Update perceiver

7 files changed, 192 insertions, 14 deletions

diff --git a/text_recognizer/models/perceiver.py b/text_recognizer/models/perceiver.py
new file mode 100644
index 0000000..c482235
--- /dev/null
+++ b/text_recognizer/models/perceiver.py
@@ -0,0 +1,76 @@
+"""Lightning model for base Perceiver."""
+from typing import Optional, Tuple, Type
+
+from omegaconf import DictConfig
+import torch
+from torch import nn, Tensor
+
+from text_recognizer.data.mappings import EmnistMapping
+from text_recognizer.models.base import LitBase
+from text_recognizer.models.metrics import CharacterErrorRate
+
+
+class LitPerceiver(LitBase):
+    """A PyTorch Lightning model for transformer networks."""
+
+    def __init__(
+        self,
+        network: Type[nn.Module],
+        loss_fn: Type[nn.Module],
+        optimizer_config: DictConfig,
+        lr_scheduler_config: Optional[DictConfig],
+        mapping: EmnistMapping,
+        max_output_len: int = 682,
+        start_token: str = "<s>",
+        end_token: str = "<e>",
+        pad_token: str = "<p>",
+    ) -> None:
+        super().__init__(
+            network, loss_fn, optimizer_config, lr_scheduler_config, mapping
+        )
+        self.max_output_len = max_output_len
+        self.start_token = start_token
+        self.end_token = end_token
+        self.pad_token = pad_token
+        self.start_index = int(self.mapping.get_index(self.start_token))
+        self.end_index = int(self.mapping.get_index(self.end_token))
+        self.pad_index = int(self.mapping.get_index(self.pad_token))
+        self.ignore_indices = set([self.start_index, self.end_index, self.pad_index])
+        self.val_cer = CharacterErrorRate(self.ignore_indices)
+        self.test_cer = CharacterErrorRate(self.ignore_indices)
+
+    def forward(self, data: Tensor) -> Tensor:
+        """Forward pass with the transformer network."""
+        return self.predict(data)
+
+    def training_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> Tensor:
+        """Training step."""
+        data, targets = batch
+        logits = self.network(data)
+        loss = self.loss_fn(logits, targets)
+        self.log("train/loss", loss)
+        return loss
+
+    def validation_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> None:
+        """Validation step."""
+        data, targets = batch
+        preds = self.predict(data)
+        self.val_acc(preds, targets)
+        self.log("val/acc", self.val_acc, on_step=False, on_epoch=True)
+        self.val_cer(preds, targets)
+        self.log("val/cer", self.val_cer, on_step=False, on_epoch=True, prog_bar=True)
+
+    def test_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> None:
+        """Test step."""
+        data, targets = batch
+
+        # Compute the text prediction.
+        pred = self(data)
+        self.test_cer(pred, targets)
+        self.log("test/cer", self.test_cer, on_step=False, on_epoch=True, prog_bar=True)
+        self.test_acc(pred, targets)
+        self.log("test/acc", self.test_acc, on_step=False, on_epoch=True)
+
+    @torch.no_grad()
+    def predict(self, x: Tensor) -> Tensor:
+        return self.network(x).argmax(dim=1)
diff --git a/text_recognizer/networks/conv_perceiver.py b/text_recognizer/networks/conv_perceiver.py
new file mode 100644
index 0000000..cda5e0a
--- /dev/null
+++ b/text_recognizer/networks/conv_perceiver.py
@@ -0,0 +1,62 @@
+"""Perceiver network module."""
+from typing import Optional, Tuple, Type
+
+from torch import nn, Tensor
+
+from text_recognizer.networks.perceiver.perceiver import PerceiverIO
+from text_recognizer.networks.transformer.embeddings.axial import (
+    AxialPositionalEmbedding,
+)
+
+
+class ConvPerceiver(nn.Module):
+    """Base transformer network."""
+
+    def __init__(
+        self,
+        input_dims: Tuple[int, int, int],
+        hidden_dim: int,
+        queries_dim: int,
+        num_classes: int,
+        pad_index: Tensor,
+        encoder: Type[nn.Module],
+        decoder: PerceiverIO,
+        max_length: int,
+        pixel_embedding: AxialPositionalEmbedding,
+    ) -> None:
+        super().__init__()
+        self.input_dims = input_dims
+        self.hidden_dim = hidden_dim
+        self.num_classes = num_classes
+        self.pad_index = pad_index
+        self.max_length = max_length
+        self.encoder = encoder
+        self.decoder = decoder
+        self.pixel_embedding = pixel_embedding
+        self.to_queries = nn.Linear(self.hidden_dim, queries_dim)
+        self.conv = nn.Conv2d(
+            in_channels=self.encoder.out_channels,
+            out_channels=self.hidden_dim,
+            kernel_size=1,
+        )
+
+    def encode(self, x: Tensor) -> Tensor:
+        z = self.encoder(x)
+        z = self.conv(z)
+        z = self.pixel_embedding(z)
+        z = z.flatten(start_dim=2)
+
+        # Permute tensor from [B, E, Ho * Wo] to [B, Sx, E]
+        z = z.permute(0, 2, 1)
+        return z
+
+    def decode(self, z: Tensor) -> Tensor:
+        queries = self.to_queries(z[:, : self.max_length, :])
+        logits = self.decoder(data=z, queries=queries)
+        logits = logits.permute(0, 2, 1)  # [B, C, Sy]
+        return logits
+
+    def forward(self, x: Tensor) -> Tensor:
+        z = self.encode(x)
+        logits = self.decode(z)
+        return logits
diff --git a/text_recognizer/networks/perceiver/__init__.py b/text_recognizer/networks/perceiver/__init__.py
index e69de29..ac2c102 100644
--- a/text_recognizer/networks/perceiver/__init__.py
+++ b/text_recognizer/networks/perceiver/__init__.py
@@ -0,0 +1 @@
+from text_recognizer.networks.perceiver.perceiver import PerceiverIO
diff --git a/text_recognizer/networks/perceiver/attention.py b/text_recognizer/networks/perceiver/attention.py
index 66aeaa8..0ee51b1 100644
--- a/text_recognizer/networks/perceiver/attention.py
+++ b/text_recognizer/networks/perceiver/attention.py
@@ -36,11 +36,11 @@ class Attention(nn.Module):
         q, v, k = map(lambda t: rearrange(t, "b n (h d) -> (b h) n d", h=h), (q, k, v))
         sim = einsum("b i d, b j d -> b i j", q, k) * self.scale
 
-        if mask is not None:
-            mask = rearrange(mask, "b ... -> b (...)")
-            max_neg_val = -torch.finfo(sim.dtype).max
-            mask = repeat(mask, "b j -> (b h) () j", h=h)
-            sim.masked_fill_(~mask, max_neg_val)
+        # if mask is not None:
+        #     mask = rearrange(mask, "b ... -> b (...)")
+        #     max_neg_val = -torch.finfo(sim.dtype).max
+        #     mask = repeat(mask, "b j -> (b h) () j", h=h)
+        #     sim.masked_fill_(~mask, max_neg_val)
 
         attn = sim.softmax(dim=-1)
         out = einsum("b i j, b j d -> b i d", attn, v)
diff --git a/text_recognizer/networks/perceiver/perceiver.py b/text_recognizer/networks/perceiver/perceiver.py
index 65ee20c..d4bca0b 100644
--- a/text_recognizer/networks/perceiver/perceiver.py
+++ b/text_recognizer/networks/perceiver/perceiver.py
@@ -2,9 +2,9 @@
 
 A copy from lucidrains.
 """
-from itertools import repeat
 from typing import Optional
 
+from einops import repeat, rearrange
 import torch
 from torch import nn, Tensor
 
@@ -44,13 +44,17 @@ class PerceiverIO(nn.Module):
 
         self.layers = nn.ModuleList(
             [
-                [
-                    PreNorm(
-                        latent_dim,
-                        Attention(latent_dim, heads=latent_heads, dim_head=latent_dim),
-                    ),
-                    PreNorm(latent_dim, FeedForward(latent_dim)),
-                ]
+                nn.ModuleList(
+                    [
+                        PreNorm(
+                            latent_dim,
+                            Attention(
+                                latent_dim, heads=latent_heads, dim_head=latent_dim
+                            ),
+                        ),
+                        PreNorm(latent_dim, FeedForward(latent_dim)),
+                    ]
+                )
                 for _ in range(depth)
             ]
         )
@@ -69,7 +73,7 @@ class PerceiverIO(nn.Module):
         self, data: Tensor, queries: Tensor, mask: Optional[Tensor] = None
     ) -> Tensor:
         b = data.shape[0]
-        x = repeat(self.latents, "nd -> bnd", b=b)
+        x = repeat(self.latents, "n d -> b n d", b=b)
 
         cross_attn, cross_ff = self.cross_attn_block
 
diff --git a/training/conf/model/lit_perceiver.yaml b/training/conf/model/lit_perceiver.yaml
new file mode 100644
index 0000000..6d1ec82
--- /dev/null
+++ b/training/conf/model/lit_perceiver.yaml
@@ -0,0 +1,5 @@
+_target_: text_recognizer.models.LitPerceiver
+max_output_len: 682
+start_token: <s>
+end_token: <e>
+pad_token: <p>
diff --git a/training/conf/network/conv_perceiver.yaml b/training/conf/network/conv_perceiver.yaml
new file mode 100644
index 0000000..e6906fa
--- /dev/null
+++ b/training/conf/network/conv_perceiver.yaml
@@ -0,0 +1,30 @@
+_target_: text_recognizer.networks.ConvPerceiver
+input_dims: [1, 1, 576, 640]
+hidden_dim: &hidden_dim 144
+num_classes: &num_classes 58
+queries_dim: &queries_dim 16
+max_length: 89
+pad_index: 3
+encoder:
+  _target_: text_recognizer.networks.EfficientNet
+  arch: b0
+  stochastic_dropout_rate: 0.2
+  bn_momentum: 0.99
+  bn_eps: 1.0e-3
+  depth: 5
+  out_channels: *hidden_dim
+decoder:
+  _target_: text_recognizer.networks.perceiver.PerceiverIO
+  dim: *hidden_dim
+  cross_heads: 1
+  cross_head_dim: 64
+  num_latents: 256
+  latent_dim: 512
+  latent_heads: 8
+  depth: 6
+  queries_dim: *queries_dim
+  logits_dim: *num_classes
+pixel_embedding:
+  _target_: text_recognizer.networks.transformer.AxialPositionalEmbedding
+  dim: *hidden_dim
+  shape: [3, 64]