg_bos 48 | target_space_idx = tr_stream.space_idx['target'] 49 | 50 | # Construct model 51 | logger.info('Building RNN encoder-decoder') 52 | 53 | encoder = BidirectionalEncoder(config['src_vocab_size'], config['enc_embed'], config['src_dgru_nhids'], 54 | config['enc_nhids'], config['src_dgru_depth'], config['bidir_encoder_depth']) 55 | 56 | decoder = Decoder(config['trg_vocab_size'], config['dec_embed'], config['trg_dgru_nhids'], config['trg_igru_nhids'], 57 | config['dec_nhids'], config['enc_nhids'] * 2, config['transition_depth'], config['trg_igru_depth'], 58 | config['trg_dgru_depth'], target_space_idx, target_bos_idx) 59 | 60 | representation = encoder.apply(source_char_seq, source_sample_matrix, source_char_aux, 61 | source_word_mask) 62 | cost = decoder.cost(representation, source_word_mask, target_char_seq, target_sample_matrix, 63 | target_resample_matrix, target_char_aux, target_char_mask, 64 | target_word_mask, target_prev_char_seq, target_prev_char_aux) 65 | 66 | logger.info('Creating computational graph') 67 | cg = ComputationGraph(cost) 68 | 69 | # Initialize model 70 | logger.info('Initializing model') 71 | encoder.weights_init = decoder.weights_init = IsotropicGaussian( 72 | config['weight_scale']) 73 | encoder.biases_init = decoder.biases_init = Constant(0) 74 | encoder.push_initialization_config() 75 | decoder.push_initialization_config() 76 | for layer_n in range(config['src_dgru_depth']): 77 | encoder.decimator.dgru.transitions[layer_n].weights_init = Orthogonal() 78 | for layer_n in range(config['bidir_encoder_depth']): 79 | encoder.children[1 + layer_n].prototype.recurrent.weights_init = Orthogonal() 80 | if config['trg_igru_depth'] == 1: 81 | decoder.interpolator.igru.weights_init = Orthogonal() 82 | else: 83 | for layer_n in range(config['trg_igru_depth']): 84 | decoder.interpolator.igru.transitions[layer_n].weights_init = Orthogonal() 85 | for layer_n in range(config['trg_dgru_depth']): 86 | decoder.interpolator.feedback_brick.dgru.transitions[layer_n].weights_init = Orthogonal() 87 | for layer_n in range(config['transition_depth']): 88 | decoder.transition.transitions[layer_n].weights_init = Orthogonal() 89 | encoder.initialize() 90 | decoder.initialize() 91 | 92 | 93 | # Print shapes 94 | shapes = [param.get_value().shape for param in cg.parameters] 95 | logger.info("Parameter shapes: ") 96 | for shape, count in Counter(shapes).most_common(): 97 | logger.info(' {:15}: {}'.format(str(shape), count)) 98 | logger.info("Total number of parameters: {}".format(len(shapes))) 99 | 100 | # Print parameter names 101 | enc_dec_param_dict = merge(Selector(encoder).get_parameters(), 102 | Selector(decoder).get_parameters()) 103 | logger.info("Parameter names: ") 104 | for name, value in enc_dec_param_dict.items(): 105 | logger.info(' {:15}: {}'.format(str(value.get_value().shape), name)) 106 | logger.info("Total number of parameters: {}" 107 | .format(len(enc_dec_param_dict))) 108 | 109 | # Set up training model 110 | logger.info("Building model") 111 | training_model = Model(cost) 112 | # Set up training algorithm 113 | logger.info("Initializing training algorithm") 114 | algorithm = GradientDescent( 115 | cost=cost, parameters=cg.parameters, 116 | step_rule=CompositeRule([StepClipping(config['step_clipping']), 117 | eval(config['step_rule'])()]) 118 | ) 119 | 120 | # Set extensions 121 | logger.info("Initializing extensions") 122 | # Extensions 123 | gradient_norm = aggregation.mean(algorithm.total_gradient_norm) 124 | step_norm = aggregation.mean(algorithm.total_step_norm) 125 | train_monitor = CostCurve([cost, gradient_norm, step_norm], config=config, after_batch=True, 126 | before_first_epoch=True, prefix='tra') 127 | extensions = [ 128 | train_monitor, Timing(), 129 | Printing(every_n_batches=config['print_freq']), 130 | FinishAfter(after_n_batches=config['finish_after']), 131 | CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])] 132 | 133 | # Set up beam search and sampling computation graphs if necessary 134 | if config['hook_samples'] >= 1 or config['bleu_script'] is not None: 135 | logger.info("Building sampling model") 136 | generated = decoder.generate(representation, source_word_mask) 137 | search_model = Model(generated) 138 | _, samples = VariableFilter( 139 | bricks=[decoder.sequence_generator], name="outputs")( 140 | ComputationGraph(generated[config['transition_depth']])) # generated[transition_depth] is next_outputs 141 | 142 | # Add sampling 143 | if config['hook_samples'] >= 1: 144 | logger.info("Building sampler") 145 | extensions.append( 146 | Sampler(model=search_model, data_stream=tr_stream, 147 | hook_samples=config['hook_samples'], transition_depth=config['transition_depth'], 148 | every_n_batches=config['sampling_freq'], src_vocab_size=config['src_vocab_size'])) 149 | 150 | # Add early stopping based on bleu 151 | if config['bleu_script'] is not None: 152 | logger.info("Building bleu validator") 153 | extensions.append( 154 | BleuValidator(source_char_seq, source_sample_matrix, source_char_aux, 155 | source_word_mask, samples=samples, config=config, 156 | model=search_model, data_stream=dev_stream, 157 | normalize=config['normalized_bleu'], 158 | every_n_batches=config['bleu_val_freq'])) 159 | 160 | # Reload model if necessary 161 | if config['reload']: 162 | extensions.append(LoadNMT(config['saveto'])) 163 | 164 | # Initialize main loop 165 | logger.info("Initializing main loop") 166 | main_loop = MainLoop( 167 | model=training_model, 168 | algorithm=algorithm, 169 | data_stream=tr_stream, 170 | extensions=extensions 171 | ) 172 | 173 | # Train! 174 | main_loop.run() 175 | 176 | 177 | if __name__ == "__main__": 178 | assert sys.version_info >= (3,4) 179 | # Get configurations for model 180 | configuration = configurations.get_config() 181 | logger.info("Model options:\n{}".format(pprint.pformat(configuration))) 182 | # Get data streams and call main 183 | main(configuration, get_tr_stream(**configuration), 184 | get_dev_stream(**configuration)) 185 | -------------------------------------------------------------------------------- /stream.py: -------------------------------------------------------------------------------- 1 | import numpy 2 | 3 | from fuel.datasets import TextFile 4 | from fuel.schemes import ConstantScheme 5 | from fuel.streams import DataStream 6 | from fuel.transformers import ( 7 | Merge, Batch, Filter, Padding, SortMapping, Unpack, Mapping) 8 | 9 | import pickle 10 | import configurations 11 | 12 | 13 | def _ensure_special_tokens(vocab, bos_idx=0, eos_idx=0, unk_idx=1): 14 | """Ensures special tokens exist in the dictionary.""" 15 | 16 | # remove tokens if they exist in some other index 17 | tokens_to_remove = [k for k, v in vocab.items() 18 | if v in [bos_idx, eos_idx, unk_idx]] 19 | for token in tokens_to_remove: 20 | vocab.pop(token) 21 | # put corresponding item 22 | vocab[''] = bos_idx 23 | vocab[''] = eos_idx 24 | vocab[''] = unk_idx 25 | return vocab 26 | 27 | 28 | def _length(sentence_pair): 29 | """Assumes target is the last element in the tuple.""" 30 | return len(sentence_pair[-1]) 31 | 32 | 33 | class TextFileWithSEOSS(TextFile): 34 | """ Add space eos space to end of source """ 35 | 36 | def __init__(self, files, dictionary, bos_token='', eos_token='', 37 | unk_token='', level='word', preprocess=None, 38 | encoding=None): 39 | super(TextFileWithSEOSS, self).__init__(files, dictionary, bos_token, eos_token, 40 | unk_token, level, preprocess, encoding) 41 | 42 | def get_data(self, state=None, request=None): 43 | if request is not None: 44 | raise ValueError 45 | sentence = next(state) 46 | if self.preprocess is not None: 47 | sentence = self.preprocess(sentence) 48 | data = [self.dictionary[self.bos_token]] if self.bos_token else [] 49 | if self.level == 'word': 50 | data.extend(self._get_from_dictionary(word) 51 | for word in sentence.split()) 52 | else: 53 | data.extend(self._get_from_dictionary(char) 54 | for char in sentence.strip()) 55 | if self.eos_token: 56 | data.append(self.dictionary[' ']) 57 | data.append(self.dictionary[self.eos_token]) 58 | data.append(self.dictionary[' ']) 59 | return (data,) 60 | 61 | 62 | class PaddingWithEOS(Padding): 63 | """Padds a stream with given end of sequence idx.""" 64 | 65 | def __init__(self, data_stream, space_idx, trg_bos, **kwargs): 66 | kwargs['data_stream'] = data_stream 67 | self.space_idx = space_idx 68 | self.trg_bos = trg_bos 69 | super(PaddingWithEOS, self).__init__(**kwargs) 70 | 71 | @property 72 | def sources(self): 73 | sources = ['source_char_seq', 'source_sample_matrix', 'source_char_aux', 'source_word_mask', 74 | 'target_char_seq', 'target_sample_matrix', 'target_char_aux', 'target_word_mask', 75 | 'target_char_mask', 'target_resample_matrix', 'target_prev_char_seq', 'target_prev_char_aux'] 76 | return tuple(sources) 77 | 78 | def transform_batch(self, batch): 79 | batch_with_masks = [] 80 | for k, (source, source_batch) in enumerate(zip(self.data_stream.sources, batch)): 81 | if source not in self.mask_sources: 82 | batch_with_masks.append(source_batch) 83 | continue 84 | word_shapes = [0] * len(source_batch) 85 | shapes = [0] * len(source_batch) 86 | for i, sample in enumerate(source_batch): 87 | np_sample = numpy.asarray(sample) 88 | word_shapes[i] = numpy.count_nonzero(np_sample == self.space_idx[source]) 89 | shapes[i] = np_sample.shape 90 | 91 | lengths = [shape[0] for shape in shapes] 92 | max_word_len = max(word_shapes) 93 | add_space_length = [] 94 | for i in range(len(lengths)): 95 | add_space_length.append(lengths[i] + max_word_len - word_shapes[i]) 96 | max_char_seq_length = max(add_space_length) 97 | rest_shape = shapes[0][1:] 98 | if not all([shape[1:] == rest_shape for shape in shapes]): 99 | raise ValueError("All dimensions except length must be equal") 100 | dtype = numpy.asarray(source_batch[0]).dtype 101 | 102 | char_seq = numpy.zeros( 103 | (len(source_batch), max_char_seq_length) + rest_shape, dtype=dtype) 104 | 105 | for i, sample in enumerate(source_batch): 106 | char_seq[i, :len(sample)] = sample 107 | char_seq[i, len(sample):add_space_length[i]] = self.space_idx[source] 108 | 109 | sample_matrix = numpy.zeros((len(source_batch), max_word_len, max_char_seq_length), 110 | dtype=self.mask_dtype) 111 | char_seq_space_index = char_seq == self.space_idx[source] 112 | 113 | for i in range(len(source_batch)): 114 | sample_matrix[i, range(max_word_len), 115 | numpy.where(char_seq_space_index[i])[0] - 1] = 1 116 | 117 | char_aux = numpy.ones((len(source_batch), max_char_seq_length), self.mask_dtype) 118 | char_aux[char_seq_space_index] = 0 119 | 120 | word_mask = numpy.zeros((len(source_batch), max_word_len), self.mask_dtype) 121 | for i, ws in enumerate(word_shapes): 122 | word_mask[i, :ws] = 1 123 | 124 | batch_with_masks.append(char_seq) 125 | batch_with_masks.append(sample_matrix) 126 | batch_with_masks.append(char_aux) 127 | batch_with_masks.append(word_mask) 128 | 129 | # target sequence 130 | if source == 'target': 131 | target_char_mask = numpy.zeros((len(source_batch), max_char_seq_length), self.mask_dtype) 132 | for i, sequence_length in enumerate(lengths): 133 | target_char_mask[i, :sequence_length] = 1 134 | target_prev_char_seq = numpy.roll(char_seq, 1) 135 | target_prev_char_seq[:, 0] = self.trg_bos 136 | target_prev_char_aux = numpy.roll(char_aux, 1) 137 | # start of sequence, must be 0 138 | target_prev_char_aux[:, 0] = 0 139 | target_resample_matrix = numpy.zeros((len(source_batch), max_char_seq_length, max_word_len), 140 | dtype=self.mask_dtype) 141 | 142 | curr_space_idx = numpy.where(char_seq_space_index) 143 | for i in range(len(source_batch)): 144 | pj = 0 145 | for cj in range(max_word_len): 146 | target_resample_matrix[i, pj:curr_space_idx[1][i * max_word_len + cj] + 1, cj] = 1 147 | pj = curr_space_idx[1][i * max_word_len + cj] + 1 148 | 149 | batch_with_masks.append(target_char_mask) 150 | batch_with_masks.append(target_resample_matrix) 151 | batch_with_masks.append(target_prev_char_seq) 152 | batch_with_masks.append(target_prev_char_aux) 153 | return tuple(batch_with_masks) 154 | 155 | 156 | class _oov_to_unk(object): 157 | """Maps out of vocabulary token index to unk token index.""" 158 | 159 | def __init__(self, src_vocab_size=120, trg_vocab_size=120, 160 | unk_id=1): 161 | self.src_vocab_size = src_vocab_size 162 | self.trg_vocab_size = trg_vocab_size 163 | self.unk_id = unk_id 164 | 165 | def __call__(self, sentence_pair): 166 | return ([x if x < self.src_vocab_size else self.unk_id 167 | for x in sentence_pair[0]], 168 | [x if x < self.trg_vocab_size else self.unk_id 169 | for x in sentence_pair[1]]) 170 | 171 | 172 | class _too_long(object): 173 | """Filters sequences longer than given sequence length.""" 174 | 175 | def __init__(self, unk_id, space_idx, max_src_seq_char_len, max_src_seq_word_len, 176 | max_trg_seq_char_len, max_trg_seq_word_len): 177 | self.unk_id = unk_id 178 | self.max_src_seq_char_len = max_src_seq_char_len 179 | self.max_src_seq_word_len = max_src_seq_word_len 180 | self.max_trg_seq_char_len = max_trg_seq_char_len 181 | self.max_trg_seq_word_len = max_trg_seq_word_len 182 | self.space_idx = space_idx 183 | 184 | def __call__(self, sentence_pair): 185 | max_unk = 5 186 | return all( 187 | [len(sentence_pair[0]) <= self.max_src_seq_char_len and sentence_pair[0].count(self.unk_id) < max_unk and 188 | sentence_pair[0].count(self.space_idx[0]) < self.max_src_seq_word_len, 189 | len(sentence_pair[1]) <= self.max_trg_seq_char_len and sentence_pair[1].count(self.unk_id) < max_unk and 190 | sentence_pair[1].count(self.space_idx[1]) < self.max_trg_seq_word_len]) 191 | 192 | 193 | def get_tr_stream(src_vocab, trg_vocab, src_data, trg_data, 194 | src_vocab_size=120, trg_vocab_size=120, unk_id=1, bos_token='', max_src_seq_char_len=300, 195 | max_src_seq_word_len=50, max_trg_seq_char_len=300, max_trg_seq_word_len=50, 196 | batch_size=80, sort_k_batches=12, **kwargs): 197 | """Prepares the training data stream.""" 198 | 199 | # Load dictionaries and ensure special tokens exist 200 | src_vocab = _ensure_special_tokens( 201 | src_vocab if isinstance(src_vocab, dict) 202 | else pickle.load(open(src_vocab, 'rb')), 203 | bos_idx=0, eos_idx=src_vocab_size - 1, unk_idx=unk_id) 204 | 205 | trg_vocab = _ensure_special_tokens( 206 | trg_vocab if isinstance(trg_vocab, dict) else 207 | pickle.load(open(trg_vocab, 'rb')), 208 | bos_idx=0, eos_idx=trg_vocab_size - 1, unk_idx=unk_id) 209 | 210 | # Get text files from both source and target 211 | src_dataset = TextFileWithSEOSS([src_data], src_vocab, None, level='character') 212 | trg_dataset = TextFileWithSEOSS([trg_data], trg_vocab, None, level='character') 213 | 214 | # Merge them to get a source, target pair 215 | stream = Merge([src_dataset.get_example_stream(), 216 | trg_dataset.get_example_stream()], 217 | ('source', 'target')) 218 | 219 | # Filter sequences that are too long 220 | stream = Filter(stream, predicate=_too_long(unk_id, [src_vocab[' '], trg_vocab[' ']], 221 | max_src_seq_char_len, max_src_seq_word_len, 222 | max_trg_seq_char_len, max_trg_seq_word_len)) 223 | 224 | # Replace out of vocabulary tokens with unk token 225 | stream = Mapping(stream, 226 | _oov_to_unk(src_vocab_size=src_vocab_size, 227 | trg_vocab_size=trg_vocab_size, 228 | unk_id=unk_id)) 229 | 230 | # Build a batched version of stream to read k batches ahead 231 | stream = Batch(stream, 232 | iteration_scheme=ConstantScheme( 233 | batch_size * sort_k_batches)) 234 | 235 | # Sort all samples in the read-ahead batch 236 | stream = Mapping(stream, SortMapping(_length)) 237 | 238 | # Convert it into a stream again 239 | stream = Unpack(stream) 240 | # Construct batches from the stream with specified batch size 241 | stream = Batch( 242 | stream, iteration_scheme=ConstantScheme(batch_size)) 243 | 244 | # Pad sequences that are short 245 | masked_stream = PaddingWithEOS(stream, {'source': src_vocab[' '], 'target': trg_vocab[' ']}, trg_vocab[bos_token], 246 | mask_dtype='int8') 247 | 248 | return masked_stream 249 | 250 | 251 | def get_dev_stream(val_set=None, src_vocab=None, src_vocab_size=120, 252 | unk_id=1, **kwargs): 253 | """Setup development set stream if necessary.""" 254 | dev_stream = None 255 | if val_set is not None and src_vocab is not None: 256 | src_vocab = _ensure_special_tokens( 257 | src_vocab if isinstance(src_vocab, dict) else 258 | pickle.load(open(src_vocab, 'rb')), 259 | bos_idx=0, eos_idx=src_vocab_size - 1, unk_idx=unk_id) 260 | dev_src_dataset = TextFileWithSEOSS([val_set], src_vocab, None, level='character') 261 | dev_stream = DataStream(dev_src_dataset) 262 | return dev_stream 263 | 264 | 265 | def get_test_stream(test_set=None, src_vocab=None, trg_vocab=None, src_vocab_size=120, trg_vocab_size=120, unk_id=1, 266 | bos_token='', **kwargs): 267 | """Setup development set stream if necessary.""" 268 | test_stream = None 269 | if test_set is not None and src_vocab is not None and trg_vocab is not None: 270 | src_vocab = _ensure_special_tokens( 271 | src_vocab if isinstance(src_vocab, dict) else 272 | pickle.load(open(src_vocab, 'rb')), 273 | bos_idx=0, eos_idx=src_vocab_size - 1, unk_idx=unk_id) 274 | 275 | trg_vocab = _ensure_special_tokens( 276 | trg_vocab if isinstance(trg_vocab, dict) else 277 | pickle.load(open(trg_vocab, 'rb')), 278 | bos_idx=0, eos_idx=trg_vocab_size - 1, unk_idx=unk_id) 279 | 280 | test_src_dataset = TextFileWithSEOSS([test_set], src_vocab, None, level='character') 281 | test_stream = DataStream(test_src_dataset) 282 | test_stream.space_idx = {'source': src_vocab[' '], 'target': trg_vocab[' ']} 283 | test_stream.trg_bos = trg_vocab[bos_token] 284 | test_stream.trg_vocab = trg_vocab 285 | test_stream.eos_token = '' 286 | 287 | return test_stream 288 | 289 | 290 | if __name__ == '__main__': 291 | # test stream 292 | configuration = configurations.get_config() 293 | tr = get_tr_stream(**configuration) 294 | total = 0 295 | # test 296 | for s in tr.get_epoch_iterator(): 297 | total += 1 298 | if total % 10000 == 0: 299 | print(total) 300 | print(total) 301 | -------------------------------------------------------------------------------- /sampling.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | 3 | import logging 4 | import operator 5 | import os 6 | import time 7 | import re 8 | import signal 9 | 10 | import numpy 11 | from blocks.extensions import SimpleExtension, TrainingExtension 12 | from blocks.extensions.monitoring import TrainingDataMonitoring 13 | from blocks.search import BeamSearch 14 | from subprocess import Popen, PIPE 15 | from checkpoint import SaveLoadUtils 16 | 17 | logger = logging.getLogger(__name__) 18 | 19 | 20 | class SamplingBase(object): 21 | """Utility class for Validator and Sampler.""" 22 | 23 | def _get_attr_rec(self, obj, attr): 24 | return self._get_attr_rec(getattr(obj, attr), attr) \ 25 | if hasattr(obj, attr) else obj 26 | 27 | def _get_true_length(self, seq, vocab): 28 | try: 29 | return seq.tolist().index(vocab['']) + 1 30 | except ValueError: 31 | return len(seq) 32 | 33 | def _oov_to_unk(self, seq, vocab_size, unk_idx): 34 | return [x if x < vocab_size else unk_idx for x in seq] 35 | 36 | def _idx_to_word(self, seq, ivocab): 37 | return "".join([ivocab.get(idx, "") for idx in seq]) 38 | 39 | def build_input_dict(self, input_, src_vocab): 40 | input_length = self._get_true_length(input_, src_vocab) + 1 41 | input_ = input_[:input_length] 42 | total_word = list(input_).count(src_vocab[' ']) 43 | 44 | source_sample_matrix = numpy.zeros((total_word, input_length), dtype='int8') 45 | source_sample_matrix[range(total_word), numpy.nonzero(input_ == src_vocab[' '])[0] - 1] = 1 46 | 47 | source_word_mask = numpy.ones(total_word, dtype='int8') 48 | 49 | source_char_aux = numpy.ones(input_length, dtype='int8') 50 | source_char_aux[input_ == src_vocab[' ']] = 0 51 | 52 | input_dict = {'source_sample_matrix': source_sample_matrix[None, :], 53 | 'source_char_aux': source_char_aux[None, :], 54 | 'source_char_seq': input_[None, :], 55 | 'source_word_mask': source_word_mask[None, :]} 56 | return input_length, input_dict 57 | 58 | def build_input_dict_tile(self, input_, src_vocab, beam_size): 59 | input_length = self._get_true_length(input_, src_vocab) + 1 60 | input_ = input_[:input_length] 61 | total_word = list(input_).count(src_vocab[' ']) 62 | 63 | source_sample_matrix = numpy.zeros((total_word, input_length), dtype='int8') 64 | source_sample_matrix[range(total_word), numpy.nonzero(input_ == src_vocab[' '])[0] - 1] = 1 65 | 66 | source_word_mask = numpy.ones(total_word, dtype='int8') 67 | 68 | source_char_aux = numpy.ones(input_length, dtype='int8') 69 | source_char_aux[input_ == src_vocab[' ']] = 0 70 | 71 | input_dict = {'source_sample_matrix': numpy.tile(source_sample_matrix, (beam_size, 1, 1)), 72 | 'source_word_mask': numpy.tile(source_word_mask, (beam_size, 1)), 73 | 'source_char_aux': numpy.tile(source_char_aux, (beam_size, 1)), 74 | 'source_char_seq': numpy.tile(input_, (beam_size, 1))} 75 | 76 | return input_dict 77 | 78 | 79 | class CostCurve(TrainingDataMonitoring): 80 | """ Record training curve """ 81 | 82 | def __init__(self, variables, config, **kwargs): 83 | super(CostCurve, self).__init__(variables, **kwargs) 84 | self.cost_curve = [] 85 | self.config = config 86 | # Create saving directory if it does not exist 87 | if not os.path.exists(self.config['saveto']): 88 | os.makedirs(self.config['saveto']) 89 | 90 | if self.config['reload']: 91 | try: 92 | self.cost_curve = numpy.load(os.path.join(self.config['saveto'], 93 | 'cost_curve.npz')) 94 | self.cost_curve = self.cost_curve['cost_curves'].tolist() 95 | logger.info("Cost Curve Reloaded") 96 | except: 97 | logger.info("Cost Curve not Found") 98 | 99 | def do(self, callback_name, *args): 100 | """Initializes the buffer or commits the values to the log. 101 | What this method does depends on from what callback it is called 102 | and with which arguments. When called within `before_training`, it 103 | initializes the aggregation buffer and instructs the training 104 | algorithm what additional computations should be carried at each 105 | step by adding corresponding updates to it. In most_other cases it 106 | writes aggregated values of the monitored variables to the log. An 107 | exception is when an argument `just_aggregate` is given: in this 108 | cases it updates the values of monitored non-Theano quantities, but 109 | does not write anything to the log. 110 | """ 111 | data, args = self.parse_args(callback_name, args) 112 | if callback_name == 'before_training': 113 | self.main_loop.algorithm.add_updates( 114 | self._variables.accumulation_updates) 115 | self.main_loop.algorithm.add_updates( 116 | self._required_for_non_variables.accumulation_updates) 117 | self._variables.initialize_aggregators() 118 | self._required_for_non_variables.initialize_aggregators() 119 | self._non_variables.initialize_quantities() 120 | else: 121 | # When called first time at any iterations, update 122 | # monitored non-Theano quantities 123 | if (self.main_loop.status['iterations_done'] > 124 | self._last_time_called): 125 | self._non_variables.aggregate_quantities( 126 | list(self._required_for_non_variables 127 | .get_aggregated_values().values())) 128 | self._required_for_non_variables.initialize_aggregators() 129 | self._last_time_called = ( 130 | self.main_loop.status['iterations_done']) 131 | # If only called to update non-Theano quantities, 132 | # do just that 133 | if args == ('just_aggregate',): 134 | return 135 | # Otherwise, also save current values of from the accumulators 136 | curr_iter = self.main_loop.status['iterations_done'] 137 | if curr_iter == 0: 138 | return 139 | 140 | curr_cost = self._variables.get_aggregated_values() 141 | curr_cost = curr_cost['decoder_cost_cost'].tolist() 142 | self.cost_curve.append({curr_iter: curr_cost}) 143 | 144 | if curr_iter % 100 == 0: 145 | numpy.savez(os.path.join(self.config['saveto'], 'cost_curve.npz'), 146 | cost_curves=self.cost_curve) 147 | 148 | self.add_records( 149 | self.main_loop.log, 150 | self._variables.get_aggregated_values().items()) 151 | self._variables.initialize_aggregators() 152 | self.add_records( 153 | self.main_loop.log, 154 | self._non_variables.get_aggregated_values().items()) 155 | self._non_variables.initialize_quantities() 156 | 157 | 158 | class Sampler(SimpleExtension, SamplingBase): 159 | """Random Sampling from model.""" 160 | 161 | def __init__(self, model, data_stream, hook_samples=1, transition_depth=1, 162 | src_vocab=None, trg_vocab=None, src_ivocab=None, 163 | trg_ivocab=None, src_vocab_size=None, **kwargs): 164 | super(Sampler, self).__init__(**kwargs) 165 | self.model = model 166 | self.hook_samples = hook_samples 167 | self.data_stream = data_stream 168 | self.src_vocab = src_vocab 169 | self.trg_vocab = trg_vocab 170 | self.src_ivocab = src_ivocab 171 | self.transition_depth = transition_depth 172 | self.trg_ivocab = trg_ivocab 173 | self.src_vocab_size = src_vocab_size 174 | self.is_synced = False 175 | self.sampling_fn = model.get_theano_function() 176 | 177 | def do(self, which_callback, *args): 178 | 179 | # Get dictionaries, this may not be the practical way 180 | sources = self._get_attr_rec(self.main_loop, 'data_stream') 181 | 182 | # Load vocabularies and invert if necessary 183 | # WARNING: Source and target indices from data stream 184 | # can be different 185 | if not self.src_vocab: 186 | self.src_vocab = sources.data_streams[0].dataset.dictionary 187 | if not self.trg_vocab: 188 | self.trg_vocab = sources.data_streams[1].dataset.dictionary 189 | if not self.src_ivocab: 190 | self.src_ivocab = {v: k for k, v in self.src_vocab.items()} 191 | if not self.trg_ivocab: 192 | self.trg_ivocab = {v: k for k, v in self.trg_vocab.items()} 193 | if not self.src_vocab_size: 194 | self.src_vocab_size = len(self.src_vocab) 195 | 196 | # Randomly select source samples from the current batch 197 | # WARNING: Source and target indices from data stream 198 | # can be different 199 | batch = args[0] 200 | batch_size = batch['source_char_seq'].shape[0] 201 | hook_samples = min(batch_size, self.hook_samples) 202 | 203 | # TODO: this is problematic for boundary conditions, eg. last batch 204 | sample_idx = numpy.random.choice( 205 | batch_size, hook_samples, replace=False) 206 | src_batch = batch['source_char_seq'] 207 | trg_batch = batch['target_char_seq'] 208 | input_ = src_batch[sample_idx, :] 209 | target_ = trg_batch[sample_idx, :] 210 | 211 | # Sample 212 | print() 213 | for i in range(hook_samples): 214 | input_length, input_dict = self.build_input_dict(input_[i], self.src_vocab) 215 | target_length = self._get_true_length(target_[i], self.trg_vocab) + 1 216 | sfn = self.sampling_fn(**input_dict) 217 | outputs = sfn[self.transition_depth] 218 | costs = sfn[-1] 219 | outputs = outputs.flatten() 220 | costs = costs.flatten() 221 | 222 | sample_length = self._get_true_length(outputs, self.trg_vocab) 223 | print("Input : ", self._idx_to_word(input_[i][:input_length], 224 | self.src_ivocab)) 225 | print("Target: ", self._idx_to_word(target_[i][:target_length], 226 | self.trg_ivocab)) 227 | print("Sample: ", self._idx_to_word(outputs[:sample_length], 228 | self.trg_ivocab)) 229 | print("Sample cost: ", costs[:sample_length].mean()) 230 | print() 231 | 232 | 233 | class BleuValidator(SimpleExtension, SamplingBase, SaveLoadUtils): 234 | # TODO: a lot has been changed in NMT, sync respectively 235 | """Implements early stopping based on BLEU score.""" 236 | 237 | def __init__(self, source_char_seq, source_sample_matrix, source_char_aux, 238 | source_word_mask, samples, model, data_stream, 239 | config, n_best=1, track_n_models=1, 240 | normalize=True, **kwargs): 241 | # TODO: change config structure 242 | super(BleuValidator, self).__init__(**kwargs) 243 | self.source_char_seq = source_char_seq 244 | self.source_sample_matrix = source_sample_matrix 245 | self.source_char_aux = source_char_aux 246 | self.source_word_mask = source_word_mask 247 | self.samples = samples 248 | self.model = model 249 | self.data_stream = data_stream 250 | self.config = config 251 | self.n_best = n_best 252 | self.track_n_models = track_n_models 253 | self.normalize = normalize 254 | self.verbose = config.get('val_set_out', None) 255 | 256 | # Helpers 257 | self.vocab = data_stream.dataset.dictionary 258 | self.src_ivocab = {v: k for k, v in self.vocab.items()} 259 | self.unk_sym = data_stream.dataset.unk_token 260 | self.eos_sym = data_stream.dataset.eos_token 261 | self.unk_idx = self.vocab[self.unk_sym] 262 | self.eos_idx = self.vocab[self.eos_sym] 263 | self.best_models = [] 264 | self.val_bleu_curve = [] 265 | self.beam_search = BeamSearch(samples=samples) 266 | self.multibleu_cmd = ['perl', self.config['bleu_script'], 267 | self.config['val_set_grndtruth'], '<'] 268 | 269 | # Create saving directory if it does not exist 270 | if not os.path.exists(self.config['saveto']): 271 | os.makedirs(self.config['saveto']) 272 | 273 | if self.config['reload']: 274 | try: 275 | bleu_score = numpy.load(os.path.join(self.config['saveto'], 276 | 'val_bleu_scores.npz')) 277 | self.val_bleu_curve = bleu_score['bleu_scores'].tolist() 278 | # Track n best previous bleu scores 279 | for i, bleu in enumerate( 280 | sorted([list(v.values())[0] for v in self.val_bleu_curve], reverse=True)): 281 | if i < self.track_n_models: 282 | self.best_models.append(ModelInfo(bleu, self.config['saveto'])) 283 | logger.info("BleuScores Reloaded") 284 | except: 285 | logger.info("BleuScores not Found") 286 | 287 | def do(self, which_callback, *args): 288 | 289 | # Track validation burn in 290 | if self.main_loop.status['iterations_done'] < \ 291 | self.config['val_burn_in']: 292 | return 293 | 294 | # Evaluate and save if necessary 295 | self._save_model(self._evaluate_model(self.main_loop)) 296 | 297 | def _evaluate_model(self, main_loop): 298 | curr_iter = main_loop.status['iterations_done'] 299 | logger.info("Started Validation: ") 300 | val_start_time = time.time() 301 | mb_subprocess = Popen(self.multibleu_cmd, stdin=PIPE, stdout=PIPE, universal_newlines=True) 302 | total_cost = 0.0 303 | 304 | # Get target vocabulary 305 | sources = self._get_attr_rec(self.main_loop, 'data_stream') 306 | trg_vocab = sources.data_streams[1].dataset.dictionary 307 | self.trg_vocab = trg_vocab 308 | self.trg_ivocab = {v: k for k, v in trg_vocab.items()} 309 | trg_eos_sym = sources.data_streams[1].dataset.eos_token 310 | self.trg_eos_idx = trg_vocab[trg_eos_sym] 311 | 312 | if self.verbose: 313 | ftrans = open(self.config['val_set_out'] + str(curr_iter), 'w') 314 | 315 | for i, line in enumerate(self.data_stream.get_epoch_iterator()): 316 | """ 317 | Load the sentence, retrieve the sample, write to file 318 | """ 319 | 320 | seq = self._oov_to_unk( 321 | line[0], self.config['src_vocab_size'], self.unk_idx) 322 | input_dict = self.build_input_dict_tile(numpy.asarray(seq), self.vocab, self.config['beam_size']) 323 | 324 | # draw sample, checking to ensure we don't get an empty string back 325 | trans, costs = \ 326 | self.beam_search.search( 327 | input_values={self.source_char_seq: input_dict['source_char_seq'], 328 | self.source_sample_matrix: input_dict['source_sample_matrix'], 329 | self.source_word_mask: input_dict['source_word_mask'], 330 | self.source_char_aux: input_dict['source_char_aux']}, 331 | max_length=3 * len(seq), eol_symbol=self.trg_eos_idx, 332 | ignore_first_eol=False) 333 | 334 | # normalize costs according to the sequence lengths 335 | if self.normalize: 336 | lengths = numpy.array([len(s) for s in trans]) 337 | costs = costs / lengths 338 | 339 | nbest_idx = numpy.argsort(costs)[:self.n_best] 340 | for j, best in enumerate(nbest_idx): 341 | try: 342 | total_cost += costs[best] 343 | trans_out = trans[best] 344 | 345 | # convert idx to words 346 | try: 347 | sample_length = trans_out.index(self.trg_vocab['']) 348 | except ValueError: 349 | sample_length = len(seq) 350 | trans_out = trans_out[:sample_length] 351 | trans_out = self._idx_to_word(trans_out, self.trg_ivocab) 352 | 353 | except ValueError: 354 | logger.info( 355 | "Can NOT find a translation for line: {}".format(i + 1)) 356 | trans_out = '' 357 | 358 | if j == 0: 359 | # Write to subprocess and file if it exists 360 | print("Line:", i) 361 | print("Input : ", self._idx_to_word(line[0], self.src_ivocab)) 362 | print("Sample: ", trans_out) 363 | print("Error:", costs[best]) 364 | print() 365 | 366 | print(trans_out, file=mb_subprocess.stdin) 367 | if self.verbose: 368 | print(trans_out, file=ftrans) 369 | 370 | if i != 0 and i % 100 == 0: 371 | logger.info( 372 | "Translated {} lines of validation set...".format(i)) 373 | 374 | mb_subprocess.stdin.flush() 375 | 376 | logger.info("Total cost of the validation: {}".format(total_cost)) 377 | self.data_stream.reset() 378 | if self.verbose: 379 | ftrans.close() 380 | 381 | # send end of file, read output. 382 | mb_subprocess.stdin.close() 383 | stdout = mb_subprocess.stdout.readline() 384 | logger.info(stdout) 385 | out_parse = re.match(r'BLEU = [-.0-9]+', stdout) 386 | logger.info("Validation Took: {} minutes".format( 387 | float(time.time() - val_start_time) / 60.)) 388 | assert out_parse is not None 389 | 390 | # extract the score 391 | bleu_score = float(out_parse.group()[6:]) 392 | self.val_bleu_curve.append({curr_iter: bleu_score}) 393 | logger.info(bleu_score) 394 | mb_subprocess.terminate() 395 | 396 | return bleu_score 397 | 398 | def _is_valid_to_save(self, bleu_score): 399 | if not self.best_models or min(self.best_models, 400 | key=operator.attrgetter('bleu_score')).bleu_score < bleu_score: 401 | return True 402 | return False 403 | 404 | def _save_model(self, bleu_score): 405 | numpy.savez( 406 | os.path.join(self.config['saveto'], 'val_bleu_scores.npz'), 407 | bleu_scores=self.val_bleu_curve) 408 | if self._is_valid_to_save(bleu_score): 409 | model = ModelInfo(bleu_score, self.config['saveto']) 410 | 411 | # Manage n-best model list first 412 | if len(self.best_models) >= self.track_n_models: 413 | old_model = self.best_models[0] 414 | if old_model.path and os.path.isfile(old_model.path): 415 | logger.info("Deleting old model %s" % old_model.path) 416 | os.remove(old_model.path) 417 | self.best_models.remove(old_model) 418 | 419 | self.best_models.append(model) 420 | self.best_models.sort(key=operator.attrgetter('bleu_score')) 421 | 422 | # Save the model here 423 | s = signal.signal(signal.SIGINT, signal.SIG_IGN) 424 | logger.info("Saving new model {}".format(model.path)) 425 | params_to_save = self.main_loop.model.get_parameter_values() 426 | self.save_parameter_values(params_to_save, model.path) 427 | signal.signal(signal.SIGINT, s) 428 | 429 | 430 | class BleuTester(TrainingExtension, SamplingBase): 431 | # TODO: a lot has been changed in NMT, sync respectively 432 | """Implements Testing BLEU score.""" 433 | 434 | def __init__(self, source_char_seq, source_sample_matrix, source_char_aux, 435 | source_word_mask, samples, model, data_stream, 436 | config, n_best=1, track_n_models=1, 437 | normalize=True, **kwargs): 438 | # TODO: change config structure 439 | super(BleuTester, self).__init__(**kwargs) 440 | self.source_char_seq = source_char_seq 441 | self.source_sample_matrix = source_sample_matrix 442 | self.source_char_aux = source_char_aux 443 | self.source_word_mask = source_word_mask 444 | self.samples = samples 445 | self.model = model 446 | self.data_stream = data_stream 447 | self.config = config 448 | self.n_best = n_best 449 | self.track_n_models = track_n_models 450 | self.normalize = normalize 451 | self.verbose = True 452 | 453 | # Helpers 454 | self.vocab = data_stream.dataset.dictionary 455 | self.src_ivocab = {v: k for k, v in self.vocab.items()} 456 | self.unk_sym = data_stream.dataset.unk_token 457 | self.eos_sym = data_stream.dataset.eos_token 458 | self.unk_idx = self.vocab[self.unk_sym] 459 | self.eos_idx = self.vocab[self.eos_sym] 460 | self.beam_search = BeamSearch(samples=samples) 461 | self.multibleu_cmd = ['perl', self.config['bleu_script'], 462 | self.config['test_set_grndtruth'], '<'] 463 | 464 | # Create saving directory if it does not exist 465 | if not os.path.exists(self.config['saveto']): 466 | os.makedirs(self.config['saveto']) 467 | 468 | def before_training(self): 469 | self._evaluate_model() 470 | 471 | def _evaluate_model(self): 472 | 473 | logger.info("Started Test: ") 474 | test_start_time = time.time() 475 | mb_subprocess = Popen(self.multibleu_cmd, stdin=PIPE, stdout=PIPE, universal_newlines=True) 476 | total_cost = 0.0 477 | 478 | # Get target vocabulary 479 | trg_vocab = self.data_stream.trg_vocab 480 | self.trg_vocab = trg_vocab 481 | self.trg_ivocab = {v: k for k, v in trg_vocab.items()} 482 | trg_eos_sym = self.data_stream.eos_token 483 | self.trg_eos_idx = trg_vocab[trg_eos_sym] 484 | 485 | if self.verbose: 486 | ftrans = open(self.config['test_set_out'], 'w') 487 | 488 | for i, line in enumerate(self.data_stream.get_epoch_iterator()): 489 | """ 490 | Load the sentence, retrieve the sample, write to file 491 | """ 492 | 493 | seq = self._oov_to_unk( 494 | line[0], self.config['src_vocab_size'], self.unk_idx) 495 | input_dict = self.build_input_dict_tile(numpy.asarray(seq), self.vocab, self.config['beam_size']) 496 | 497 | # draw sample, checking to ensure we don't get an empty string back 498 | trans, costs = \ 499 | self.beam_search.search( 500 | input_values={self.source_char_seq: input_dict['source_char_seq'], 501 | self.source_sample_matrix: input_dict['source_sample_matrix'], 502 | self.source_word_mask: input_dict['source_word_mask'], 503 | self.source_char_aux: input_dict['source_char_aux']}, 504 | max_length=3 * len(seq), eol_symbol=self.trg_eos_idx, 505 | ignore_first_eol=False) 506 | 507 | # normalize costs according to the sequence lengths 508 | if self.normalize: 509 | lengths = numpy.array([len(s) for s in trans]) 510 | costs = costs / lengths 511 | 512 | nbest_idx = numpy.argsort(costs)[:self.n_best] 513 | for j, best in enumerate(nbest_idx): 514 | try: 515 | total_cost += costs[best] 516 | trans_out = trans[best] 517 | 518 | # convert idx to words 519 | try: 520 | sample_length = trans_out.index(self.trg_vocab['']) 521 | except ValueError: 522 | sample_length = len(seq) 523 | trans_out = trans_out[:sample_length] 524 | trans_out = self._idx_to_word(trans_out, self.trg_ivocab) 525 | 526 | except ValueError: 527 | logger.info( 528 | "Can NOT find a translation for line: {}".format(i + 1)) 529 | trans_out = '' 530 | 531 | if j == 0: 532 | # Write to subprocess and file if it exists 533 | print("Line:", i) 534 | print("Input : ", self._idx_to_word(line[0], self.src_ivocab)) 535 | print("Sample: ", trans_out) 536 | print("Error:", costs[best]) 537 | print() 538 | 539 | print(trans_out, file=mb_subprocess.stdin) 540 | if self.verbose: 541 | print(trans_out, file=ftrans) 542 | 543 | if i != 0 and i % 100 == 0: 544 | logger.info( 545 | "Translated {} lines of test set...".format(i)) 546 | 547 | mb_subprocess.stdin.flush() 548 | 549 | logger.info("Total cost of the test: {}".format(total_cost)) 550 | self.data_stream.reset() 551 | if self.verbose: 552 | ftrans.close() 553 | 554 | # send end of file, read output. 555 | mb_subprocess.stdin.close() 556 | stdout = mb_subprocess.stdout.readline() 557 | logger.info(stdout) 558 | out_parse = re.match(r'BLEU = [-.0-9]+', stdout) 559 | logger.info("Test Took: {} minutes".format( 560 | float(time.time() - test_start_time) / 60.)) 561 | assert out_parse is not None 562 | 563 | # extract the score 564 | bleu_score = float(out_parse.group()[6:]) 565 | logger.info(bleu_score) 566 | mb_subprocess.terminate() 567 | 568 | return bleu_score 569 | 570 | 571 | 572 | class ModelInfo: 573 | """Utility class to keep track of evaluated models.""" 574 | 575 | def __init__(self, bleu_score, path=''): 576 | self.bleu_score = bleu_score 577 | 578 | self.path = self._generate_path(path) 579 | 580 | def _generate_path(self, path): 581 | gen_path = os.path.join( 582 | path, 'best_bleu_params_BLEU%.2f.npz' % 583 | (self.bleu_score) if path else None) 584 | return gen_path 585 | -------------------------------------------------------------------------------- /model.py: -------------------------------------------------------------------------------- 1 | import copy 2 | 3 | from blocks.bricks import (Tanh, Linear, FeedforwardSequence, Identity, 4 | Initializable, MLP) 5 | from blocks.bricks.attention import SequenceContentAttention, AttentionRecurrent 6 | from blocks.bricks.base import application, lazy 7 | from blocks.bricks.lookup import LookupTable 8 | from blocks.bricks.parallel import Fork 9 | from blocks.bricks.recurrent import GatedRecurrent, Bidirectional, recurrent, RecurrentStack 10 | from blocks.bricks.sequence_generators import ( 11 | Readout, SoftmaxEmitter, BaseSequenceGenerator) 12 | from blocks.roles import add_role, WEIGHT 13 | from blocks.utils import shared_floatx_nans, dict_subset, dict_union 14 | from theano import tensor 15 | from toolz import merge 16 | from blocks.bricks.recurrent.misc import RECURRENTSTACK_SEPARATOR 17 | 18 | # Helper class 19 | class InitializableFeedforwardSequence(FeedforwardSequence, Initializable): 20 | pass 21 | 22 | 23 | class DGRU(GatedRecurrent): 24 | """DGRU in Decimator""" 25 | 26 | def __init__(self, dim, activation=None, gate_activation=None, 27 | **kwargs): 28 | super(DGRU, self).__init__(dim, activation, gate_activation, **kwargs) 29 | 30 | @recurrent(sequences=['mask', 'inputs', 'gate_inputs'], 31 | states=['states'], outputs=['states'], contexts=[]) 32 | def apply(self, inputs, gate_inputs, states, mask=None): 33 | """Apply the gated recurrent transition. 34 | Parameters 35 | ---------- 36 | states : :class:`~tensor.TensorVariable` 37 | The 2 dimensional matrix of current states in the shape 38 | (batch_size, dim). Required for `one_step` usage. 39 | inputs : :class:`~tensor.TensorVariable` 40 | The 2 dimensional matrix of inputs in the shape (batch_size, 41 | dim) 42 | gate_inputs : :class:`~tensor.TensorVariable` 43 | The 2 dimensional matrix of inputs to the gates in the 44 | shape (batch_size, 2 * dim). 45 | mask : :class:`~tensor.TensorVariable` 46 | A 1D binary array in the shape (batch,) which is 1 if there is 47 | the charater available, 0 if there is the delimiter. 48 | Returns 49 | ------- 50 | output : :class:`~tensor.TensorVariable` 51 | Next states of the network. 52 | """ 53 | gate_values = self.gate_activation.apply( 54 | states.dot(self.state_to_gates) + gate_inputs) 55 | update_values = gate_values[:, :self.dim] 56 | reset_values = gate_values[:, self.dim:] 57 | states_reset = states * reset_values 58 | next_states = self.activation.apply( 59 | states_reset.dot(self.state_to_state) + inputs) 60 | next_states = (next_states * update_values + 61 | states * (1 - update_values)) 62 | if mask: 63 | next_states = (mask[:, None] * next_states + 64 | (1 - mask[:, None]) * self.initial_states(mask.shape[0])) 65 | return next_states 66 | 67 | 68 | class Decimator(Initializable): 69 | """Char encoder, mapping a charater-level word to a vector""" 70 | 71 | def __init__(self, vocab_size, embedding_dim, dgru_state_dim, dgru_depth, **kwargs): 72 | super(Decimator, self).__init__(**kwargs) 73 | 74 | self.vocab_size = vocab_size 75 | self.embedding_dim = embedding_dim 76 | self.dgru_state_dim = dgru_state_dim 77 | self.embedding_dim = embedding_dim 78 | self.lookup = LookupTable(name='embeddings') 79 | self.dgru_depth = dgru_depth 80 | self.dgru = RecurrentStack([DGRU(activation=Tanh(), dim=self.dgru_state_dim) for _ in range(dgru_depth)], 81 | skip_connections=True) 82 | 83 | self.gru_fork = Fork([name for name in self.dgru.apply.sequences 84 | if name != 'mask'], prototype=Linear(), name='gru_fork') 85 | 86 | self.children = [self.lookup, self.dgru, self.gru_fork] 87 | 88 | def _push_allocation_config(self): 89 | self.lookup.length = self.vocab_size 90 | self.lookup.dim = self.embedding_dim 91 | 92 | self.gru_fork.input_dim = self.embedding_dim 93 | self.gru_fork.output_dims = [self.dgru.get_dim(name) 94 | for name in self.gru_fork.output_names] 95 | 96 | @application(inputs=['char_seq', 'sample_matrix', 'char_aux'], 97 | outputs=['representation']) 98 | def apply(self, char_seq, sample_matrix, char_aux): 99 | # Time as first dimension 100 | embeddings = self.lookup.apply(char_seq) 101 | gru_out = self.dgru.apply( 102 | **merge(self.gru_fork.apply(embeddings, as_dict=True), 103 | {'mask': char_aux})) 104 | if self.dgru_depth > 1: 105 | gru_out = gru_out[-1] 106 | sampled_representation = tensor.batched_dot(sample_matrix, gru_out.dimshuffle([1, 0, 2])) 107 | return sampled_representation.dimshuffle([1, 0, 2]) 108 | 109 | @application(inputs=['target_single_char']) 110 | def single_emit(self, target_single_char, batch_size, mask, states=None): 111 | # Time as first dimension 112 | # only one batch 113 | embeddings = self.lookup.apply(target_single_char) 114 | if states is None: 115 | states = self.dgru.initial_states(batch_size) 116 | states_dict = {'states':states[0]} 117 | for i in range(1,self.dgru_depth): 118 | states_dict['states'+RECURRENTSTACK_SEPARATOR+str(i)] = states[i] 119 | gru_out = self.dgru.apply(**merge(self.gru_fork.apply(embeddings, as_dict=True), states_dict, 120 | {'mask': mask, 'iterate': False})) 121 | return gru_out 122 | 123 | @single_emit.property('outputs') 124 | def single_emit_outputs(self): 125 | return ['gru_out' + RECURRENTSTACK_SEPARATOR + str(i) for i in range(self.dgru_depth)] 126 | 127 | def get_dim(self, name): 128 | if name in ['output', 'feedback']: 129 | return self.dgru_state_dim 130 | super(Decimator, self).get_dim(name) 131 | 132 | 133 | class RecurrentWithFork(Initializable): 134 | @lazy(allocation=['input_dim']) 135 | def __init__(self, proto, input_dim, **kwargs): 136 | super(RecurrentWithFork, self).__init__(**kwargs) 137 | self.recurrent = proto 138 | self.input_dim = input_dim 139 | self.fork = Fork( 140 | [name for name in self.recurrent.apply.sequences 141 | if name != 'mask'], 142 | prototype=Linear()) 143 | self.children = [self.recurrent, self.fork] 144 | 145 | def _push_allocation_config(self): 146 | self.fork.input_dim = self.input_dim 147 | self.fork.output_dims = [self.recurrent.get_dim(name) 148 | for name in self.fork.output_names] 149 | 150 | @application(inputs=['input_', 'mask']) 151 | def apply(self, input_, mask=None, **kwargs): 152 | return self.recurrent.apply( 153 | mask=mask, **dict_union(self.fork.apply(input_, as_dict=True), 154 | kwargs)) 155 | 156 | @apply.property('outputs') 157 | def apply_outputs(self): 158 | return self.recurrent.states 159 | 160 | 161 | class BidirectionalEncoder(Initializable): 162 | """Encoder of model.""" 163 | 164 | def __init__(self, src_vocab_size, embedding_dim, dgru_state_dim, state_dim, src_dgru_depth, 165 | bidir_encoder_depth, **kwargs): 166 | super(BidirectionalEncoder, self).__init__(**kwargs) 167 | self.state_dim = state_dim 168 | self.dgru_state_dim = dgru_state_dim 169 | self.decimator = Decimator(src_vocab_size, embedding_dim, dgru_state_dim, src_dgru_depth) 170 | self.bidir = Bidirectional( 171 | RecurrentWithFork(GatedRecurrent(activation=Tanh(), dim=state_dim), dgru_state_dim, name='with_fork'), 172 | name='bidir0') 173 | 174 | self.children = [self.decimator, self.bidir] 175 | for layer_n in range(1, bidir_encoder_depth): 176 | self.children.append(copy.deepcopy(self.bidir)) 177 | for child in self.children[-1].children: 178 | child.input_dim = 2 * state_dim 179 | self.children[-1].name = 'bidir{}'.format(layer_n) 180 | 181 | @application(inputs=['source_char_seq', 'source_sample_matrix', 'source_char_aux', 'source_word_mask'], 182 | outputs=['representation']) 183 | def apply(self, source_char_seq, source_sample_matrix, source_char_aux, source_word_mask): 184 | # Time as first dimension 185 | source_char_seq = source_char_seq.T 186 | source_char_aux = source_char_aux.T 187 | source_word_mask = source_word_mask.T 188 | source_word_representation = self.decimator.apply(source_char_seq, source_sample_matrix, source_char_aux) 189 | representation = source_word_representation 190 | for bidir in self.children[1:]: 191 | representation = bidir.apply(representation, source_word_mask) 192 | return representation 193 | 194 | 195 | class IGRU(GatedRecurrent): 196 | """IGRU in interpolator """ 197 | 198 | def __init__(self, dim, activation=None, gate_activation=None, 199 | **kwargs): 200 | super(IGRU, self).__init__(dim, activation, gate_activation, **kwargs) 201 | 202 | @recurrent(sequences=['mask', 'inputs', 'gate_inputs', 'input_states'], 203 | states=['states'], outputs=['states'], contexts=[]) 204 | def apply(self, inputs, gate_inputs, states, input_states, mask=None): 205 | """Apply the gated recurrent transition. 206 | Parameters 207 | ---------- 208 | states : :class:`~tensor.TensorVariable` 209 | The 2 dimensional matrix of current states in the shape 210 | (batch_size, dim). Required for `one_step` usage. 211 | inputs : :class:`~tensor.TensorVariable` 212 | The 2 dimensional matrix of inputs in the shape (batch_size, 213 | dim) 214 | gate_inputs : :class:`~tensor.TensorVariable` 215 | The 2 dimensional matrix of inputs to the gates in the 216 | shape (batch_size, 2 * dim). 217 | input_states : :class:`~tensor.TensorVariable` 218 | The 2 dimensional matrix of outputs of decoder in the shape 219 | (batch_size, dim), which generated by decoder. 220 | mask : :class:`~tensor.TensorVariable` 221 | A 1D binary array in the shape (batch,) which is 1 if there is 222 | the charater available, 0 if there is the delimiter. 223 | Returns 224 | ------- 225 | output : :class:`~tensor.TensorVariable` 226 | Next states of the network. 227 | """ 228 | # put masked states at last may be possible 229 | if mask: 230 | states = (mask[:, None] * states + (1 - mask[:, None]) * input_states) 231 | gate_values = self.gate_activation.apply( 232 | states.dot(self.state_to_gates) + gate_inputs) 233 | update_values = gate_values[:, :self.dim] 234 | reset_values = gate_values[:, self.dim:] 235 | #states_reset = (states + input_states) * reset_values / 2 236 | states_reset = states * reset_values 237 | next_states = self.activation.apply( 238 | states_reset.dot(self.state_to_state) + inputs) 239 | next_states = (next_states * update_values + 240 | states * (1 - update_values)) 241 | return next_states 242 | 243 | # using constant initial_states 244 | def _allocate(self): 245 | self.parameters.append(shared_floatx_nans((self.dim, self.dim), 246 | name='state_to_state')) 247 | self.parameters.append(shared_floatx_nans((self.dim, 2 * self.dim), 248 | name='state_to_gates')) 249 | for i in range(2): 250 | if self.parameters[i]: 251 | add_role(self.parameters[i], WEIGHT) 252 | 253 | @application(outputs=apply.states) 254 | def initial_states(self, batch_size, *args, **kwargs): 255 | return tensor.zeros((batch_size, self.dim)) 256 | 257 | 258 | class UpperIGRU(GatedRecurrent): 259 | """ Upper IGRU in interpolator """ 260 | 261 | def __init__(self, dim, activation=None, gate_activation=None, 262 | **kwargs): 263 | super(UpperIGRU, self).__init__(dim, activation, gate_activation, **kwargs) 264 | 265 | @recurrent(sequences=['mask', 'inputs', 'gate_inputs'], 266 | states=['states'], outputs=['states'], contexts=[]) 267 | def apply(self, inputs, gate_inputs, states, mask=None): 268 | """Apply the gated recurrent transition. 269 | Parameters 270 | ---------- 271 | states : :class:`~tensor.TensorVariable` 272 | The 2 dimensional matrix of current states in the shape 273 | (batch_size, dim). Required for `one_step` usage. 274 | inputs : :class:`~tensor.TensorVariable` 275 | The 2 dimensional matrix of inputs in the shape (batch_size, 276 | dim) 277 | gate_inputs : :class:`~tensor.TensorVariable` 278 | The 2 dimensional matrix of inputs to the gates in the 279 | shape (batch_size, 2 * dim). 280 | input_states : :class:`~tensor.TensorVariable` 281 | The 2 dimensional matrix of outputs of decoder in the shape 282 | (batch_size, dim), which generated by decoder. 283 | mask : :class:`~tensor.TensorVariable` 284 | A 1D binary array in the shape (batch,) which is 1 if there is 285 | the charater available, 0 if there is the delimiter. 286 | Returns 287 | ------- 288 | output : :class:`~tensor.TensorVariable` 289 | Next states of the network. 290 | """ 291 | if mask: 292 | states = (mask[:, None] * states + (1 - mask[:, None]) * self.initial_states(mask.shape[0])) 293 | gate_values = self.gate_activation.apply( 294 | states.dot(self.state_to_gates) + gate_inputs) 295 | update_values = gate_values[:, :self.dim] 296 | reset_values = gate_values[:, self.dim:] 297 | states_reset = states * reset_values 298 | next_states = self.activation.apply( 299 | states_reset.dot(self.state_to_state) + inputs) 300 | next_states = (next_states * update_values + 301 | states * (1 - update_values)) 302 | 303 | return next_states 304 | 305 | 306 | class Interpolator(Readout): 307 | def __init__(self, vocab_size, embedding_dim, igru_state_dim, igru_depth, trg_dgru_depth, emitter=None, feedback_brick=None, 308 | merge=None, merge_prototype=None, post_merge=None, merged_dim=None, igru=None, **kwargs): 309 | # for compatible 310 | if igru_depth == 1: 311 | self.igru = IGRU(dim=igru_state_dim) 312 | else: 313 | self.igru = RecurrentStack([IGRU(dim=igru_state_dim, name='igru')] + 314 | [UpperIGRU(dim=igru_state_dim, activation=Tanh(), name='upper_igru' + str(i)) 315 | for i in range(1, igru_depth)], 316 | skip_connections=True) 317 | self.igru_depth = igru_depth 318 | self.trg_dgru_depth = trg_dgru_depth 319 | self.lookup = LookupTable(name='embeddings') 320 | self.vocab_size = vocab_size 321 | self.igru_state_dim = igru_state_dim 322 | self.gru_to_softmax = Linear(input_dim=igru_state_dim, output_dim=vocab_size) 323 | self.embedding_dim = embedding_dim 324 | self.gru_fork = Fork([name for name in self.igru.apply.sequences 325 | if name != 'mask' and name != 'input_states'], prototype=Linear(), name='gru_fork') 326 | kwargs['children'] = [self.igru, self.lookup, self.gru_to_softmax, self.gru_fork] 327 | super(Interpolator, self).__init__(emitter=emitter, feedback_brick=feedback_brick, merge=merge, 328 | merge_prototype=merge_prototype, post_merge=post_merge, 329 | merged_dim=merged_dim, **kwargs) 330 | 331 | @application 332 | def initial_igru_outputs(self, batch_size): 333 | return self.igru.initial_states(batch_size) 334 | 335 | def _push_allocation_config(self): 336 | self.lookup.length = self.vocab_size 337 | self.lookup.dim = self.embedding_dim 338 | self.emitter.readout_dim = self.get_dim('readouts') 339 | self.merge.input_names = self.source_names 340 | self.merge.input_dims = self.source_dims 341 | self.merge.output_dim = self.merged_dim 342 | self.post_merge.input_dim = self.merged_dim 343 | self.post_merge.output_dim = self.igru_state_dim 344 | self.gru_fork.input_dim = self.embedding_dim 345 | self.gru_fork.output_dims = [self.igru.get_dim(name) 346 | for name in self.gru_fork.output_names] 347 | 348 | @application(outputs=['feedback']) 349 | def feedback_apply(self, target_char_seq, target_sample_matrix, target_char_aux): 350 | return self.feedback_brick.apply(target_char_seq, target_sample_matrix, target_char_aux) 351 | 352 | @application 353 | def single_feedback(self, target_single_char, batch_size, mask=None, states=None): 354 | return self.feedback_brick.single_emit(target_single_char, batch_size, mask, states) 355 | 356 | @single_feedback.property('outputs') 357 | def single_feedback_outputs(self): 358 | return ['single_feedback' + RECURRENTSTACK_SEPARATOR + str(i) for i in range(self.trg_dgru_depth)] 359 | 360 | @application(outputs=['gru_out', 'readout_chars']) 361 | def single_readout_gru(self, target_prev_char, target_prev_char_aux, input_states, states): 362 | embeddings = self.lookup.apply(target_prev_char) 363 | states_dict = {'states':states[0]} 364 | if self.igru_depth > 1: 365 | for i in range(1, self.igru_depth): 366 | states_dict['states' + RECURRENTSTACK_SEPARATOR +str(i)] = states[i] 367 | gru_out = self.igru.apply( 368 | **merge(self.gru_fork.apply(embeddings, as_dict=True), states_dict, 369 | {'mask': target_prev_char_aux, 'input_states': input_states, 370 | 'iterate': False})) 371 | if self.igru_depth > 1: 372 | readout_chars = self.gru_to_softmax.apply(gru_out[-1]) 373 | else: 374 | readout_chars = self.gru_to_softmax.apply(gru_out) 375 | return gru_out, readout_chars 376 | 377 | @application(outputs=['readout_chars']) 378 | def readout_gru(self, target_prev_char_seq, target_prev_char_aux, input_states): 379 | embeddings = self.lookup.apply(target_prev_char_seq) 380 | gru_out = self.igru.apply( 381 | **merge(self.gru_fork.apply(embeddings, as_dict=True), 382 | {'mask': target_prev_char_aux, 'input_states': input_states})) 383 | if self.igru_depth > 1: 384 | gru_out = gru_out[-1] 385 | readout_chars = self.gru_to_softmax.apply(gru_out) 386 | return readout_chars 387 | 388 | 389 | class SequenceGeneratorDCNMT(BaseSequenceGenerator): 390 | """A more user-friendly interface for :class:`BaseSequenceGenerator`. """ 391 | 392 | def __init__(self, trg_space_idx, readout, transition, attention=None, transition_depth=1, igru_depth=1, trg_dgru_depth=1, 393 | add_contexts=True, **kwargs): 394 | self.trg_space_idx = trg_space_idx 395 | self.transition_depth = transition_depth 396 | self.igru_depth = igru_depth 397 | self.trg_dgru_depth = trg_dgru_depth 398 | self.igru_states_name = ['igru_states'+RECURRENTSTACK_SEPARATOR+str(i) for i in range(self.igru_depth)] 399 | self.feedback_name = ['feedback'+RECURRENTSTACK_SEPARATOR+str(i) for i in range(self.trg_dgru_depth)] 400 | 401 | normal_inputs = [name for name in transition.apply.sequences 402 | if 'mask' not in name] 403 | kwargs.setdefault('fork', Fork(normal_inputs)) 404 | transition = AttentionRecurrent( 405 | transition, attention, 406 | add_contexts=add_contexts, name="att_trans") 407 | super(SequenceGeneratorDCNMT, self).__init__( 408 | readout, transition, **kwargs) 409 | 410 | @application 411 | def cost_matrix_nmt(self, application_call, target_char_seq, target_sample_matrix, target_resample_matrix, 412 | target_word_mask, target_char_aux, target_prev_char_seq, target_prev_char_aux, **kwargs): 413 | """Returns generation costs for output sequences. 414 | 415 | See Also 416 | -------- 417 | :meth:`cost` : Scalar cost. 418 | 419 | """ 420 | # We assume the data has axes (time, batch, features, ...) 421 | batch_size = target_char_seq.shape[1] 422 | 423 | # Prepare input for the iterative part 424 | states = dict_subset(kwargs, self._state_names, must_have=False) 425 | # masks in context are optional (e.g. `attended_mask`) 426 | contexts = dict_subset(kwargs, self._context_names, must_have=False) 427 | feedback = self.readout.feedback_apply(target_char_seq, target_sample_matrix, target_char_aux) 428 | inputs = self.fork.apply(feedback, as_dict=True) 429 | 430 | # Run the recurrent network 431 | results = self.transition.apply( 432 | mask=target_word_mask, return_initial_states=True, as_dict=True, 433 | **dict_union(inputs, states, contexts)) 434 | 435 | # Separate the deliverables. The last states are discarded: they 436 | # are not used to predict any output symbol. The initial glimpses 437 | # are discarded because they are not used for prediction. 438 | # Remember, glimpses are computed _before_ output stage, states are 439 | # computed after. 440 | states = {name: results[name][:-1] for name in self._state_names} 441 | glimpses = {name: results[name][1:] for name in self._glimpse_names} 442 | 443 | feedback = tensor.roll(feedback, 1, 0) 444 | if self.trg_dgru_depth == 1: 445 | feedback = tensor.set_subtensor( 446 | feedback[0], 447 | self.readout.single_feedback(self.readout.initial_outputs(batch_size), batch_size)) 448 | else: 449 | feedback = tensor.set_subtensor( 450 | feedback[0], 451 | self.readout.single_feedback(self.readout.initial_outputs(batch_size), batch_size)[-1]) 452 | 453 | decoder_readout_outputs = self.readout.readout( 454 | feedback=feedback, **dict_union(states, glimpses, contexts)) 455 | resampled_representation = tensor.batched_dot(target_resample_matrix, 456 | decoder_readout_outputs.dimshuffle([1, 0, 2])) 457 | resampled_readouts = resampled_representation.dimshuffle([1, 0, 2]) 458 | readouts_chars = self.readout.readout_gru(target_prev_char_seq, target_prev_char_aux, resampled_readouts) 459 | 460 | # Compute the cost 461 | costs = self.readout.cost(readouts_chars, target_char_seq) 462 | 463 | for name, variable in list(glimpses.items()) + list(states.items()): 464 | application_call.add_auxiliary_variable( 465 | variable.copy(), name=name) 466 | 467 | # This variables can be used to initialize the initial states of the 468 | # next batch using the last states of the current batch. 469 | for name in self._state_names + self._glimpse_names: 470 | application_call.add_auxiliary_variable( 471 | results[name][-1].copy(), name=name + "_final_value") 472 | 473 | return costs 474 | 475 | @recurrent 476 | def generate(self, outputs, **kwargs): 477 | """A sequence generation step. 478 | 479 | Parameters 480 | ---------- 481 | outputs : :class:`~tensor.TensorVariable` 482 | The outputs from the previous step. 483 | 484 | Notes 485 | ----- 486 | The contexts, previous states and glimpses are expected as keyword 487 | arguments. 488 | 489 | """ 490 | states = dict_subset(kwargs, self._state_names) 491 | # masks in context are optional (e.g. `attended_mask`) 492 | contexts = dict_subset(kwargs, self._context_names, must_have=False) 493 | glimpses = dict_subset(kwargs, self._glimpse_names) 494 | feedback = dict_subset(kwargs, self.feedback_name) 495 | readout_feedback = dict_subset(kwargs, ['readout_feedback'])['readout_feedback'] 496 | batch_size = outputs.shape[0] 497 | igru_states = dict_subset(kwargs, self.igru_states_name) 498 | 499 | next_glimpses = self.transition.take_glimpses( 500 | as_dict=True, **dict_union(states, glimpses, contexts)) 501 | 502 | next_readouts = self.readout.readout( 503 | feedback=readout_feedback, 504 | **dict_union(states, next_glimpses, contexts)) 505 | 506 | next_char_aux = 1 - tensor.eq(outputs, 0) - tensor.eq(outputs, self.trg_space_idx) 507 | next_igru_states, readout_chars = self.readout.single_readout_gru(outputs, next_char_aux, next_readouts, 508 | [igru_states[self.igru_states_name[i]] for i in range(self.igru_depth)]) 509 | next_outputs = self.readout.emit(readout_chars) 510 | next_costs = self.readout.cost(readout_chars, next_outputs) 511 | 512 | update_next = tensor.eq(next_outputs, self.trg_space_idx) 513 | next_char_mask = 1 - update_next 514 | update_next = update_next[:, None] 515 | next_readout_feedback = (1 - update_next) * readout_feedback + update_next * feedback[self.feedback_name[-1]] 516 | 517 | next_feedback = self.readout.single_feedback(next_outputs, batch_size, next_char_mask, 518 | [feedback[self.feedback_name[i]] for i in range(self.trg_dgru_depth)]) 519 | 520 | next_inputs = (self.fork.apply(next_readout_feedback, as_dict=True) 521 | if self.fork else {'feedback': next_readout_feedback}) 522 | next_states = self.transition.compute_states( 523 | as_list=True, 524 | **dict_union(next_inputs, states, next_glimpses, contexts)) 525 | 526 | next_states[0] = update_next * next_states[0] + (1 - update_next) * states['states'] 527 | for i in range(1, self.transition_depth): 528 | next_states[i] = update_next * next_states[i] + (1 - update_next) * states['states' + RECURRENTSTACK_SEPARATOR + str(i)] 529 | 530 | next_glimpses['weights'] = update_next * next_glimpses['weights'] + (1 - update_next) * glimpses['weights'] 531 | next_glimpses['weighted_averages'] = update_next * next_glimpses['weighted_averages'] + \ 532 | (1 - update_next) * glimpses['weighted_averages'] 533 | # combine all updates 534 | next_all = list(next_states) + [next_outputs] + list(next_glimpses.values()) 535 | if self.trg_dgru_depth > 1: 536 | next_all += list(next_feedback) 537 | else: 538 | next_all += [next_feedback] 539 | if self.igru_depth > 1: 540 | next_all += list(next_igru_states) 541 | else: 542 | next_all += [next_igru_states] 543 | next_all += [next_readout_feedback] + [next_costs] 544 | 545 | return (next_all) 546 | 547 | @generate.delegate 548 | def generate_delegate(self): 549 | return self.transition.apply 550 | 551 | @generate.property('outputs') 552 | def generate_outputs(self): 553 | return self._state_names + ['outputs'] + self._glimpse_names + \ 554 | self.feedback_name + self.igru_states_name + ['readout_feedback', 'cost'] 555 | 556 | @generate.property('states') 557 | def generate_states(self): 558 | return self._state_names + ['outputs'] + self._glimpse_names + \ 559 | self.feedback_name + self.igru_states_name + ['readout_feedback'] 560 | 561 | def get_dim(self, name): 562 | if name in (self._state_names + self._context_names + self._glimpse_names): 563 | return self.transition.get_dim(name) 564 | elif name == 'outputs' or name in self.feedback_name or name == 'readout_feedback' or name in self.igru_states_name: 565 | return self.readout.get_dim('outputs') 566 | return super(BaseSequenceGenerator, self).get_dim(name) 567 | 568 | @application 569 | def initial_states(self, batch_size, *args, **kwargs): 570 | # TODO: support dict of outputs for application methods 571 | # to simplify this code. 572 | igru_initial_states = self.readout.initial_igru_outputs(batch_size) 573 | if self.igru_depth == 1: 574 | igru_initial_states_dict = {self.igru_states_name[0]:igru_initial_states} 575 | else: 576 | igru_initial_states_dict = {self.igru_states_name[i]:igru_initial_states[i] 577 | for i in range(self.igru_depth)} 578 | 579 | initial_outputs=self.readout.initial_outputs(batch_size) 580 | feedback = self.readout.single_feedback(initial_outputs, batch_size) 581 | if self.trg_dgru_depth == 1: 582 | feedback_dict = {self.feedback_name[0]:feedback, 'readout_feedback':feedback} 583 | else: 584 | feedback_dict = {'readout_feedback':feedback[-1]} 585 | feedback_dict.update({self.feedback_name[i]:feedback[i] for i in range(self.trg_dgru_depth)}) 586 | 587 | state_dict = dict( 588 | self.transition.initial_states( 589 | batch_size, as_dict=True, *args, **kwargs), 590 | outputs=initial_outputs) 591 | state_dict.update(feedback_dict) 592 | state_dict.update(igru_initial_states_dict) 593 | return [state_dict[state_name] 594 | for state_name in self.generate.states] 595 | 596 | @initial_states.property('outputs') 597 | def initial_states_outputs(self): 598 | return self.generate.states 599 | 600 | 601 | class GRUInitialState(GatedRecurrent): 602 | """Gated Recurrent with special initial state. 603 | 604 | Initial state of Gated Recurrent is set by an MLP that conditions on the 605 | first hidden state of the bidirectional encoder, applies an affine 606 | transformation followed by a tanh non-linearity to set initial state. 607 | 608 | """ 609 | 610 | def __init__(self, attended_dim, **kwargs): 611 | super(GRUInitialState, self).__init__(**kwargs) 612 | self.attended_dim = attended_dim 613 | self.initial_transformer = MLP(activations=[Tanh()], 614 | dims=[attended_dim, self.dim], 615 | name='state_initializer') 616 | self.children.append(self.initial_transformer) 617 | 618 | @application 619 | def initial_states(self, batch_size, *args, **kwargs): 620 | attended = kwargs['attended'] 621 | initial_state = self.initial_transformer.apply( 622 | attended[0, :, -self.attended_dim:]) 623 | return initial_state 624 | 625 | def _allocate(self): 626 | self.parameters.append(shared_floatx_nans((self.dim, self.dim), 627 | name='state_to_state')) 628 | self.parameters.append(shared_floatx_nans((self.dim, 2 * self.dim), 629 | name='state_to_gates')) 630 | for i in range(2): 631 | if self.parameters[i]: 632 | add_role(self.parameters[i], WEIGHT) 633 | 634 | 635 | class Decoder(Initializable): 636 | """Decoder of dcnmt model.""" 637 | 638 | def __init__(self, vocab_size, embedding_dim, dgru_state_dim, igru_state_dim, state_dim, 639 | representation_dim, transition_depth, trg_igru_depth, trg_dgru_depth, trg_space_idx, trg_bos, 640 | theano_seed=None, **kwargs): 641 | super(Decoder, self).__init__(**kwargs) 642 | self.vocab_size = vocab_size 643 | self.embedding_dim = embedding_dim 644 | self.dgru_state_dim = dgru_state_dim 645 | self.igru_state_dim = igru_state_dim 646 | self.state_dim = state_dim 647 | self.trg_space_idx = trg_space_idx 648 | self.representation_dim = representation_dim 649 | self.theano_seed = theano_seed 650 | 651 | # Initialize gru with special initial state 652 | self.transition = RecurrentStack( 653 | [GRUInitialState(attended_dim=state_dim, dim=state_dim, activation=Tanh(), name='decoder_gru_withinit')] + 654 | [GatedRecurrent(dim=state_dim, activation=Tanh(), name='decoder_gru' + str(i)) 655 | for i in range(1, transition_depth)], skip_connections=True) 656 | 657 | # Initialize the attention mechanism 658 | self.attention = SequenceContentAttention( 659 | state_names=self.transition.apply.states, 660 | attended_dim=representation_dim, 661 | match_dim=state_dim, name="attention") 662 | 663 | self.interpolator = Interpolator( 664 | vocab_size=vocab_size, 665 | embedding_dim=embedding_dim, 666 | igru_state_dim=igru_state_dim, 667 | igru_depth=trg_igru_depth, 668 | trg_dgru_depth=trg_dgru_depth, 669 | source_names=['states', 'feedback', self.attention.take_glimpses.outputs[0]], 670 | readout_dim=self.vocab_size, 671 | emitter=SoftmaxEmitter(initial_output=trg_bos, theano_seed=theano_seed), 672 | feedback_brick=Decimator(vocab_size, embedding_dim, self.dgru_state_dim, trg_dgru_depth), 673 | post_merge=InitializableFeedforwardSequence([Identity().apply]), 674 | merged_dim=igru_state_dim) 675 | 676 | # Build sequence generator accordingly 677 | self.sequence_generator = SequenceGeneratorDCNMT( 678 | trg_space_idx=self.trg_space_idx, 679 | readout=self.interpolator, 680 | transition=self.transition, 681 | attention=self.attention, 682 | transition_depth=transition_depth, 683 | igru_depth=trg_igru_depth, 684 | trg_dgru_depth=trg_dgru_depth, 685 | fork=Fork([name for name in self.transition.apply.sequences 686 | if name != 'mask'], prototype=Linear()) 687 | ) 688 | self.children = [self.sequence_generator] 689 | 690 | @application(inputs=['representation', 'source_word_mask', 'target_char_seq', 'target_sample_matrix', 691 | 'target_resample_matrix', 'target_char_aux', 'target_char_mask', 'target_word_mask', 692 | 'target_prev_char_seq', 'target_prev_char_aux'], 693 | outputs=['cost']) 694 | def cost(self, representation, source_word_mask, target_char_seq, target_sample_matrix, 695 | target_resample_matrix, target_char_aux, target_char_mask, target_word_mask, 696 | target_prev_char_seq, target_prev_char_aux): 697 | source_word_mask = source_word_mask.T 698 | target_char_seq = target_char_seq.T 699 | target_prev_char_seq = target_prev_char_seq.T 700 | target_char_mask = target_char_mask.T 701 | target_char_aux = target_char_aux.T 702 | target_prev_char_aux = target_prev_char_aux.T 703 | target_word_mask = target_word_mask.T 704 | 705 | # Get the cost matrix 706 | cost = self.sequence_generator.cost_matrix_nmt(**{ 707 | 'target_char_seq': target_char_seq, 708 | 'target_sample_matrix': target_sample_matrix, 709 | 'target_resample_matrix': target_resample_matrix, 710 | 'target_word_mask': target_word_mask, 711 | 'target_char_aux': target_char_aux, 712 | 'target_prev_char_seq': target_prev_char_seq, 713 | 'target_prev_char_aux': target_prev_char_aux, 714 | 'attended': representation, 715 | 'attended_mask': source_word_mask}) 716 | 717 | return (cost * target_char_mask).sum() / target_char_mask.shape[1] 718 | 719 | @application 720 | def generate(self, representation, attended_mask, **kwargs): 721 | return self.sequence_generator.generate( 722 | n_steps=10 * attended_mask.shape[1], 723 | batch_size=attended_mask.shape[0], 724 | attended=representation, 725 | attended_mask=attended_mask.T, 726 | **kwargs) 727 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | {one line to give the program's name and a brief idea of what it does.} 635 | Copyright (C) {year} {name of author} 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | {project} Copyright (C) {year} {fullname} 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | --------------------------------------------------------------------------------