Correct spelling of PARSeq

keras-team · Jan 14, 2025 · 7e5794e · 7e5794e
1 parent 055bddc
commit 7e5794e
Show file tree

Hide file tree

Showing 5 changed files with 59 additions and 25 deletions.
diff --git a/keras_hub/api/layers/__init__.py b/keras_hub/api/layers/__init__.py
@@ -53,7 +53,7 @@
     PaliGemmaImageConverter,
 )
 from keras_hub.src.models.parseq.parseq_image_converter import (
-    ParseQImageConverter,
+    PARSeqImageConverter,
 )
 from keras_hub.src.models.resnet.resnet_image_converter import (
     ResNetImageConverter,

diff --git a/keras_hub/api/models/__init__.py b/keras_hub/api/models/__init__.py
@@ -250,9 +250,9 @@
 from keras_hub.src.models.pali_gemma.pali_gemma_tokenizer import (
     PaliGemmaTokenizer,
 )
-from keras_hub.src.models.parseq.parseq_backbone import ParseQBackbone
-from keras_hub.src.models.parseq.parseq_ocr import ParseQOCR
-from keras_hub.src.models.parseq.parseq_preprocessor import ParseQPreprocessor
+from keras_hub.src.models.parseq.parseq_backbone import PARSeqBackbone
+from keras_hub.src.models.parseq.parseq_ocr import PARSeqOCR
+from keras_hub.src.models.parseq.parseq_preprocessor import PARSeqPreprocessor
 from keras_hub.src.models.phi3.phi3_backbone import Phi3Backbone
 from keras_hub.src.models.phi3.phi3_causal_lm import Phi3CausalLM
 from keras_hub.src.models.phi3.phi3_causal_lm_preprocessor import (

diff --git a/keras_hub/src/models/parseq/parseq_backbone.py b/keras_hub/src/models/parseq/parseq_backbone.py
@@ -46,7 +46,7 @@ def __init__(
             drop_rate,
             attn_drop_rate,
             drop_path_rate,
-            name="parseq_vit"
+            name="parseq_vit",
         )
 
         # Configure the decoder layer parameters
@@ -65,7 +65,9 @@ def __init__(
         )
 
         # Embedding layer for input tokens
-        text_embed = TokenEmbedding(out_channels, embed_dim, name="parseq_embed")
+        text_embed = TokenEmbedding(
+            out_channels, embed_dim, name="parseq_embed"
+        )
 
         dropout_layer = layers.Dropout(dropout, name="parseq_dropout")
         # Output head to project decoder outputs to token probabilities

diff --git a/keras_hub/src/models/parseq/parseq_head.py b/keras_hub/src/models/parseq/parseq_head.py
@@ -15,18 +15,34 @@ def __init__(
     ):
         super(DecoderLayer, self).__init__()
         self.self_attn = layers.MultiHeadAttention(
-            num_heads=nhead, key_dim=d_model//nhead, dropout=dropout, name=f"{name}_sattn"
+            num_heads=nhead,
+            key_dim=d_model // nhead,
+            dropout=dropout,
+            name=f"{name}_sattn",
         )
         self.cross_attn = layers.MultiHeadAttention(
-            num_heads=nhead, key_dim=d_model//nhead, dropout=dropout, name=f"{name}_xattn"
+            num_heads=nhead,
+            key_dim=d_model // nhead,
+            dropout=dropout,
+            name=f"{name}_xattn",
+        )
+        self.linear1 = layers.Dense(
+            dim_feedforward, activation=activation, name=f"{name}_dense1"
         )
-        self.linear1 = layers.Dense(dim_feedforward, activation=activation, name=f"{name}_dense1")
         self.dropout = layers.Dropout(dropout, name=f"{name}_dropout")
         self.linear2 = layers.Dense(d_model, name=f"{name}_dense2")
-        self.norm1 = layers.LayerNormalization(epsilon=LAYERNORM_EPSILON, name=f"{name}_norm1")
-        self.norm2 = layers.LayerNormalization(epsilon=LAYERNORM_EPSILON, name=f"{name}_norm2")
-        self.norm_q = layers.LayerNormalization(epsilon=LAYERNORM_EPSILON, name=f"{name}_normq")
-        self.norm_c = layers.LayerNormalization(epsilon=LAYERNORM_EPSILON, name=f"{name}_normc")
+        self.norm1 = layers.LayerNormalization(
+            epsilon=LAYERNORM_EPSILON, name=f"{name}_norm1"
+        )
+        self.norm2 = layers.LayerNormalization(
+            epsilon=LAYERNORM_EPSILON, name=f"{name}_norm2"
+        )
+        self.norm_q = layers.LayerNormalization(
+            epsilon=LAYERNORM_EPSILON, name=f"{name}_normq"
+        )
+        self.norm_c = layers.LayerNormalization(
+            epsilon=LAYERNORM_EPSILON, name=f"{name}_normc"
+        )
         self.dropout1 = layers.Dropout(dropout, name=f"{name}_dropout1")
         self.dropout2 = layers.Dropout(dropout, name=f"{name}_dropout2")
         self.dropout3 = layers.Dropout(dropout, name=f"{name}_dropout3")
@@ -116,11 +132,13 @@ def __init__(
                 dim_feedforward=dim_feedforward,
                 dropout=dropout,
                 activation=activation,
-                name=f"{name}_layer{i}"
+                name=f"{name}_layer{i}",
             )
             for i in range(num_layers)
         ]
-        self.norm = layers.LayerNormalization(epsilon=LAYERNORM_EPSILON, name=f"{name}_norm")
+        self.norm = layers.LayerNormalization(
+            epsilon=LAYERNORM_EPSILON, name=f"{name}_norm"
+        )
 
     def call(
         self,

diff --git a/keras_hub/src/models/parseq/parseq_vit.py b/keras_hub/src/models/parseq/parseq_vit.py
@@ -61,7 +61,9 @@ def __init__(
             hidden_features if hidden_features is not None else in_features
         )
 
-        self.fc1 = layers.Dense(hidden_features, activation=act, name=f"{name}_dense1")
+        self.fc1 = layers.Dense(
+            hidden_features, activation=act, name=f"{name}_dense1"
+        )
         self.drop1 = layers.Dropout(drop, name=f"{name}_dropout1")
         self.fc2 = layers.Dense(out_features, name=f"{name}_dense2")
         self.drop2 = layers.Dropout(drop, name=f"{name}_dropout2")
@@ -85,7 +87,13 @@ class PatchEmbed(layers.Layer):
     """
 
     def __init__(
-        self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, name="patchembed", **kwargs
+        self,
+        img_size=224,
+        patch_size=16,
+        in_chans=3,
+        embed_dim=768,
+        name="patchembed",
+        **kwargs,
     ):
         super().__init__(**kwargs, name=name)
         img_size = to_2tuple(img_size)
@@ -102,7 +110,7 @@ def __init__(
             kernel_size=patch_size,
             strides=patch_size,
             padding="valid",
-            name=f"{name}_conv"
+            name=f"{name}_conv",
         )
 
     def call(self, x):
@@ -146,7 +154,7 @@ def __init__(
             key_dim=dim // num_heads,
             dropout=attn_drop,
             use_bias=qkv_bias,
-            name=f"{name}_mha"
+            name=f"{name}_mha",
         )
         self.proj_drop = layers.Dropout(proj_drop, name=f"{name}_dropout")
 
@@ -174,27 +182,31 @@ def __init__(
     ):
         super().__init__(**kwargs, name=name)
 
-        self.norm1 = layers.LayerNormalization(epsilon=LAYERNORM_EPSILON, name=f"{name}_norm1")
+        self.norm1 = layers.LayerNormalization(
+            epsilon=LAYERNORM_EPSILON, name=f"{name}_norm1"
+        )
         self.attn = Attention(
             dim=dim,
             num_heads=num_heads,
             qkv_bias=qkv_bias,
             attn_drop=attn_drop,
             proj_drop=drop,
-            name=f"{name}_attn"
+            name=f"{name}_attn",
         )
         self.drop_path = (
             DropPath(drop_path) if drop_path > 0.0 else (lambda x: x)
         )
 
-        self.norm2 = layers.LayerNormalization(epsilon=LAYERNORM_EPSILON, name=f"{name}_norm2")
+        self.norm2 = layers.LayerNormalization(
+            epsilon=LAYERNORM_EPSILON, name=f"{name}_norm2"
+        )
         mlp_hidden_dim = int(dim * mlp_ratio)
         self.mlp = Mlp(
             in_features=dim,
             hidden_features=mlp_hidden_dim,
             act=act,
             drop=drop,
-            name=f"{name}_mlp"
+            name=f"{name}_mlp",
         )
 
     def call(self, x, training=False):
@@ -264,11 +276,13 @@ def __init__(
                 drop=drop_rate,
                 attn_drop=attn_drop_rate,
                 drop_path=dpr_values[i],
-                name=f"{name}_block{i}"
+                name=f"{name}_block{i}",
             )
             self.blocks.append(block)
 
-        self.norm = layers.LayerNormalization(epsilon=LAYERNORM_EPSILON, name=f"{name}_norm")
+        self.norm = layers.LayerNormalization(
+            epsilon=LAYERNORM_EPSILON, name=f"{name}_norm"
+        )
 
         # Classifier head
         if class_num: