8. (666cc456) · Commits · wesenberg / rl-sentence-compression-wesenberg

Baseline.py

+18 −22

Original line number	Diff line number	Diff line
		@@ -14,7 +14,7 @@ from transformers import AutoTokenizer

		from wesenberg.Konstanten import MODEL_PATH, DEVICE_MODEL_PREDICTION, TOKENIZER, ROOT_PATH, \
		NAME_SOURCE_DATA, NAME_GOLD_DATA, LIST_MODEL_NAME, ALL_RESULTS_PATH, NAME_RESULTS_TXT, HUGGINGFACE_PATH, \
		LATEST_MODEL, BATCH_SIZE_ELEMENTS, CALC_ELEMENTS, SAFE_C_PROFILE_BASELINE
		LATEST_MODEL, BATCH_SIZE_ELEMENTS, CALC_ELEMENTS, SAFE_C_PROFILE_BASELINE, USE_AMR

		from rouge_score import rouge_scorer

		@@ -37,12 +37,11 @@ def get_sub_lists(list_input, n):
		lists = []
		tmp = []
		for i, data in enumerate(list_input):
		if i % n == 0:
		if i != 0:
		if i % n == 0 and i != 0:
		lists.append(tmp)
		tmp = []
		tmp.append(data)
		if len(list_input) % n != 0:
		if tmp not in lists:
		lists.append(tmp)
		return lists

		@@ -59,8 +58,8 @@ if __name__ == '__main__':
		cp = cProfile.Profile()
		cp.enable()

		if DEVICE_MODEL_PREDICTION == "cuda":
		torch.cuda.empty_cache()
		# if DEVICE_MODEL_PREDICTION == "cuda":
		# torch.cuda.empty_cache()
		# loads huggingface dataset

		sources = read(HUGGINGFACE_PATH + NAME_SOURCE_DATA)
		@@ -70,10 +69,6 @@ if __name__ == '__main__':
		sources = sources[0:CALC_ELEMENTS]
		gold = gold[0:CALC_ELEMENTS]

		list_rouge = []
		for (s, g) in zip(sources, gold):
		list_rouge.append(rouge_scorer.score(g, s))

		list_rouge_model_list = {}
		list_string_results_model = {}
		time_model_list = {}
		@@ -85,7 +80,7 @@ if __name__ == '__main__':
		# 1. Textzusammenfassung
		lists_sources = get_sub_lists(sources, BATCH_SIZE_ELEMENTS)
		summaries = []
		print("Start to predict")
		print("Start to predict: " + model_name)
		for i, tmp_sources in enumerate(lists_sources):
		summaries = summaries + predict_summaries(model_name, tmp_sources)
		print(str((BATCH_SIZE_ELEMENTS * i) + len(tmp_sources)) + " / " + str(len(sources)))
		@@ -94,18 +89,19 @@ if __name__ == '__main__':
		time_model_list[model_name] = str(time_model)

		list_rouge_model = []
		for (so, su) in zip(sources, summaries):
		list_rouge_model.append(rouge_scorer.score(su, so))
		for (go, su) in zip(gold, summaries):
		list_rouge_model.append(rouge_scorer.score(target=go, prediction=su))
		list_rouge_model_list[model_name] = list_rouge_model

		if USE_AMR:
		amr_class = AmrClass()
		list_string_results_stog = amr_class.calc_all_amr(sources=sources, summaries=summaries, gold=gold,
		model_count=model_count, model_name=model_name)

		list_string_results_model[model_name] = list_string_results_stog

		util.write_string_to_file(path=ROOT_PATH + ALL_RESULTS_PATH, name=NAME_RESULTS_TXT,
		list_string_results_model=list_string_results_model,
		time_model_list=time_model_list, list_rouge_model_list=list_rouge_model_list, list_rouge=list_rouge)
		time_model_list=time_model_list, list_rouge_model_list=list_rouge_model_list)

		if SAFE_C_PROFILE_BASELINE:
		cp.disable()

config/gigaword-L10_test.json

0 → 100644

+37 −0

Original line number	Diff line number	Diff line
		{
		"loader": "loaders/gigaword.py",
		"dataset": "data/train-data/gigaword",
		"indices": "data/train-data/gigaword/indices.npy",
		"model_dir": "data/models/gigaword-L10_test",
		"verbose": true,
		"print_every": 1,
		"eval_every": 50,
		"save_every": 50,
		"max_val_steps": 8000,
		"max_train_seconds": null,
		"max_train_steps": 8000,
		"batch_size": 4,
		"learning_rate": 1e-05,
		"k_samples": 100,
		"sample_aggregation": "max",
		"loss": "pgb",
		"encoder_model_id": "distilroberta-base",
		"rewards": {
		"Fluency": {
		"weight": 1,
		"type": "masked",
		"model_id": "distilroberta-base",
		"max_score": 40.0,
		"norm": "max"
		},
		"BiEncoderSimilarity": {
		"weight": 1,
		"model_id": "all-distilroberta-v1"
		},
		"GaussianLength": {
		"weight": 1,
		"mean": 10,
		"std": 3.2
		}
		}
		}

config/gigaword-L8_neu.json

+2 −2

Original line number	Diff line number	Diff line
		@@ -7,9 +7,9 @@
		"print_every": 1,
		"eval_every": 50,
		"save_every": 50,
		"max_val_steps": 8000,
		"max_val_steps": 10000,
		"max_train_seconds": null,
		"max_train_steps": 8000,
		"max_train_steps": 10000,
		"batch_size": 4,
		"learning_rate": 1e-05,
		"k_samples": 100,

config/gigaword-L8_test.json

0 → 100644

+37 −0

Original line number	Diff line number	Diff line
		{
		"loader": "loaders/gigaword.py",
		"dataset": "data/train-data/gigaword",
		"indices": "data/train-data/gigaword/indices.npy",
		"model_dir": "data/models/gigaword-L8_test",
		"verbose": true,
		"print_every": 1,
		"eval_every": 50,
		"save_every": 50,
		"max_val_steps": 8000,
		"max_train_seconds": null,
		"max_train_steps": 8000,
		"batch_size": 4,
		"learning_rate": 1e-05,
		"k_samples": 100,
		"sample_aggregation": "max",
		"loss": "pgb",
		"encoder_model_id": "distilroberta-base",
		"rewards": {
		"Fluency": {
		"weight": 1,
		"type": "masked",
		"model_id": "distilroberta-base",
		"max_score": 40.0,
		"norm": "max"
		},
		"BiEncoderSimilarity": {
		"weight": 1,
		"model_id": "all-distilroberta-v1"
		},
		"GaussianLength": {
		"weight": 1,
		"mean": 8,
		"std": 3.2
		}
		}
		}

scrl/rewards.py

+17 −17

Original line number	Diff line number	Diff line
		@@ -11,9 +11,9 @@ from pprint import pprint
		from rouge_score import rouge_scorer

		from wesenberg import util
		from wesenberg.Konstanten import STOG_MODEL_GSII, BACKUP_REWARD_REGION_PATH, SAFE_REWARDS, STOG_MODEL_TRAIN
		from wesenberg.Konstanten import BACKUP_REWARD_REGION_PATH, SAFE_REWARDS, STOG_MODEL_TRAIN
		from wesenberg.SmatchClass import calc_smatch_to_r
		from wesenberg.util import safe_txt_list
		from wesenberg.util import safe_reward_list, safe_fluency_rewards

		ROUGE_TYPES = ["rouge1", "rouge2", "rougeL"]
		rouge_scorer = rouge_scorer.RougeScorer(
		@@ -24,14 +24,14 @@ rouge_scorer = rouge_scorer.RougeScorer(
		safe_score_region = {}


		def do_safe_score_region():
		def do_safe_score_region(step, reward_lists):
		if SAFE_REWARDS:
		safe_fluency_rewards(path=BACKUP_REWARD_REGION_PATH, name="mean_", reward_lists=reward_lists, step=step)
		keys = safe_score_region.keys()
		for key in keys:
		"""old = load_txt_list(BACKUP_REWARD_REGION_PATH, key)
		if old is not None:
		safe_score_region[key] = safe_score_region[key] + old"""
		safe_txt_list(path=BACKUP_REWARD_REGION_PATH, name=key, safe_list=safe_score_region[key], n=100)
		safe_reward_list(path=BACKUP_REWARD_REGION_PATH, name=key, safe_list=safe_score_region[key], step=step, n=50)
		safe_score_region[key] = []



		def load_rewards(args):
		@@ -139,7 +139,7 @@ class Fluency:
		i_score = i_score_minmax
		scores.append(i_score)
		if SAFE_REWARDS:
		safe_score_region[self.name].append(scores)
		safe_score_region[self.name].append((sources, summaries, scores))
		return scores


		@@ -167,7 +167,7 @@ class BiEncoderSimilarity:
		)[0, 0].item()
		scores.append(score)
		if SAFE_REWARDS:
		safe_score_region[self.name].append(scores)
		safe_score_region[self.name].append((sources, summaries, scores))
		return scores


		@@ -190,7 +190,7 @@ class CrossEncoderSimilarity:
		])

		if SAFE_REWARDS:
		safe_score_region[self.name].append(scores.tolist())
		safe_score_region[self.name].append((sources, summaries, scores.tolist()))
		return scores.tolist()


		@@ -252,7 +252,7 @@ class SelectedTokenSimilarity:
		scores.append(score)

		if SAFE_REWARDS:
		safe_score_region[self.name].append(scores)
		safe_score_region[self.name].append((sources, summaries, scores))
		return scores


		@@ -284,7 +284,7 @@ class NLIReward:
		for i, r in enumerate(rewards)
		]
		if SAFE_REWARDS:
		safe_score_region[self.name].append(rewards)
		safe_score_region[self.name].append((sources, summaries, rewards))
		return rewards


		@@ -307,7 +307,7 @@ class GaussianLength:
		for l in lens
		]
		if SAFE_REWARDS:
		safe_score_region[self.name].append(scores)
		safe_score_region[self.name].append((sources, summaries, scores))
		return scores


		@@ -333,7 +333,7 @@ class GaussianCR:
		for ratio in ratios
		]
		if SAFE_REWARDS:
		safe_score_region[self.name].append(scores)
		safe_score_region[self.name].append((sources, summaries, scores))
		return scores


		@@ -359,7 +359,7 @@ class NoDaysReward:
		score = 1.
		scores.append(score)
		if SAFE_REWARDS:
		safe_score_region[self.name].append(scores)
		safe_score_region[self.name].append((sources, summaries, scores))
		return scores


		@@ -383,7 +383,7 @@ class RougeReward:
		score = rouge_scores[self.rouge_type].fmeasure
		scores.append(score)
		if SAFE_REWARDS:
		safe_score_region[self.name].append(scores)
		safe_score_region[self.name].append((sources, summaries, scores))
		return scores


		@@ -406,5 +406,5 @@ class AmrReward:
		score = 0
		scores.append(score)
		if SAFE_REWARDS:
		safe_score_region[self.name].append(scores)
		safe_score_region[self.name].append((sources, summaries, scores))
		return scores