[dino] update code for inverse sigmoid (#7715)

ppdet/modeling/transformers/dino_transformer.py

@@ -257,7 +257,7 @@ class DINOTransformerDecoder(nn.Layer):
 
     def forward(self,
                 tgt,
-                reference_points,
+                ref_points_unact,
                 memory,
                 memory_spatial_shapes,
                 memory_level_start_index,
@@ -272,9 +272,9 @@ class DINOTransformerDecoder(nn.Layer):
 
         output = tgt
         intermediate = []
-        inter_ref_bboxes = []
+        inter_ref_bboxes_unact = []
         for i, layer in enumerate(self.layers):
-            reference_points_input = reference_points.unsqueeze(
+            reference_points_input = F.sigmoid(ref_points_unact).unsqueeze(
                 2) * valid_ratios.tile([1, 1, 2]).unsqueeze(1)
             query_pos_embed = get_sine_pos_embed(
                 reference_points_input[..., 0, :], self.hidden_dim // 2)
@@ -284,19 +284,19 @@ class DINOTransformerDecoder(nn.Layer):
                            memory_spatial_shapes, memory_level_start_index,
                            attn_mask, memory_mask, query_pos_embed)
 
-            inter_ref_bbox = F.sigmoid(bbox_head[i](output) + inverse_sigmoid(
-                reference_points))
+            inter_ref_bbox_unact = bbox_head[i](output) + ref_points_unact
 
             if self.return_intermediate:
                 intermediate.append(self.norm(output))
-                inter_ref_bboxes.append(inter_ref_bbox)
+                inter_ref_bboxes_unact.append(inter_ref_bbox_unact)
 
-            reference_points = inter_ref_bbox.detach()
+            ref_points_unact = inter_ref_bbox_unact.detach()
 
         if self.return_intermediate:
-            return paddle.stack(intermediate), paddle.stack(inter_ref_bboxes)
+            return paddle.stack(intermediate), paddle.stack(
+                inter_ref_bboxes_unact)
 
-        return output, reference_points
+        return output, ref_points_unact
 
 
 @register
@@ -417,7 +417,8 @@ class DINOTransformer(nn.Layer):
         linear_init_(self.enc_output[0])
         xavier_uniform_(self.enc_output[0].weight)
         normal_(self.level_embed.weight)
-        xavier_uniform_(self.tgt_embed.weight)
+        if self.learnt_init_query:
+            xavier_uniform_(self.tgt_embed.weight)
         xavier_uniform_(self.query_pos_head.layers[0].weight)
         xavier_uniform_(self.query_pos_head.layers[1].weight)
         for l in self.input_proj:
@@ -523,7 +524,7 @@ class DINOTransformer(nn.Layer):
 
         # prepare denoising training
         if self.training:
-            denoising_class, denoising_bbox, attn_mask, dn_meta = \
+            denoising_class, denoising_bbox_unact, attn_mask, dn_meta = \
                 get_contrastive_denoising_training_group(gt_meta,
                                             self.num_classes,
                                             self.num_queries,
@@ -532,18 +533,18 @@ class DINOTransformer(nn.Layer):
                                             self.label_noise_ratio,
                                             self.box_noise_scale)
         else:
-            denoising_class, denoising_bbox, attn_mask, dn_meta = None, None, None, None
+            denoising_class, denoising_bbox_unact, attn_mask, dn_meta = None, None, None, None
 
-        target, init_ref_points, enc_topk_bboxes, enc_topk_logits = \
+        target, init_ref_points_unact, enc_topk_bboxes, enc_topk_logits = \
             self._get_decoder_input(
             memory, spatial_shapes, mask_flatten, denoising_class,
-            denoising_bbox)
+            denoising_bbox_unact)
 
         # decoder
-        inter_feats, inter_ref_bboxes = self.decoder(
-            target, init_ref_points, memory, spatial_shapes, level_start_index,
-            self.dec_bbox_head, self.query_pos_head, valid_ratios, attn_mask,
-            mask_flatten)
+        inter_feats, inter_ref_bboxes_unact = self.decoder(
+            target, init_ref_points_unact, memory, spatial_shapes,
+            level_start_index, self.dec_bbox_head, self.query_pos_head,
+            valid_ratios, attn_mask, mask_flatten)
         out_bboxes = []
         out_logits = []
         for i in range(self.num_decoder_layers):
@@ -551,11 +552,11 @@ class DINOTransformer(nn.Layer):
             if i == 0:
                 out_bboxes.append(
                     F.sigmoid(self.dec_bbox_head[i](inter_feats[i]) +
-                              inverse_sigmoid(init_ref_points)))
+                              init_ref_points_unact))
             else:
                 out_bboxes.append(
                     F.sigmoid(self.dec_bbox_head[i](inter_feats[i]) +
-                              inverse_sigmoid(inter_ref_bboxes[i - 1])))
+                              inter_ref_bboxes_unact[i - 1]))
 
         out_bboxes = paddle.stack(out_bboxes)
         out_logits = paddle.stack(out_logits)
@@ -611,7 +612,7 @@ class DINOTransformer(nn.Layer):
                            spatial_shapes,
                            memory_mask=None,
                            denoising_class=None,
-                           denoising_bbox=None):
+                           denoising_bbox_unact=None):
         bs, _, _ = memory.shape
         # prepare input for decoder
         output_memory, output_anchors = self._get_encoder_output_anchors(
@@ -626,12 +627,12 @@ class DINOTransformer(nn.Layer):
         batch_ind = paddle.arange(end=bs, dtype=topk_ind.dtype)
         batch_ind = batch_ind.unsqueeze(-1).tile([1, self.num_queries])
         topk_ind = paddle.stack([batch_ind, topk_ind], axis=-1)
-        topk_coords_unact = paddle.gather_nd(enc_outputs_coord_unact,
-                                             topk_ind)  # unsigmoided.
-        reference_points = enc_topk_bboxes = F.sigmoid(topk_coords_unact)
-        if denoising_bbox is not None:
-            reference_points = paddle.concat([denoising_bbox, enc_topk_bboxes],
-                                             1)
+        reference_points_unact = paddle.gather_nd(enc_outputs_coord_unact,
+                                                  topk_ind)  # unsigmoided.
+        enc_topk_bboxes = F.sigmoid(reference_points_unact)
+        if denoising_bbox_unact is not None:
+            reference_points_unact = paddle.concat(
+                [denoising_bbox_unact, reference_points_unact], 1)
         enc_topk_logits = paddle.gather_nd(enc_outputs_class, topk_ind)
 
         # extract region features
@@ -642,5 +643,5 @@ class DINOTransformer(nn.Layer):
         if denoising_class is not None:
             target = paddle.concat([denoising_class, target], 1)
 
-        return target, reference_points.detach(
+        return target, reference_points_unact.detach(
         ), enc_topk_bboxes, enc_topk_logits

ppdet/modeling/transformers/utils.py

@@ -194,7 +194,7 @@ def get_contrastive_denoising_training_group(targets,
         known_bbox += rand_part * diff
         known_bbox.clip_(min=0.0, max=1.0)
         input_query_bbox = bbox_xyxy_to_cxcywh(known_bbox)
-        input_query_bbox.clip_(min=0.0, max=1.0)
+        input_query_bbox = inverse_sigmoid(input_query_bbox)
 
     class_embed = paddle.concat(
         [class_embed, paddle.zeros([1, class_embed.shape[-1]])])