add functional ctc_loss and CTCLoss class. (#26384)

* add functional ctc_loss and CTCLoss class.

* modified docstring of ctc_loss and CTCLoss

python/paddle/fluid/tests/unittests/test_warpctc_op.py

@@ -21,25 +21,25 @@ from op_test import OpTest
 from test_softmax_op import stable_softmax
 import paddle.fluid as fluid
 from paddle.fluid import Program, program_guard
+import paddle
+import paddle.nn.functional as F
 
 CUDA_BLOCK_SIZE = 512
 
 
 class CTCForward(object):
-    def __init__(self, softmax, softmax_lod, labels, labels_lod, blank,
-                 norm_by_times):
+    def __init__(self, softmax, softmax_lod, labels, labels_lod, num_classes,
+                 batch_size, blank, norm_by_times):
         self.softmax = softmax
         self.softmax_lod = softmax_lod
-        assert labels.shape[1] == 1
         self.labels = labels
         self.labels_lod = labels_lod
         self.blank = blank
         self.norm_by_times = norm_by_times
 
         self.level = 0
-        self.num_classes = softmax.shape[1]
-        self.batch_size = len(softmax_lod[self.level])
-        assert self.batch_size == len(labels_lod[self.level])
+        self.num_classes = num_classes
+        self.batch_size = batch_size
 
         self.loss = np.zeros([self.batch_size, 1], dtype="float32")
         self.gradient = np.zeros(self.softmax.shape, dtype="float32")
@@ -163,17 +163,25 @@ class CTCForward(object):
         softmax_offset = 0
         labels_offset = 0
         for i in range(self.batch_size):
+            if self.labels.shape[1] == 1:
                 softmax_start_i = softmax_offset
                 softmax_end_i = softmax_offset + self.softmax_lod[self.level][i]
                 labels_start_i = labels_offset
                 labels_end_i = labels_offset + self.labels_lod[self.level][i]
 
-            softmax_a_sequence = self.softmax[softmax_start_i:softmax_end_i, :]
+                softmax_a_sequence = self.softmax[softmax_start_i:
+                                                  softmax_end_i, :]
                 labels_a_sequence = self.labels[labels_start_i:labels_end_i, :]
                 self.loss[i] = self.forward_a_sequence(softmax_a_sequence,
                                                        labels_a_sequence)
                 softmax_offset += self.softmax_lod[self.level][i]
                 labels_offset += self.labels_lod[self.level][i]
+            else:
+                softmax_a_sequence = self.softmax[:self.softmax_lod[i], i, :]
+                labels_a_sequence = self.labels[:self.labels_lod[i], :]
+                self.loss[i] = self.forward_a_sequence(softmax_a_sequence,
+                                                       labels_a_sequence)
+
         return self.loss
 
 
@@ -201,7 +209,8 @@ class TestWarpCTCOp(OpTest):
             dtype="int32")
 
         ctc = CTCForward(softmax, self.logits_lod, labels, self.labels_lod,
-                         self.blank, self.norm_by_times)
+                         self.num_classes, self.batch_size, self.blank,
+                         self.norm_by_times)
         loss = ctc.forward()
 
         max_sequence_length = 0
@@ -223,7 +232,7 @@ class TestWarpCTCOp(OpTest):
         }
 
     def test_check_output(self):
-        self.check_output(check_dygraph=False)
+        self.check_output()
 
     def test_check_grad(self):
         self.outputs['WarpCTCGrad'] = self.gradient
@@ -237,7 +246,7 @@ class TestWarpCTCOpCase1(TestWarpCTCOp):
         self.num_classes = CUDA_BLOCK_SIZE + 2
         self.logits_lod = [[4, 1, 3, 3]]
         self.labels_lod = [[3, 1, 4, 4]]
-        self.blank = 0
+        self.blank = self.num_classes - 1
         self.norm_by_times = False
 
 
@@ -267,7 +276,8 @@ class TestWarpCTCOpWithPadding(OpTest):
             dtype="int32")
 
         ctc = CTCForward(softmax, self.logits_lod, labels, self.labels_lod,
-                         self.blank, self.norm_by_times)
+                         self.num_classes, self.batch_size, self.blank,
+                         self.norm_by_times)
         loss = ctc.forward()
 
         max_sequence_length = 0
@@ -317,7 +327,7 @@ class TestWarpCTCOpWithPadding(OpTest):
         }
 
     def test_check_output(self):
-        self.check_output(check_dygraph=False)
+        self.check_output()
 
     def test_check_grad(self):
         self.outputs['WarpCTCGrad'] = self.gradient
@@ -333,7 +343,7 @@ class TestWarpCTCOpWithPaddingCase1(TestWarpCTCOpWithPadding):
         self.labels_lod = [[3, 1, 4, 4]]
         self.logits_length = np.array([4, 1, 3, 3], dtype=np.int64)
         self.labels_length = np.array([3, 1, 4, 4], dtype=np.int64)
-        self.blank = 0
+        self.blank = self.num_classes - 1
         self.norm_by_times = False
 
 
@@ -389,5 +399,97 @@ class TestWarpCTCOpError(unittest.TestCase):
             self.assertRaises(TypeError, test_label_len_Variable)
 
 
+class TestCTCLossAPICase(unittest.TestCase):
+    def test_functinal_api(self):
+        self.batch_size = 4
+        self.num_classes = CUDA_BLOCK_SIZE + 2
+        self.logits_length = np.array([4, 1, 3, 3], dtype=np.int64)
+        self.labels_length = np.array([3, 1, 4, 4], dtype=np.int64)
+        self.blank = self.num_classes - 1
+        self.norm_by_times = False
+
+        logits = np.random.uniform(0.1, 1.0, [
+            max(self.logits_length), self.batch_size, self.num_classes
+        ]).astype("float32")
+        softmax = np.apply_along_axis(stable_softmax, -1, logits)
+        # labels should not be blank
+        labels = np.random.randint(
+            0,
+            self.num_classes - 1, [self.batch_size, max(self.labels_length)],
+            dtype="int32")
+
+        ctc = CTCForward(softmax, self.logits_length, labels,
+                         self.labels_length, self.num_classes, self.batch_size,
+                         self.blank, self.norm_by_times)
+        loss_np = ctc.forward()
+
+        paddle.disable_static()
+        softmax = paddle.to_variable(logits)
+        labels = paddle.to_variable(labels)
+        logits_length = paddle.to_variable(self.logits_length)
+        labels_length = paddle.to_variable(self.labels_length)
+        loss_pd_mean = F.ctc_loss(
+            softmax,
+            labels,
+            logits_length,
+            labels_length,
+            blank=self.blank,
+            reduction='mean')
+        loss_pd_mean = loss_pd_mean.numpy()
+
+        loss_pd_sum = F.ctc_loss(
+            softmax,
+            labels,
+            logits_length,
+            labels_length,
+            blank=self.blank,
+            reduction='sum')
+        loss_pd_sum = loss_pd_sum.numpy()
+        paddle.enable_static()
+        loss_np = np.squeeze(loss_np, axis=-1)
+        loss_np_mean = (loss_np / labels_length.numpy()).mean()
+        loss_np_sum = loss_np.sum()
+
+        self.assertTrue(np.allclose(loss_pd_mean, loss_np_mean, atol=1))
+        self.assertTrue(np.allclose(loss_pd_sum, loss_np_sum, atol=1))
+
+    def test_class_api(self):
+        self.batch_size = 3
+        self.num_classes = 15
+        self.logits_length = np.array([3, 3, 3], dtype=np.int64)
+        self.labels_length = np.array([0, 1, 2], dtype=np.int64)
+        self.blank = 0
+        self.norm_by_times = False
+
+        logits = np.random.uniform(0.1, 1.0, [
+            max(self.logits_length), self.batch_size, self.num_classes
+        ]).astype("float32")
+        softmax = np.apply_along_axis(stable_softmax, -1, logits)
+        # labels should not be blank
+        labels = np.random.randint(
+            1,
+            self.num_classes, [self.batch_size, max(self.labels_length)],
+            dtype="int32")
+
+        ctc = CTCForward(softmax, self.logits_length, labels,
+                         self.labels_length, self.num_classes, self.batch_size,
+                         self.blank, self.norm_by_times)
+        loss_np = ctc.forward()
+
+        paddle.disable_static()
+        softmax = paddle.to_variable(logits)
+        labels = paddle.to_variable(labels)
+        logits_length = paddle.to_variable(self.logits_length)
+        labels_length = paddle.to_variable(self.labels_length)
+
+        loss_pd = paddle.nn.CTCLoss(self.blank, 'none')(
+            softmax, labels, logits_length, labels_length)
+        loss_pd = loss_pd.numpy()
+        paddle.enable_static()
+        loss_np = np.squeeze(loss_np, axis=-1)
+
+        self.assertTrue(np.allclose(loss_pd, loss_np, atol=1))
+
+
 if __name__ == "__main__":
     unittest.main()

python/paddle/nn/__init__.py

@@ -111,6 +111,7 @@ from .layer.loss import NLLLoss  #DEFINE_ALIAS
 from .layer.loss import BCELoss  #DEFINE_ALIAS
 from .layer.loss import KLDivLoss  #DEFINE_ALIAS
 from .layer.loss import MarginRankingLoss  #DEFINE_ALIAS
+from .layer.loss import CTCLoss  #DEFINE_ALIAS
 from .layer.loss import SmoothL1Loss  #DEFINE_ALIAS
 from .layer.norm import BatchNorm  #DEFINE_ALIAS
 from .layer.norm import SyncBatchNorm  #DEFINE_ALIAS

python/paddle/nn/functional/__init__.py

@@ -25,6 +25,8 @@ from . import extension
 __all__ += extension.__all__
 from . import common
 __all__ += common.__all__
+from . import loss
+__all__ += loss.__all__
 from .activation import brelu  #DEFINE_ALIAS
 from .activation import elu  #DEFINE_ALIAS
 from .activation import erf  #DEFINE_ALIAS
@@ -147,6 +149,7 @@ from .loss import softmax_with_cross_entropy  #DEFINE_ALIAS
 from .loss import square_error_cost  #DEFINE_ALIAS
 from .loss import ssd_loss  #DEFINE_ALIAS
 from .loss import teacher_student_sigmoid_loss  #DEFINE_ALIAS
+from .loss import ctc_loss  #DEFINE_ALIAS
 # from .norm import batch_norm        #DEFINE_ALIAS
 # from .norm import data_norm        #DEFINE_ALIAS
 # from .norm import group_norm        #DEFINE_ALIAS

python/paddle/nn/functional/loss.py

@@ -13,6 +13,9 @@
 # limitations under the License.
 
 import paddle
+from ...fluid.layer_helper import LayerHelper
+from ...fluid.data_feeder import check_variable_and_dtype
+import paddle.fluid as fluid
 
 # TODO: define loss functions of neural network
 import numpy as np
@@ -70,7 +73,8 @@ __all__ = [
     'softmax_with_cross_entropy',
     'square_error_cost',
     'ssd_loss',
-    'teacher_student_sigmoid_loss'
+    'teacher_student_sigmoid_loss',
+    'ctc_loss',
 ]
 
 
@@ -791,6 +795,102 @@ def mse_loss(input, label, reduction='mean', name=None):
                           name=name)
 
 
+def ctc_loss(log_probs,
+             labels,
+             input_lengths,
+             label_lengths,
+             blank=0,
+             reduction='mean'):
+    """
+
+    An operator integrating the open source Warp-CTC library (https://github.com/baidu-research/warp-ctc) 
+    to compute Connectionist Temporal Classification (CTC) loss. 
+    It can be aliased as softmax with CTC, since a native softmax activation 
+    is interated to the Warp-CTC library to normalize values for each row of the input tensor.
+
+    Parameters:
+        log_probs (Tensor): The unscaled probability sequence with padding, which is a 3-D Tensor. The tensor shape is [max_logit_length, batch_size, num_classes + 1], where max_logit_length is the longest length of input logit sequence. The data type must be float32.
+        labels (Tensor): The ground truth sequence with padding, which must be a 3-D Tensor. The tensor shape is [batch_size, max_label_length], where max_label_length is the longest length of label sequence. The data type must be int32.
+        input_lengths (Tensor): The length for each input sequence, it should have shape [batch_size] and dtype int64.
+        label_lengths (Tensor): The length for each label sequence, it should have shape [batch_size] and dtype int64.
+        blank (int, optional): The blank label index of Connectionist Temporal Classification (CTC) loss, which is in the half-opened interval [0, num_classes + 1). The data type must be int32. Default is 0.
+        reduction (string, optional): Indicate how to average the loss, the candicates are ``'none'`` | ``'mean'`` | ``'sum'``. If :attr:`reduction` is ``'mean'``, the output loss will be divided by the label_lengths, and then return the mean of quotient; If :attr:`reduction` is ``'sum'``, return the sum of loss; If :attr:`reduction` is ``'none'``, no reduction will be applied. Default is ``'mean'``.
+
+    Returns:
+        Tensor, The Connectionist Temporal Classification (CTC) loss between ``log_probs`` and  ``labels``. If attr:`reduction` is ``'none'``, the shape of loss is [batch_size], otherwise, the shape of loss is [1]. Data type is the same as ``log_probs``.
+    
+    Examples:
+
+        .. code-block:: python
+
+            # declarative mode
+            import paddle.nn.functional as F
+            import numpy as np
+            import paddle
+
+            # length of the longest logit sequence
+            max_seq_length = 4
+            #length of the longest label sequence
+            max_label_length = 3
+            # number of logit sequences
+            batch_size = 2
+            # class num
+            class_num = 3
+
+            np.random.seed(1)
+            log_probs = np.array([[[4.17021990e-01, 7.20324516e-01, 1.14374816e-04],
+                                    [3.02332580e-01, 1.46755889e-01, 9.23385918e-02]],
+
+                                    [[1.86260208e-01, 3.45560730e-01, 3.96767467e-01],
+                                    [5.38816750e-01, 4.19194520e-01, 6.85219526e-01]],
+
+                                    [[2.04452246e-01, 8.78117442e-01, 2.73875929e-02],
+                                    [6.70467496e-01, 4.17304814e-01, 5.58689833e-01]],
+
+                                    [[1.40386939e-01, 1.98101491e-01, 8.00744593e-01],
+                                    [9.68261600e-01, 3.13424170e-01, 6.92322612e-01]],
+
+                                    [[8.76389146e-01, 8.94606650e-01, 8.50442126e-02],
+                                    [3.90547849e-02, 1.69830427e-01, 8.78142476e-01]]]).astype("float32")
+            labels = np.array([[1, 2, 2],
+                            [1, 2, 2]]).astype("int32")
+            input_lengths = np.array([5, 5]).astype("int64")
+            label_lengths = np.array([3, 3]).astype("int64")
+
+            paddle.disable_static()
+            log_probs = paddle.to_variable(log_probs)
+            labels = paddle.to_variable(labels)
+            input_lengths = paddle.to_variable(input_lengths)
+            label_lengths = paddle.to_variable(label_lengths)
+
+            loss = F.ctc_loss(log_probs, labels, 
+                input_lengths, 
+                label_lengths, 
+                blank=0, 
+                reduction='none')
+            print(loss.numpy())  #[3.9179852 2.9076521]
+
+            loss = F.ctc_loss(log_probs, labels, 
+                input_lengths, 
+                label_lengths, 
+                blank=0, 
+                reduction='mean') 
+            print(loss.numpy())  #[1.1376063]
+
+    """
+
+    loss_out = fluid.layers.warpctc(log_probs, labels, blank, False,
+                                    input_lengths, label_lengths)
+
+    loss_out = fluid.layers.squeeze(loss_out, [-1])
+    assert reduction in ['mean', 'sum', 'none']
+    if reduction == 'mean':
+        loss_out = paddle.mean(loss_out / label_lengths)
+    elif reduction == 'sum':
+        loss_out = paddle.sum(loss_out)
+    return loss_out
+
+
 def cross_entropy(input,
                   label,
                   weight=None,

python/paddle/nn/layer/__init__.py

@@ -76,6 +76,7 @@ from .loss import NLLLoss  #DEFINE_ALIAS
 from .loss import BCELoss  #DEFINE_ALIAS
 from .loss import KLDivLoss  #DEFINE_ALIAS
 from .loss import MarginRankingLoss  #DEFINE_ALIAS
+from .loss import CTCLoss  #DEFINE_ALIAS
 from .loss import SmoothL1Loss  #DEFINE_ALIAS
 from .norm import BatchNorm  #DEFINE_ALIAS
 from .norm import SyncBatchNorm  #DEFINE_ALIAS

python/paddle/nn/layer/loss.py

@@ -28,6 +28,7 @@ __all__ = [
     'BCELoss',
     'KLDivLoss',
     'MarginRankingLoss',
+    'CTCLoss',
     'SmoothL1Loss',
 ]
 
@@ -672,6 +673,94 @@ class MarginRankingLoss(fluid.dygraph.Layer):
         return out
 
 
+class CTCLoss(fluid.dygraph.Layer):
+    """
+	:alias_main: paddle.nn.CTCLoss
+	:alias: paddle.nn.CTCLoss, paddle.nn.layer.CTCLoss, paddle.nn.layer.loss.CTCLoss
+
+    An operator integrating the open source Warp-CTC library (https://github.com/baidu-research/warp-ctc) 
+    to compute Connectionist Temporal Classification (CTC) loss. 
+    It can be aliased as softmax with CTC, since a native softmax activation 
+    is interated to the Warp-CTC library to normalize values for each row of the input tensor.
+
+    Parameters:
+        blank (int, optional): The blank label index of Connectionist Temporal Classification (CTC) loss, which is in the half-opened interval [0, num_classes + 1). The data type must be int32. Default is 0.
+        reduction (string, optional): Indicate how to average the loss, the candicates are ``'none'`` | ``'mean'`` | ``'sum'``. If :attr:`reduction` is ``'mean'``, the output loss will be divided by the label_lengths, and then return the mean of quotient; If :attr:`reduction` is ``'sum'``, return the sum of loss; If :attr:`reduction` is ``'none'``, no reduction will be applied. Default is ``'mean'``.
+
+    Shape:
+        log_probs (Tensor): The unscaled probability sequence with padding, which is a 3-D Tensor. The tensor shape is [max_logit_length, batch_size, num_classes + 1], where max_logit_length is the longest length of input logit sequence. The data type must be float32.
+        labels (Tensor): The ground truth sequence with padding, which must be a 3-D Tensor. The tensor shape is [batch_size, max_label_length], where max_label_length is the longest length of label sequence. The data type must be int32.
+        input_lengths (Tensor): The length for each input sequence, it should have shape [batch_size] and dtype int64.
+        label_lengths (Tensor): The length for each label sequence, it should have shape [batch_size] and dtype int64.
+
+    Returns:
+        Tensor, The Connectionist Temporal Classification (CTC) loss between ``log_probs`` and  ``labels``. If attr:`reduction` is ``'none'``, the shape of loss is [batch_size], otherwise, the shape of loss is [1]. Data type is the same as ``log_probs``.
+    
+    Examples:
+
+        .. code-block:: python
+
+            # declarative mode
+            import numpy as np
+            import paddle
+
+            # length of the longest logit sequence
+            max_seq_length = 4
+            #length of the longest label sequence
+            max_label_length = 3
+            # number of logit sequences
+            batch_size = 2
+            # class num
+            class_num = 3
+
+            np.random.seed(1)
+            log_probs = np.array([[[4.17021990e-01, 7.20324516e-01, 1.14374816e-04],
+                                    [3.02332580e-01, 1.46755889e-01, 9.23385918e-02]],
+
+                                    [[1.86260208e-01, 3.45560730e-01, 3.96767467e-01],
+                                    [5.38816750e-01, 4.19194520e-01, 6.85219526e-01]],
+
+                                    [[2.04452246e-01, 8.78117442e-01, 2.73875929e-02],
+                                    [6.70467496e-01, 4.17304814e-01, 5.58689833e-01]],
+
+                                    [[1.40386939e-01, 1.98101491e-01, 8.00744593e-01],
+                                    [9.68261600e-01, 3.13424170e-01, 6.92322612e-01]],
+
+                                    [[8.76389146e-01, 8.94606650e-01, 8.50442126e-02],
+                                    [3.90547849e-02, 1.69830427e-01, 8.78142476e-01]]]).astype("float32")
+            labels = np.array([[1, 2, 2],
+                            [1, 2, 2]]).astype("int32")
+            input_lengths = np.array([5, 5]).astype("int64")
+            label_lengths = np.array([3, 3]).astype("int64")
+
+            paddle.disable_static()
+            log_probs = paddle.to_variable(log_probs)
+            labels = paddle.to_variable(labels)
+            input_lengths = paddle.to_variable(input_lengths)
+            label_lengths = paddle.to_variable(label_lengths)
+
+            loss = paddle.nn.CTCLoss(blank=0, reduction='none')(log_probs, labels, 
+                input_lengths, 
+                label_lengths)
+            print(loss.numpy())  #[3.9179852 2.9076521]
+
+            loss = paddle.nn.CTCLoss(blank=0, reduction='mean')(log_probs, labels, 
+                input_lengths, 
+                label_lengths)
+            print(loss.numpy())  #[1.1376063]
+    """
+
+    def __init__(self, blank=0, reduction='mean'):
+        super(CTCLoss, self).__init__()
+        self.blank = blank
+        self.reduction = reduction
+
+    def forward(self, log_probs, labels, input_lengths, label_lengths):
+        return paddle.nn.functional.ctc_loss(log_probs, labels, input_lengths,
+                                             label_lengths, self.blank,
+                                             self.reduction)
+
+
 class SmoothL1Loss(fluid.dygraph.Layer):
     """
     This operator calculates smooth_l1_loss. Creates a criterion that uses a squared