Merge pull request #1719 from qingen/cluster

[vec][loss] add NCE Loss from RNNLM

paddlespeech/vector/modules/loss.py

@@ -91,3 +91,137 @@ class LogSoftmaxWrapper(nn.Layer):
         predictions = F.log_softmax(predictions, axis=1)
         loss = self.criterion(predictions, targets) / targets.sum()
         return loss
+
+
+class NCELoss(nn.Layer):
+    """Noise Contrastive Estimation loss funtion
+
+    Noise Contrastive Estimation (NCE) is an approximation method that is used to
+    work around the huge computational cost of large softmax layer.
+    The basic idea is to convert the prediction problem into classification problem
+    at training stage. It has been proved that these two criterions converges to
+    the same minimal point as long as noise distribution is close enough to real one.
+
+    NCE bridges the gap between generative models and discriminative models,
+    rather than simply speedup the softmax layer.
+    With NCE, you can turn almost anything into posterior with less effort (I think).
+
+    Refs:
+    NCE：http://www.cs.helsinki.fi/u/ahyvarin/papers/Gutmann10AISTATS.pdf
+    Thanks: https://github.com/mingen-pan/easy-to-use-NCE-RNN-for-Pytorch/blob/master/nce.py
+
+    Examples:
+    Q = Q_from_tokens(output_dim)
+    NCELoss(Q)
+    """
+
+    def __init__(self, Q, noise_ratio=100, Z_offset=9.5):
+        """Noise Contrastive Estimation loss funtion
+
+        Args:
+            Q (tensor): prior model, uniform or guassian
+            noise_ratio (int, optional): noise sampling times. Defaults to 100.
+            Z_offset (float, optional): scale of post processing the score. Defaults to 9.5.
+        """
+        super(NCELoss, self).__init__()
+        assert type(noise_ratio) is int
+        self.Q = paddle.to_tensor(Q, stop_gradient=False)
+        self.N = self.Q.shape[0]
+        self.K = noise_ratio
+        self.Z_offset = Z_offset
+
+    def forward(self, output, target):
+        """Forward inference
+        
+        Args:
+            output (tensor): the model output, which is the input of loss function
+        """
+        output = paddle.reshape(output, [-1, self.N])
+        B = output.shape[0]
+        noise_idx = self.get_noise(B)
+        idx = self.get_combined_idx(target, noise_idx)
+        P_target, P_noise = self.get_prob(idx, output, sep_target=True)
+        Q_target, Q_noise = self.get_Q(idx)
+        loss = self.nce_loss(P_target, P_noise, Q_noise, Q_target)
+        return loss.mean()
+
+    def get_Q(self, idx, sep_target=True):
+        """Get prior model of batchsize data
+        """
+        idx_size = idx.size
+        prob_model = paddle.to_tensor(
+            self.Q.numpy()[paddle.reshape(idx, [-1]).numpy()])
+        prob_model = paddle.reshape(prob_model, [idx.shape[0], idx.shape[1]])
+        if sep_target:
+            return prob_model[:, 0], prob_model[:, 1:]
+        else:
+            return prob_model
+
+    def get_prob(self, idx, scores, sep_target=True):
+        """Post processing the score of post model(output of nn) of batchsize data
+        """
+        scores = self.get_scores(idx, scores)
+        scale = paddle.to_tensor([self.Z_offset], dtype='float32')
+        scores = paddle.add(scores, -scale)
+        prob = paddle.exp(scores)
+        if sep_target:
+            return prob[:, 0], prob[:, 1:]
+        else:
+            return prob
+
+    def get_scores(self, idx, scores):
+        """Get the score of post model(output of nn) of batchsize data
+        """
+        B, N = scores.shape
+        K = idx.shape[1]
+        idx_increment = paddle.to_tensor(
+            N * paddle.reshape(paddle.arange(B), [B, 1]) * paddle.ones([1, K]),
+            dtype="int64",
+            stop_gradient=False)
+        new_idx = idx_increment + idx
+        new_scores = paddle.index_select(
+            paddle.reshape(scores, [-1]), paddle.reshape(new_idx, [-1]))
+
+        return paddle.reshape(new_scores, [B, K])
+
+    def get_noise(self, batch_size, uniform=True):
+        """Select noise sample
+        """
+        if uniform:
+            noise = np.random.randint(self.N, size=self.K * batch_size)
+        else:
+            noise = np.random.choice(
+                self.N, self.K * batch_size, replace=True, p=self.Q.data)
+        noise = paddle.to_tensor(noise, dtype='int64', stop_gradient=False)
+        noise_idx = paddle.reshape(noise, [batch_size, self.K])
+        return noise_idx
+
+    def get_combined_idx(self, target_idx, noise_idx):
+        """Combined target and noise
+        """
+        target_idx = paddle.reshape(target_idx, [-1, 1])
+        return paddle.concat((target_idx, noise_idx), 1)
+
+    def nce_loss(self, prob_model, prob_noise_in_model, prob_noise,
+                 prob_target_in_noise):
+        """Combined the loss of target and noise
+        """
+
+        def safe_log(tensor):
+            """Safe log
+            """
+            EPSILON = 1e-10
+            return paddle.log(EPSILON + tensor)
+
+        model_loss = safe_log(prob_model /
+                              (prob_model + self.K * prob_target_in_noise))
+        model_loss = paddle.reshape(model_loss, [-1])
+
+        noise_loss = paddle.sum(
+            safe_log((self.K * prob_noise) /
+                     (prob_noise_in_model + self.K * prob_noise)), -1)
+        noise_loss = paddle.reshape(noise_loss, [-1])
+
+        loss = -(model_loss + noise_loss)
+
+        return loss

paddlespeech/vector/utils/vector_utils.py

@@ -30,3 +30,11 @@ def get_chunks(seg_dur, audio_id, audio_duration):
         for i in range(num_chunks)
     ]
     return chunk_lst
+
+
+def Q_from_tokens(token_num):
+    """Get prior model, data from uniform, would support others(guassian) in future
+    """
+    freq = [1] * token_num
+    Q = paddle.to_tensor(freq, dtype='float64')
+    return Q / Q.sum()