The code only works on num_queries=1 about dpgn HOT 1 OPEN

DPGN/main.py

Line 316 in b940111

query_node_pred_loss = [

self.pred_loss(query_node_pred_generation, query_label.long()).mean()

For example, 5 way 1 shot, num_queries=1,
query_node_pred_generation has shape of [batch_size, 5, 5]
query_label has shape of [batch_size, 5]

5 way 1 shot, num_queries=2,
query_node_pred_generation have shape of [batch_size, 10, 5]
query_label has shape of [batch_size, 10]

In query_node_pred_generation, which dimension is the class (i.e., N ways)?

Dear Jing Li,

It is very unfortunate that the codebase is designed only for the setting of num_queries = 1 in the beginning.

We are sorry that the implementation of auxiliary loss (query_node_pred_loss) confuses many readers. Therefore, we decided to propose a modified version that follows most people's intuition. The accuracy for the result of 5-way 1-shot, resnet12, mini-imagenet is still around 67.70±0.52 (almost the same as reported accuracy on the paper, 67.77).

Please replace the content at line 316-319 of main.py by the code below:

        query_node_pred_loss = []
        for query_node_pred_generation in query_node_pred_generations_:
            temp_query_node_pred_generation = query_node_pred_generation.contiguous().view(-1, query_node_pred_generation.shape[-1])
            temp_query_label = query_label.long().contiguous().view(-1, )
            temp_loss = self.pred_loss(temp_query_node_pred_generation, temp_query_label).mean()
            query_node_pred_loss.append(temp_loss)

Also, some configs need to be revised:

config['num_loss_generation'] = 7
train_opt['dec_lr'] = 18000

Let us know if you have any other confusion when using the DPGN.

Yours,
DPGN Team

from dpgn.