update README for CNN model

README.md

@@ -10,6 +10,9 @@ Follow the steps,
 run the code,
 and it works!
 
+the easiest model requires only 2 convolutional layers.
+accuracy is 90%+
+
 There are several more steps to put this prototype on production.
 
 **Ping me for paid technical supports**.
@@ -102,36 +105,117 @@ Navigate your web browser to `http://127.0.0.1:6006/` to view the TensorBoard.
 
 ## Solve Captcha Using CNN Model
 
-#### 1-char captcha -- also covered in MNIST tutorial.
+#### 1-char Captcha
+
+the same 2-layer CNN model as talked in MNIST tutorial
 
-1. 2 Convolutional layer.
-2. Accuracy: 99%
+1. 2 Convolutional layers.
+2. Accuracy: 97%
 3. Training data: 20000 images.
-4. Time cost: <5min. using GTX-1080.
+4. Time cost: <4min. using GTX-1080.
+
+```bash
+# generating test data
+$ python datasets/gen_captcha.py -d --npi 1 -n 2000
+# run the model
+$ python cnn_1_char.py --data_dir images/char-1-epoch-2000/
+```
 
+Output While Training:
 
 ```bash
-$ time python conv_captcha.py
+$ $ time python cnn_1_char.py --data_dir images/char-1-epoch-2000/
 data loaded
-train images: 10000. test images: 2000
+train images: 20000. test images: 4000
 label_size: 10, image_size: 6000
 ...
-step 9100, training accuracy = 100.00%, testing accuracy = 98.90%
-step 9200, training accuracy = 100.00%, testing accuracy = 98.80%
-step 9300, training accuracy = 100.00%, testing accuracy = 98.90%
-step 9400, training accuracy = 100.00%, testing accuracy = 98.80%
-step 9500, training accuracy = 100.00%, testing accuracy = 98.90%
-step 9600, training accuracy = 100.00%, testing accuracy = 98.60%
-step 9700, training accuracy = 100.00%, testing accuracy = 98.65%
-step 9800, training accuracy = 100.00%, testing accuracy = 98.95%
-step 9900, training accuracy = 100.00%, testing accuracy = 98.95%
-testing accuracy = 99.15%
+step 9100, training accuracy = 100.00%, testing accuracy = 96.60%
+step 9200, training accuracy = 100.00%, testing accuracy = 97.80%
+step 9300, training accuracy = 100.00%, testing accuracy = 97.65%
+step 9400, training accuracy = 100.00%, testing accuracy = 97.95%
+step 9500, training accuracy = 100.00%, testing accuracy = 97.25%
+step 9600, training accuracy = 98.00%, testing accuracy = 96.80%
+step 9700, training accuracy = 100.00%, testing accuracy = 97.70%
+step 9800, training accuracy = 100.00%, testing accuracy = 97.25%
+step 9900, training accuracy = 100.00%, testing accuracy = 95.40%
+testing accuracy = 97.85%
+
+real	3m35.217s
+user	3m4.168s
+sys	0m31.036s
+```
+
+#### 2-char Captcha
+
+Add 1 more layer, reshape the output before calculating Loss.
+
+1. Accuracy: 97%
+3. Training data: 45000 images.
+4. Time cost: <5min. using GTX-1080.
+
+```bash
+# generating test data
+$ python datasets/gen_captcha.py -d --npi=2 -n 500
+# run the model
+$ python cnn_n_char.py --data_dir=images/char-2-epoch-500/
+```
+
+Output While Training:
+
+```bash
+$ time python cnn_n_char.py --data_dir=images/char-2-epoch-500/
+data loaded
+train images: 45000. test images: 9000
+label_size: 10, image_size: 8000
+...
+step 9100, training accuracy = 99.00%, testing accuracy = 96.45%
+step 9200, training accuracy = 100.00%, testing accuracy = 97.28%
+step 9300, training accuracy = 100.00%, testing accuracy = 96.75%
+step 9400, training accuracy = 100.00%, testing accuracy = 96.68%
+step 9500, training accuracy = 100.00%, testing accuracy = 96.55%
+step 9600, training accuracy = 100.00%, testing accuracy = 96.40%
+step 9700, training accuracy = 100.00%, testing accuracy = 95.58%
+step 9800, training accuracy = 99.00%, testing accuracy = 96.70%
+step 9900, training accuracy = 99.00%, testing accuracy = 96.68%
+testing accuracy = 97.33%
+
+real	4m49.509s
+user	4m10.980s
+sys	0m34.164s
+```
+
+#### 4-char Captcha
+
+the same model as 2-char captcha.
+
+only change the training dataset to 4-char ones.
 
-real	4m44.143s
-user	3m40.896s
-sys	0m33.492s
+```bash
+# generating test data
+$ python datasets/gen_captcha.py -d --npi=4 -n 6
+# run the model
+$ python cnn_n_char.py --data_dir images/char-4-epoch-6/
 ```
 
+Output While Training:
+
+```bash
+$ python cnn_n_char.py --data_dir images/char-4-epoch-6/
+data loaded
+train images: 30240. test images: 5040
+label_size: 10, image_size: 12000
+...
+step 9100, training accuracy = 100.00%, testing accuracy = 92.73%
+step 9200, training accuracy = 98.00%, testing accuracy = 92.86%
+step 9300, training accuracy = 97.50%, testing accuracy = 91.74%
+step 9400, training accuracy = 99.00%, testing accuracy = 92.84%
+step 9500, training accuracy = 99.00%, testing accuracy = 92.99%
+step 9600, training accuracy = 99.00%, testing accuracy = 93.78%
+step 9700, training accuracy = 100.00%, testing accuracy = 92.56%
+step 9800, training accuracy = 99.50%, testing accuracy = 93.01%
+step 9900, training accuracy = 99.00%, testing accuracy = 92.84%
+testing accuracy = 93.24%
+```
 
 ## Generate DataSet for Training
 
@@ -222,3 +306,27 @@ test images: (7564, 100, 80), labels: (7564, 124)
 1 epoch has `10*9*8*7=5040` images, generate 10 epoches for training.
 
 generating the dataset:
+
+```bash
+$ python datasets/gen_captcha.py -d --npi=4 -n 6
+10 choices: 0123456789
+generating 6 epoches of captchas in ./images/char-4-epoch-6/train
+generating 1 epoches of captchas in ./images/char-4-epoch-6/test
+write meta info in ./images/char-4-epoch-6/meta.json
+```
+
+preview the dataset:
+
+```bash
+$ python datasets/base.py images/char-4-epoch-6/
+========== Meta Info ==========
+num_per_image: 4
+label_choices: 0123456789
+height: 100
+width: 120
+n_epoch: 6
+label_size: 10
+==============================
+train images: (30240, 100, 120), labels: (30240, 40)
+test images: (5040, 100, 120), labels: (5040, 40)
+```