eliminate shape and slice

paddlespeech/s2t/modules/conformer_convolution.py

@@ -127,11 +127,11 @@ class ConvolutionModule(nn.Layer):
         x = x.transpose([0, 2, 1])  # [B, C, T]
 
         # mask batch padding
-        if paddle.shape(mask_pad)[2] > 0:  # time > 0
+        if mask_pad.shape[2] > 0:  # time > 0
             x = x.masked_fill(mask_pad, 0.0)
 
         if self.lorder > 0:
-            if paddle.shape(cache)[2] == 0:  # cache_t == 0
+            if cache.shape[2] == 0:  # cache_t == 0
                 x = nn.functional.pad(
                     x, [self.lorder, 0], 'constant', 0.0, data_format='NCL')
             else:
@@ -161,7 +161,7 @@ class ConvolutionModule(nn.Layer):
         x = self.pointwise_conv2(x)
 
         # mask batch padding
-        if paddle.shape(mask_pad)[2] > 0:  # time > 0
+        if mask_pad.shape[2] > 0:  # time > 0
             x = x.masked_fill(mask_pad, 0.0)
 
         x = x.transpose([0, 2, 1])  # [B, T, C]

paddlespeech/s2t/modules/decoder.py

@@ -243,7 +243,7 @@ class TransformerDecoder(BatchScorerInterface, nn.Layer):
             ]
 
         # batch decoding
-        ys_mask = subsequent_mask(paddle.shape(ys)[-1]).unsqueeze(0)  # (B,L,L)
+        ys_mask = subsequent_mask(ys.shape[-1]).unsqueeze(0)  # (B,L,L)
         xs_mask = make_xs_mask(xs).unsqueeze(1)  # (B,1,T)
         logp, states = self.forward_one_step(
             xs, xs_mask, ys, ys_mask, cache=batch_state)

paddlespeech/s2t/modules/encoder.py

@@ -246,7 +246,7 @@ class BaseEncoder(nn.Layer):
             # tensor zeros([0,0,0,0]) support [i:i+1] slice, will return zeros([0,0,0,0]) tensor
             # raw code as below:
             #   att_cache=att_cache[i:i+1] if elayers > 0 else att_cache,
-            #   cnn_cache=cnn_cache[i:i+1] if paddle.shape(cnn_cache)[0] > 0 else cnn_cache,
+            #   cnn_cache=cnn_cache[i:i+1] if cnn_cache.shape[0] > 0 else cnn_cache,
             xs, _, new_att_cache, new_cnn_cache = layer(
                 xs,
                 att_mask,

paddlespeech/s2t/modules/loss.py

@@ -85,7 +85,7 @@ class CTCLoss(nn.Layer):
         Returns:
             [paddle.Tensor]: scalar. If reduction is 'none', then (N), where N = \text{batch size}.
         """
-        B = paddle.shape(logits)[0]
+        B = logits.shape[0]
         # warp-ctc need logits, and do softmax on logits by itself
         # warp-ctc need activation with shape [T, B, V + 1]
         # logits: (B, L, D) -> (L, B, D)
@@ -158,7 +158,7 @@ class LabelSmoothingLoss(nn.Layer):
         Returns:
             loss (paddle.Tensor) : The KL loss, scalar float value
         """
-        B, T, D = paddle.shape(x)
+        B, T, D = x.shape
         assert D == self.size
         x = x.reshape((-1, self.size))
         target = target.reshape([-1])

paddlespeech/s2t/modules/subsampling.py

@@ -192,8 +192,8 @@ class Conv2dSubsampling6(Conv2dSubsampling):
         """
         x = x.unsqueeze(1)  # (b, c, t, f)
         x = self.conv(x)
-        b, c, t, f = paddle.shape(x)
-        x = self.linear(x.transpose([0, 2, 1, 3]).reshape([b, t, c * f]))
+        b, c, t, f = x.shape
+        x = self.linear(x.transpose([0, 2, 1, 3]).reshape([b, -1, c * f]))
         x, pos_emb = self.pos_enc(x, offset)
         return x, pos_emb, x_mask[:, :, :-2:2][:, :, :-4:3]
 
@@ -245,6 +245,7 @@ class Conv2dSubsampling8(Conv2dSubsampling):
         """
         x = x.unsqueeze(1)  # (b, c, t, f)
         x = self.conv(x)
-        x = self.linear(x.transpose([0, 2, 1, 3]).reshape([b, t, c * f]))
+        b, c, t, f = x.shape
+        x = self.linear(x.transpose([0, 2, 1, 3]).reshape([b, -1, c * f]))
         x, pos_emb = self.pos_enc(x, offset)
         return x, pos_emb, x_mask[:, :, :-2:2][:, :, :-2:2][:, :, :-2:2]