update pretraining

Author: frankxu2004

datasets/conala/dataset.py

@@ -4,6 +4,7 @@
 import pickle
 
 from components.action_info import get_action_infos
+from datasets.conala.evaluator import ConalaEvaluator
 from datasets.conala.util import *
 from asdl.lang.py3.py3_transition_system import python_ast_to_asdl_ast, asdl_ast_to_python_ast, Python3TransitionSystem
 
@@ -16,7 +17,8 @@
 from components.action_info import ActionInfo
 
 
-def preprocess_conala_dataset(train_file, test_file, grammar_file, src_freq=3, code_freq=3):
+def preprocess_conala_dataset(train_file, test_file, grammar_file, src_freq=3, code_freq=3,
+                              mined_data_file=None, num_mined=0):
     np.random.seed(1234)
 
     asdl_text = open(grammar_file).read()
@@ -32,19 +34,12 @@ def preprocess_conala_dataset(train_file, test_file, grammar_file, src_freq=3, c
     dev_examples = train_examples[:200]
     train_examples = train_examples[200:]
 
-    # full_train_examples = train_examples[:]
-    # np.random.shuffle(train_examples)
-    # dev_examples = []
-    # dev_questions = set()
-    # dev_examples_id = []
-    # for i, example in enumerate(full_train_examples):
-    #     qid = example.meta['example_dict']['question_id']
-    #     if qid not in dev_questions and len(dev_examples) < 200:
-    #         dev_questions.add(qid)
-    #         dev_examples.append(example)
-    #         dev_examples_id.append(i)
-
-    # train_examples = [e for i, e in enumerate(full_train_examples) if i not in dev_examples_id]
+    if mined_data_file and num_mined > 0:
+        print("use mined data: ", num_mined)
+        mined_examples = preprocess_dataset(mined_data_file, name='mined', transition_system=transition_system,
+                                            firstk=num_mined)
+        train_examples += mined_examples
+
     print(f'{len(train_examples)} training instances', file=sys.stderr)
     print(f'{len(dev_examples)} dev instances', file=sys.stderr)
 
@@ -71,58 +66,65 @@ def preprocess_conala_dataset(train_file, test_file, grammar_file, src_freq=3, c
     print('Avg action len: %d' % np.average(action_lens), file=sys.stderr)
     print('Actions larger than 100: %d' % len(list(filter(lambda x: x > 100, action_lens))), file=sys.stderr)
 
-    pickle.dump(train_examples, open('data/conala/train.var_str_sep.bin', 'wb'))
-    pickle.dump(full_train_examples, open('data/conala/train.var_str_sep.full.bin', 'wb'))
-    pickle.dump(dev_examples, open('data/conala/dev.var_str_sep.bin', 'wb'))
-    pickle.dump(test_examples, open('data/conala/test.var_str_sep.bin', 'wb'))
-    pickle.dump(vocab, open('data/conala/vocab.var_str_sep.new_dev.src_freq%d.code_freq%d.bin' % (src_freq, code_freq), 'wb'))
+    pickle.dump(train_examples, open('data/conala/train.var_str_sep.mined_{}.bin'.format(num_mined), 'wb'))
+    pickle.dump(full_train_examples, open('data/conala/train.var_str_sep.full.mined_{}.bin'.format(num_mined), 'wb'))
+    pickle.dump(dev_examples, open('data/conala/dev.var_str_sep.mined_{}.bin'.format(num_mined), 'wb'))
+    pickle.dump(test_examples, open('data/conala/test.var_str_sep.mined_{}.bin'.format(num_mined), 'wb'))
+    pickle.dump(vocab, open('data/conala/vocab.var_str_sep.new_dev.src_freq%d.code_freq%d.mined_%s.bin' % (src_freq, code_freq, num_mined), 'wb'))
 
 
-def preprocess_dataset(file_path, transition_system, name='train'):
-    dataset = json.load(open(file_path))
+def preprocess_dataset(file_path, transition_system, name='train', firstk=None):
+    try:
+        dataset = json.load(open(file_path))
+    except:
+        dataset = [json.loads(jline) for jline in open(file_path).readlines()]
     examples = []
     evaluator = ConalaEvaluator(transition_system)
 
     f = open(file_path + '.debug', 'w')
 
     for i, example_json in enumerate(dataset):
-        example_dict = preprocess_example(example_json)
-        if example_json['question_id'] in (18351951, 9497290, 19641579, 32283692):
-            print(example_json['question_id'])
+        if firstk and i >= firstk:
+            break
+        try:
+            example_dict = preprocess_example(example_json)
+            if example_json['question_id'] in (18351951, 9497290, 19641579, 32283692):
+                print(example_json['question_id'])
+                continue
+
+            python_ast = ast.parse(example_dict['canonical_snippet'])
+            canonical_code = astor.to_source(python_ast).strip()
+            tgt_ast = python_ast_to_asdl_ast(python_ast, transition_system.grammar)
+            tgt_actions = transition_system.get_actions(tgt_ast)
+
+            # sanity check
+            hyp = Hypothesis()
+            for t, action in enumerate(tgt_actions):
+                assert action.__class__ in transition_system.get_valid_continuation_types(hyp)
+                if isinstance(action, ApplyRuleAction):
+                    assert action.production in transition_system.get_valid_continuating_productions(hyp)
+
+                p_t = -1
+                f_t = None
+                if hyp.frontier_node:
+                    p_t = hyp.frontier_node.created_time
+                    f_t = hyp.frontier_field.field.__repr__(plain=True)
+
+                # print('\t[%d] %s, frontier field: %s, parent: %d' % (t, action, f_t, p_t))
+                hyp = hyp.clone_and_apply_action(action)
+
+            assert hyp.frontier_node is None and hyp.frontier_field is None
+            hyp.code = code_from_hyp = astor.to_source(asdl_ast_to_python_ast(hyp.tree, transition_system.grammar)).strip()
+            assert code_from_hyp == canonical_code
+
+            decanonicalized_code_from_hyp = decanonicalize_code(code_from_hyp, example_dict['slot_map'])
+            assert compare_ast(ast.parse(example_json['snippet']), ast.parse(decanonicalized_code_from_hyp))
+            assert transition_system.compare_ast(transition_system.surface_code_to_ast(decanonicalized_code_from_hyp),
+                                                 transition_system.surface_code_to_ast(example_json['snippet']))
+
+            tgt_action_infos = get_action_infos(example_dict['intent_tokens'], tgt_actions)
+        except:
             continue
-
-        python_ast = ast.parse(example_dict['canonical_snippet'])
-        canonical_code = astor.to_source(python_ast).strip()
-        tgt_ast = python_ast_to_asdl_ast(python_ast, transition_system.grammar)
-        tgt_actions = transition_system.get_actions(tgt_ast)
-
-        # sanity check
-        hyp = Hypothesis()
-        for t, action in enumerate(tgt_actions):
-            assert action.__class__ in transition_system.get_valid_continuation_types(hyp)
-            if isinstance(action, ApplyRuleAction):
-                assert action.production in transition_system.get_valid_continuating_productions(hyp)
-
-            p_t = -1
-            f_t = None
-            if hyp.frontier_node:
-                p_t = hyp.frontier_node.created_time
-                f_t = hyp.frontier_field.field.__repr__(plain=True)
-
-            # print('\t[%d] %s, frontier field: %s, parent: %d' % (t, action, f_t, p_t))
-            hyp = hyp.clone_and_apply_action(action)
-
-        assert hyp.frontier_node is None and hyp.frontier_field is None
-        hyp.code = code_from_hyp = astor.to_source(asdl_ast_to_python_ast(hyp.tree, transition_system.grammar)).strip()
-        assert code_from_hyp == canonical_code
-
-        decanonicalized_code_from_hyp = decanonicalize_code(code_from_hyp, example_dict['slot_map'])
-        assert compare_ast(ast.parse(example_json['snippet']), ast.parse(decanonicalized_code_from_hyp))
-        assert transition_system.compare_ast(transition_system.surface_code_to_ast(decanonicalized_code_from_hyp),
-                                             transition_system.surface_code_to_ast(example_json['snippet']))
-
-        tgt_action_infos = get_action_infos(example_dict['intent_tokens'], tgt_actions)
-
         example = Example(idx=f'{i}-{example_json["question_id"]}',
                           src_sent=example_dict['intent_tokens'],
                           tgt_actions=tgt_action_infos,
@@ -136,7 +138,10 @@ def preprocess_dataset(file_path, transition_system, name='train'):
 
         # log!
         f.write(f'Example: {example.idx}\n')
-        f.write(f"Original Utterance: {example.meta['example_dict']['rewritten_intent']}\n")
+        if 'rewritten_intent' in example.meta['example_dict']:
+            f.write(f"Original Utterance: {example.meta['example_dict']['rewritten_intent']}\n")
+        else:
+            f.write(f"Original Utterance: {example.meta['example_dict']['intent']}\n")
         f.write(f"Original Snippet: {example.meta['example_dict']['snippet']}\n")
         f.write(f"\n")
         f.write(f"Utterance: {' '.join(example.src_sent)}\n")
@@ -150,9 +155,11 @@ def preprocess_dataset(file_path, transition_system, name='train'):
 
 def preprocess_example(example_json):
     intent = example_json['intent']
-    rewritten_intent = example_json['rewritten_intent']
+    if 'rewritten_intent' in example_json:
+        rewritten_intent = example_json['rewritten_intent']
+    else:
+        rewritten_intent = None
     snippet = example_json['snippet']
-    question_id = example_json['question_id']
 
     if rewritten_intent is None:
         rewritten_intent = intent
@@ -190,8 +197,11 @@ def generate_vocab_for_paraphrase_model(vocab_path, save_path):
 
 if __name__ == '__main__':
     # the json files can be download from http://conala-corpus.github.io
-    preprocess_conala_dataset(train_file='data/conala/conala-train.json',
+    for num in (10000, 20000):
+        preprocess_conala_dataset(train_file='data/conala/conala-train.json',
                               test_file='data/conala/conala-test.json',
-                              grammar_file='asdl/lang/py3/py3_asdl.simplified.txt', src_freq=3, code_freq=3)
+                              mined_data_file='data/conala/conala-mined.jsonl',
+                              grammar_file='asdl/lang/py3/py3_asdl.simplified.txt',
+                              src_freq=3, code_freq=3, num_mined=num)
 
     # generate_vocab_for_paraphrase_model('data/conala/vocab.src_freq3.code_freq3.bin', 'data/conala/vocab.para.src_freq3.code_freq3.bin')

datasets/conala/evaluator.py

@@ -1,3 +1,5 @@
+import csv
+
 from components.evaluator import Evaluator
 from common.registerable import Registrable
 from components.dataset import Dataset
@@ -32,7 +34,10 @@ def get_sentence_bleu(self, example, hyp):
                              tokenize_for_bleu_eval(hyp.decanonical_code),
                              smoothing_function=SmoothingFunction().method3)
 
-    def evaluate_dataset(self, dataset, decode_results, fast_mode=False):
+
+    def evaluate_dataset(self, dataset, decode_results, fast_mode=False, args=None):
+        if args.save_decode_to:
+            csv_writer = csv.writer(open(args.save_decode_to + '.csv', 'w'))
         examples = dataset.examples if isinstance(dataset, Dataset) else dataset
         assert len(examples) == len(decode_results)
 
@@ -88,7 +93,11 @@ def evaluate_dataset(self, dataset, decode_results, fast_mode=False):
 
                     top_decanonical_code_tokens = hyp_list[0].decanonical_code_tokens
                     sent_bleu_score = hyp_list[0].bleu_score
-
+                    # write results to file
+                    if args.save_decode_to:
+                        csv_writer.writerow([" ".join(example.src_sent),
+                                             " ".join(example.reference_code_tokens),
+                                             " ".join(top_decanonical_code_tokens)])
                     best_hyp_idx = np.argmax(example_hyp_bleu_scores)
                     oracle_sent_bleu = example_hyp_bleu_scores[best_hyp_idx]
                     _best_hyp_code_tokens = hyp_list[best_hyp_idx].decanonical_code_tokens

evaluation.py

@@ -58,7 +58,7 @@ def decode(examples, model, args, verbose=False, **kwargs):
 def evaluate(examples, parser, evaluator, args, verbose=False, return_decode_result=False, eval_top_pred_only=False):
     decode_results = decode(examples, parser, args, verbose=verbose)
 
-    eval_result = evaluator.evaluate_dataset(examples, decode_results, fast_mode=eval_top_pred_only)
+    eval_result = evaluator.evaluate_dataset(examples, decode_results, fast_mode=eval_top_pred_only, args=args)
 
     if return_decode_result:
         return eval_result, decode_results

exp.py

@@ -145,6 +145,7 @@ def train(args):
             model.save(model_file)
 
         # perform validation
+        is_better = False
         if args.dev_file:
             if epoch % args.valid_every_epoch == 0:
                 print('[Epoch %d] begin validation' % epoch, file=sys.stderr)

model/nn_utils.py

@@ -87,8 +87,11 @@ def to_input_variable(sequences, vocab, cuda=False, training=True, append_bounda
 
     word_ids = word2id(sequences, vocab)
     sents_t = input_transpose(word_ids, vocab['<pad>'])
-
-    sents_var = Variable(torch.LongTensor(sents_t), volatile=(not training), requires_grad=False)
+    if training:
+        sents_var = Variable(torch.LongTensor(sents_t), requires_grad=False)
+    else:
+        with torch.no_grad():
+            sents_var = Variable(torch.LongTensor(sents_t), requires_grad=False)
     if cuda:
         sents_var = sents_var.cuda()
 

model/parser.py

@@ -503,7 +503,8 @@ def parse(self, src_sent, context=None, beam_size=5, debug=False):
 
         zero_action_embed = Variable(self.new_tensor(args.action_embed_size).zero_())
 
-        hyp_scores = Variable(self.new_tensor([0.]), volatile=True)
+        with torch.no_grad():
+            hyp_scores = Variable(self.new_tensor([0.]))
 
         # For computing copy probabilities, we marginalize over tokens with the same surface form
         # `aggregated_primitive_tokens` stores the position of occurrence of each source token
@@ -525,7 +526,8 @@ def parse(self, src_sent, context=None, beam_size=5, debug=False):
             exp_src_encodings_att_linear = src_encodings_att_linear.expand(hyp_num, src_encodings_att_linear.size(1), src_encodings_att_linear.size(2))
 
             if t == 0:
-                x = Variable(self.new_tensor(1, self.decoder_lstm.input_size).zero_(), volatile=True)
+                with torch.no_grad():
+                    x = Variable(self.new_tensor(1, self.decoder_lstm.input_size).zero_())
                 if args.no_parent_field_type_embed is False:
                     offset = args.action_embed_size  # prev_action
                     offset += args.att_vec_size * (not args.no_input_feed)

scripts/atis/train.sh

@@ -1,10 +1,11 @@
 #!/bin/bash
 set -e
 
-seed=${1:-0}
-vocab="data/conala/vocab.var_str_sep.src_freq3.code_freq3.bin"
-train_file="data/conala/train.var_str_sep.bin"
-dev_file="data/conala/dev.var_str_sep.bin"
+seed=0
+mined_num=$1
+vocab="data/conala/vocab.var_str_sep.new_dev.src_freq3.code_freq3.mined_${mined_num}.bin"
+train_file="data/conala/train.var_str_sep.mined_${mined_num}.bin"
+dev_file="data/conala/dev.var_str_sep.mined_${mined_num}.bin"
 dropout=0.3
 hidden_size=256
 embed_size=128
@@ -17,7 +18,7 @@ lr_decay=0.5
 beam_size=15
 lstm='lstm'  # lstm
 lr_decay_after_epoch=15
-model_name=model.sup.conala.${lstm}.hidden${hidden_size}.embed${embed_size}.action${action_embed_size}.field${field_embed_size}.type${type_embed_size}.dr${dropout}.lr${lr}.lr_de${lr_decay}.lr_da${lr_decay_after_epoch}.beam${beam_size}.$(basename ${vocab}).$(basename ${train_file}).glorot.par_state.seed${seed}
+model_name=model.sup.conala.${lstm}.hidden${hidden_size}.embed${embed_size}.action${action_embed_size}.field${field_embed_size}.type${type_embed_size}.dr${dropout}.lr${lr}.lr_de${lr_decay}.lr_da${lr_decay_after_epoch}.beam${beam_size}.$(basename ${vocab}).$(basename ${train_file}).glorot.par_state.seed${seed}.mined_${mined_num}
 
 echo "**** Writing results to logs/conala/${model_name}.log ****"
 mkdir -p logs/conala