python-doctr 0.9.0__py3-none-any.whl → 0.11.0__py3-none-any.whl

doctr/models/recognition/master/tensorflow.py

@@ -1,10 +1,10 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, Mindee.
 
 # This program is licensed under the Apache License 2.0.
 # See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
 
 from copy import deepcopy
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any
 
 import tensorflow as tf
 from tensorflow.keras import Model, layers
@@ -13,19 +13,19 @@ from doctr.datasets import VOCABS
 from doctr.models.classification import magc_resnet31
 from doctr.models.modules.transformer import Decoder, PositionalEncoding
 
-from ...utils.tensorflow import _bf16_to_float32, load_pretrained_params
+from ...utils.tensorflow import _bf16_to_float32, _build_model, load_pretrained_params
 from .base import _MASTER, _MASTERPostProcessor
 
 __all__ = ["MASTER", "master"]
 
 
-default_cfgs: Dict[str, Dict[str, Any]] = {
+default_cfgs: dict[str, dict[str, Any]] = {
     "master": {
         "mean": (0.694, 0.695, 0.693),
         "std": (0.299, 0.296, 0.301),
         "input_shape": (32, 128, 3),
         "vocab": VOCABS["french"],
-        "url": "https://doctr-static.mindee.com/models?id=v0.6.0/master-a8232e9f.zip&src=0",
+        "url": "https://doctr-static.mindee.com/models?id=v0.9.0/master-d7fdaeff.weights.h5&src=0",
     },
 }
 
@@ -35,7 +35,6 @@ class MASTER(_MASTER, Model):
     Implementation based on the official TF implementation: <https://github.com/jiangxiluning/MASTER-TF>`_.
 
     Args:
-    ----
         feature_extractor: the backbone serving as feature extractor
         vocab: vocabulary, (without EOS, SOS, PAD)
         d_model: d parameter for the transformer decoder
@@ -51,7 +50,7 @@ class MASTER(_MASTER, Model):
 
     def __init__(
         self,
-        feature_extractor: tf.keras.Model,
+        feature_extractor: Model,
         vocab: str,
         d_model: int = 512,
         dff: int = 2048,
@@ -59,9 +58,9 @@ class MASTER(_MASTER, Model):
         num_layers: int = 3,
         max_length: int = 50,
         dropout: float = 0.2,
-        input_shape: Tuple[int, int, int] = (32, 128, 3),  # different from the paper
+        input_shape: tuple[int, int, int] = (32, 128, 3),  # different from the paper
         exportable: bool = False,
-        cfg: Optional[Dict[str, Any]] = None,
+        cfg: dict[str, Any] | None = None,
     ) -> None:
         super().__init__()
 
@@ -89,7 +88,7 @@ class MASTER(_MASTER, Model):
         self.postprocessor = MASTERPostProcessor(vocab=self.vocab)
 
     @tf.function
-    def make_source_and_target_mask(self, source: tf.Tensor, target: tf.Tensor) -> Tuple[tf.Tensor, tf.Tensor]:
+    def make_source_and_target_mask(self, source: tf.Tensor, target: tf.Tensor) -> tuple[tf.Tensor, tf.Tensor]:
         # [1, 1, 1, ..., 0, 0, 0] -> 0 is masked
         # (N, 1, 1, max_length)
         target_pad_mask = tf.cast(tf.math.not_equal(target, self.vocab_size + 2), dtype=tf.uint8)
@@ -109,19 +108,17 @@ class MASTER(_MASTER, Model):
     def compute_loss(
         model_output: tf.Tensor,
         gt: tf.Tensor,
-        seq_len: List[int],
+        seq_len: list[int],
     ) -> tf.Tensor:
         """Compute categorical cross-entropy loss for the model.
         Sequences are masked after the EOS character.
 
         Args:
-        ----
             gt: the encoded tensor with gt labels
             model_output: predicted logits of the model
             seq_len: lengths of each gt word inside the batch
 
         Returns:
-        -------
             The loss of the model on the batch
         """
         # Input length : number of timesteps
@@ -144,15 +141,14 @@ class MASTER(_MASTER, Model):
     def call(
         self,
         x: tf.Tensor,
-        target: Optional[List[str]] = None,
+        target: list[str] | None = None,
         return_model_output: bool = False,
         return_preds: bool = False,
         **kwargs: Any,
-    ) -> Dict[str, Any]:
+    ) -> dict[str, Any]:
         """Call function for training
 
         Args:
-        ----
             x: images
             target: list of str labels
             return_model_output: if True, return logits
@@ -160,7 +156,6 @@ class MASTER(_MASTER, Model):
             **kwargs: keyword arguments passed to the decoder
 
         Returns:
-        -------
             A dictionnary containing eventually loss, logits and predictions.
         """
         # Encode
@@ -171,7 +166,7 @@ class MASTER(_MASTER, Model):
         # add positional encoding to features
         encoded = self.positional_encoding(feature, **kwargs)
 
-        out: Dict[str, tf.Tensor] = {}
+        out: dict[str, tf.Tensor] = {}
 
         if kwargs.get("training", False) and target is None:
             raise ValueError("Need to provide labels during training")
@@ -209,13 +204,11 @@ class MASTER(_MASTER, Model):
         """Decode function for prediction
 
         Args:
-        ----
             encoded: encoded features
             **kwargs: keyword arguments passed to the decoder
 
         Returns:
-        -------
-            A Tuple of tf.Tensor: predictions, logits
+            A tuple of tf.Tensor: predictions, logits
         """
         b = encoded.shape[0]
 
@@ -247,14 +240,13 @@ class MASTERPostProcessor(_MASTERPostProcessor):
     """Post processor for MASTER architectures
 
     Args:
-    ----
         vocab: string containing the ordered sequence of supported characters
     """
 
     def __call__(
         self,
         logits: tf.Tensor,
-    ) -> List[Tuple[str, float]]:
+    ) -> list[tuple[str, float]]:
         # compute pred with argmax for attention models
         out_idxs = tf.math.argmax(logits, axis=2)
         # N x L
@@ -290,9 +282,14 @@ def _master(arch: str, pretrained: bool, backbone_fn, pretrained_backbone: bool
         cfg=_cfg,
         **kwargs,
     )
+    _build_model(model)
+
     # Load pretrained parameters
     if pretrained:
-        load_pretrained_params(model, default_cfgs[arch]["url"])
+        # The given vocab differs from the pretrained model => skip the mismatching layers for fine tuning
+        load_pretrained_params(
+            model, default_cfgs[arch]["url"], skip_mismatch=kwargs["vocab"] != default_cfgs[arch]["vocab"]
+        )
 
     return model
 
@@ -307,12 +304,10 @@ def master(pretrained: bool = False, **kwargs: Any) -> MASTER:
     >>> out = model(input_tensor)
 
     Args:
-    ----
         pretrained (bool): If True, returns a model pre-trained on our text recognition dataset
         **kwargs: keywoard arguments passed to the MASTER architecture
 
     Returns:
-    -------
         text recognition architecture
     """
     return _master("master", pretrained, magc_resnet31, **kwargs)

doctr/models/recognition/parseq/__init__.py

@@ -1,6 +1,6 @@
 from doctr.file_utils import is_tf_available, is_torch_available
 
-if is_tf_available():
+if is_torch_available():
+    from .pytorch import *
+elif is_tf_available():
     from .tensorflow import *
-elif is_torch_available():
-    from .pytorch import *  # type: ignore[assignment]

doctr/models/recognition/parseq/base.py

@@ -1,9 +1,8 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, Mindee.
 
 # This program is licensed under the Apache License 2.0.
 # See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
 
-from typing import List, Tuple
 
 import numpy as np
 
@@ -17,17 +16,15 @@ class _PARSeq:
 
     def build_target(
         self,
-        gts: List[str],
-    ) -> Tuple[np.ndarray, List[int]]:
+        gts: list[str],
+    ) -> tuple[np.ndarray, list[int]]:
         """Encode a list of gts sequences into a np array and gives the corresponding*
         sequence lengths.
 
         Args:
-        ----
             gts: list of ground-truth labels
 
         Returns:
-        -------
             A tuple of 2 tensors: Encoded labels and sequence lengths (for each entry of the batch)
         """
         encoded = encode_sequences(
@@ -46,7 +43,6 @@ class _PARSeqPostProcessor(RecognitionPostProcessor):
     """Abstract class to postprocess the raw output of the model
 
     Args:
-    ----
         vocab: string containing the ordered sequence of supported characters
     """
 

doctr/models/recognition/parseq/pytorch.py

@@ -1,12 +1,13 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, Mindee.
 
 # This program is licensed under the Apache License 2.0.
 # See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
 
 import math
+from collections.abc import Callable
 from copy import deepcopy
 from itertools import permutations
-from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+from typing import Any
 
 import numpy as np
 import torch
@@ -23,7 +24,7 @@ from .base import _PARSeq, _PARSeqPostProcessor
 
 __all__ = ["PARSeq", "parseq"]
 
-default_cfgs: Dict[str, Dict[str, Any]] = {
+default_cfgs: dict[str, dict[str, Any]] = {
     "parseq": {
         "mean": (0.694, 0.695, 0.693),
         "std": (0.299, 0.296, 0.301),
@@ -38,7 +39,6 @@ class CharEmbedding(nn.Module):
     """Implements the character embedding module
 
     Args:
-    ----
         vocab_size: size of the vocabulary
         d_model: dimension of the model
     """
@@ -56,7 +56,6 @@ class PARSeqDecoder(nn.Module):
     """Implements decoder module of the PARSeq model
 
     Args:
-    ----
         d_model: dimension of the model
         num_heads: number of attention heads
         ffd: dimension of the feed forward layer
@@ -92,7 +91,7 @@ class PARSeqDecoder(nn.Module):
         target,
         content,
         memory,
-        target_mask: Optional[torch.Tensor] = None,
+        target_mask: torch.Tensor | None = None,
     ):
         query_norm = self.query_norm(target)
         content_norm = self.content_norm(content)
@@ -112,7 +111,6 @@ class PARSeq(_PARSeq, nn.Module):
     Slightly modified implementation based on the official Pytorch implementation: <https://github.com/baudm/parseq/tree/main`_.
 
     Args:
-    ----
         feature_extractor: the backbone serving as feature extractor
         vocab: vocabulary used for encoding
         embedding_units: number of embedding units
@@ -136,9 +134,9 @@ class PARSeq(_PARSeq, nn.Module):
         dec_num_heads: int = 12,
         dec_ff_dim: int = 384,  # we use it from the original implementation instead of 2048
         dec_ffd_ratio: int = 4,
-        input_shape: Tuple[int, int, int] = (3, 32, 128),
+        input_shape: tuple[int, int, int] = (3, 32, 128),
         exportable: bool = False,
-        cfg: Optional[Dict[str, Any]] = None,
+        cfg: dict[str, Any] | None = None,
     ) -> None:
         super().__init__()
         self.vocab = vocab
@@ -212,12 +210,12 @@ class PARSeq(_PARSeq, nn.Module):
 
         sos_idx = torch.zeros(len(final_perms), 1, device=seqlen.device)
         eos_idx = torch.full((len(final_perms), 1), max_num_chars + 1, device=seqlen.device)
-        combined = torch.cat([sos_idx, final_perms + 1, eos_idx], dim=1).int()
+        combined = torch.cat([sos_idx, final_perms + 1, eos_idx], dim=1).int()  # type: ignore[list-item]
         if len(combined) > 1:
             combined[1, 1:] = max_num_chars + 1 - torch.arange(max_num_chars + 1, device=seqlen.device)
         return combined
 
-    def generate_permutations_attention_masks(self, permutation: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    def generate_permutations_attention_masks(self, permutation: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
         # Generate source and target mask for the decoder attention.
         sz = permutation.shape[0]
         mask = torch.ones((sz, sz), device=permutation.device)
@@ -236,8 +234,8 @@ class PARSeq(_PARSeq, nn.Module):
         self,
         target: torch.Tensor,
         memory: torch.Tensor,
-        target_mask: Optional[torch.Tensor] = None,
-        target_query: Optional[torch.Tensor] = None,
+        target_mask: torch.Tensor | None = None,
+        target_query: torch.Tensor | None = None,
     ) -> torch.Tensor:
         """Add positional information to the target sequence and pass it through the decoder."""
         batch_size, sequence_length = target.shape
@@ -250,7 +248,7 @@ class PARSeq(_PARSeq, nn.Module):
         target_query = self.dropout(target_query)
         return self.decoder(target_query, content, memory, target_mask)
 
-    def decode_autoregressive(self, features: torch.Tensor, max_len: Optional[int] = None) -> torch.Tensor:
+    def decode_autoregressive(self, features: torch.Tensor, max_len: int | None = None) -> torch.Tensor:
         """Generate predictions for the given features."""
         max_length = max_len if max_len is not None else self.max_length
         max_length = min(max_length, self.max_length) + 1
@@ -283,7 +281,7 @@ class PARSeq(_PARSeq, nn.Module):
 
                 # Stop decoding if all sequences have reached the EOS token
                 # NOTE: `break` isn't correctly translated to Onnx so we don't break here if we want to export
-                if not self.exportable and max_len is None and (ys == self.vocab_size).any(dim=-1).all():
+                if not self.exportable and max_len is None and (ys == self.vocab_size).any(dim=-1).all():  # type: ignore[attr-defined]
                     break
 
         logits = torch.cat(pos_logits, dim=1)  # (N, max_length, vocab_size + 1)
@@ -298,7 +296,7 @@ class PARSeq(_PARSeq, nn.Module):
 
         # Create padding mask for refined target input maskes all behind EOS token as False
         # (N, 1, 1, max_length)
-        target_pad_mask = ~((ys == self.vocab_size).int().cumsum(-1) > 0).unsqueeze(1).unsqueeze(1)
+        target_pad_mask = ~((ys == self.vocab_size).int().cumsum(-1) > 0).unsqueeze(1).unsqueeze(1)  # type: ignore[attr-defined]
         mask = (target_pad_mask.bool() & query_mask[:, : ys.shape[1]].bool()).int()
         logits = self.head(self.decode(ys, features, mask, target_query=pos_queries))
 
@@ -307,10 +305,10 @@ class PARSeq(_PARSeq, nn.Module):
     def forward(
         self,
         x: torch.Tensor,
-        target: Optional[List[str]] = None,
+        target: list[str] | None = None,
         return_model_output: bool = False,
         return_preds: bool = False,
-    ) -> Dict[str, Any]:
+    ) -> dict[str, Any]:
         features = self.feat_extractor(x)["features"]  # (batch_size, patches_seqlen, d_model)
         # remove cls token
         features = features[:, 1:, :]
@@ -337,7 +335,7 @@ class PARSeq(_PARSeq, nn.Module):
                 ).unsqueeze(1).unsqueeze(1)  # (N, 1, 1, seq_len)
 
                 loss = torch.tensor(0.0, device=features.device)
-                loss_numel: Union[int, float] = 0
+                loss_numel: int | float = 0
                 n = (gt_out != self.vocab_size + 2).sum().item()
                 for i, perm in enumerate(tgt_perms):
                     _, target_mask = self.generate_permutations_attention_masks(perm)  # (seq_len, seq_len)
@@ -365,7 +363,7 @@ class PARSeq(_PARSeq, nn.Module):
 
         logits = _bf16_to_float32(logits)
 
-        out: Dict[str, Any] = {}
+        out: dict[str, Any] = {}
         if self.exportable:
             out["logits"] = logits
             return out
@@ -374,8 +372,13 @@ class PARSeq(_PARSeq, nn.Module):
             out["out_map"] = logits
 
         if target is None or return_preds:
+            # Disable for torch.compile compatibility
+            @torch.compiler.disable  # type: ignore[attr-defined]
+            def _postprocess(logits: torch.Tensor) -> list[tuple[str, float]]:
+                return self.postprocessor(logits)
+
             # Post-process boxes
-            out["preds"] = self.postprocessor(logits)
+            out["preds"] = _postprocess(logits)
 
         if target is not None:
             out["loss"] = loss
@@ -387,14 +390,13 @@ class PARSeqPostProcessor(_PARSeqPostProcessor):
     """Post processor for PARSeq architecture
 
     Args:
-    ----
         vocab: string containing the ordered sequence of supported characters
     """
 
     def __call__(
         self,
         logits: torch.Tensor,
-    ) -> List[Tuple[str, float]]:
+    ) -> list[tuple[str, float]]:
         # compute pred with argmax for attention models
         out_idxs = logits.argmax(-1)
         preds_prob = torch.softmax(logits, -1).max(dim=-1)[0]
@@ -417,7 +419,7 @@ def _parseq(
     pretrained: bool,
     backbone_fn: Callable[[bool], nn.Module],
     layer: str,
-    ignore_keys: Optional[List[str]] = None,
+    ignore_keys: list[str] | None = None,
     **kwargs: Any,
 ) -> PARSeq:
     # Patch the config
@@ -462,12 +464,10 @@ def parseq(pretrained: bool = False, **kwargs: Any) -> PARSeq:
     >>> out = model(input_tensor)
 
     Args:
-    ----
         pretrained (bool): If True, returns a model pre-trained on our text recognition dataset
         **kwargs: keyword arguments of the PARSeq architecture
 
     Returns:
-    -------
         text recognition architecture
     """
     return _parseq(

doctr/models/recognition/parseq/tensorflow.py

@@ -1,4 +1,4 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, Mindee.
 
 # This program is licensed under the Apache License 2.0.
 # See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
@@ -6,7 +6,7 @@
 import math
 from copy import deepcopy
 from itertools import permutations
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any
 
 import numpy as np
 import tensorflow as tf
@@ -16,18 +16,18 @@ from doctr.datasets import VOCABS
 from doctr.models.modules.transformer import MultiHeadAttention, PositionwiseFeedForward
 
 from ...classification import vit_s
-from ...utils.tensorflow import _bf16_to_float32, load_pretrained_params
+from ...utils.tensorflow import _bf16_to_float32, _build_model, load_pretrained_params
 from .base import _PARSeq, _PARSeqPostProcessor
 
 __all__ = ["PARSeq", "parseq"]
 
-default_cfgs: Dict[str, Dict[str, Any]] = {
+default_cfgs: dict[str, dict[str, Any]] = {
     "parseq": {
         "mean": (0.694, 0.695, 0.693),
         "std": (0.299, 0.296, 0.301),
         "input_shape": (32, 128, 3),
         "vocab": VOCABS["french"],
-        "url": "https://doctr-static.mindee.com/models?id=v0.6.0/parseq-24cf693e.zip&src=0",
+        "url": "https://doctr-static.mindee.com/models?id=v0.9.0/parseq-4152a87e.weights.h5&src=0",
     },
 }
 
@@ -36,14 +36,14 @@ class CharEmbedding(layers.Layer):
     """Implements the character embedding module
 
     Args:
-    ----
+    -
         vocab_size: size of the vocabulary
         d_model: dimension of the model
     """
 
     def __init__(self, vocab_size: int, d_model: int):
         super(CharEmbedding, self).__init__()
-        self.embedding = tf.keras.layers.Embedding(vocab_size, d_model)
+        self.embedding = layers.Embedding(vocab_size, d_model)
         self.d_model = d_model
 
     def call(self, x: tf.Tensor, **kwargs: Any) -> tf.Tensor:
@@ -54,7 +54,6 @@ class PARSeqDecoder(layers.Layer):
     """Implements decoder module of the PARSeq model
 
     Args:
-    ----
         d_model: dimension of the model
         num_heads: number of attention heads
         ffd: dimension of the feed forward layer
@@ -115,7 +114,6 @@ class PARSeq(_PARSeq, Model):
     Modified implementation based on the official Pytorch implementation: <https://github.com/baudm/parseq/tree/main`_.
 
     Args:
-    ----
         feature_extractor: the backbone serving as feature extractor
         vocab: vocabulary used for encoding
         embedding_units: number of embedding units
@@ -129,7 +127,7 @@ class PARSeq(_PARSeq, Model):
         cfg: dictionary containing information about the model
     """
 
-    _children_names: List[str] = ["feat_extractor", "postprocessor"]
+    _children_names: list[str] = ["feat_extractor", "postprocessor"]
 
     def __init__(
         self,
@@ -141,9 +139,9 @@ class PARSeq(_PARSeq, Model):
         dec_num_heads: int = 12,
         dec_ff_dim: int = 384,  # we use it from the original implementation instead of 2048
         dec_ffd_ratio: int = 4,
-        input_shape: Tuple[int, int, int] = (32, 128, 3),
+        input_shape: tuple[int, int, int] = (32, 128, 3),
         exportable: bool = False,
-        cfg: Optional[Dict[str, Any]] = None,
+        cfg: dict[str, Any] | None = None,
     ) -> None:
         super().__init__()
         self.vocab = vocab
@@ -167,7 +165,6 @@ class PARSeq(_PARSeq, Model):
 
         self.postprocessor = PARSeqPostProcessor(vocab=self.vocab)
 
-    @tf.function
     def generate_permutations(self, seqlen: tf.Tensor) -> tf.Tensor:
         # Generates permutations of the target sequence.
         # Translated from https://github.com/baudm/parseq/blob/main/strhub/models/parseq/system.py
@@ -214,8 +211,7 @@ class PARSeq(_PARSeq, Model):
             )
         return combined
 
-    @tf.function
-    def generate_permutations_attention_masks(self, permutation: tf.Tensor) -> Tuple[tf.Tensor, tf.Tensor]:
+    def generate_permutations_attention_masks(self, permutation: tf.Tensor) -> tuple[tf.Tensor, tf.Tensor]:
         # Generate source and target mask for the decoder attention.
         sz = permutation.shape[0]
         mask = tf.ones((sz, sz), dtype=tf.float32)
@@ -234,13 +230,12 @@ class PARSeq(_PARSeq, Model):
         target_mask = mask[1:, :-1]
         return tf.cast(source_mask, dtype=tf.bool), tf.cast(target_mask, dtype=tf.bool)
 
-    @tf.function
     def decode(
         self,
         target: tf.Tensor,
-        memory: tf,
-        target_mask: Optional[tf.Tensor] = None,
-        target_query: Optional[tf.Tensor] = None,
+        memory: tf.Tensor,
+        target_mask: tf.Tensor | None = None,
+        target_query: tf.Tensor | None = None,
         **kwargs: Any,
     ) -> tf.Tensor:
         batch_size, sequence_length = target.shape
@@ -253,8 +248,7 @@ class PARSeq(_PARSeq, Model):
         target_query = self.dropout(target_query, **kwargs)
         return self.decoder(target_query, content, memory, target_mask, **kwargs)
 
-    @tf.function
-    def decode_autoregressive(self, features: tf.Tensor, max_len: Optional[int] = None, **kwargs) -> tf.Tensor:
+    def decode_autoregressive(self, features: tf.Tensor, max_len: int | None = None, **kwargs) -> tf.Tensor:
         """Generate predictions for the given features."""
         max_length = max_len if max_len is not None else self.max_length
         max_length = min(max_length, self.max_length) + 1
@@ -321,11 +315,11 @@ class PARSeq(_PARSeq, Model):
     def call(
         self,
         x: tf.Tensor,
-        target: Optional[List[str]] = None,
+        target: list[str] | None = None,
         return_model_output: bool = False,
         return_preds: bool = False,
         **kwargs: Any,
-    ) -> Dict[str, Any]:
+    ) -> dict[str, Any]:
         features = self.feat_extractor(x, **kwargs)  # (batch_size, patches_seqlen, d_model)
         # remove cls token
         features = features[:, 1:, :]
@@ -396,7 +390,7 @@ class PARSeq(_PARSeq, Model):
 
         logits = _bf16_to_float32(logits)
 
-        out: Dict[str, tf.Tensor] = {}
+        out: dict[str, tf.Tensor] = {}
         if self.exportable:
             out["logits"] = logits
             return out
@@ -418,14 +412,13 @@ class PARSeqPostProcessor(_PARSeqPostProcessor):
     """Post processor for PARSeq architecture
 
     Args:
-    ----
         vocab: string containing the ordered sequence of supported characters
     """
 
     def __call__(
         self,
         logits: tf.Tensor,
-    ) -> List[Tuple[str, float]]:
+    ) -> list[tuple[str, float]]:
         # compute pred with argmax for attention models
         out_idxs = tf.math.argmax(logits, axis=2)
         preds_prob = tf.math.reduce_max(tf.nn.softmax(logits, axis=-1), axis=-1)
@@ -451,7 +444,7 @@ def _parseq(
     arch: str,
     pretrained: bool,
     backbone_fn,
-    input_shape: Optional[Tuple[int, int, int]] = None,
+    input_shape: tuple[int, int, int] | None = None,
     **kwargs: Any,
 ) -> PARSeq:
     # Patch the config
@@ -476,9 +469,14 @@ def _parseq(
 
     # Build the model
     model = PARSeq(feat_extractor, cfg=_cfg, **kwargs)
+    _build_model(model)
+
     # Load pretrained parameters
     if pretrained:
-        load_pretrained_params(model, default_cfgs[arch]["url"])
+        # The given vocab differs from the pretrained model => skip the mismatching layers for fine tuning
+        load_pretrained_params(
+            model, default_cfgs[arch]["url"], skip_mismatch=kwargs["vocab"] != default_cfgs[arch]["vocab"]
+        )
 
     return model
 
@@ -494,12 +492,10 @@ def parseq(pretrained: bool = False, **kwargs: Any) -> PARSeq:
     >>> out = model(input_tensor)
 
     Args:
-    ----
         pretrained (bool): If True, returns a model pre-trained on our text recognition dataset
         **kwargs: keyword arguments of the PARSeq architecture
 
     Returns:
-    -------
         text recognition architecture
     """
     return _parseq(

doctr/models/recognition/predictor/__init__.py

@@ -1,6 +1,6 @@
-from doctr.file_utils import is_tf_available
+from doctr.file_utils import is_tf_available, is_torch_available
 
-if is_tf_available():
-    from .tensorflow import *
-else:
-    from .pytorch import *  # type: ignore[assignment]
+if is_torch_available():
+    from .pytorch import *
+elif is_tf_available():
+    from .tensorflow import *  # type: ignore[assignment]