remove size,test=asr

paddlespeech/s2t/__init__.py

@@ -189,25 +189,6 @@ if not hasattr(paddle.Tensor, 'contiguous'):
     paddle.static.Variable.contiguous = contiguous
 
 
-def size(xs: paddle.Tensor, *args: int) -> paddle.Tensor:
-    nargs = len(args)
-    assert (nargs <= 1)
-    s = paddle.shape(xs)
-    if nargs == 1:
-        return s[args[0]]
-    else:
-        return s
-
-
-#`to_static` do not process `size` property, maybe some `paddle` api dependent on it.
-logger.debug(
-    "override size of paddle.Tensor "
-    "(`to_static` do not process `size` property, maybe some `paddle` api dependent on it), remove this when fixed!"
-)
-paddle.Tensor.size = size
-paddle.static.Variable.size = size
-
-
 def view(xs: paddle.Tensor, *args: int) -> paddle.Tensor:
     return xs.reshape(args)
 

paddlespeech/s2t/decoders/beam_search/beam_search.py

@@ -194,7 +194,7 @@ class BeamSearch(paddle.nn.Layer):
 
         Args:
             hyp (Hypothesis): Hypothesis with prefix tokens to score
-            ids (paddle.Tensor): 1D tensor of new partial tokens to score, 
+            ids (paddle.Tensor): 1D tensor of new partial tokens to score,
                 len(ids) < n_vocab
             x (paddle.Tensor): Corresponding input feature, (T, D)
 
@@ -224,14 +224,14 @@ class BeamSearch(paddle.nn.Layer):
             ids (paddle.Tensor): The partial token ids(Global) to compute topk.
 
         Returns:
-            Tuple[paddle.Tensor, paddle.Tensor]: 
+            Tuple[paddle.Tensor, paddle.Tensor]:
                 The topk full token ids and partial token ids.
                 Their shapes are `(self.beam_size,)`.
                 i.e. (global ids, global relative local ids).
 
         """
         # no pre beam performed, `ids` equal to `weighted_scores`
-        if weighted_scores.size(0) == ids.size(0):
+        if weighted_scores.shape[0] == ids.shape[0]:
             top_ids = weighted_scores.topk(
                 self.beam_size)[1]  # index in n_vocab
             return top_ids, top_ids
@@ -374,8 +374,8 @@ class BeamSearch(paddle.nn.Layer):
         elif maxlenratio < 0:
             maxlen = -1 * int(maxlenratio)
         else:
-            maxlen = max(1, int(maxlenratio * x.size(0)))
-        minlen = int(minlenratio * x.size(0))
+            maxlen = max(1, int(maxlenratio * x.shape[0]))
+        minlen = int(minlenratio * x.shape[0])
         logger.info("decoder input length: " + str(x.shape[0]))
         logger.info("max output length: " + str(maxlen))
         logger.info("min output length: " + str(minlen))

paddlespeech/s2t/decoders/scorers/ctc.py

@@ -69,7 +69,7 @@ class CTCPrefixScorer(BatchPartialScorerInterface):
                 return sc[i], st[i]
             else:  # for CTCPrefixScorePD (need new_id > 0)
                 r, log_psi, f_min, f_max, scoring_idmap = state
-                s = log_psi[i, new_id].expand(log_psi.size(1))
+                s = log_psi[i, new_id].expand(log_psi.shape[1])
                 if scoring_idmap is not None:
                     return r[:, :, i, scoring_idmap[i, new_id]], s, f_min, f_max
                 else:
@@ -107,7 +107,7 @@ class CTCPrefixScorer(BatchPartialScorerInterface):
 
         """
         logp = self.ctc.log_softmax(x.unsqueeze(0))  # assuming batch_size = 1
-        xlen = paddle.to_tensor([logp.size(1)])
+        xlen = paddle.to_tensor([logp.shape[1]])
         self.impl = CTCPrefixScorePD(logp, xlen, 0, self.eos)
         return None
 

paddlespeech/s2t/decoders/scorers/ctc_prefix_score.py

@@ -33,9 +33,9 @@ class CTCPrefixScorePD():
         self.logzero = -10000000000.0
         self.blank = blank
         self.eos = eos
-        self.batch = x.size(0)
-        self.input_length = x.size(1)
-        self.odim = x.size(2)
+        self.batch = x.shape[0]
+        self.input_length = x.shape[1]
+        self.odim = x.shape[2]
         self.dtype = x.dtype
 
         # Pad the rest of posteriors in the batch
@@ -76,8 +76,7 @@ class CTCPrefixScorePD():
         last_ids = [yi[-1] for yi in y]  # last output label ids
         n_bh = len(last_ids)  # batch * hyps
         n_hyps = n_bh // self.batch  # assuming each utterance has the same # of hyps
-        self.scoring_num = scoring_ids.size(
-            -1) if scoring_ids is not None else 0
+        self.scoring_num = scoring_ids.shape[-1] if scoring_ids is not None else 0
         # prepare state info
         if state is None:
             r_prev = paddle.full(
@@ -153,7 +152,7 @@ class CTCPrefixScorePD():
 
         # compute forward probabilities log(r_t^n(h)) and log(r_t^b(h))
         for t in range(start, end):
-            rp = r[t - 1]  # (2 x BW x O') 
+            rp = r[t - 1]  # (2 x BW x O')
             rr = paddle.stack([rp[0], log_phi[t - 1], rp[0], rp[1]]).view(
                 2, 2, n_bh, snum)  # (2,2,BW,O')
             r[t] = paddle.logsumexp(rr, 1) + x_[:, t]
@@ -227,7 +226,7 @@ class CTCPrefixScorePD():
         if self.x.shape[1] < x.shape[1]:  # self.x (2,T,B,O); x (B,T,O)
             # Pad the rest of posteriors in the batch
             # TODO(takaaki-hori): need a better way without for-loops
-            xlens = [x.size(1)]
+            xlens = [x.shape[1]]
             for i, l in enumerate(xlens):
                 if l < self.input_length:
                     x[i, l:, :] = self.logzero
@@ -237,7 +236,7 @@ class CTCPrefixScorePD():
             xb = xn[:, :, self.blank].unsqueeze(2).expand(-1, -1, self.odim)
             self.x = paddle.stack([xn, xb])  # (2, T, B, O)
             self.x[:, :tmp_x.shape[1], :, :] = tmp_x
-            self.input_length = x.size(1)
+            self.input_length = x.shape[1]
             self.end_frames = paddle.to_tensor(xlens) - 1
 
     def extend_state(self, state):
@@ -318,16 +317,16 @@ class CTCPrefixScore():
             r[0, 0] = xs[0]
             r[0, 1] = self.logzero
         else:
-            # Although the code does not exactly follow Algorithm 2, 
-            # we don't have to change it because we can assume 
-            # r_t(h)=0 for t < |h| in CTC forward computation 
+            # Although the code does not exactly follow Algorithm 2,
+            # we don't have to change it because we can assume
+            # r_t(h)=0 for t < |h| in CTC forward computation
             # (Note: we assume here that index t starts with 0).
             # The purpose of this difference is to reduce the number of for-loops.
             # https://github.com/espnet/espnet/pull/3655
-            # where we start to accumulate r_t(h) from t=|h| 
-            # and iterate r_t(h) = (r_{t-1}(h) + ...) to T-1, 
+            # where we start to accumulate r_t(h) from t=|h|
+            # and iterate r_t(h) = (r_{t-1}(h) + ...) to T-1,
             # avoiding accumulating zeros for t=1~|h|-1.
-            # Thus, we need to set r_{|h|-1}(h) = 0, 
+            # Thus, we need to set r_{|h|-1}(h) = 0,
             # i.e., r[output_length-1] = logzero, for initialization.
             # This is just for reducing the computation.
             r[output_length - 1] = self.logzero

paddlespeech/s2t/models/u2/u2.py

@@ -775,7 +775,7 @@ class U2DecodeModel(U2BaseModel):
         """
         self.eval()
         x = paddle.to_tensor(x).unsqueeze(0)
-        ilen = x.size(1)
+        ilen = x.shape[1]
         enc_output, _ = self._forward_encoder(x, ilen)
         return enc_output.squeeze(0)
 

paddlespeech/s2t/modules/decoder.py

@@ -242,7 +242,7 @@ class TransformerDecoder(BatchScorerInterface, nn.Layer):
             ]
 
         # batch decoding
-        ys_mask = subsequent_mask(ys.size(-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/embedding.py

@@ -115,7 +115,7 @@ class PositionalEncoding(nn.Layer, PositionalEncodingInterface):
         assert offset + x.shape[
             1] < self.max_len, "offset: {} + x.shape[1]: {} is larger than the max_len: {}".format(
                 offset, x.shape[1], self.max_len)
-        #TODO(Hui Zhang): using T = x.size(1), __getitem__ not support Tensor
+        #TODO(Hui Zhang): using T = x.shape[1], __getitem__ not support Tensor
         pos_emb = self.pe[:, offset:offset + T]
         x = x * self.xscale + pos_emb
         return self.dropout(x), self.dropout(pos_emb)
@@ -165,6 +165,6 @@ class RelPositionalEncoding(PositionalEncoding):
             1] < self.max_len, "offset: {} + x.shape[1]: {} is larger than the max_len: {}".format(
                 offset, x.shape[1], self.max_len)
         x = x * self.xscale
-        #TODO(Hui Zhang): using x.size(1), __getitem__ not support Tensor
+        #TODO(Hui Zhang): using x.shape[1], __getitem__ not support Tensor
         pos_emb = self.pe[:, offset:offset + x.shape[1]]
         return self.dropout(x), self.dropout(pos_emb)

paddlespeech/s2t/utils/tensor_utils.py

@@ -58,8 +58,8 @@ def pad_sequence(sequences: List[paddle.Tensor],
         >>> a = paddle.ones(25, 300)
         >>> b = paddle.ones(22, 300)
         >>> c = paddle.ones(15, 300)
-        >>> pad_sequence([a, b, c]).size()
-        paddle.Tensor([25, 3, 300])
+        >>> pad_sequence([a, b, c]).shape
+        [25, 3, 300]
 
     Note:
         This function returns a Tensor of size ``T x B x *`` or ``B x T x *``
@@ -79,7 +79,7 @@ def pad_sequence(sequences: List[paddle.Tensor],
 
     # assuming trailing dimensions and type of all the Tensors
     # in sequences are same and fetching those from sequences[0]
-    max_size = sequences[0].size()
+    max_size = sequences[0].shape
     # (TODO Hui Zhang): slice not supprot `end==start`
     # trailing_dims = max_size[1:]
     trailing_dims = tuple(max_size[1:].numpy().tolist()) if sequences[0].ndim >= 2 else ()