The issue of query loss not converging

[JiaJia075]

MD-DETR/engine.py

Lines 95 to 115 in 125e771

	if self.args.use_prompts:
	with torch.no_grad():
	outputs = self.model(pixel_values=pixel_values, pixel_mask=pixel_mask, labels=labels, train=False, task_id=self.task_id)
	if not self.args.local_query:
	query = outputs.last_hidden_state.mean(dim=1)
	else:
	query = outputs.last_hidden_state
	outputs_without_aux = {k: v for k, v in outputs.items() if k != "auxiliary_outputs" and k != "enc_outputs"}
	# Retrieve the matching between the outputs of the last layer and the targets
	indices = self.model.matcher(outputs_without_aux, labels)
	one_hot_proposals = torch.zeros((len(labels),300)).to(self.device)
	for i,ind in enumerate(indices):
	for j in ind[0]:
	one_hot_proposals[i][j] = 1
	query_wt = self.model.model.prompts.query_tf(query.view(query.shape[0],-1))
	query_loss = F.cross_entropy(query_wt, one_hot_proposals)

I found that the query loss is within the torch.no_grad(): block, which prevents the query loss from being included in the computation graph. As a result, the query loss cannot converge.

[JiaJia075]

We also found that the query tf fully connected layer is quite large, with nearly 90M parameters, which is more than 10 times the size of the prompt, and undoubtedly unacceptable.

[JiaJia075]

I tried modifying the query_tf used for weighted averaging of queries, but encountered some strange phenomena. This module was not trained, yet deleting it caused a drop in precision. Furthermore, training this module also led to a decrease in precision.

The original output precision was: task123map0.5 = 0.52, task4map0.5 = 0.52.

1、After directly deleting it, the precision of the old tasks dropped: task123map0.5 = 0.33, task4map0.5 = 0.5.

2、Modifying the training logic of query_tf so that it gets recorded in the computation graph and has gradients caused the precision of both new and old tasks to drop: task123map0.5 = 0.18, task4map0.5 = 0.43.

3、Detaching the input of query_tf and only training query_tf itself improved the training for the new task, suggesting that this loss affected the optimization of other losses. Moreover, more precise weight prediction did not help maintain the precision for the old tasks and actually had a negative impact: task123map0.5 = 0.29, task4map0.5 = 0.5.

4、Keeping query_tf but removing the code that calculates query_loss resulted in almost no change in the published precision: task123map0.5 = 0.52, task4map0.5 = 0.48.

This is quite perplexing. A module that wasn't trained still causes a significant drop in precision when deleted.