Spaces:
Runtime error
Runtime error
File size: 8,476 Bytes
01bb3bb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 |
# Scene Text Recognition Model Hub
# Copyright 2022 Darwin Bautista
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
from itertools import permutations
from typing import Any, Optional, Sequence
import numpy as np
import torch
import torch.nn.functional as F
from torch import Tensor
from pytorch_lightning.utilities.types import STEP_OUTPUT
from IndicPhotoOCR.utils.strhub.models.base import CrossEntropySystem
from .model import PARSeq as Model
class PARSeq(CrossEntropySystem):
def __init__(
self,
charset_train: str,
charset_test: str,
max_label_length: int,
batch_size: int,
lr: float,
warmup_pct: float,
weight_decay: float,
img_size: Sequence[int],
patch_size: Sequence[int],
embed_dim: int,
enc_num_heads: int,
enc_mlp_ratio: int,
enc_depth: int,
dec_num_heads: int,
dec_mlp_ratio: int,
dec_depth: int,
perm_num: int,
perm_forward: bool,
perm_mirrored: bool,
decode_ar: bool,
refine_iters: int,
dropout: float,
**kwargs: Any,
) -> None:
super().__init__(charset_train, charset_test, batch_size, lr, warmup_pct, weight_decay)
self.save_hyperparameters()
self.model = Model(
len(self.tokenizer),
max_label_length,
img_size,
patch_size,
embed_dim,
enc_num_heads,
enc_mlp_ratio,
enc_depth,
dec_num_heads,
dec_mlp_ratio,
dec_depth,
decode_ar,
refine_iters,
dropout,
)
# Perm/attn mask stuff
self.rng = np.random.default_rng()
self.max_gen_perms = perm_num // 2 if perm_mirrored else perm_num
self.perm_forward = perm_forward
self.perm_mirrored = perm_mirrored
def forward(self, images: Tensor, max_length: Optional[int] = None) -> Tensor:
return self.model.forward(self.tokenizer, images, max_length)
def gen_tgt_perms(self, tgt):
"""Generate shared permutations for the whole batch.
This works because the same attention mask can be used for the shorter sequences
because of the padding mask.
"""
# We don't permute the position of BOS, we permute EOS separately
max_num_chars = tgt.shape[1] - 2
# Special handling for 1-character sequences
if max_num_chars == 1:
return torch.arange(3, device=self._device).unsqueeze(0)
perms = [torch.arange(max_num_chars, device=self._device)] if self.perm_forward else []
# Additional permutations if needed
max_perms = math.factorial(max_num_chars)
if self.perm_mirrored:
max_perms //= 2
num_gen_perms = min(self.max_gen_perms, max_perms)
# For 4-char sequences and shorter, we generate all permutations and sample from the pool to avoid collisions
# Note that this code path might NEVER get executed since the labels in a mini-batch typically exceed 4 chars.
if max_num_chars < 5:
# Pool of permutations to sample from. We only need the first half (if complementary option is selected)
# Special handling for max_num_chars == 4 which correctly divides the pool into the flipped halves
if max_num_chars == 4 and self.perm_mirrored:
selector = [0, 3, 4, 6, 9, 10, 12, 16, 17, 18, 19, 21]
else:
selector = list(range(max_perms))
perm_pool = torch.as_tensor(
list(permutations(range(max_num_chars), max_num_chars)),
device=self._device,
)[selector]
# If the forward permutation is always selected, no need to add it to the pool for sampling
if self.perm_forward:
perm_pool = perm_pool[1:]
perms = torch.stack(perms)
if len(perm_pool):
i = self.rng.choice(len(perm_pool), size=num_gen_perms - len(perms), replace=False)
perms = torch.cat([perms, perm_pool[i]])
else:
perms.extend(
[torch.randperm(max_num_chars, device=self._device) for _ in range(num_gen_perms - len(perms))]
)
perms = torch.stack(perms)
if self.perm_mirrored:
# Add complementary pairs
comp = perms.flip(-1)
# Stack in such a way that the pairs are next to each other.
perms = torch.stack([perms, comp]).transpose(0, 1).reshape(-1, max_num_chars)
# NOTE:
# The only meaningful way of permuting the EOS position is by moving it one character position at a time.
# However, since the number of permutations = T! and number of EOS positions = T + 1, the number of possible EOS
# positions will always be much less than the number of permutations (unless a low perm_num is set).
# Thus, it would be simpler to just train EOS using the full and null contexts rather than trying to evenly
# distribute it across the chosen number of permutations.
# Add position indices of BOS and EOS
bos_idx = perms.new_zeros((len(perms), 1))
eos_idx = perms.new_full((len(perms), 1), max_num_chars + 1)
perms = torch.cat([bos_idx, perms + 1, eos_idx], dim=1)
# Special handling for the reverse direction. This does two things:
# 1. Reverse context for the characters
# 2. Null context for [EOS] (required for learning to predict [EOS] in NAR mode)
if len(perms) > 1:
perms[1, 1:] = max_num_chars + 1 - torch.arange(max_num_chars + 1, device=self._device)
return perms
def generate_attn_masks(self, perm):
"""Generate attention masks given a sequence permutation (includes pos. for bos and eos tokens)
:param perm: the permutation sequence. i = 0 is always the BOS
:return: lookahead attention masks
"""
sz = perm.shape[0]
mask = torch.zeros((sz, sz), dtype=torch.bool, device=self._device)
for i in range(sz):
query_idx = perm[i]
masked_keys = perm[i + 1 :]
mask[query_idx, masked_keys] = True
content_mask = mask[:-1, :-1].clone()
mask[torch.eye(sz, dtype=torch.bool, device=self._device)] = True # mask "self"
query_mask = mask[1:, :-1]
return content_mask, query_mask
def training_step(self, batch, batch_idx) -> STEP_OUTPUT:
images, labels = batch
tgt = self.tokenizer.encode(labels, self._device)
# Encode the source sequence (i.e. the image codes)
memory = self.model.encode(images)
# Prepare the target sequences (input and output)
tgt_perms = self.gen_tgt_perms(tgt)
tgt_in = tgt[:, :-1]
tgt_out = tgt[:, 1:]
# The [EOS] token is not depended upon by any other token in any permutation ordering
tgt_padding_mask = (tgt_in == self.pad_id) | (tgt_in == self.eos_id)
loss = 0
loss_numel = 0
n = (tgt_out != self.pad_id).sum().item()
for i, perm in enumerate(tgt_perms):
tgt_mask, query_mask = self.generate_attn_masks(perm)
out = self.model.decode(tgt_in, memory, tgt_mask, tgt_padding_mask, tgt_query_mask=query_mask)
logits = self.model.head(out).flatten(end_dim=1)
loss += n * F.cross_entropy(logits, tgt_out.flatten(), ignore_index=self.pad_id)
loss_numel += n
# After the second iteration (i.e. done with canonical and reverse orderings),
# remove the [EOS] tokens for the succeeding perms
if i == 1:
tgt_out = torch.where(tgt_out == self.eos_id, self.pad_id, tgt_out)
n = (tgt_out != self.pad_id).sum().item()
loss /= loss_numel
self.log('loss', loss)
return loss
|