def main(_): tf.logging.set_verbosity(tf.logging.INFO) processors = { "cola": classifier_utils.ColaProcessor, "mnli": classifier_utils.MnliProcessor, "mismnli": classifier_utils.MisMnliProcessor, "mrpc": classifier_utils.MrpcProcessor, "rte": classifier_utils.RteProcessor, "sst-2": classifier_utils.Sst2Processor, "sts-b": classifier_utils.StsbProcessor, "qqp": classifier_utils.QqpProcessor, "qnli": classifier_utils.QnliProcessor, "wnli": classifier_utils.WnliProcessor, } tokenization.validate_case_matches_checkpoint(FLAGS.do_lower_case, FLAGS.init_checkpoint) if not FLAGS.do_train and not FLAGS.do_eval and not FLAGS.do_predict: raise ValueError( "At least one of `do_train`, `do_eval` or `do_predict' must be True." ) if not FLAGS.albert_config_file and not FLAGS.albert_hub_module_handle: raise ValueError("At least one of `--albert_config_file` and " "`--albert_hub_module_handle` must be set") if FLAGS.albert_config_file: albert_config = modeling.AlbertConfig.from_json_file( FLAGS.albert_config_file) if FLAGS.max_seq_length > albert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length %d because the ALBERT model " "was only trained up to sequence length %d" % (FLAGS.max_seq_length, albert_config.max_position_embeddings)) else: albert_config = None # Get the config from TF-Hub. tf.gfile.MakeDirs(FLAGS.output_dir) task_name = FLAGS.task_name.lower() if task_name not in processors: raise ValueError("Task not found: %s" % (task_name)) processor = processors[task_name]( use_spm=True if FLAGS.spm_model_file else False, do_lower_case=FLAGS.do_lower_case) label_list = processor.get_labels() tokenizer = fine_tuning_utils.create_vocab( vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case, spm_model_file=FLAGS.spm_model_file, hub_module=FLAGS.albert_hub_module_handle) tpu_cluster_resolver = None if FLAGS.use_tpu and FLAGS.tpu_name: tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver( FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2 if FLAGS.do_train: iterations_per_loop = int( min(FLAGS.iterations_per_loop, FLAGS.save_checkpoints_steps)) else: iterations_per_loop = FLAGS.iterations_per_loop run_config = contrib_tpu.RunConfig( cluster=tpu_cluster_resolver, master=FLAGS.master, model_dir=FLAGS.output_dir, save_checkpoints_steps=int(FLAGS.save_checkpoints_steps), keep_checkpoint_max=0, tpu_config=contrib_tpu.TPUConfig( iterations_per_loop=iterations_per_loop, num_shards=FLAGS.num_tpu_cores, per_host_input_for_training=is_per_host)) train_examples = None if FLAGS.do_train: train_examples = processor.get_train_examples(FLAGS.data_dir) model_fn = classifier_utils.model_fn_builder( albert_config=albert_config, num_labels=len(label_list), init_checkpoint=FLAGS.init_checkpoint, learning_rate=FLAGS.learning_rate, num_train_steps=FLAGS.train_step, num_warmup_steps=FLAGS.warmup_step, use_tpu=FLAGS.use_tpu, use_one_hot_embeddings=FLAGS.use_tpu, task_name=task_name, hub_module=FLAGS.albert_hub_module_handle, optimizer=FLAGS.optimizer) # If TPU is not available, this will fall back to normal Estimator on CPU # or GPU. estimator = contrib_tpu.TPUEstimator( use_tpu=FLAGS.use_tpu, model_fn=model_fn, config=run_config, train_batch_size=FLAGS.train_batch_size, eval_batch_size=FLAGS.eval_batch_size, predict_batch_size=FLAGS.predict_batch_size) if FLAGS.do_train: cached_dir = FLAGS.cached_dir if not cached_dir: cached_dir = FLAGS.output_dir train_file = os.path.join(cached_dir, task_name + "_train.tf_record") if not tf.gfile.Exists(train_file): classifier_utils.file_based_convert_examples_to_features( train_examples, label_list, FLAGS.max_seq_length, tokenizer, train_file, task_name) tf.logging.info("***** Running training *****") tf.logging.info(" Num examples = %d", len(train_examples)) tf.logging.info(" Batch size = %d", FLAGS.train_batch_size) tf.logging.info(" Num steps = %d", FLAGS.train_step) train_input_fn = classifier_utils.file_based_input_fn_builder( input_file=train_file, seq_length=FLAGS.max_seq_length, is_training=True, drop_remainder=True, task_name=task_name, use_tpu=FLAGS.use_tpu, bsz=FLAGS.train_batch_size) estimator.train(input_fn=train_input_fn, max_steps=FLAGS.train_step) if FLAGS.do_eval: eval_examples = processor.get_dev_examples(FLAGS.data_dir) num_actual_eval_examples = len(eval_examples) if FLAGS.use_tpu: # TPU requires a fixed batch size for all batches, therefore the number # of examples must be a multiple of the batch size, or else examples # will get dropped. So we pad with fake examples which are ignored # later on. These do NOT count towards the metric (all tf.metrics # support a per-instance weight, and these get a weight of 0.0). while len(eval_examples) % FLAGS.eval_batch_size != 0: eval_examples.append(classifier_utils.PaddingInputExample()) cached_dir = FLAGS.cached_dir if not cached_dir: cached_dir = FLAGS.output_dir eval_file = os.path.join(cached_dir, task_name + "_eval.tf_record") if not tf.gfile.Exists(eval_file): classifier_utils.file_based_convert_examples_to_features( eval_examples, label_list, FLAGS.max_seq_length, tokenizer, eval_file, task_name) tf.logging.info("***** Running evaluation *****") tf.logging.info(" Num examples = %d (%d actual, %d padding)", len(eval_examples), num_actual_eval_examples, len(eval_examples) - num_actual_eval_examples) tf.logging.info(" Batch size = %d", FLAGS.eval_batch_size) # This tells the estimator to run through the entire set. eval_steps = None # However, if running eval on the TPU, you will need to specify the # number of steps. if FLAGS.use_tpu: assert len(eval_examples) % FLAGS.eval_batch_size == 0 eval_steps = int(len(eval_examples) // FLAGS.eval_batch_size) eval_drop_remainder = True if FLAGS.use_tpu else False eval_input_fn = classifier_utils.file_based_input_fn_builder( input_file=eval_file, seq_length=FLAGS.max_seq_length, is_training=False, drop_remainder=eval_drop_remainder, task_name=task_name, use_tpu=FLAGS.use_tpu, bsz=FLAGS.eval_batch_size) best_trial_info_file = os.path.join(FLAGS.output_dir, "best_trial.txt") def _best_trial_info(): """Returns information about which checkpoints have been evaled so far.""" if tf.gfile.Exists(best_trial_info_file): with tf.gfile.GFile(best_trial_info_file, "r") as best_info: global_step, best_metric_global_step, metric_value = ( best_info.read().split(":")) global_step = int(global_step) best_metric_global_step = int(best_metric_global_step) metric_value = float(metric_value) else: metric_value = -1 best_metric_global_step = -1 global_step = -1 tf.logging.info( "Best trial info: Step: %s, Best Value Step: %s, " "Best Value: %s", global_step, best_metric_global_step, metric_value) return global_step, best_metric_global_step, metric_value def _remove_checkpoint(checkpoint_path): for ext in ["meta", "data-00000-of-00001", "index"]: src_ckpt = checkpoint_path + ".{}".format(ext) tf.logging.info("removing {}".format(src_ckpt)) tf.gfile.Remove(src_ckpt) def _find_valid_cands(curr_step): filenames = tf.gfile.ListDirectory(FLAGS.output_dir) candidates = [] for filename in filenames: if filename.endswith(".index"): ckpt_name = filename[:-6] idx = ckpt_name.split("-")[-1] if int(idx) > curr_step: candidates.append(filename) return candidates output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt") if task_name == "sts-b": key_name = "pearson" elif task_name == "cola": key_name = "matthew_corr" else: key_name = "eval_accuracy" global_step, best_perf_global_step, best_perf = _best_trial_info() writer = tf.gfile.GFile(output_eval_file, "w") while global_step < FLAGS.train_step: steps_and_files = {} filenames = tf.gfile.ListDirectory(FLAGS.output_dir) for filename in filenames: if filename.endswith(".index"): ckpt_name = filename[:-6] cur_filename = os.path.join(FLAGS.output_dir, ckpt_name) if cur_filename.split("-")[-1] == "best": continue gstep = int(cur_filename.split("-")[-1]) if gstep not in steps_and_files: tf.logging.info( "Add {} to eval list.".format(cur_filename)) steps_and_files[gstep] = cur_filename tf.logging.info("found {} files.".format(len(steps_and_files))) if not steps_and_files: tf.logging.info( "found 0 file, global step: {}. Sleeping.".format( global_step)) time.sleep(60) else: for checkpoint in sorted(steps_and_files.items()): step, checkpoint_path = checkpoint if global_step >= step: if (best_perf_global_step != step and len(_find_valid_cands(step)) > 1): _remove_checkpoint(checkpoint_path) continue result = estimator.evaluate( input_fn=eval_input_fn, steps=eval_steps, checkpoint_path=checkpoint_path) global_step = result["global_step"] tf.logging.info("***** Eval results *****") for key in sorted(result.keys()): tf.logging.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) writer.write("best = {}\n".format(best_perf)) if result[key_name] > best_perf: best_perf = result[key_name] best_perf_global_step = global_step elif len(_find_valid_cands(global_step)) > 1: _remove_checkpoint(checkpoint_path) writer.write("=" * 50 + "\n") writer.flush() with tf.gfile.GFile(best_trial_info_file, "w") as best_info: best_info.write("{}:{}:{}".format( global_step, best_perf_global_step, best_perf)) writer.close() for ext in ["meta", "data-00000-of-00001", "index"]: src_ckpt = "model.ckpt-{}.{}".format(best_perf_global_step, ext) tgt_ckpt = "model.ckpt-best.{}".format(ext) tf.logging.info("saving {} to {}".format(src_ckpt, tgt_ckpt)) tf.io.gfile.rename(os.path.join(FLAGS.output_dir, src_ckpt), os.path.join(FLAGS.output_dir, tgt_ckpt), overwrite=True) if FLAGS.do_predict: predict_examples = processor.get_test_examples(FLAGS.data_dir) num_actual_predict_examples = len(predict_examples) if FLAGS.use_tpu: # TPU requires a fixed batch size for all batches, therefore the number # of examples must be a multiple of the batch size, or else examples # will get dropped. So we pad with fake examples which are ignored # later on. while len(predict_examples) % FLAGS.predict_batch_size != 0: predict_examples.append(classifier_utils.PaddingInputExample()) predict_file = os.path.join(FLAGS.output_dir, "predict.tf_record") classifier_utils.file_based_convert_examples_to_features( predict_examples, label_list, FLAGS.max_seq_length, tokenizer, predict_file, task_name) tf.logging.info("***** Running prediction*****") tf.logging.info(" Num examples = %d (%d actual, %d padding)", len(predict_examples), num_actual_predict_examples, len(predict_examples) - num_actual_predict_examples) tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size) predict_drop_remainder = True if FLAGS.use_tpu else False predict_input_fn = classifier_utils.file_based_input_fn_builder( input_file=predict_file, seq_length=FLAGS.max_seq_length, is_training=False, drop_remainder=predict_drop_remainder, task_name=task_name, use_tpu=FLAGS.use_tpu, bsz=FLAGS.predict_batch_size) checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best") result = estimator.predict(input_fn=predict_input_fn, checkpoint_path=checkpoint_path) output_predict_file = os.path.join(FLAGS.output_dir, "test_results.tsv") output_submit_file = os.path.join(FLAGS.output_dir, "submit_results.tsv") with tf.gfile.GFile(output_predict_file, "w") as pred_writer,\ tf.gfile.GFile(output_submit_file, "w") as sub_writer: sub_writer.write("index" + "\t" + "prediction\n") num_written_lines = 0 tf.logging.info("***** Predict results *****") for (i, (example, prediction)) in\ enumerate(zip(predict_examples, result)): probabilities = prediction["probabilities"] if i >= num_actual_predict_examples: break output_line = "\t".join( str(class_probability) for class_probability in probabilities) + "\n" pred_writer.write(output_line) if task_name != "sts-b": actual_label = label_list[int(prediction["predictions"])] else: actual_label = str(prediction["predictions"]) sub_writer.write(example.guid + "\t" + actual_label + "\n") num_written_lines += 1 assert num_written_lines == num_actual_predict_examples
def main(_): tf.logging.set_verbosity(tf.logging.INFO) processors = { "cola": classifier_utils.ColaProcessor, "mnli": classifier_utils.MnliProcessor, "mismnli": classifier_utils.MisMnliProcessor, "mrpc": classifier_utils.MrpcProcessor, "rte": classifier_utils.RteProcessor, "sst-2": classifier_utils.Sst2Processor, "sts-b": classifier_utils.StsbProcessor, "qqp": classifier_utils.QqpProcessor, "qnli": classifier_utils.QnliProcessor, "wnli": classifier_utils.WnliProcessor, } tokenization.validate_case_matches_checkpoint(FLAGS.do_lower_case, FLAGS.init_checkpoint) if not FLAGS.albert_config_file and not FLAGS.albert_hub_module_handle: raise ValueError("At least one of `--albert_config_file` and " "`--albert_hub_module_handle` must be set") if FLAGS.albert_config_file: albert_config = modeling.AlbertConfig.from_json_file( FLAGS.albert_config_file) if FLAGS.max_seq_length > albert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length %d because the ALBERT model " "was only trained up to sequence length %d" % (FLAGS.max_seq_length, albert_config.max_position_embeddings)) else: albert_config = None # Get the config from TF-Hub. tf.gfile.MakeDirs(FLAGS.output_dir) task_name = FLAGS.task_name.lower() if task_name not in processors: raise ValueError("Task not found: %s" % (task_name)) processor = processors[task_name]( use_spm=True if FLAGS.spm_model_file else False, do_lower_case=FLAGS.do_lower_case) label_list = processor.get_labels() tokenizer = fine_tuning_utils.create_vocab( vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case, spm_model_file=FLAGS.spm_model_file, hub_module=FLAGS.albert_hub_module_handle) tpu_cluster_resolver = None if FLAGS.use_tpu and FLAGS.tpu_name: tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver( FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2 iterations_per_loop = FLAGS.iterations_per_loop run_config = contrib_tpu.RunConfig( cluster=tpu_cluster_resolver, master=FLAGS.master, model_dir=FLAGS.output_dir, save_checkpoints_steps=int(FLAGS.save_checkpoints_steps), keep_checkpoint_max=0, tpu_config=contrib_tpu.TPUConfig( iterations_per_loop=iterations_per_loop, num_shards=FLAGS.num_tpu_cores, per_host_input_for_training=is_per_host)) model_fn = classifier_utils.model_fn_builder( albert_config=albert_config, num_labels=len(label_list), init_checkpoint=FLAGS.init_checkpoint, learning_rate=FLAGS.learning_rate, num_train_steps=FLAGS.train_step, num_warmup_steps=FLAGS.warmup_step, use_tpu=FLAGS.use_tpu, use_one_hot_embeddings=FLAGS.use_tpu, task_name=task_name, hub_module=FLAGS.albert_hub_module_handle, optimizer=FLAGS.optimizer) # If TPU is not available, this will fall back to normal Estimator on CPU # or GPU. estimator = contrib_tpu.TPUEstimator( use_tpu=FLAGS.use_tpu, model_fn=model_fn, config=run_config, train_batch_size=FLAGS.train_batch_size, eval_batch_size=FLAGS.eval_batch_size, predict_batch_size=FLAGS.predict_batch_size) # if FLAGS.do_predict: eval_examples = processor.get_dev_examples(FLAGS.data_dir) num_actual_eval_examples = len(eval_examples) if FLAGS.use_tpu: # TPU requires a fixed batch size for all batches, therefore the number # of examples must be a multiple of the batch size, or else examples # will get dropped. So we pad with fake examples which are ignored # later on. while len(eval_examples) % FLAGS.predict_batch_size != 0: eval_examples.append(classifier_utils.PaddingInputExample()) error_analysis_file = os.path.join(FLAGS.output_dir, "error_analysis.tf_record") classifier_utils.file_based_convert_examples_to_features( eval_examples, label_list, FLAGS.max_seq_length, tokenizer, error_analysis_file, task_name) tf.logging.info("***** Running error analysis on dev set*****") tf.logging.info(" Num examples = %d (%d actual, %d padding)", len(eval_examples), num_actual_eval_examples, len(eval_examples) - num_actual_eval_examples) tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size) error_analysis_drop_remainder = True if FLAGS.use_tpu else False error_analysis_input_fn = classifier_utils.file_based_input_fn_builder( input_file=error_analysis_file, seq_length=FLAGS.max_seq_length, is_training=False, drop_remainder=error_analysis_drop_remainder, task_name=task_name, use_tpu=FLAGS.use_tpu, bsz=FLAGS.predict_batch_size) checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best") result = estimator.predict(input_fn=error_analysis_input_fn, checkpoint_path=checkpoint_path) output_error_analysis_predict_file = os.path.join( FLAGS.output_dir, "error_analysis_test_results.tsv") output_error_analysis_submit_file = os.path.join( FLAGS.output_dir, "error_analysis_submit_results.tsv") with tf.gfile.GFile(output_error_analysis_predict_file, "w") as pred_writer,\ tf.gfile.GFile(output_error_analysis_submit_file, "w") as sub_writer: sub_writer.write("index" + "\t" + "text_a" + "\t" + "text_b" + "\t" + "prediction" + "\t" + "label" + "\n") num_written_lines = 0 tf.logging.info("***** Error analysis results *****") for (i, (example, prediction)) in\ enumerate(zip(eval_examples, result)): probabilities = prediction["probabilities"] if i >= num_actual_eval_examples: break output_line = "\t".join( str(class_probability) for class_probability in probabilities) + "\n" pred_writer.write(output_line) if task_name != "sts-b": actual_label = label_list[int(prediction["predictions"])] else: actual_label = str(prediction["predictions"]) sub_writer.write(example.guid + "\t" + str(example.text_a) + "\t" + str(example.text_b) + "\t" + str(actual_label) + "\t" + str(example.label) + "\n") num_written_lines += 1 assert num_written_lines == num_actual_eval_examples
def main(_): tf.logging.set_verbosity(tf.logging.INFO) processors = {"race": race_utils.RaceProcessor} tokenization.validate_case_matches_checkpoint(FLAGS.do_lower_case, FLAGS.init_checkpoint) if not FLAGS.do_train and not FLAGS.do_eval and not FLAGS.do_predict: raise ValueError( "At least one of `do_train`, `do_eval` or `do_predict' must be True." ) albert_config = modeling.AlbertConfig.from_json_file( FLAGS.albert_config_file) if FLAGS.max_seq_length > albert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length %d because the ALBERT model " "was only trained up to sequence length %d" % (FLAGS.max_seq_length, albert_config.max_position_embeddings)) tf.gfile.MakeDirs(FLAGS.output_dir) task_name = FLAGS.task_name.lower() if task_name not in processors: raise ValueError("Task not found: %s" % (task_name)) processor = processors[task_name]( use_spm=True if FLAGS.spm_model_file else False, do_lower_case=FLAGS.do_lower_case, high_only=FLAGS.high_only, middle_only=FLAGS.middle_only) label_list = processor.get_labels() tokenizer = fine_tuning_utils.create_vocab( vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case, spm_model_file=FLAGS.spm_model_file, hub_module=FLAGS.albert_hub_module_handle) tpu_cluster_resolver = None if FLAGS.use_tpu and FLAGS.tpu_name: tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver( FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2 if FLAGS.do_train: iterations_per_loop = int( min(FLAGS.iterations_per_loop, FLAGS.save_checkpoints_steps)) else: iterations_per_loop = FLAGS.iterations_per_loop run_config = contrib_tpu.RunConfig( cluster=tpu_cluster_resolver, master=FLAGS.master, model_dir=FLAGS.output_dir, save_checkpoints_steps=int(FLAGS.save_checkpoints_steps), keep_checkpoint_max=0, tpu_config=contrib_tpu.TPUConfig( iterations_per_loop=iterations_per_loop, num_shards=FLAGS.num_tpu_cores, per_host_input_for_training=is_per_host)) train_examples = None if FLAGS.do_train: train_examples = processor.get_train_examples(FLAGS.data_dir) model_fn = race_utils.model_fn_builder( albert_config=albert_config, num_labels=len(label_list), init_checkpoint=FLAGS.init_checkpoint, learning_rate=FLAGS.learning_rate, num_train_steps=FLAGS.train_step, num_warmup_steps=FLAGS.warmup_step, use_tpu=FLAGS.use_tpu, use_one_hot_embeddings=FLAGS.use_tpu, max_seq_length=FLAGS.max_seq_length, dropout_prob=FLAGS.dropout_prob, hub_module=FLAGS.albert_hub_module_handle) # If TPU is not available, this will fall back to normal Estimator on CPU # or GPU. estimator = contrib_tpu.TPUEstimator( use_tpu=FLAGS.use_tpu, model_fn=model_fn, config=run_config, train_batch_size=FLAGS.train_batch_size, eval_batch_size=FLAGS.eval_batch_size, predict_batch_size=FLAGS.predict_batch_size) if FLAGS.do_train: if not tf.gfile.Exists(FLAGS.train_file): race_utils.file_based_convert_examples_to_features( train_examples, label_list, FLAGS.max_seq_length, tokenizer, FLAGS.train_file, FLAGS.max_qa_length) tf.logging.info("***** Running training *****") tf.logging.info(" Num examples = %d", len(train_examples)) tf.logging.info(" Batch size = %d", FLAGS.train_batch_size) tf.logging.info(" Num steps = %d", FLAGS.train_step) train_input_fn = classifier_utils.file_based_input_fn_builder( input_file=FLAGS.train_file, seq_length=FLAGS.max_seq_length, is_training=True, drop_remainder=True, task_name=task_name, use_tpu=FLAGS.use_tpu, bsz=FLAGS.train_batch_size, multiple=len(label_list)) estimator.train(input_fn=train_input_fn, max_steps=FLAGS.train_step) if FLAGS.do_eval: eval_examples = processor.get_dev_examples(FLAGS.data_dir) num_actual_eval_examples = len(eval_examples) if FLAGS.use_tpu: # TPU requires a fixed batch size for all batches, therefore the number # of examples must be a multiple of the batch size, or else examples # will get dropped. So we pad with fake examples which are ignored # later on. These do NOT count towards the metric (all tf.metrics # support a per-instance weight, and these get a weight of 0.0). while len(eval_examples) % FLAGS.eval_batch_size != 0: eval_examples.append(classifier_utils.PaddingInputExample()) if not tf.gfile.Exists(FLAGS.eval_file): race_utils.file_based_convert_examples_to_features( eval_examples, label_list, FLAGS.max_seq_length, tokenizer, FLAGS.eval_file, FLAGS.max_qa_length) tf.logging.info("***** Running evaluation *****") tf.logging.info(" Num examples = %d (%d actual, %d padding)", len(eval_examples), num_actual_eval_examples, len(eval_examples) - num_actual_eval_examples) tf.logging.info(" Batch size = %d", FLAGS.eval_batch_size) # This tells the estimator to run through the entire set. eval_steps = None # However, if running eval on the TPU, you will need to specify the # number of steps. if FLAGS.use_tpu: assert len(eval_examples) % FLAGS.eval_batch_size == 0 eval_steps = int(len(eval_examples) // FLAGS.eval_batch_size) eval_drop_remainder = True if FLAGS.use_tpu else False eval_input_fn = classifier_utils.file_based_input_fn_builder( input_file=FLAGS.eval_file, seq_length=FLAGS.max_seq_length, is_training=False, drop_remainder=eval_drop_remainder, task_name=task_name, use_tpu=FLAGS.use_tpu, bsz=FLAGS.eval_batch_size, multiple=len(label_list)) def _find_valid_cands(curr_step): filenames = tf.gfile.ListDirectory(FLAGS.output_dir) candidates = [] for filename in filenames: if filename.endswith(".index"): ckpt_name = filename[:-6] idx = ckpt_name.split("-")[-1] if idx != "best" and int(idx) > curr_step: candidates.append(filename) return candidates output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt") checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best") key_name = "eval_accuracy" if tf.gfile.Exists(checkpoint_path + ".index"): result = estimator.evaluate(input_fn=eval_input_fn, steps=eval_steps, checkpoint_path=checkpoint_path) best_perf = result[key_name] global_step = result["global_step"] else: global_step = -1 best_perf = -1 checkpoint_path = None writer = tf.gfile.GFile(output_eval_file, "w") while global_step < FLAGS.train_step: steps_and_files = {} filenames = tf.gfile.ListDirectory(FLAGS.output_dir) for filename in filenames: if filename.endswith(".index"): ckpt_name = filename[:-6] cur_filename = os.path.join(FLAGS.output_dir, ckpt_name) if cur_filename.split("-")[-1] == "best": continue gstep = int(cur_filename.split("-")[-1]) if gstep not in steps_and_files: tf.logging.info( "Add {} to eval list.".format(cur_filename)) steps_and_files[gstep] = cur_filename tf.logging.info("found {} files.".format(len(steps_and_files))) # steps_and_files = sorted(steps_and_files, key=lambda x: x[0]) if not steps_and_files: tf.logging.info( "found 0 file, global step: {}. Sleeping.".format( global_step)) time.sleep(1) else: for ele in sorted(steps_and_files.items()): step, checkpoint_path = ele if global_step >= step: if len(_find_valid_cands(step)) > 1: for ext in [ "meta", "data-00000-of-00001", "index" ]: src_ckpt = checkpoint_path + ".{}".format(ext) tf.logging.info("removing {}".format(src_ckpt)) tf.gfile.Remove(src_ckpt) continue result = estimator.evaluate( input_fn=eval_input_fn, steps=eval_steps, checkpoint_path=checkpoint_path) global_step = result["global_step"] tf.logging.info("***** Eval results *****") for key in sorted(result.keys()): tf.logging.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) writer.write("best = {}\n".format(best_perf)) if result[key_name] > best_perf: best_perf = result[key_name] for ext in ["meta", "data-00000-of-00001", "index"]: src_ckpt = checkpoint_path + ".{}".format(ext) tgt_ckpt = checkpoint_path.rsplit( "-", 1)[0] + "-best.{}".format(ext) tf.logging.info("saving {} to {}".format( src_ckpt, tgt_ckpt)) tf.gfile.Copy(src_ckpt, tgt_ckpt, overwrite=True) writer.write("saved {} to {}\n".format( src_ckpt, tgt_ckpt)) if len(_find_valid_cands(global_step)) > 1: for ext in ["meta", "data-00000-of-00001", "index"]: src_ckpt = checkpoint_path + ".{}".format(ext) tf.logging.info("removing {}".format(src_ckpt)) tf.gfile.Remove(src_ckpt) writer.write("=" * 50 + "\n") writer.close() if FLAGS.do_predict: predict_examples = processor.get_test_examples(FLAGS.data_dir) num_actual_predict_examples = len(predict_examples) if FLAGS.use_tpu: # TPU requires a fixed batch size for all batches, therefore the number # of examples must be a multiple of the batch size, or else examples # will get dropped. So we pad with fake examples which are ignored # later on. while len(predict_examples) % FLAGS.predict_batch_size != 0: predict_examples.append(classifier_utils.PaddingInputExample()) assert len(predict_examples) % FLAGS.predict_batch_size == 0 predict_steps = int( len(predict_examples) // FLAGS.predict_batch_size) predict_file = os.path.join(FLAGS.output_dir, "predict.tf_record") race_utils.file_based_convert_examples_to_features( predict_examples, label_list, FLAGS.max_seq_length, tokenizer, predict_file, FLAGS.max_qa_length) tf.logging.info("***** Running prediction*****") tf.logging.info(" Num examples = %d (%d actual, %d padding)", len(predict_examples), num_actual_predict_examples, len(predict_examples) - num_actual_predict_examples) tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size) predict_drop_remainder = True if FLAGS.use_tpu else False predict_input_fn = classifier_utils.file_based_input_fn_builder( input_file=predict_file, seq_length=FLAGS.max_seq_length, is_training=False, drop_remainder=predict_drop_remainder, task_name=task_name, use_tpu=FLAGS.use_tpu, bsz=FLAGS.predict_batch_size, multiple=len(label_list)) checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best") result = estimator.evaluate(input_fn=predict_input_fn, steps=predict_steps, checkpoint_path=checkpoint_path) output_predict_file = os.path.join(FLAGS.output_dir, "predict_results.txt") with tf.gfile.GFile(output_predict_file, "w") as pred_writer: # num_written_lines = 0 tf.logging.info("***** Predict results *****") pred_writer.write("***** Predict results *****\n") for key in sorted(result.keys()): tf.logging.info(" %s = %s", key, str(result[key])) pred_writer.write("%s = %s\n" % (key, str(result[key]))) pred_writer.write("best = {}\n".format(best_perf))
def main(_): tf.logging.set_verbosity(tf.logging.INFO) tf.get_logger().propagate = False albert_config = modeling.AlbertConfig.from_json_file( FLAGS.albert_config_file) validate_flags_or_throw(albert_config) tf.gfile.MakeDirs(FLAGS.output_dir) print("Output:", FLAGS.output_dir) tokenizer = fine_tuning_utils.create_vocab( vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case, spm_model_file=FLAGS.spm_model_file, hub_module=FLAGS.albert_hub_module_handle) tpu_cluster_resolver = None if FLAGS.use_tpu and FLAGS.tpu_name: tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver( FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2 if FLAGS.do_train: iterations_per_loop = int( min(FLAGS.iterations_per_loop, FLAGS.save_checkpoints_steps)) else: iterations_per_loop = FLAGS.iterations_per_loop run_config = contrib_tpu.RunConfig( cluster=tpu_cluster_resolver, master=FLAGS.master, model_dir=FLAGS.output_dir, keep_checkpoint_max=0, save_checkpoints_steps=FLAGS.save_checkpoints_steps, tpu_config=contrib_tpu.TPUConfig( iterations_per_loop=iterations_per_loop, num_shards=FLAGS.num_tpu_cores, per_host_input_for_training=is_per_host)) train_examples = None num_train_steps = None num_warmup_steps = None train_examples = squad_utils.read_squad_examples( input_file=FLAGS.train_file, is_training=True) num_train_steps = int( len(train_examples) / FLAGS.train_batch_size * FLAGS.num_train_epochs) if FLAGS.do_train: num_warmup_steps = int(num_train_steps * FLAGS.warmup_proportion) # Pre-shuffle the input to avoid having to make a very large shuffle # buffer in in the `input_fn`. rng = random.Random(12345) rng.shuffle(train_examples) model_fn = squad_utils.v2_model_fn_builder( albert_config=albert_config, init_checkpoint=FLAGS.init_checkpoint, learning_rate=FLAGS.learning_rate, num_train_steps=num_train_steps, num_warmup_steps=num_warmup_steps, use_tpu=FLAGS.use_tpu, use_one_hot_embeddings=FLAGS.use_tpu, max_seq_length=FLAGS.max_seq_length, start_n_top=FLAGS.start_n_top, end_n_top=FLAGS.end_n_top, dropout_prob=FLAGS.dropout_prob, hub_module=FLAGS.albert_hub_module_handle) # If TPU is not available, this will fall back to normal Estimator on CPU # or GPU. estimator = contrib_tpu.TPUEstimator( use_tpu=FLAGS.use_tpu, model_fn=model_fn, config=run_config, train_batch_size=FLAGS.train_batch_size, predict_batch_size=FLAGS.predict_batch_size) if FLAGS.do_train: # We write to a temporary file to avoid storing very large constant tensors # in memory. if not tf.gfile.Exists(FLAGS.train_feature_file): train_writer = squad_utils.FeatureWriter(filename=os.path.join( FLAGS.train_feature_file), is_training=True) squad_utils.convert_examples_to_features( examples=train_examples, tokenizer=tokenizer, max_seq_length=FLAGS.max_seq_length, doc_stride=FLAGS.doc_stride, max_query_length=FLAGS.max_query_length, is_training=True, output_fn=train_writer.process_feature, do_lower_case=FLAGS.do_lower_case) train_writer.close() tf.logging.info("***** Running training *****") tf.logging.info(" Num orig examples = %d", len(train_examples)) # tf.logging.info(" Num split examples = %d", train_writer.num_features) tf.logging.info(" Batch size = %d", FLAGS.train_batch_size) tf.logging.info(" Num steps = %d", num_train_steps) del train_examples train_input_fn = squad_utils.input_fn_builder( input_file=FLAGS.train_feature_file, seq_length=FLAGS.max_seq_length, is_training=True, drop_remainder=True, use_tpu=FLAGS.use_tpu, bsz=FLAGS.train_batch_size, is_v2=True) estimator.train(input_fn=train_input_fn, max_steps=num_train_steps) if FLAGS.do_predict: with tf.gfile.Open(FLAGS.predict_file) as predict_file: prediction_json = json.load(predict_file)["data"] eval_examples = squad_utils.read_squad_examples( input_file=FLAGS.predict_file, is_training=False) if (tf.gfile.Exists(FLAGS.predict_feature_file) and tf.gfile.Exists(FLAGS.predict_feature_left_file)): tf.logging.info("Loading eval features from {}".format( FLAGS.predict_feature_left_file)) with tf.gfile.Open(FLAGS.predict_feature_left_file, "rb") as fin: eval_features = pickle.load(fin) else: eval_writer = squad_utils.FeatureWriter( filename=FLAGS.predict_feature_file, is_training=False) eval_features = [] def append_feature(feature): eval_features.append(feature) eval_writer.process_feature(feature) squad_utils.convert_examples_to_features( examples=eval_examples, tokenizer=tokenizer, max_seq_length=FLAGS.max_seq_length, doc_stride=FLAGS.doc_stride, max_query_length=FLAGS.max_query_length, is_training=False, output_fn=append_feature, do_lower_case=FLAGS.do_lower_case) eval_writer.close() with tf.gfile.Open(FLAGS.predict_feature_left_file, "wb") as fout: pickle.dump(eval_features, fout) tf.logging.info("***** Running predictions *****") tf.logging.info(" Num orig examples = %d", len(eval_examples)) tf.logging.info(" Num split examples = %d", len(eval_features)) tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size) predict_input_fn = squad_utils.input_fn_builder( input_file=FLAGS.predict_feature_file, seq_length=FLAGS.max_seq_length, is_training=False, drop_remainder=False, use_tpu=FLAGS.use_tpu, bsz=FLAGS.predict_batch_size, is_v2=True) def get_result(checkpoint): """Evaluate the checkpoint on SQuAD v2.0.""" # If running eval on the TPU, you will need to specify the number of # steps. reader = tf.train.NewCheckpointReader(checkpoint) global_step = reader.get_tensor(tf.GraphKeys.GLOBAL_STEP) all_results = [] for result in estimator.predict(predict_input_fn, yield_single_examples=True, checkpoint_path=checkpoint): if len(all_results) % 1000 == 0: tf.logging.info("Processing example: %d" % (len(all_results))) unique_id = int(result["unique_ids"]) start_top_log_probs = ([ float(x) for x in result["start_top_log_probs"].flat ]) start_top_index = [ int(x) for x in result["start_top_index"].flat ] end_top_log_probs = ([ float(x) for x in result["end_top_log_probs"].flat ]) end_top_index = [int(x) for x in result["end_top_index"].flat] cls_logits = float(result["cls_logits"].flat[0]) all_results.append( squad_utils.RawResultV2( unique_id=unique_id, start_top_log_probs=start_top_log_probs, start_top_index=start_top_index, end_top_log_probs=end_top_log_probs, end_top_index=end_top_index, cls_logits=cls_logits)) output_prediction_file = os.path.join(FLAGS.output_dir, "predictions.json") output_nbest_file = os.path.join(FLAGS.output_dir, "nbest_predictions.json") output_null_log_odds_file = os.path.join(FLAGS.output_dir, "null_odds.json") result_dict = {} cls_dict = {} squad_utils.accumulate_predictions_v2( result_dict, cls_dict, eval_examples, eval_features, all_results, FLAGS.n_best_size, FLAGS.max_answer_length, FLAGS.start_n_top, FLAGS.end_n_top) return squad_utils.evaluate_v2( result_dict, cls_dict, prediction_json, eval_examples, eval_features, all_results, FLAGS.n_best_size, FLAGS.max_answer_length, output_prediction_file, output_nbest_file, output_null_log_odds_file), int(global_step) def _find_valid_cands(curr_step): filenames = tf.gfile.ListDirectory(FLAGS.output_dir) candidates = [] for filename in filenames: if filename.endswith(".index"): ckpt_name = filename[:-6] idx = ckpt_name.split("-")[-1] if idx != "best" and int(idx) > curr_step: candidates.append(filename) return candidates output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt") checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best") key_name = "f1" writer = tf.gfile.GFile(output_eval_file, "w") if tf.gfile.Exists(checkpoint_path + ".index"): result = get_result(checkpoint_path) best_perf = result[0][key_name] global_step = result[1] else: global_step = -1 best_perf = -1 checkpoint_path = None while global_step < num_train_steps: steps_and_files = {} filenames = tf.gfile.ListDirectory(FLAGS.output_dir) for filename in filenames: if filename.endswith(".index"): ckpt_name = filename[:-6] cur_filename = os.path.join(FLAGS.output_dir, ckpt_name) if cur_filename.split("-")[-1] == "best": continue gstep = int(cur_filename.split("-")[-1]) if gstep not in steps_and_files: tf.logging.info( "Add {} to eval list.".format(cur_filename)) steps_and_files[gstep] = cur_filename tf.logging.info("found {} files.".format(len(steps_and_files))) if not steps_and_files: tf.logging.info( "found 0 file, global step: {}. Sleeping.".format( global_step)) time.sleep(60) else: for ele in sorted(steps_and_files.items()): step, checkpoint_path = ele print("GS: ", global_step, step) if global_step >= step: if len(_find_valid_cands(step)) > 1: for ext in [ "meta", "data-00000-of-00001", "index" ]: src_ckpt = checkpoint_path + ".{}".format(ext) tf.logging.info("removing {}".format(src_ckpt)) tf.gfile.Remove(src_ckpt) continue result, global_step = get_result(checkpoint_path) print("EVAL RESULTS") tf.logging.info("***** Eval results *****") for key in sorted(result.keys()): tf.logging.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) if result[key_name] > best_perf: best_perf = result[key_name] for ext in ["meta", "data-00000-of-00001", "index"]: src_ckpt = checkpoint_path + ".{}".format(ext) tgt_ckpt = checkpoint_path.rsplit( "-", 1)[0] + "-best.{}".format(ext) tf.logging.info("saving {} to {}".format( src_ckpt, tgt_ckpt)) tf.gfile.Copy(src_ckpt, tgt_ckpt, overwrite=True) writer.write("saved {} to {}\n".format( src_ckpt, tgt_ckpt)) writer.write("best {} = {}\n".format(key_name, best_perf)) tf.logging.info(" best {} = {}\n".format( key_name, best_perf)) if len(_find_valid_cands(global_step)) > 2: for ext in ["meta", "data-00000-of-00001", "index"]: src_ckpt = checkpoint_path + ".{}".format(ext) tf.logging.info("removing {}".format(src_ckpt)) tf.gfile.Remove(src_ckpt) writer.write("=" * 50 + "\n") print("Sleeping") time.sleep(10) checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best") result, global_step = get_result(checkpoint_path) tf.logging.info("***** Final Eval results *****") for key in sorted(result.keys()): tf.logging.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) writer.write("best perf happened at step: {}".format(global_step))
def main(_): tf.logging.set_verbosity(tf.logging.INFO) albert_config = modeling.AlbertConfig.from_json_file( FLAGS.albert_config_file) validate_flags_or_throw(albert_config) tf.gfile.MakeDirs(FLAGS.output_dir) tokenizer = fine_tuning_utils.create_vocab( vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case, spm_model_file=FLAGS.spm_model_file, hub_module=FLAGS.albert_hub_module_handle) tpu_cluster_resolver = None if FLAGS.use_tpu and FLAGS.tpu_name: tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver( FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2 if FLAGS.do_train: iterations_per_loop = int( min(FLAGS.iterations_per_loop, FLAGS.save_checkpoints_steps)) else: iterations_per_loop = FLAGS.iterations_per_loop run_config = contrib_tpu.RunConfig( cluster=tpu_cluster_resolver, master=FLAGS.master, model_dir=FLAGS.output_dir, keep_checkpoint_max=0, save_checkpoints_steps=FLAGS.save_checkpoints_steps, tpu_config=contrib_tpu.TPUConfig( iterations_per_loop=iterations_per_loop, num_shards=FLAGS.num_tpu_cores, per_host_input_for_training=is_per_host)) train_examples = None num_train_steps = None num_warmup_steps = None # if FLAGS.do_train: # train_examples = squad_utils.read_squad_examples( # input_file=FLAGS.train_file, is_training=True) # num_train_steps = int( # len(train_examples) / FLAGS.train_batch_size * FLAGS.num_train_epochs) # num_warmup_steps = int(num_train_steps * FLAGS.warmup_proportion) # # # Pre-shuffle the input to avoid having to make a very large shuffle # # buffer in in the `input_fn`. # rng = random.Random(12345) # rng.shuffle(train_examples) model_fn = squad_utils.v2_model_fn_builder( albert_config=albert_config, init_checkpoint=FLAGS.init_checkpoint, learning_rate=FLAGS.learning_rate, num_train_steps=num_train_steps, num_warmup_steps=num_warmup_steps, use_tpu=FLAGS.use_tpu, use_one_hot_embeddings=FLAGS.use_tpu, max_seq_length=FLAGS.max_seq_length, start_n_top=FLAGS.start_n_top, end_n_top=FLAGS.end_n_top, dropout_prob=FLAGS.dropout_prob, hub_module=FLAGS.albert_hub_module_handle) # If TPU is not available, this will fall back to normal Estimator on CPU # or GPU. estimator = contrib_tpu.TPUEstimator( use_tpu=FLAGS.use_tpu, model_fn=model_fn, config=run_config, train_batch_size=FLAGS.train_batch_size, predict_batch_size=FLAGS.predict_batch_size) if FLAGS.do_train: # We write to a temporary file to avoid storing very large constant tensors # in memory. if not tf.gfile.Exists(FLAGS.train_feature_file): train_writer = squad_utils.FeatureWriter(filename=os.path.join( FLAGS.train_feature_file), is_training=True) squad_utils.convert_examples_to_features( examples=train_examples, tokenizer=tokenizer, max_seq_length=FLAGS.max_seq_length, doc_stride=FLAGS.doc_stride, max_query_length=FLAGS.max_query_length, is_training=True, output_fn=train_writer.process_feature, do_lower_case=FLAGS.do_lower_case) train_writer.close() tf.logging.info("***** Running training *****") tf.logging.info(" Num orig examples = %d", len(train_examples)) # tf.logging.info(" Num split examples = %d", train_writer.num_features) tf.logging.info(" Batch size = %d", FLAGS.train_batch_size) tf.logging.info(" Num steps = %d", num_train_steps) del train_examples train_input_fn = squad_utils.input_fn_builder( input_file=FLAGS.train_feature_file, seq_length=FLAGS.max_seq_length, is_training=True, drop_remainder=True, use_tpu=FLAGS.use_tpu, bsz=FLAGS.train_batch_size, is_v2=True) estimator.train(input_fn=train_input_fn, max_steps=num_train_steps) if FLAGS.do_predict: with tf.gfile.Open(FLAGS.predict_file) as predict_file: prediction_json = json.load(predict_file)["data"] eval_examples = squad_utils.read_squad_examples( input_file=FLAGS.predict_file, is_training=False) if (tf.gfile.Exists(FLAGS.predict_feature_file) and tf.gfile.Exists(FLAGS.predict_feature_left_file)): tf.logging.info("Loading eval features from {}".format( FLAGS.predict_feature_left_file)) with tf.gfile.Open(FLAGS.predict_feature_left_file, "rb") as fin: eval_features = pickle.load(fin) else: eval_writer = squad_utils.FeatureWriter( filename=FLAGS.predict_feature_file, is_training=False) eval_features = [] def append_feature(feature): eval_features.append(feature) eval_writer.process_feature(feature) squad_utils.convert_examples_to_features( examples=eval_examples, tokenizer=tokenizer, max_seq_length=FLAGS.max_seq_length, doc_stride=FLAGS.doc_stride, max_query_length=FLAGS.max_query_length, is_training=False, output_fn=append_feature, do_lower_case=FLAGS.do_lower_case) eval_writer.close() with tf.gfile.Open(FLAGS.predict_feature_left_file, "wb") as fout: pickle.dump(eval_features, fout) tf.logging.info("***** Running predictions *****") tf.logging.info(" Num orig examples = %d", len(eval_examples)) tf.logging.info(" Num split examples = %d", len(eval_features)) tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size) predict_input_fn = squad_utils.input_fn_builder( input_file=FLAGS.predict_feature_file, seq_length=FLAGS.max_seq_length, is_training=False, drop_remainder=False, use_tpu=FLAGS.use_tpu, bsz=FLAGS.predict_batch_size, is_v2=True) def get_result(checkpoint): """Evaluate the checkpoint on SQuAD v2.0.""" # If running eval on the TPU, you will need to specify the number of # steps. reader = tf.train.NewCheckpointReader(checkpoint) global_step = reader.get_tensor(tf.GraphKeys.GLOBAL_STEP) all_results = [] for result in estimator.predict(predict_input_fn, yield_single_examples=True, checkpoint_path=checkpoint): if len(all_results) % 1000 == 0: tf.logging.info("Processing example: %d" % (len(all_results))) unique_id = int(result["unique_ids"]) cls_logits = float(result["cls_logits"].flat[0]) all_results.append( squad_utils.RawResultV2(unique_id=unique_id, cls_logits=cls_logits)) output_prediction_file = os.path.join(FLAGS.output_dir, "predictions.json") output_nbest_file = os.path.join(FLAGS.output_dir, "nbest_predictions.json") output_null_log_odds_file = os.path.join(FLAGS.output_dir, "null_odds.json") result_dict = {} cls_dict = {} squad_utils.accumulate_predictions_v2( result_dict, cls_dict, eval_examples, eval_features, all_results, FLAGS.n_best_size, FLAGS.max_answer_length, FLAGS.start_n_top, FLAGS.end_n_top) from squad_utils import make_qid_to_has_ans import numpy as np qid_to_has_ans = make_qid_to_has_ans( prediction_json) # maps qid to True/False has_ans_qids = [k for k, v in qid_to_has_ans.items() if v] no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v] print("has_ans", len(has_ans_qids)) print("no_ans", len(no_ans_qids)) def compute_metrics_with_threshold(threshold): nonlocal result_dict result_dict = {} tp = 0 tn = 0 fp = 0 fn = 0 for example_index, example in enumerate(eval_examples): m = np.min(cls_dict[example_index]) predict_is_impossible = 1 / (1 + np.exp(-m)) > threshold # predict_is_impossible = m > threshold result_dict[example.qas_id] = m if example.is_impossible: if predict_is_impossible: tp += 1 else: fn += 1 else: if predict_is_impossible: fp += 1 else: tn += 1 precision = tp / (tp + fp) recall = tp / (fn + tp) f1 = 2 * tp / (2 * tp + fp + fn) tf.logging.info(f"precision: {precision}" f"recall: {recall}" f"f1: {f1}") return precision, recall, f1 # precision, recall, f1 = compute_metrics_with_threshold(0.4) precision, recall, f1 = compute_metrics_with_threshold(0.5) # precision, recall, f1 = compute_metrics_with_threshold(0.6) with tf.gfile.GFile(output_prediction_file, "w") as writer: writer.write(json.dumps(result_dict, indent=4) + "\n") return { "precision": precision, "recall": recall, "f1": f1, "total": len(eval_examples) }, int(global_step) def _find_valid_cands(curr_step): filenames = tf.gfile.ListDirectory(FLAGS.output_dir) candidates = [] for filename in filenames: if filename.endswith(".index"): ckpt_name = filename[:-6] idx = ckpt_name.split("-")[-1] if idx != "best" and int(idx) > curr_step: candidates.append(filename) return candidates # output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt") # checkpoint_path = os.path.join(FLAGS.output_dir, "model.ckpt-best") # key_name = "f1" # writer = tf.gfile.GFile(output_eval_file, "w") # if tf.gfile.Exists(checkpoint_path + ".index"): # result = get_result(checkpoint_path) # best_perf = result[0][key_name] # global_step = result[1] # else: # global_step = -1 # best_perf = -1 # checkpoint_path = None # while global_step < num_train_steps: # steps_and_files = {} # filenames = tf.gfile.ListDirectory(FLAGS.output_dir) # for filename in filenames: # if filename.endswith(".index"): # ckpt_name = filename[:-6] # cur_filename = os.path.join(FLAGS.output_dir, ckpt_name) # if cur_filename.split("-")[-1] == "best": # continue # gstep = int(cur_filename.split("-")[-1]) # if gstep not in steps_and_files: # tf.logging.info("Add {} to eval list.".format(cur_filename)) # steps_and_files[gstep] = cur_filename # tf.logging.info("found {} files.".format(len(steps_and_files))) # if not steps_and_files: # tf.logging.info("found 0 file, global step: {}. Sleeping." # .format(global_step)) # time.sleep(60) # else: # for ele in sorted(steps_and_files.items()): # step, checkpoint_path = ele # if global_step >= step: # if len(_find_valid_cands(step)) > 1: # for ext in ["meta", "data-00000-of-00001", "index"]: # src_ckpt = checkpoint_path + ".{}".format(ext) # tf.logging.info("removing {}".format(src_ckpt)) # tf.gfile.Remove(src_ckpt) # continue # result, global_step = get_result(checkpoint_path) # tf.logging.info("***** Eval results *****") # for key in sorted(result.keys()): # tf.logging.info(" %s = %s", key, str(result[key])) # writer.write("%s = %s\n" % (key, str(result[key]))) # if result[key_name] > best_perf: # best_perf = result[key_name] # for ext in ["meta", "data-00000-of-00001", "index"]: # src_ckpt = checkpoint_path + ".{}".format(ext) # tgt_ckpt = checkpoint_path.rsplit( # "-", 1)[0] + "-best.{}".format(ext) # tf.logging.info("saving {} to {}".format(src_ckpt, tgt_ckpt)) # tf.gfile.Copy(src_ckpt, tgt_ckpt, overwrite=True) # writer.write("saved {} to {}\n".format(src_ckpt, tgt_ckpt)) # writer.write("best {} = {}\n".format(key_name, best_perf)) # tf.logging.info(" best {} = {}\n".format(key_name, best_perf)) # # if len(_find_valid_cands(global_step)) > 2: # for ext in ["meta", "data-00000-of-00001", "index"]: # src_ckpt = checkpoint_path + ".{}".format(ext) # tf.logging.info("removing {}".format(src_ckpt)) # tf.gfile.Remove(src_ckpt) # writer.write("=" * 50 + "\n") result, global_step = get_result(FLAGS.init_checkpoint)