Merge pull request #697 from 9173860/patch-3

Vectorized infer_beam_batch for improved performance

manga_translator/ocr/model_48px.py

@@ -99,7 +99,7 @@ async def _infer(self, image: np.ndarray, textlines: List[Quadrilateral], args:
             if self.use_gpu:
                 image_tensor = image_tensor.to(self.device)
             with torch.no_grad():
-                ret = self.model.infer_beam_batch(image_tensor, widths, beams_k = 5, max_seq_length = 255)
+                ret = self.model.infer_beam_batch_tensor(image_tensor, widths, beams_k = 5, max_seq_length = 255)
             for i, (pred_chars_index, prob, fg_pred, bg_pred, fg_ind_pred, bg_ind_pred) in enumerate(ret):
                 if prob < 0.2:
                     continue
@@ -497,16 +497,21 @@ def __init__(self, dictionary, max_len):
         self.backbone = ConvNext_FeatureExtractor(48, 3, embd_dim)
         self.encoders = nn.ModuleList()
         self.decoders = nn.ModuleList()
+
         for i in range(4) :
             encoder = nn.TransformerEncoderLayer(embd_dim, nhead, dropout = 0, batch_first = True, norm_first = True)
             encoder.self_attn = XposMultiheadAttention(embd_dim, nhead, self_attention = True)
             encoder.forward = transformer_encoder_forward
             self.encoders.append(encoder)
+        self.encoders.forward = self.encoder_forward
+
         for i in range(5) :
             decoder = nn.TransformerDecoderLayer(embd_dim, nhead, dropout = 0, batch_first = True, norm_first = True)
             decoder.self_attn = XposMultiheadAttention(embd_dim, nhead, self_attention = True)
             decoder.multihead_attn = XposMultiheadAttention(embd_dim, nhead, encoder_decoder_attention = True)
             self.decoders.append(decoder)
+        self.decoders.forward = self.decoder_forward
+
         self.embd = nn.Embedding(self.dict_size, embd_dim)
         self.pred1 = nn.Sequential(nn.Linear(embd_dim, embd_dim), nn.GELU(), nn.Dropout(0.15))
         self.pred = nn.Linear(embd_dim, self.dict_size)
@@ -517,6 +522,37 @@ def __init__(self, dictionary, max_len):
         self.color_pred_fg_ind = nn.Linear(64, 2)
         self.color_pred_bg_ind = nn.Linear(64, 2)
 
+    def encoder_forward(self, memory, encoder_mask):
+        for layer in self.encoders :
+            memory = layer(layer, src = memory, src_key_padding_mask = encoder_mask)
+        return memory
+
+    def decoder_forward(
+        self,
+        embd: torch.Tensor,
+        cached_activations: torch.Tensor,  # Shape [N, L, T, E] where L=num_layers, T=sequence length, E=embedding size
+        memory: torch.Tensor,  # Shape [N, H, W, C] (Encoder memory output)
+        memory_mask: torch.BoolTensor,
+        step: int
+    ):
+
+        layer: nn.TransformerDecoderLayer
+        tgt = embd  # N, 1, E for the last token embedding
+
+        for l, layer in enumerate(self.decoders):
+            combined_activations = cached_activations[:, l, :step, :]  # N, T, E
+            combined_activations = torch.cat([combined_activations, tgt], dim=1)  # N, T+1, E
+            cached_activations[:, l, step, :] = tgt.squeeze(1)
+
+            # Update cache and perform self attention
+            tgt = tgt + layer.self_attn(layer.norm1(tgt), layer.norm1(combined_activations), layer.norm1(combined_activations), q_offset=step)[0]
+            tgt = tgt + layer.multihead_attn(layer.norm2(tgt), memory, memory, key_padding_mask=memory_mask, q_offset=step)[0]
+            tgt = tgt + layer._ff_block(layer.norm3(tgt))
+
+        cached_activations[:, l+1, step, :] = tgt.squeeze(1) # Append the new activations
+
+        return tgt.squeeze_(1), cached_activations
+
     def forward(self,
         img: torch.FloatTensor,
         char_idx: torch.LongTensor,
@@ -621,6 +657,119 @@ def infer_beam_batch(self, img: torch.FloatTensor, img_widths: List[int], beams_
             result.append((cur_hypo.out_idx[1:], cur_hypo.prob(), fg_pred[0], bg_pred[0], fg_ind_pred[0], bg_ind_pred[0]))
         return result
 
+    def infer_beam_batch_tensor(self, img: torch.FloatTensor, img_widths: List[int], beams_k: int = 5, start_tok = 1, end_tok = 2, pad_tok = 0, max_finished_hypos: int = 2, max_seq_length = 384):
+        N, C, H, W = img.shape
+        assert H == 48 and C == 3
+
+        memory = self.backbone(img)
+        memory = einops.rearrange(memory, 'N C 1 W -> N W C')
+        valid_feats_length = [(x + 3) // 4 + 2 for x in img_widths]
+        input_mask = torch.zeros(N, memory.size(1), dtype = torch.bool).to(img.device)
+
+        for i, l in enumerate(valid_feats_length):
+            input_mask[i, l:] = True
+        memory = self.encoders(memory, input_mask) # N, W, Dim
+
+        out_idx = torch.full((N, 1), start_tok, dtype=torch.long, device=img.device)  # Shape [N, 1]
+        cached_activations = torch.zeros(N, len(self.decoders)+1, max_seq_length, 320, device=img.device)  # [N, L, S, E]
+        log_probs = torch.zeros(N, 1, device=img.device)  # Shape [N, 1]        # N, E
+        idx_embedded = self.embd(out_idx[:, -1:]) 
+
+        decoded, cached_activations = self.decoders(idx_embedded, cached_activations, memory, input_mask, 0)
+        pred_char_logprob = self.pred(self.pred1(decoded)).log_softmax(-1)   # N, n_chars
+        pred_chars_values, pred_chars_index = torch.topk(pred_char_logprob, beams_k, dim = 1)  # N, k
+
+        out_idx = torch.cat([out_idx.unsqueeze(1).expand(-1, beams_k, -1), pred_chars_index.unsqueeze(-1)], dim=-1).reshape(-1, 2)  # Shape [N * k, 2]
+        log_probs = pred_chars_values.view(-1, 1)  # Shape [N * k, 1]
+        memory = memory.repeat_interleave(beams_k, dim=0)
+        input_mask = input_mask.repeat_interleave(beams_k, dim=0)
+        cached_activations = cached_activations.repeat_interleave(beams_k, dim=0)
+        batch_index = torch.arange(N).repeat_interleave(beams_k, dim=0).to(img.device)
+
+        finished_hypos = defaultdict(list)
+        N_remaining = N
+
+        for step in range(1, max_seq_length):
+            idx_embedded = self.embd(out_idx[:, -1:])
+            decoded, cached_activations = self.decoders(idx_embedded, cached_activations, memory, input_mask, step)
+            pred_char_logprob = self.pred(self.pred1(decoded)).log_softmax(-1)  # Shape [N * k, dict_size]
+            pred_chars_values, pred_chars_index = torch.topk(pred_char_logprob, beams_k, dim=1)  # [N * k, k]
+
+            finished = out_idx[:, -1] == end_tok
+            pred_chars_values[finished] = 0
+            pred_chars_index[finished] = end_tok
+
+            # Extend hypotheses
+            new_out_idx = out_idx.unsqueeze(1).expand(-1, beams_k, -1)  # Shape [N * k, k, seq_len]
+            new_out_idx = torch.cat([new_out_idx, pred_chars_index.unsqueeze(-1)], dim=-1)  # Shape [N * k, k, seq_len + 1]
+            new_out_idx = new_out_idx.view(-1, step + 2)  # Reshape to [N * k^2, seq_len + 1]
+            new_log_probs = log_probs.unsqueeze(1).expand(-1, beams_k, -1) + pred_chars_values.unsqueeze(-1)  # Shape [N * k^2, 1]
+            new_log_probs = new_log_probs.view(-1, 1)  # [N * k^2, 1]
+
+            # Sort and select top-k hypotheses per sample
+            new_out_idx = new_out_idx.view(N_remaining, -1, step + 2)  # [N, k^2, seq_len + 1]
+            new_log_probs = new_log_probs.view(N_remaining, -1)  # [N, k^2]
+            batch_topk_log_probs, batch_topk_indices = new_log_probs.topk(beams_k, dim=1)  # [N, k]
+
+            # Gather the top-k hypotheses based on log probabilities
+            expanded_topk_indices = batch_topk_indices.unsqueeze(-1).expand(-1, -1, new_out_idx.shape[-1])  # Shape [N, k, seq_len + 1]
+            out_idx = torch.gather(new_out_idx, 1, expanded_topk_indices).reshape(-1, step + 2)  # [N * k, seq_len + 1]
+            log_probs = batch_topk_log_probs.view(-1, 1)  # Reshape to [N * k, 1]
+
+            # Check for finished sequences
+            finished = (out_idx[:, -1] == end_tok)  # Check if the last token is the end token
+            finished = finished.view(N_remaining, beams_k)  # Reshape to [N, k]
+            finished_counts = finished.sum(dim=1)  # Count the number of finished hypotheses per sample
+            finished_batch_indices = (finished_counts >= max_finished_hypos).nonzero(as_tuple=False).squeeze()
+
+            if finished_batch_indices.numel() == 0:
+                continue
+
+            if finished_batch_indices.dim() == 0:
+                finished_batch_indices = finished_batch_indices.unsqueeze(0)
+
+            for idx in finished_batch_indices:
+                batch_log_probs = batch_topk_log_probs[idx]
+                best_beam_idx = batch_log_probs.argmax()
+                finished_hypos[batch_index[beams_k * idx].item()] = \
+                    out_idx[idx * beams_k + best_beam_idx], \
+                    torch.exp(batch_log_probs[best_beam_idx]).item(), \
+                    cached_activations[idx * beams_k + best_beam_idx] 
+
+            remaining_indexs = []
+            for i in range(N_remaining):
+                if i not in finished_batch_indices:
+                    for j in range(beams_k):
+                        remaining_indexs.append(i * beams_k + j)
+
+            if not remaining_indexs:
+                break
+
+            N_remaining = int(len(remaining_indexs) / beams_k)
+            out_idx = out_idx.index_select(0, torch.tensor(remaining_indexs, device=img.device))
+            log_probs = log_probs.index_select(0, torch.tensor(remaining_indexs, device=img.device))
+            memory = memory.index_select(0, torch.tensor(remaining_indexs, device=img.device))
+            cached_activations = cached_activations.index_select(0, torch.tensor(remaining_indexs, device=img.device))
+            input_mask = input_mask.index_select(0, torch.tensor(remaining_indexs, device=img.device))
+            batch_index = batch_index.index_select(0, torch.tensor(remaining_indexs, device=img.device))
+
+        # Ensure we have the correct number of finished hypotheses for each sample
+        assert len(finished_hypos) == N
+
+        # Final output processing and color predictions
+        result = []
+        for i in range(N):
+            final_idx, prob, decoded = finished_hypos[i] 
+            color_feats = self.color_pred1(decoded[-1].unsqueeze(0))
+            fg_pred, bg_pred, fg_ind_pred, bg_ind_pred = \
+                self.color_pred_fg(color_feats), \
+                self.color_pred_bg(color_feats), \
+                self.color_pred_fg_ind(color_feats), \
+                self.color_pred_bg_ind(color_feats)
+            result.append((final_idx[1:], prob, fg_pred[0], bg_pred[0], fg_ind_pred[0], bg_ind_pred[0]))
+
+        return result
+
 import numpy as np
 
 def convert_pl_model(filename: str) :