def predict_for_evalai(self, dataset_type):
reporter = self.dataset_loader.get_test_reporter(dataset_type)
with torch.no_grad():
self.model.eval()
message = "Starting {} inference for evalai".format(dataset_type)
self.writer.write(message)
while reporter.next_dataset():
dataloader = reporter.get_dataloader()
for batch in tqdm(dataloader):
prepared_batch = reporter.prepare_batch(batch)
model_output = self.model(prepared_batch)
prepared_batch.update(model_output)
if visualization_flag:
self._summarize_visualization(prepared_batch)
report = Report(prepared_batch, model_output)
reporter.add_to_report(report)
self.writer.write("Finished predicting")
if visualization_flag:
# with open('textcap_visualization.json', 'w+') as f:
# json.dump(val_predictions, f)
torch.save(val_predictions, 'textcap_visualization.json')
print('success save the visulaization file, Break')
self.model.train()
def _forward_pass(self, batch):
prepared_batch = self.dataset_loader.prepare_batch(batch)
self.profile("Batch prepare time")
# Arguments should be a dict at this point
model_output = self.model(prepared_batch)
report = Report(prepared_batch, model_output)
self.profile("Forward time")
return report
def _forward_pass(self, batch):
prepared_batch = self.task_loader.prepare_batch(batch)
self.profile("Batch prepare time")
model_output = self.model(prepared_batch)
report = Report(prepared_batch, model_output)
self.profile("Forward time")
return report
def _forward_pass(self, batch):
prepared_batch = self.task_loader.prepare_batch(batch)
self.profile("Batch prepare time")
# Arguments should be a dict at this point
model_output = self.model(prepared_batch) # a dict of losses, metrics, scores, and att
report = Report(prepared_batch, model_output)
self.profile("Forward time")
return report, (model_output["att"] if "att" in model_output.keys() else None)
def _forward_pass(self, batch):
prepared_batch = self.task_loader.prepare_batch(batch)
self.profile("Batch prepare time")
# Arguments should be a dict at this point
model_output = self.model(prepared_batch)
print('model_output["captions"]', model_output['captions'])
report = Report(prepared_batch, model_output)
self.profile("Forward time")
return report
def predict_for_evalai(self, dataset_type):
reporter = self.task_loader.get_test_reporter(dataset_type)
with torch.no_grad():
self.model.eval()
message = "Starting {} inference for evalai".format(dataset_type)
self.writer.write(message)
while reporter.next_dataset():
dataloader = reporter.get_dataloader()
for batch in tqdm(dataloader):
prepared_batch = reporter.prepare_batch(batch)
model_output = self.model(prepared_batch)
report = Report(prepared_batch, model_output)
reporter.add_to_report(report)
self.writer.write("Finished predicting")
self.model.train()