# Copyright (C) 2021-2024, Mindee. # This program is licensed under the Apache License 2.0. # See LICENSE or go to for full license details. from copy import deepcopy from typing import Any, Dict, List, Optional, Tuple import tensorflow as tf from tensorflow.keras import Model, Sequential, layers from doctr.datasets import VOCABS from doctr.utils.repr import NestedObject from ...classification import resnet31 from ...utils.tensorflow import _bf16_to_float32, load_pretrained_params from ..core import RecognitionModel, RecognitionPostProcessor __all__ = ["SAR", "sar_resnet31"] default_cfgs: Dict[str, Dict[str, Any]] = { "sar_resnet31": { "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/sar_resnet31-c41e32a5.zip&src=0", }, } class SAREncoder(layers.Layer, NestedObject): """Implements encoder module of the SAR model Args: ---- rnn_units: number of hidden rnn units dropout_prob: dropout probability """ def __init__(self, rnn_units: int, dropout_prob: float = 0.0) -> None: super().__init__() self.rnn = Sequential([ layers.LSTM(units=rnn_units, return_sequences=True, recurrent_dropout=dropout_prob), layers.LSTM(units=rnn_units, return_sequences=False, recurrent_dropout=dropout_prob), ]) def call( self, x: tf.Tensor, **kwargs: Any, ) -> tf.Tensor: # (N, C) return self.rnn(x, **kwargs) class AttentionModule(layers.Layer, NestedObject): """Implements attention module of the SAR model Args: ---- attention_units: number of hidden attention units """ def __init__(self, attention_units: int) -> None: super().__init__() self.hidden_state_projector = layers.Conv2D( attention_units, 1, strides=1, use_bias=False, padding="same", kernel_initializer="he_normal", ) self.features_projector = layers.Conv2D( attention_units, 3, strides=1, use_bias=True, padding="same", kernel_initializer="he_normal", ) self.attention_projector = layers.Conv2D( 1, 1, strides=1, use_bias=False, padding="same", kernel_initializer="he_normal", ) self.flatten = layers.Flatten() def call( self, features: tf.Tensor, hidden_state: tf.Tensor, **kwargs: Any, ) -> tf.Tensor: [H, W] = features.get_shape().as_list()[1:3] # shape (N, H, W, vgg_units) -> (N, H, W, attention_units) features_projection = self.features_projector(features, **kwargs) # shape (N, 1, 1, rnn_units) -> (N, 1, 1, attention_units) hidden_state = tf.expand_dims(tf.expand_dims(hidden_state, axis=1), axis=1) hidden_state_projection = self.hidden_state_projector(hidden_state, **kwargs) projection = tf.math.tanh(hidden_state_projection + features_projection) # shape (N, H, W, attention_units) -> (N, H, W, 1) attention = self.attention_projector(projection, **kwargs) # shape (N, H, W, 1) -> (N, H * W) attention = self.flatten(attention) attention = tf.nn.softmax(attention) # shape (N, H * W) -> (N, H, W, 1) attention_map = tf.reshape(attention, [-1, H, W, 1]) glimpse = tf.math.multiply(features, attention_map) # shape (N, H * W) -> (N, C) return tf.reduce_sum(glimpse, axis=[1, 2]) class SARDecoder(layers.Layer, NestedObject): """Implements decoder module of the SAR model Args: ---- rnn_units: number of hidden units in recurrent cells max_length: maximum length of a sequence vocab_size: number of classes in the model alphabet embedding_units: number of hidden embedding units attention_units: number of hidden attention units num_decoder_cells: number of LSTMCell layers to stack dropout_prob: dropout probability """ def __init__( self, rnn_units: int, max_length: int, vocab_size: int, embedding_units: int, attention_units: int, num_decoder_cells: int = 2, dropout_prob: float = 0.0, ) -> None: super().__init__() self.vocab_size = vocab_size self.max_length = max_length self.embed = layers.Dense(embedding_units, use_bias=False) self.embed_tgt = layers.Embedding(embedding_units, self.vocab_size + 1) self.lstm_cells = layers.StackedRNNCells([ layers.LSTMCell(rnn_units, implementation=1) for _ in range(num_decoder_cells) ]) self.attention_module = AttentionModule(attention_units) self.output_dense = layers.Dense(self.vocab_size + 1, use_bias=True) self.dropout = layers.Dropout(dropout_prob) def call( self, features: tf.Tensor, holistic: tf.Tensor, gt: Optional[tf.Tensor] = None, **kwargs: Any, ) -> tf.Tensor: if gt is not None: gt_embedding = self.embed_tgt(gt, **kwargs) logits_list: List[tf.Tensor] = [] for t in range(self.max_length + 1): # 32 if t == 0: # step to init the first states of the LSTMCell states = self.lstm_cells.get_initial_state( inputs=None, batch_size=features.shape[0], dtype=features.dtype ) prev_symbol = holistic elif t == 1: # step to init a 'blank' sequence of length vocab_size + 1 filled with zeros # (N, vocab_size + 1) --> (N, embedding_units) prev_symbol = tf.zeros([features.shape[0], self.vocab_size + 1], dtype=features.dtype) prev_symbol = self.embed(prev_symbol, **kwargs) else: if gt is not None and kwargs.get("training", False): # (N, embedding_units) -2 because of and (same) prev_symbol = self.embed(gt_embedding[:, t - 2], **kwargs) else: # -1 to start at timestep where prev_symbol was initialized index = tf.argmax(logits_list[t - 1], axis=-1) # update prev_symbol with ones at the index of the previous logit vector prev_symbol = self.embed(self.embed_tgt(index, **kwargs), **kwargs) # (N, C), (N, C) take the last hidden state and cell state from current timestep _, states = self.lstm_cells(prev_symbol, states, **kwargs) # states = (hidden_state, cell_state) hidden_state = states[0][0] # (N, H, W, C), (N, C) --> (N, C) glimpse = self.attention_module(features, hidden_state, **kwargs) # (N, C), (N, C) --> (N, 2 * C) logits = tf.concat([hidden_state, glimpse], axis=1) logits = self.dropout(logits, **kwargs) # (N, vocab_size + 1) logits_list.append(self.output_dense(logits, **kwargs)) # (max_length + 1, N, vocab_size + 1) --> (N, max_length + 1, vocab_size + 1) return tf.transpose(tf.stack(logits_list[1:]), (1, 0, 2)) class SAR(Model, RecognitionModel): """Implements a SAR architecture as described in `"Show, Attend and Read:A Simple and Strong Baseline for Irregular Text Recognition" `_. Args: ---- feature_extractor: the backbone serving as feature extractor vocab: vocabulary used for encoding rnn_units: number of hidden units in both encoder and decoder LSTM embedding_units: number of embedding units attention_units: number of hidden units in attention module max_length: maximum word length handled by the model num_decoder_cells: number of LSTMCell layers to stack dropout_prob: dropout probability for the encoder and decoder exportable: onnx exportable returns only logits cfg: dictionary containing information about the model """ _children_names: List[str] = ["feat_extractor", "encoder", "decoder", "postprocessor"] def __init__( self, feature_extractor, vocab: str, rnn_units: int = 512, embedding_units: int = 512, attention_units: int = 512, max_length: int = 30, num_decoder_cells: int = 2, dropout_prob: float = 0.0, exportable: bool = False, cfg: Optional[Dict[str, Any]] = None, ) -> None: super().__init__() self.vocab = vocab self.exportable = exportable self.cfg = cfg self.max_length = max_length + 1 # Add 1 timestep for EOS after the longest word self.feat_extractor = feature_extractor self.encoder = SAREncoder(rnn_units, dropout_prob) self.decoder = SARDecoder( rnn_units, self.max_length, len(vocab), embedding_units, attention_units, num_decoder_cells, dropout_prob, ) self.postprocessor = SARPostProcessor(vocab=vocab) @staticmethod def compute_loss( model_output: tf.Tensor, gt: tf.Tensor, seq_len: tf.Tensor, ) -> 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 input_len = tf.shape(model_output)[1] # Add one for additional token seq_len = seq_len + 1 # One-hot gt labels oh_gt = tf.one_hot(gt, depth=model_output.shape[2]) # Compute loss cce = tf.nn.softmax_cross_entropy_with_logits(oh_gt, model_output) # Compute mask mask_values = tf.zeros_like(cce) mask_2d = tf.sequence_mask(seq_len, input_len) masked_loss = tf.where(mask_2d, cce, mask_values) ce_loss = tf.math.divide(tf.reduce_sum(masked_loss, axis=1), tf.cast(seq_len, model_output.dtype)) return tf.expand_dims(ce_loss, axis=1) def call( self, x: tf.Tensor, target: Optional[List[str]] = None, return_model_output: bool = False, return_preds: bool = False, **kwargs: Any, ) -> Dict[str, Any]: features = self.feat_extractor(x, **kwargs) # vertical max pooling --> (N, C, W) pooled_features = tf.reduce_max(features, axis=1) # holistic (N, C) encoded = self.encoder(pooled_features, **kwargs) if target is not None: gt, seq_len = self.build_target(target) seq_len = tf.cast(seq_len, tf.int32) if kwargs.get("training", False) and target is None: raise ValueError("Need to provide labels during training for teacher forcing") decoded_features = _bf16_to_float32( self.decoder(features, encoded, gt=None if target is None else gt, **kwargs) ) out: Dict[str, tf.Tensor] = {} if self.exportable: out["logits"] = decoded_features return out if return_model_output: out["out_map"] = decoded_features if target is None or return_preds: # Post-process boxes out["preds"] = self.postprocessor(decoded_features) if target is not None: out["loss"] = self.compute_loss(decoded_features, gt, seq_len) return out class SARPostProcessor(RecognitionPostProcessor): """Post processor for SAR architectures Args: ---- vocab: string containing the ordered sequence of supported characters """ def __call__( self, logits: tf.Tensor, ) -> List[Tuple[str, float]]: # compute pred with argmax for attention models out_idxs = tf.math.argmax(logits, axis=2) # N x L probs = tf.gather(tf.nn.softmax(logits, axis=-1), out_idxs, axis=-1, batch_dims=2) # Take the minimum confidence of the sequence probs = tf.math.reduce_min(probs, axis=1) # decode raw output of the model with tf_label_to_idx out_idxs = tf.cast(out_idxs, dtype="int32") embedding = tf.constant(self._embedding, dtype=tf.string) decoded_strings_pred = tf.strings.reduce_join(inputs=tf.nn.embedding_lookup(embedding, out_idxs), axis=-1) decoded_strings_pred = tf.strings.split(decoded_strings_pred, "") decoded_strings_pred = tf.sparse.to_dense(decoded_strings_pred.to_sparse(), default_value="not valid")[:, 0] word_values = [word.decode() for word in decoded_strings_pred.numpy().tolist()] return list(zip(word_values, probs.numpy().clip(0, 1).tolist())) def _sar( arch: str, pretrained: bool, backbone_fn, pretrained_backbone: bool = True, input_shape: Optional[Tuple[int, int, int]] = None, **kwargs: Any, ) -> SAR: pretrained_backbone = pretrained_backbone and not pretrained # Patch the config _cfg = deepcopy(default_cfgs[arch]) _cfg["input_shape"] = input_shape or _cfg["input_shape"] _cfg["vocab"] = kwargs.get("vocab", _cfg["vocab"]) # Feature extractor feat_extractor = backbone_fn( pretrained=pretrained_backbone, input_shape=_cfg["input_shape"], include_top=False, ) kwargs["vocab"] = _cfg["vocab"] # Build the model model = SAR(feat_extractor, cfg=_cfg, **kwargs) # Load pretrained parameters if pretrained: load_pretrained_params(model, default_cfgs[arch]["url"]) return model def sar_resnet31(pretrained: bool = False, **kwargs: Any) -> SAR: """SAR with a resnet-31 feature extractor as described in `"Show, Attend and Read:A Simple and Strong Baseline for Irregular Text Recognition" `_. >>> import tensorflow as tf >>> from doctr.models import sar_resnet31 >>> model = sar_resnet31(pretrained=False) >>> input_tensor = tf.random.uniform(shape=[1, 64, 256, 3], maxval=1, dtype=tf.float32) >>> out = model(input_tensor) Args: ---- pretrained (bool): If True, returns a model pre-trained on our text recognition dataset **kwargs: keyword arguments of the SAR architecture Returns: ------- text recognition architecture """ return _sar("sar_resnet31", pretrained, resnet31, **kwargs)