def main(_): tf.logging.set_verbosity(tf.logging.INFO) news_config = GroverConfig.from_json_file(FLAGS.config_file) tf.gfile.MakeDirs(FLAGS.output_dir) input_files = [] for input_pattern in FLAGS.input_file.split(","): input_files.extend(tf.gfile.Glob(input_pattern)) tf.logging.info("*** Input Files ***") for input_file in input_files: tf.logging.info(" %s" % input_file) tpu_cluster_resolver = None if FLAGS.use_tpu and FLAGS.tpu_name: tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver( FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2 run_config = tf.contrib.tpu.RunConfig( cluster=tpu_cluster_resolver, master=FLAGS.master, model_dir=FLAGS.output_dir, save_checkpoints_steps=FLAGS.save_checkpoints_steps, keep_checkpoint_max=None, tpu_config=tf.contrib.tpu.TPUConfig( iterations_per_loop=FLAGS.iterations_per_loop, num_shards=FLAGS.num_tpu_cores, per_host_input_for_training=is_per_host)) model_fn = model_fn_builder( news_config, init_checkpoint=FLAGS.init_checkpoint, learning_rate=FLAGS.learning_rate, num_train_steps=FLAGS.num_train_steps, num_warmup_steps=FLAGS.num_warmup_steps, use_tpu=FLAGS.use_tpu, ) # If TPU is not available, this will fall back to normal Estimator on CPU # or GPU. estimator = tf.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.train_batch_size, params={'model_dir': FLAGS.output_dir}) tf.logging.info("***** Running training *****") tf.logging.info(" Batch size = %d", FLAGS.train_batch_size) train_input_fn = input_fn_builder(input_files=input_files, seq_length=FLAGS.max_seq_length, is_training=True) estimator.train(input_fn=train_input_fn, max_steps=FLAGS.num_train_steps)
def main(_): tf.logging.set_verbosity(tf.logging.INFO) news_config = GroverConfig.from_json_file(FLAGS.config_file) tf.gfile.MakeDirs(FLAGS.output_dir) input_files = [] for input_pattern in FLAGS.input_file.split(","): input_files.extend(tf.gfile.Glob(input_pattern)) tf.logging.info("*** Input Files ***") for input_file in input_files: tf.logging.info(" %s" % input_file) tpu_cluster_resolver = None if FLAGS.use_tpu and FLAGS.tpu_name: tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver( FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2 run_config = tf.contrib.tpu.RunConfig( cluster=tpu_cluster_resolver, master=FLAGS.master, model_dir=FLAGS.output_dir, save_checkpoints_steps=FLAGS.iterations_per_loop, keep_checkpoint_max=None, tpu_config=tf.contrib.tpu.TPUConfig( iterations_per_loop=FLAGS.iterations_per_loop, num_shards=FLAGS.num_tpu_cores, per_host_input_for_training=is_per_host)) model_fn = model_fn_builder(news_config, init_checkpoint=FLAGS.init_checkpoint, learning_rate=1e-4, num_train_steps=0, num_warmup_steps=0, use_tpu=FLAGS.use_tpu, ) # If TPU is not available, this will fall back to normal Estimator on CPU # or GPU. estimator = tf.contrib.tpu.TPUEstimator( use_tpu=FLAGS.use_tpu, model_fn=model_fn, config=run_config, train_batch_size=FLAGS.batch_size, eval_batch_size=FLAGS.batch_size, predict_batch_size=FLAGS.batch_size, params={'model_dir': FLAGS.output_dir} ) eval_input_fn = input_fn_builder( input_files=input_files, seq_length=FLAGS.max_seq_length, evaluate_for_fixed_number_of_steps=False, num_cpu_threads=1, is_training=False) result = [x for x in estimator.predict(input_fn=eval_input_fn, yield_single_examples=True)] cats = sorted(result[0].keys()) result_stack = {cat: np.stack([x[cat] for x in result]) for cat in cats} with gcloudwriter(os.path.join(FLAGS.output_dir, FLAGS.validation_name)) as tempfile_name: with h5py.File(tempfile_name, 'w') as h5: for cat, data in result_stack.items(): dtype2use = np.float16 if cat.endswith(('logprobs', 'top_p_required')) else np.uint16 h5.create_dataset(cat, data=data.astype(dtype2use)) h5.create_dataset('model', data=FLAGS.config_file) h5.create_dataset('ckpt', data=FLAGS.init_checkpoint) h5.create_dataset('input_file', data=FLAGS.input_file) # This gives the perplexity of the entire article. if you want to replicate the results of the paper you # might need to do something different to extract the ppl of just the body in particular. ppl_ex = [] for logprobs_i, ids_i in zip(result_stack['gt_logprobs'], result_stack['labels']): # Omit the first token. Keep in mind input_ids is shifted by 1 start_ind = ind_where(ids_i, target=50265, default_value=0) end_ind = ind_where(ids_i, target=50266, default_value=ids_i.shape[0] - 1) ppl_ex.append(logprobs_i[start_ind:end_ind]) ppl_ex = np.concatenate(ppl_ex, 0) print("Article perplexity is {:.3f}".format(np.exp(-np.mean(ppl_ex))), flush=True)