def visualize(vectorizer, data=None):
data = get_data(p.VAL_FILE,
vectorizer,
with_oov=p.POINTER_GEN,
aspect_file=p.ASPECT_FILE) if data is None else data
model = new_model(vectorizer, data.dataset.aspects).eval()
with open(p.MODEL_FILE, 'rb') as modelfile:
model.load_state_dict(pkl.load(modelfile))
batch = data[271:271 + p.DECODING_BATCH_SIZE]
aspect_results = summarize(batch, model, beam_size=p.BEAM_SIZE)
print_batch(batch, [r[0] for r in aspect_results], vectorizer,
model.aspects)
vis(p.VISUALIZATION_FILE,
batch, [r[0] for r in aspect_results],
vectorizer,
model.aspects,
0,
0,
pointer_gen=p.POINTER_GEN)
def evaluate(vectorizer, data=None):
data = get_data(p.VAL_FILE,
vectorizer,
with_oov=p.POINTER_GEN,
aspect_file=p.ASPECT_FILE) if data is None else data
model = new_model(vectorizer, data.dataset.aspects).eval()
with open(p.MODEL_FILE, 'rb') as modelfile:
model.load_state_dict(pkl.load(modelfile))
text_path = p.TEXT_PATH
produce_summary_files(data,
p.DECODING_BATCH_SIZE,
vectorizer,
model,
text_path,
beam_size=p.BEAM_SIZE,
max_num_batch=None)
#run_rouge_2(save_to=os.path.join(p.CHECKPOINT_PATH, 'rouge_scores.txt') if p.CHECKPOINT_PATH is not None else None)
run_rouge_1(os.path.join(text_path, 'system'),
os.path.join(text_path, 'reference'),
save_to=os.path.join(text_path, 'rouge_scores.txt'))
def train(vectorizer, data=None, val=None):
data = get_data(p.DATA_FILE,
vectorizer,
with_oov=p.POINTER_GEN,
aspect_file=p.ASPECT_FILE) if data is None else data
val = get_data(p.VAL_FILE,
vectorizer,
with_oov=p.POINTER_GEN,
aspect_file=p.ASPECT_FILE) if val is None else val
if p.CONTINUE_FROM_CHECKPOINT:
# check if all of the proper files exist
if not TrainingTracker.valid_checkpoint(p.CHECKPOINT_PATH):
print(
"Cannot continue from checkpoint in \"" + p.CHECKPOINT_PATH +
"\" because not all of the proper files exist; restarting training."
)
p.CONTINUE_FROM_CHECKPOINT = False
print("Saving parameters to file again.")
p.save_params(
os.path.join(p.CHECKPOINT_PATH, 'train_param_info.txt'))
else:
print("Loading from the last checkpoint in \"" +
p.CHECKPOINT_PATH +
"\"; model parameters must match the saved model state.")
if p.NEW_EPOCH:
print("Starting from a new epoch.")
else:
print(
"Continuing from the same place in the epoch; this expects the same datafile."
)
model = new_model(vectorizer, data.dataset.aspects).train()
if p.CONTINUE_FROM_CHECKPOINT:
TrainingTracker.load_model_state_(model, p.CHECKPOINT_PATH)
# optimizer = torch.optim.Adam(model.parameters(), lr=p.LEARNING_RATE)
optimizer = torch.optim.Adagrad(
model.parameters(),
lr=p.LEARNING_RATE,
initial_accumulator_value=p.INITIAL_ACCUMULATOR_VALUE,
)
if p.CONTINUE_FROM_CHECKPOINT:
TrainingTracker.load_optimizer_state(optimizer, p.CHECKPOINT_PATH)
model_manip = ModelManipulator(model,
optimizer,
aspect_summarizer_loss,
aspect_summarizer_error,
grad_mod=clip_grad_norm,
no_nan_grad=True)
with torch.autograd.set_detect_anomaly(p.DETECT_ANOMALY):
train_stats, val_stats = model_manip.train(
data,
p.BATCH_SIZE,
p.NUM_EPOCHS,
dataset_val=val,
stats_every=10,
verbose_every=10,
checkpoint_every=10,
checkpoint_path=p.CHECKPOINT_PATH,
restart=not p.CONTINUE_FROM_CHECKPOINT,
new_epoch=p.NEW_EPOCH,
max_steps=p.MAX_TRAINING_STEPS,
save_whole_model=False)
if p.CHECKPOINT_PATH is not None:
plot_checkpoint(p.CHECKPOINT_PATH,
figure_name='plot',
show=False,
average_over=p.AVERAGE_OVER)
else:
plot_learning_curves(training_values=train_stats,
validation_values=val_stats,
figure_name=os.path.join(p.TRAINING_PLOTS_PATH,
'plot'),
show=False,
average_over=p.AVERAGE_OVER)
with_coverage=True),
dict(max_training_steps=233000,
max_text_length=400,
max_summary_length=100,
with_coverage=True),
]
if __name__ == '__main__':
checkpoint_path = os.path.join(CHECKPOINTS_FOLDER, 'checkpoint')
vectorizer = setup(checkpoint_path=checkpoint_path,
device=DEVICE,
pointer_gen=POINTER_GEN,
use_transformer=USE_TRANSFORMER,
mode='train')
data = get_data(p.DATA_FILE,
vectorizer,
with_oov=p.POINTER_GEN,
aspect_file=p.ASPECT_FILE)
val = get_data(p.VAL_FILE,
vectorizer,
with_oov=p.POINTER_GEN,
aspect_file=p.ASPECT_FILE)
for i, params in enumerate(sections):
print(('starting section %i:\n' % (i + 1)) + str(params))
set_params(**params, continue_from_checkpoint=True)
try:
train(vectorizer, data=data, val=val)
subprocess.run([
'cp', '-r', checkpoint_path,
os.path.join(CHECKPOINTS_FOLDER, 'checkpoint%i' % (i + 1))
])
except StopEarlyWithoutSavingException: