strexp / model_trba.py
markytools's picture
added strexp
d61b9c7
raw
history blame
16.4 kB
"""
Copyright (c) 2019-present NAVER Corp.
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
http://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 torch.nn as nn
from modules_trba.transformation import TPS_SpatialTransformerNetwork
from modules_trba.feature_extraction import VGG_FeatureExtractor, RCNN_FeatureExtractor, ResNet_FeatureExtractor
from modules_trba.sequence_modeling import BidirectionalLSTM
from modules_trba.prediction import Attention
import numpy as np
import torch
import torch.nn.functional as F
import random
import copy
# from torch_edit_distance import levenshtein_distance
class STRScore(nn.Module):
def __init__(self, opt, converter, device, gtStr="", enableSingleCharAttrAve=False):
super(STRScore, self).__init__()
self.opt = opt
self.converter = converter
self.device = device
self.gtStr = gtStr
self.enableSingleCharAttrAve = enableSingleCharAttrAve
self.blank = torch.tensor([-1], dtype=torch.float).to(self.device)
self.separator = torch.tensor([-2], dtype=torch.float).to(self.device)
# singleChar - if >=0, then the output of STRScore will only be a single character
# instead of a whole. The char index will be equal to the parameter "singleChar".
def setSingleCharOutput(self, singleChar):
self.singleChar = singleChar
def forward(self, preds):
bs = preds.shape[0]
# text_for_loss, length_for_loss = self.converter.encode(labels, batch_max_length=self.opt.batch_max_length)
text_for_loss_length = self.opt.batch_max_length + 1
length_for_pred = torch.IntTensor([self.opt.batch_max_length] * bs).to(self.device)
if 'CTC' in self.opt.Prediction:
# Calculate evaluation loss for CTC decoder.
preds_size = torch.FloatTensor([preds.size(1)] * bs)
if self.opt.baiduCTC:
_, preds_index = preds.max(2)
preds_index = preds_index.view(-1)
else:
_, preds_index = preds.max(2)
# print("preds_index shape: ", preds_index.shape)
preds_str = self.converter.decode(preds_index.data, preds_size.data)
# preds_str = self.converter.decode(preds_index, length_for_pred)
preds = preds.log_softmax(2).permute(1, 0, 2)
else:
preds = preds[:, :text_for_loss_length, :]
# select max probabilty (greedy decoding) then decode index to character
_, preds_index = preds.max(2)
# print("preds shape: ", preds.shape)
# print("preds_index: ", preds_index)
preds_str = self.converter.decode(preds_index, length_for_pred)
# print("preds_str: ", preds_str)
# Confidence score
# ARGMAX calculation
sum = torch.FloatTensor([0]*bs).to(self.device)
if self.enableSingleCharAttrAve:
sum = torch.zeros((bs, preds.shape[2])).to(self.device)
if self.opt.confidence_mode == 0:
preds_prob = F.softmax(preds, dim=2)
# print("preds_prob shape: ", preds_prob.shape)
preds_max_prob, _ = preds_prob.max(dim=2)
# print("preds_max_prob shape: ", preds_max_prob.shape)
confidence_score_list = []
count = 0
for one_hot_preds, pred, pred_max_prob in zip(preds_prob, preds_str, preds_max_prob):
if 'Attn' in self.opt.Prediction:
if self.enableSingleCharAttrAve:
one_hot = one_hot_preds[self.singleChar, :]
sum[count] = one_hot
else:
pred_EOS = pred.find('[s]')
pred = pred[:pred_EOS]
pred_max_prob = pred_max_prob[:pred_EOS] ### Use score of all letters
# pred_max_prob = pred_max_prob[0:1] ### Use score of first letter only
if pred_max_prob.shape[0] == 0: continue
if self.opt.scorer == "cumprod":
confidence_score = pred_max_prob.cumprod(dim=0)[-1] ### Maximum is 1
elif self.opt.scorer == "mean":
confidence_score = torch.mean(pred_max_prob) ### Maximum is 1
sum[count] += confidence_score
sum = sum.unsqueeze(1)
elif 'CTC' in self.opt.Prediction:
if self.opt.scorer == "cumprod":
confidence_score = pred_max_prob.cumprod(dim=0)[-1] ### Maximum is 1
elif self.opt.scorer == "mean":
confidence_score = torch.mean(pred_max_prob) ### Maximum is 1
sum[count] += confidence_score
sum = sum.unsqueeze(1)
count += 1
# return sum.detach().cpu().numpy()
# print("sumshape: ", sum.shape)
elif self.opt.confidence_mode == 1:
preds_prob = F.softmax(preds, dim=2)
### Predicted indices
preds_max_prob = torch.argmax(preds_prob, 2)
# print("preds_max_prob shape: ", preds_max_prob.shape)
### Ground truth indices
gtIndices, _ = self.converter.encode([self.gtStr for i in range(0,preds_prob.shape[0])], batch_max_length=self.opt.batch_max_length-1)
# print("gtIndices shape: ", gtIndices.shape)
### Acquire levenstein distance
m = torch.tensor([preds_prob.shape[1] for i in range(0, gtIndices.shape[0])], dtype=torch.float).to(self.device)
n = torch.tensor([preds_prob.shape[1] for i in range(0, gtIndices.shape[0])], dtype=torch.float).to(self.device)
# print("m: ", m)
# print("preds_max_prob dtype: ", preds_max_prob.dtype)
# print("gtIndices dtype: ", gtIndices.dtype)
preds_max_prob = preds_max_prob.type(torch.float)
gtIndices = gtIndices.type(torch.float)
r = levenshtein_distance(preds_max_prob.to(self.device), gtIndices.to(self.device), n, m, torch.cat([self.blank, self.separator]), torch.empty([], dtype=torch.float).to(self.device))
# print("r shape: ", r.shape)
# confidence_score_list = []
# count = 0
# for pred, pred_max_prob in zip(preds_str, preds_max_prob):
# if 'Attn' in self.opt.Prediction:
# pred_EOS = pred.find('[s]')
# pred = pred[:pred_EOS]
# pred_max_prob = pred_max_prob[:pred_EOS] ### Use score of all letters
# # pred_max_prob = pred_max_prob[0:1] ### Use score of first letter only
# if pred_max_prob.shape[0] == 0: continue
# confidence_score = pred_max_prob.cumprod(dim=0)[-1]
# sum[count] += confidence_score
# count += 1
# return sum.detach().cpu().numpy()
# print("sumshape: ", sum.shape)
# sum = sum.unsqueeze(1)
rSoft = F.softmax(r[:,2].type(torch.float))
# rSoft = rSoft.contiguous()
rNorm = rSoft.max()-rSoft
sum = rNorm.unsqueeze(1)
print("sum shape: ", sum.shape)
return sum
class SuperPixler(nn.Module):
def __init__(self, n_super_pixel, imageList, super_pixel_width, super_pixel_height, opt):
super(SuperPixler, self).__init__()
self.opt = opt
self.imageList = imageList
self.n_super_pixel = n_super_pixel
# self.image = image
# self.image = image.transpose(2, 0, 1) # model expects images in BRG, not RGB, so transpose color channels
# self.mean_color = self.image.mean()
# self.image = np.expand_dims(self.image, axis=0)
self.super_pixel_width = super_pixel_width
self.super_pixel_height = super_pixel_height
# def setImage(self, image):
# self.image = image
# self.image_height = image.shape[2]
# self.image_width = image.shape[3]
def sampleImages(self, size):
newImgList = []
for i in range(0, size):
randIdx = random.randint(0, len(self.imageList)-1)
newImgList.append(copy.deepcopy(self.imageList[randIdx]))
return np.array(newImgList)
def forward(self, x):
"""
In the forward step we accept the super pixel masks and transform them to a batch of images
"""
# x = self.sampleMasks(image.shape[0])
image = self.sampleImages(x.shape[0])
self.image = image
self.image_height = image.shape[2]
self.image_width = image.shape[3]
self.mean_color = self.image.mean()
# self.mean_color = self.image.mean(axis=(1,2,3))
# pixeled_image = np.repeat(self.image.copy(), x.shape[0], axis=0)# WARNING:
pixeled_image = self.image.copy()
# print("pixeled_image shape: ", pixeled_image.shape)
# print("x shape: ", x.shape)
for i, super_pixel in enumerate(x.T):
images_to_pixelate = [bool(p) for p in super_pixel]
# print("super_pixel shape: ", super_pixel.shape)
# print("images_to_pixelate len: ", len(images_to_pixelate))
# print("i: {}, superPix: {}, images_to_pixelate: {}".format(i, super_pixel, images_to_pixelate))
x = (i*self.super_pixel_height//self.image_height)*self.super_pixel_width
y = i*self.super_pixel_height%self.image_height
### Reshape image means since it has n-dim size, not a single number. Need to repeat sideways.
# origShapeToApply = pixeled_image[images_to_pixelate,:,y:y+self.super_pixel_height,x:x+self.super_pixel_width].shape
# print("origShapeToApply: ", origShapeToApply)
# mean_color_spec = np.tile(self.mean_color, origShapeToApply[1:]) #
# mean_color_spec = np.reshape(mean_color_spec, origShapeToApply[::-1]).T ### reshape to reversed
### Apply image means
pixeled_image[images_to_pixelate,:,y:y+self.super_pixel_height,x:x+self.super_pixel_width] = self.mean_color
return pixeled_image
class CastNumpy(nn.Module):
def __init__(self, device):
super(CastNumpy, self).__init__()
self.device = device
def forward(self, image):
"""
In the forward function we accept the inputs and cast them to a pytorch tensor
"""
image = np.ascontiguousarray(image)
image = torch.from_numpy(image).to(self.device)
if image.ndimension() == 3:
image = image.unsqueeze(0)
image_half = image.half()
return image_half.float()
class Model(nn.Module):
def __init__(self, opt, device, feature_ext_outputs=None):
super(Model, self).__init__()
self.opt = opt
self.device = device
self.gtText = None
self.stages = {'Trans': opt.Transformation, 'Feat': opt.FeatureExtraction,
'Seq': opt.SequenceModeling, 'Pred': opt.Prediction}
""" Transformation """
if opt.Transformation == 'TPS':
self.Transformation = TPS_SpatialTransformerNetwork(
F=opt.num_fiducial, I_size=(opt.imgH, opt.imgW), I_r_size=(opt.imgH, opt.imgW), I_channel_num=opt.input_channel)
else:
print('No Transformation module specified')
""" FeatureExtraction """
if opt.FeatureExtraction == 'VGG':
self.FeatureExtraction = VGG_FeatureExtractor(opt.input_channel, opt.output_channel)
elif opt.FeatureExtraction == 'RCNN':
self.FeatureExtraction = RCNN_FeatureExtractor(opt.input_channel, opt.output_channel)
elif opt.FeatureExtraction == 'ResNet':
self.FeatureExtraction = ResNet_FeatureExtractor(opt.input_channel, opt.output_channel)
else:
raise Exception('No FeatureExtraction module specified')
self.FeatureExtraction_output = opt.output_channel # int(imgH/16-1) * 512
self.AdaptiveAvgPool = nn.AdaptiveAvgPool2d((None, 1)) # Transform final (imgH/16-1) -> 1
""" Sequence modeling"""
if opt.SequenceModeling == 'BiLSTM':
self.SequenceModeling = nn.Sequential(
BidirectionalLSTM(self.FeatureExtraction_output, opt.hidden_size, opt.hidden_size),
BidirectionalLSTM(opt.hidden_size, opt.hidden_size, opt.hidden_size))
self.SequenceModeling_output = opt.hidden_size
else:
print('No SequenceModeling module specified')
self.SequenceModeling_output = self.FeatureExtraction_output
""" Prediction """
if opt.Prediction == 'CTC':
self.Prediction = nn.Linear(self.SequenceModeling_output, opt.num_class)
elif opt.Prediction == 'Attn':
self.Prediction = Attention(self.SequenceModeling_output, opt.hidden_size, opt.num_class)
else:
raise Exception('Prediction is neither CTC or Attn')
### Set feature map outputter modules
if opt.output_feat_maps:
feature_ext_outputs.set_feature_ext(self.FeatureExtraction)
### Define hooks
feature_ext_outputs = feature_ext_outputs
totalCNNLayers = 0
idxToOutput = []
layersList = []
layerCount = 0
# print("list(self.FeatureExtraction._modules.items()): ", list(self.FeatureExtraction._modules.items()))
# print("list(self.FeatureExtraction.ConvNet_modules.items())[0][1]: ", list(self.FeatureExtraction.ConvNet._modules.items())[0][1])
first_layer = list(self.FeatureExtraction.ConvNet._modules.items())[0][1]
first_layer.register_backward_hook(feature_ext_outputs.append_first_grads)
for layer in self.FeatureExtraction.modules():
if isinstance(layer, nn.Conv2d):
layerCount += 1
if layerCount >= opt.min_layer_out and layerCount <= opt.max_layer_out:
layer.register_forward_hook(feature_ext_outputs.append_layer_out)
layer.register_backward_hook(feature_ext_outputs.append_grad_out)
# def get_feature_ext(self):
# return self.FeatureExtraction
def setGTText(self, text):
self.gtText = text
def forward(self, input, text="", is_train=True):
if self.opt.is_shap:
text = torch.LongTensor(input.shape[0], self.opt.batch_max_length + 1).fill_(0).to(self.device)
elif self.gtText is not None:
text = self.gtText
else:
text = torch.LongTensor(input.shape[0], self.opt.batch_max_length + 1).fill_(0).to(self.device)
# print("text shape: ", text.shape) (1,26)
tpsOut = input.contiguous()
""" Transformation stage """
if not self.stages['Trans'] == "None":
tpsOut = self.Transformation(tpsOut)
# print("Transformation feature shape: ", input.shape)
""" Feature extraction stage """
visual_feature = self.FeatureExtraction(tpsOut)
visual_feature = self.AdaptiveAvgPool(visual_feature.permute(0, 3, 1, 2)) # [b, c, h, w] -> [b, w, c, h]
visual_feature = visual_feature.squeeze(3)
# print("visual feature shape: ", visual_feature.shape)
""" Sequence modeling stage """
if self.stages['Seq'] == 'BiLSTM':
contextual_feature = self.SequenceModeling(visual_feature)
else:
contextual_feature = visual_feature # for convenience. this is NOT contextually modeled by BiLSTM
# print("Sequence feature shape: ", contextual_feature.shape)
""" Prediction stage """
if self.stages['Pred'] == 'CTC':
prediction = self.Prediction(contextual_feature.contiguous())
else:
prediction = self.Prediction(contextual_feature.contiguous(), text, is_train, batch_max_length=self.opt.batch_max_length)
# print("prediction feature shape: ", prediction.shape)
# return prediction, tpsOut
return prediction