diff --git a/.gitignore b/.gitignore
index e58673f6545c14befec20c231c84f9f8daada26c..61969afaf401bed5fa1a554a03ae522df6da4eb5 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,8 +1,6 @@
data/words*
data/words.txt
src/__pycache__/
-model/checkpoint
-model/snapshot-*
notes/
*.so
*.pyc
diff --git a/README.md b/README.md
index a22c6ccf143e26033c13eadd3aedeff761b1fa2f..c919f348aa685b2fd1a2f8c847802b0b89c84238 100644
--- a/README.md
+++ b/README.md
@@ -21,7 +21,6 @@ The input image and the expected output is shown below.
```
> python main.py
-Validation character error rate of saved model: 11.118344571029994%
Init with stored values from ../model/snapshot-76
Recognized: "Hello"
Probability: 0.8462573289871216
diff --git a/data/pixelRelevance.npy b/data/pixelRelevance.npy
deleted file mode 100644
index 5d1a525240ba5ef3690f42e0d45ee7c032654662..0000000000000000000000000000000000000000
Binary files a/data/pixelRelevance.npy and /dev/null differ
diff --git a/data/translationInvariance.npy b/data/translationInvariance.npy
deleted file mode 100644
index 72c21ed61205d0d076c863f03026f6933e5376ad..0000000000000000000000000000000000000000
Binary files a/data/translationInvariance.npy and /dev/null differ
diff --git a/data/translationInvarianceTexts.pickle b/data/translationInvarianceTexts.pickle
deleted file mode 100644
index 05bf9cf0a5989e7b38f4e34bb761e27019d3766c..0000000000000000000000000000000000000000
Binary files a/data/translationInvarianceTexts.pickle and /dev/null differ
diff --git a/model/.gitignore b/model/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..86d0cb2726c6c7c179b99520c452dd1b68e7a813
--- /dev/null
+++ b/model/.gitignore
@@ -0,0 +1,4 @@
+# Ignore everything in this directory
+*
+# Except this file
+!.gitignore
\ No newline at end of file
diff --git a/model/accuracy.txt b/model/accuracy.txt
deleted file mode 100644
index 24c9310a87b90eb930fae9bc0c4e5ec5d1899210..0000000000000000000000000000000000000000
--- a/model/accuracy.txt
+++ /dev/null
@@ -1 +0,0 @@
-Validation character error rate of saved model: 10.89210233592881%
\ No newline at end of file
diff --git a/src/SamplePreprocessor.py b/src/SamplePreprocessor.py
index c82ff0a4c929f85ea2a9bedbcf611209f2072306..b58a02c2a7aa17fa0dc749c63d7afec02dc3a10c 100644
--- a/src/SamplePreprocessor.py
+++ b/src/SamplePreprocessor.py
@@ -23,9 +23,9 @@ def preprocess(img, imgSize, dataAugmentation=False):
if random.random() < 0.25:
img = cv2.erode(img, np.ones((3, 3)))
if random.random() < 0.5:
- img = img * (0.5 + random.random() * 0.5)
+ img = img * (0.1 + random.random() * 0.9)
if random.random() < 0.25:
- img = np.clip(img + (np.random.random(img.shape) - 0.5) * random.randint(1, 50), 0, 255)
+ img = np.clip(img + (np.random.random(img.shape) - 0.5) * random.randint(1, 25), 0, 255)
if random.random() < 0.1:
img = 255 - img
@@ -33,20 +33,20 @@ def preprocess(img, imgSize, dataAugmentation=False):
wt, ht = imgSize
h, w = img.shape
f = min(wt / w, ht / h)
- fx = f * np.random.uniform(0.75, 1.25)
- fy = f * np.random.uniform(0.75, 1.25)
+ fx = f * np.random.uniform(0.75, 1.5)
+ fy = f * np.random.uniform(0.75, 1.5)
# random position around center
txc = (wt - w * fx) / 2
tyc = (ht - h * fy) / 2
- freedom_x = wt // 10
- freedom_y = ht // 10
- tx = txc + np.random.randint(-freedom_x, freedom_x)
- ty = tyc + np.random.randint(-freedom_y, freedom_y)
+ freedom_x = wt / 5
+ freedom_y = ht / 5
+ tx = txc + np.random.uniform(-freedom_x, freedom_x)
+ ty = tyc + np.random.uniform(-freedom_y, freedom_y)
# map image into target image
M = np.float32([[fx, 0, tx], [0, fy, ty]])
- target = np.ones(imgSize[::-1]) * 255 / 2
+ target = np.ones(imgSize[::-1]) * np.random.uniform(0, 255)
img = cv2.warpAffine(img, M, dsize=imgSize, dst=target, borderMode=cv2.BORDER_TRANSPARENT)
# no data augmentation
diff --git a/src/main.py b/src/main.py
index 72a6be074ef714573f4e18420c552441dfae30de..d28dacca3971688089c02472b8139bce51baac2f 100644
--- a/src/main.py
+++ b/src/main.py
@@ -1,4 +1,5 @@
import argparse
+import json
import cv2
import editdistance
@@ -12,14 +13,20 @@ from SamplePreprocessor import preprocess
class FilePaths:
"filenames and paths to data"
fnCharList = '../model/charList.txt'
- fnAccuracy = '../model/accuracy.txt'
+ fnSummary = '../model/summary.json'
fnInfer = '../data/test.png'
fnCorpus = '../data/corpus.txt'
+def write_summary(charErrorRates):
+ with open(FilePaths.fnSummary, 'w') as f:
+ json.dump(charErrorRates, f)
+
+
def train(model, loader):
"train NN"
epoch = 0 # number of training epochs since start
+ summaryCharErrorRates = []
bestCharErrorRate = float('inf') # best valdiation character error rate
noImprovementSince = 0 # number of epochs no improvement of character error rate occured
earlyStopping = 25 # stop training after this number of epochs without improvement
@@ -39,14 +46,16 @@ def train(model, loader):
# validate
charErrorRate = validate(model, loader)
+ # write summary
+ summaryCharErrorRates.append(charErrorRate)
+ write_summary(summaryCharErrorRates)
+
# if best validation accuracy so far, save model parameters
if charErrorRate < bestCharErrorRate:
print('Character error rate improved, save model')
bestCharErrorRate = charErrorRate
noImprovementSince = 0
model.save()
- open(FilePaths.fnAccuracy, 'w').write(
- f'Validation character error rate of saved model: {charErrorRate * 100.0}%')
else:
print(f'Character error rate not improved, best so far: {charErrorRate * 100.0}%')
noImprovementSince += 1
@@ -140,7 +149,6 @@ def main():
# infer text on test image
else:
- print(open(FilePaths.fnAccuracy).read())
model = Model(open(FilePaths.fnCharList).read(), decoderType, mustRestore=True, dump=args.dump)
infer(model, FilePaths.fnInfer)