fix bugs

ppdet/modeling/vl/embedder/__init__.py

@@ -21,6 +21,7 @@ import paddle.nn as nn
 import paddle.nn.functional as F
 
 from ppdet.core.workspace import register
+from .clip import *
 
 __all__ = ['ClipImageTextEmbedder']
 

ppdet/modeling/vl/embedder/clip/__init__.py

@@ -12,6 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .models import ModifiedResNet, TextEncoder, VisionTransformer
+from .models import ModifiedResNet, TextEncoder, ViT
 from .layers import LayerNorm, QuickGELU, AttentionPool2D
 from .clip import CLIP

ppdet/modeling/vl/embedder/clip/clip.py

@@ -31,7 +31,7 @@ from ppdet.modeling.layers import MultiHeadAttention
 from ppdet.modeling.initializer import zeros_, normal_
 from ppdet.core.workspace import register
 
-from .models import ModifiedResNet, VisionTransformer, TextEncoder
+from .models import ModifiedResNet, ViT, TextEncoder
 
 
 @register

ppdet/modeling/vl/embedder/clip/layers.py

@@ -84,7 +84,7 @@ class Bottleneck(nn.Layer):
         return out
 
 
-class AttentionPool2D(nn.Module):
+class AttentionPool2D(nn.Layer):
     def __init__(self, spacial_dim, embed_dim, num_heads, output_dim):
         super().__init__()
         # TODO: need check whether it is consistent with torch or not
@@ -151,10 +151,9 @@ class ResidualAttentionBlock(nn.Layer):
 
         self.attn = MultiHeadAttention(d_model, n_head)
         self.ln_1 = LayerNorm(d_model)
-        self.mlp = nn.Sequential(
-            OrderedDict([("c_fc", nn.Linear(d_model, d_model * 4)), (
-                "gelu", QuickGELU()), ("c_proj", nn.Linear(d_model * 4, d_model)
-                                       )]))
+        self.mlp = nn.Sequential(("c_fc", nn.Linear(d_model, d_model * 4)),
+                                 ("gelu", QuickGELU()),
+                                 ("c_proj", nn.Linear(d_model * 4, d_model)))
         self.ln_2 = LayerNorm(d_model)
         self.attn_mask = attn_mask
         self.droplayer_p = droplayer_p
@@ -192,6 +191,7 @@ class Transformer(nn.Layer):
         super().__init__()
         self.width = width
         self.layers = layers
+        self.stochastic_droplayer_rate = stochastic_droplayer_rate
         blocks = []
         for i in range(self.layers):
             droplayer_p = (i / max(self.layers - 1,

ppdet/modeling/vl/embedder/clip/models.py

@@ -32,7 +32,7 @@ from ppdet.core.workspace import register
 
 from .layers import *
 
-__all__ = ['ModifiedResNet', 'VisionTransformer', 'TextEncoder']
+__all__ = ['ModifiedResNet', 'ViT', 'TextEncoder']
 
 
 @register
@@ -105,7 +105,7 @@ class ModifiedResNet(nn.Layer):
 
 
 @register
-class VisionTransformer(nn.Layer):
+class ViT(nn.Layer):
     def __init__(self,
                  input_resolution,
                  patch_size,
@@ -115,6 +115,7 @@ class VisionTransformer(nn.Layer):
                  output_dim=None,
                  stochastic_droplayer_rate=0.0):
         super().__init__()
+        self.width = width
         self.input_resolution = input_resolution
         self.output_dim = output_dim
         self.conv1 = nn.Conv2D(
@@ -122,7 +123,7 @@ class VisionTransformer(nn.Layer):
             out_channels=width,
             kernel_size=patch_size,
             stride=patch_size,
-            bias=False)
+            bias_attr=False)
         scale = width**-0.5
         self.class_embedding = self.create_parameter(
             shape=[width], attr=ParamAttr(initializer=Normal(std=scale)))
@@ -157,9 +158,14 @@ class VisionTransformer(nn.Layer):
 
 @register
 class TextEncoder(nn.Layer):
-    def __init__(self, context_length, vocab_size, transformer_width,
-                 transformer_heads, transformer_layers,
-                 stochastic_droplayer_rate):
+    def __init__(self,
+                 embed_dim,
+                 context_length,
+                 vocab_size,
+                 transformer_width,
+                 transformer_heads,
+                 transformer_layers,
+                 stochastic_droplayer_rate=0.0):
         super().__init__()
         self.context_length = context_length
 
@@ -178,8 +184,6 @@ class TextEncoder(nn.Layer):
         self.ln_final = LayerNorm(transformer_width)
         self.text_projection = nn.Linear(
             transformer_width, embed_dim, bias_attr=False)
-        self.logit_scale = self.create_parameter(
-            shape=[], attr=ParamAttr(initializer=Constant(np.log(1. / 0.07))))
 
     def build_attention_mask(self):
         # lazily create causal attention mask, with full attention between the vision tokens

ppdet/modeling/vl/head/__init__.py

@@ -11,3 +11,5 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
+from .owl_vit_head import *
\ No newline at end of file

ppdet/modeling/vl/head/owl_vit_head.py

@@ -22,6 +22,7 @@ import paddle
 import paddle.nn as nn
 import paddle.nn.functional as F
 from ppdet.modeling.ops import get_act_fn
+from ppdet.core.workspace import register
 
 from ..utils import compute_box_bias
 
@@ -46,12 +47,13 @@ class PredictorMLP(nn.Layer):
                  in_dim,
                  out_dim,
                  num_layers,
-                 mlp_dim,
-                 hidden_activation,
+                 mlp_dim=None,
+                 hidden_activation='gelu',
                  out_activation=None):
         super().__init__()
 
         layers = []
+        mlp_dim = in_dim if mlp_dim is None else mlp_dim
         for _ in range(num_layers - 1):
             layers.append(nn.Linear(in_dim, mlp_dim))
             in_dim = mlp_dim
@@ -138,7 +140,6 @@ class OWLViTHead(nn.Layer):
         self.class_head = class_head
         self.bbox_head = bbox_head
         self.box_bias = box_bias
-        self.matcher = matcher
         self.loss = loss
 
     def box_predictor(self, image_features, feature_map):

ppdet/modeling/vl/loss/__init__.py

@@ -11,3 +11,5 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
+from .owl_vit_loss import *
\ No newline at end of file

ppdet/modeling/vl/loss/owl_vit_loss.py

@@ -32,7 +32,7 @@ class OWLViTLoss(nn.Layer):
     __inject__ = ['HungarianMatcher']
 
     def __init__(self,
-                 num_classes,
+                 num_classes=80,
                  matcher='HungarianMatcher',
                  normalization='per_example',
                  loss_coeff=None,

ppdet/modeling/vl/tokenizer/simple_tokenizer.py

@@ -21,6 +21,7 @@ from __future__ import print_function
 import gzip
 import html
 import os
+import functools
 from functools import lru_cache
 
 import ftfy