diff --git a/examples/vision/captcha_ocr.py b/examples/vision/captcha_ocr.py
index fc4430b1f3..740fdebc6b 100644
--- a/examples/vision/captcha_ocr.py
+++ b/examples/vision/captcha_ocr.py
@@ -22,6 +22,10 @@
## Setup
"""
+import os
+
+os.environ["KERAS_BACKEND"] = "tensorflow"
+
import os
import numpy as np
import matplotlib.pyplot as plt
@@ -30,9 +34,8 @@
from collections import Counter
import tensorflow as tf
-from tensorflow import keras
-from tensorflow.keras import layers
-
+import keras
+from keras import layers
"""
## Load the data: [Captcha Images](https://www.kaggle.com/fournierp/captcha-version-2-images)
@@ -180,10 +183,64 @@ def encode_single_sample(img_path, label):
"""
+def ctc_batch_cost(y_true, y_pred, input_length, label_length):
+ label_length = tf.cast(tf.squeeze(label_length, axis=-1), tf.int32)
+ input_length = tf.cast(tf.squeeze(input_length, axis=-1), tf.int32)
+ sparse_labels = tf.cast(ctc_label_dense_to_sparse(y_true, label_length), tf.int32)
+
+ y_pred = tf.math.log(tf.transpose(y_pred, perm=[1, 0, 2]) + keras.backend.epsilon())
+
+ return tf.expand_dims(
+ tf.compat.v1.nn.ctc_loss(
+ inputs=y_pred, labels=sparse_labels, sequence_length=input_length
+ ),
+ 1,
+ )
+
+
+def ctc_label_dense_to_sparse(labels, label_lengths):
+ label_shape = tf.shape(labels)
+ num_batches_tns = tf.stack([label_shape[0]])
+ max_num_labels_tns = tf.stack([label_shape[1]])
+
+ def range_less_than(old_input, current_input):
+ return tf.expand_dims(tf.range(tf.shape(old_input)[1]), 0) < tf.fill(
+ max_num_labels_tns, current_input
+ )
+
+ init = tf.cast(tf.fill([1, label_shape[1]], 0), tf.bool)
+ dense_mask = tf.compat.v1.scan(
+ range_less_than, label_lengths, initializer=init, parallel_iterations=1
+ )
+ dense_mask = dense_mask[:, 0, :]
+
+ label_array = tf.reshape(
+ tf.tile(tf.range(0, label_shape[1]), num_batches_tns), label_shape
+ )
+ label_ind = tf.compat.v1.boolean_mask(label_array, dense_mask)
+
+ batch_array = tf.transpose(
+ tf.reshape(
+ tf.tile(tf.range(0, label_shape[0]), max_num_labels_tns),
+ tf.reverse(label_shape, [0]),
+ )
+ )
+ batch_ind = tf.compat.v1.boolean_mask(batch_array, dense_mask)
+ indices = tf.transpose(
+ tf.reshape(tf.concat([batch_ind, label_ind], axis=0), [2, -1])
+ )
+
+ vals_sparse = tf.compat.v1.gather_nd(labels, indices)
+
+ return tf.SparseTensor(
+ tf.cast(indices, tf.int64), vals_sparse, tf.cast(label_shape, tf.int64)
+ )
+
+
class CTCLayer(layers.Layer):
def __init__(self, name=None):
super().__init__(name=name)
- self.loss_fn = keras.backend.ctc_batch_cost
+ self.loss_fn = ctc_batch_cost
def call(self, y_true, y_pred):
# Compute the training-time loss value and add it
@@ -272,7 +329,8 @@ def build_model():
"""
-epochs = 100
+# TODO restore epoch count.
+epochs = 2
early_stopping_patience = 10
# Add early stopping
early_stopping = keras.callbacks.EarlyStopping(
@@ -296,9 +354,33 @@ def build_model():
"""
+def ctc_decode(y_pred, input_length, greedy=True, beam_width=100, top_paths=1):
+ input_shape = tf.shape(y_pred)
+ num_samples, num_steps = input_shape[0], input_shape[1]
+ y_pred = tf.math.log(tf.transpose(y_pred, perm=[1, 0, 2]) + keras.backend.epsilon())
+ input_length = tf.cast(input_length, tf.int32)
+
+ if greedy:
+ (decoded, log_prob) = tf.nn.ctc_greedy_decoder(
+ inputs=y_pred, sequence_length=input_length
+ )
+ else:
+ (decoded, log_prob) = tf.compat.v1.nn.ctc_beam_search_decoder(
+ inputs=y_pred,
+ sequence_length=input_length,
+ beam_width=beam_width,
+ top_paths=top_paths,
+ )
+ decoded_dense = []
+ for st in decoded:
+ st = tf.SparseTensor(st.indices, st.values, (num_samples, num_steps))
+ decoded_dense.append(tf.sparse.to_dense(sp_input=st, default_value=-1))
+ return (decoded_dense, log_prob)
+
+
# Get the prediction model by extracting layers till the output layer
prediction_model = keras.models.Model(
- model.get_layer(name="image").input, model.get_layer(name="dense2").output
+ model.input[0], model.get_layer(name="dense2").output
)
prediction_model.summary()
@@ -307,7 +389,7 @@ def build_model():
def decode_batch_predictions(pred):
input_len = np.ones(pred.shape[0]) * pred.shape[1]
# Use greedy search. For complex tasks, you can use beam search
- results = keras.backend.ctc_decode(pred, input_length=input_len, greedy=True)[0][0][
+ results = ctc_decode(pred, input_length=input_len, greedy=True)[0][0][
:, :max_length
]
# Iterate over the results and get back the text
diff --git a/examples/vision/img/captcha_ocr/captcha_ocr_13_0.png b/examples/vision/img/captcha_ocr/captcha_ocr_13_0.png
index d8f4b5e135..946d8b72bd 100644
Binary files a/examples/vision/img/captcha_ocr/captcha_ocr_13_0.png and b/examples/vision/img/captcha_ocr/captcha_ocr_13_0.png differ
diff --git a/examples/vision/img/captcha_ocr/captcha_ocr_19_6.png b/examples/vision/img/captcha_ocr/captcha_ocr_19_6.png
new file mode 100644
index 0000000000..ea2588e453
Binary files /dev/null and b/examples/vision/img/captcha_ocr/captcha_ocr_19_6.png differ
diff --git a/examples/vision/ipynb/captcha_ocr.ipynb b/examples/vision/ipynb/captcha_ocr.ipynb
index b34e3d9f37..5cbb2c5a92 100644
--- a/examples/vision/ipynb/captcha_ocr.ipynb
+++ b/examples/vision/ipynb/captcha_ocr.ipynb
@@ -41,12 +41,16 @@
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
+ "import os\n",
+ "\n",
+ "os.environ[\"KERAS_BACKEND\"] = \"tensorflow\"\n",
+ "\n",
"import os\n",
"import numpy as np\n",
"import matplotlib.pyplot as plt\n",
@@ -55,8 +59,8 @@
"from collections import Counter\n",
"\n",
"import tensorflow as tf\n",
- "from tensorflow import keras\n",
- "from tensorflow.keras import layers\n"
+ "import keras\n",
+ "from keras import layers"
]
},
{
@@ -71,7 +75,7 @@
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 0,
"metadata": {
"colab_type": "code"
},
@@ -97,7 +101,7 @@
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 0,
"metadata": {
"colab_type": "code"
},
@@ -133,7 +137,8 @@
"downsample_factor = 4\n",
"\n",
"# Maximum length of any captcha in the dataset\n",
- "max_length = max([len(label) for label in labels])\n"
+ "max_length = max([len(label) for label in labels])\n",
+ ""
]
},
{
@@ -147,7 +152,7 @@
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 0,
"metadata": {
"colab_type": "code"
},
@@ -155,9 +160,7 @@
"source": [
"\n",
"# Mapping characters to integers\n",
- "char_to_num = layers.StringLookup(\n",
- " vocabulary=list(characters), mask_token=None\n",
- ")\n",
+ "char_to_num = layers.StringLookup(vocabulary=list(characters), mask_token=None)\n",
"\n",
"# Mapping integers back to original characters\n",
"num_to_char = layers.StringLookup(\n",
@@ -199,7 +202,8 @@
" # 6. Map the characters in label to numbers\n",
" label = char_to_num(tf.strings.unicode_split(label, input_encoding=\"UTF-8\"))\n",
" # 7. Return a dict as our model is expecting two inputs\n",
- " return {\"image\": img, \"label\": label}\n"
+ " return {\"image\": img, \"label\": label}\n",
+ ""
]
},
{
@@ -213,7 +217,7 @@
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 0,
"metadata": {
"colab_type": "code"
},
@@ -222,18 +226,14 @@
"\n",
"train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))\n",
"train_dataset = (\n",
- " train_dataset.map(\n",
- " encode_single_sample, num_parallel_calls=tf.data.AUTOTUNE\n",
- " )\n",
+ " train_dataset.map(encode_single_sample, num_parallel_calls=tf.data.AUTOTUNE)\n",
" .batch(batch_size)\n",
" .prefetch(buffer_size=tf.data.AUTOTUNE)\n",
")\n",
"\n",
"validation_dataset = tf.data.Dataset.from_tensor_slices((x_valid, y_valid))\n",
"validation_dataset = (\n",
- " validation_dataset.map(\n",
- " encode_single_sample, num_parallel_calls=tf.data.AUTOTUNE\n",
- " )\n",
+ " validation_dataset.map(encode_single_sample, num_parallel_calls=tf.data.AUTOTUNE)\n",
" .batch(batch_size)\n",
" .prefetch(buffer_size=tf.data.AUTOTUNE)\n",
")"
@@ -250,7 +250,7 @@
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 0,
"metadata": {
"colab_type": "code"
},
@@ -281,17 +281,71 @@
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
+ "\n",
+ "def ctc_batch_cost(y_true, y_pred, input_length, label_length):\n",
+ " label_length = tf.cast(tf.squeeze(label_length, axis=-1), tf.int32)\n",
+ " input_length = tf.cast(tf.squeeze(input_length, axis=-1), tf.int32)\n",
+ " sparse_labels = tf.cast(ctc_label_dense_to_sparse(y_true, label_length), tf.int32)\n",
+ "\n",
+ " y_pred = tf.math.log(tf.transpose(y_pred, perm=[1, 0, 2]) + keras.backend.epsilon())\n",
+ "\n",
+ " return tf.expand_dims(\n",
+ " tf.compat.v1.nn.ctc_loss(\n",
+ " inputs=y_pred, labels=sparse_labels, sequence_length=input_length\n",
+ " ),\n",
+ " 1,\n",
+ " )\n",
+ "\n",
+ "\n",
+ "def ctc_label_dense_to_sparse(labels, label_lengths):\n",
+ " label_shape = tf.shape(labels)\n",
+ " num_batches_tns = tf.stack([label_shape[0]])\n",
+ " max_num_labels_tns = tf.stack([label_shape[1]])\n",
+ "\n",
+ " def range_less_than(old_input, current_input):\n",
+ " return tf.expand_dims(tf.range(tf.shape(old_input)[1]), 0) < tf.fill(\n",
+ " max_num_labels_tns, current_input\n",
+ " )\n",
+ "\n",
+ " init = tf.cast(tf.fill([1, label_shape[1]], 0), tf.bool)\n",
+ " dense_mask = tf.compat.v1.scan(\n",
+ " range_less_than, label_lengths, initializer=init, parallel_iterations=1\n",
+ " )\n",
+ " dense_mask = dense_mask[:, 0, :]\n",
+ "\n",
+ " label_array = tf.reshape(\n",
+ " tf.tile(tf.range(0, label_shape[1]), num_batches_tns), label_shape\n",
+ " )\n",
+ " label_ind = tf.compat.v1.boolean_mask(label_array, dense_mask)\n",
+ "\n",
+ " batch_array = tf.transpose(\n",
+ " tf.reshape(\n",
+ " tf.tile(tf.range(0, label_shape[0]), max_num_labels_tns),\n",
+ " tf.reverse(label_shape, [0]),\n",
+ " )\n",
+ " )\n",
+ " batch_ind = tf.compat.v1.boolean_mask(batch_array, dense_mask)\n",
+ " indices = tf.transpose(\n",
+ " tf.reshape(tf.concat([batch_ind, label_ind], axis=0), [2, -1])\n",
+ " )\n",
+ "\n",
+ " vals_sparse = tf.compat.v1.gather_nd(labels, indices)\n",
+ "\n",
+ " return tf.SparseTensor(\n",
+ " tf.cast(indices, tf.int64), vals_sparse, tf.cast(label_shape, tf.int64)\n",
+ " )\n",
+ "\n",
"\n",
"class CTCLayer(layers.Layer):\n",
" def __init__(self, name=None):\n",
" super().__init__(name=name)\n",
- " self.loss_fn = keras.backend.ctc_batch_cost\n",
+ " self.loss_fn = ctc_batch_cost\n",
"\n",
" def call(self, y_true, y_pred):\n",
" # Compute the training-time loss value and add it\n",
@@ -387,14 +441,15 @@
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"\n",
- "epochs = 100\n",
+ "# TODO restore epoch count.\n",
+ "epochs = 2\n",
"early_stopping_patience = 10\n",
"# Add early stopping\n",
"early_stopping = keras.callbacks.EarlyStopping(\n",
@@ -407,7 +462,8 @@
" validation_data=validation_dataset,\n",
" epochs=epochs,\n",
" callbacks=[early_stopping],\n",
- ")\n"
+ ")\n",
+ ""
]
},
{
@@ -418,30 +474,55 @@
"source": [
"## Inference\n",
"\n",
- "You can use the trained model hosted on [Hugging Face Hub](https://huggingface.co/keras-io/ocr-for-captcha) \n",
+ "You can use the trained model hosted on [Hugging Face Hub](https://huggingface.co/keras-io/ocr-for-captcha)\n",
"and try the demo on [Hugging Face Spaces](https://huggingface.co/spaces/keras-io/ocr-for-captcha)."
]
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
+ "\n",
+ "def ctc_decode(y_pred, input_length, greedy=True, beam_width=100, top_paths=1):\n",
+ " input_shape = tf.shape(y_pred)\n",
+ " num_samples, num_steps = input_shape[0], input_shape[1]\n",
+ " y_pred = tf.math.log(tf.transpose(y_pred, perm=[1, 0, 2]) + keras.backend.epsilon())\n",
+ " input_length = tf.cast(input_length, tf.int32)\n",
+ "\n",
+ " if greedy:\n",
+ " (decoded, log_prob) = tf.nn.ctc_greedy_decoder(\n",
+ " inputs=y_pred, sequence_length=input_length\n",
+ " )\n",
+ " else:\n",
+ " (decoded, log_prob) = tf.compat.v1.nn.ctc_beam_search_decoder(\n",
+ " inputs=y_pred,\n",
+ " sequence_length=input_length,\n",
+ " beam_width=beam_width,\n",
+ " top_paths=top_paths,\n",
+ " )\n",
+ " decoded_dense = []\n",
+ " for st in decoded:\n",
+ " st = tf.SparseTensor(st.indices, st.values, (num_samples, num_steps))\n",
+ " decoded_dense.append(tf.sparse.to_dense(sp_input=st, default_value=-1))\n",
+ " return (decoded_dense, log_prob)\n",
+ "\n",
"\n",
"# Get the prediction model by extracting layers till the output layer\n",
"prediction_model = keras.models.Model(\n",
- " model.get_layer(name=\"image\").input, model.get_layer(name=\"dense2\").output\n",
+ " model.input[0], model.get_layer(name=\"dense2\").output\n",
")\n",
"prediction_model.summary()\n",
"\n",
+ "\n",
"# A utility function to decode the output of the network\n",
"def decode_batch_predictions(pred):\n",
" input_len = np.ones(pred.shape[0]) * pred.shape[1]\n",
" # Use greedy search. For complex tasks, you can use beam search\n",
- " results = keras.backend.ctc_decode(pred, input_length=input_len, greedy=True)[0][0][\n",
+ " results = ctc_decode(pred, input_length=input_len, greedy=True)[0][0][\n",
" :, :max_length\n",
" ]\n",
" # Iterate over the results and get back the text\n",
@@ -506,4 +587,4 @@
},
"nbformat": 4,
"nbformat_minor": 0
-}
+}
\ No newline at end of file
diff --git a/examples/vision/md/captcha_ocr.md b/examples/vision/md/captcha_ocr.md
index 75329fe5f2..b9e41ecf5b 100644
--- a/examples/vision/md/captcha_ocr.md
+++ b/examples/vision/md/captcha_ocr.md
@@ -25,6 +25,10 @@ in the developer guides.
```python
+import os
+
+os.environ["KERAS_BACKEND"] = "tensorflow"
+
import os
import numpy as np
import matplotlib.pyplot as plt
@@ -33,9 +37,8 @@ from pathlib import Path
from collections import Counter
import tensorflow as tf
-from tensorflow import keras
-from tensorflow.keras import layers
-
+import keras
+from keras import layers
```
---
@@ -52,8 +55,8 @@ Let's download the data.
```
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
-100 159 100 159 0 0 164 0 --:--:-- --:--:-- --:--:-- 164
-100 8863k 100 8863k 0 0 4882k 0 0:00:01 0:00:01 --:--:-- 33.0M
+ 0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0
+100 8863k 100 8863k 0 0 19.6M 0 --:--:-- --:--:-- --:--:-- 19.6M
```
@@ -105,7 +108,7 @@ max_length = max([len(label) for label in labels])
Number of images found: 1040
Number of labels found: 1040
Number of unique characters: 19
-Characters present: {'d', 'w', 'y', '4', 'f', '6', 'g', 'e', '3', '5', 'p', 'x', '2', 'c', '7', 'n', 'b', '8', 'm'}
+Characters present: ['2', '3', '4', '5', '6', '7', '8', 'b', 'c', 'd', 'e', 'f', 'g', 'm', 'n', 'p', 'w', 'x', 'y']
```
@@ -116,9 +119,7 @@ Characters present: {'d', 'w', 'y', '4', 'f', '6', 'g', 'e', '3', '5', 'p', 'x'
```python
# Mapping characters to integers
-char_to_num = layers.StringLookup(
- vocabulary=list(characters), mask_token=None
-)
+char_to_num = layers.StringLookup(vocabulary=list(characters), mask_token=None)
# Mapping integers back to original characters
num_to_char = layers.StringLookup(
@@ -172,18 +173,14 @@ def encode_single_sample(img_path, label):
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = (
- train_dataset.map(
- encode_single_sample, num_parallel_calls=tf.data.AUTOTUNE
- )
+ train_dataset.map(encode_single_sample, num_parallel_calls=tf.data.AUTOTUNE)
.batch(batch_size)
.prefetch(buffer_size=tf.data.AUTOTUNE)
)
validation_dataset = tf.data.Dataset.from_tensor_slices((x_valid, y_valid))
validation_dataset = (
- validation_dataset.map(
- encode_single_sample, num_parallel_calls=tf.data.AUTOTUNE
- )
+ validation_dataset.map(encode_single_sample, num_parallel_calls=tf.data.AUTOTUNE)
.batch(batch_size)
.prefetch(buffer_size=tf.data.AUTOTUNE)
)
@@ -209,7 +206,9 @@ plt.show()
```
+
+
---
@@ -218,10 +217,64 @@ plt.show()
```python
+def ctc_batch_cost(y_true, y_pred, input_length, label_length):
+ label_length = tf.cast(tf.squeeze(label_length, axis=-1), tf.int32)
+ input_length = tf.cast(tf.squeeze(input_length, axis=-1), tf.int32)
+ sparse_labels = tf.cast(ctc_label_dense_to_sparse(y_true, label_length), tf.int32)
+
+ y_pred = tf.math.log(tf.transpose(y_pred, perm=[1, 0, 2]) + keras.backend.epsilon())
+
+ return tf.expand_dims(
+ tf.compat.v1.nn.ctc_loss(
+ inputs=y_pred, labels=sparse_labels, sequence_length=input_length
+ ),
+ 1,
+ )
+
+
+def ctc_label_dense_to_sparse(labels, label_lengths):
+ label_shape = tf.shape(labels)
+ num_batches_tns = tf.stack([label_shape[0]])
+ max_num_labels_tns = tf.stack([label_shape[1]])
+
+ def range_less_than(old_input, current_input):
+ return tf.expand_dims(tf.range(tf.shape(old_input)[1]), 0) < tf.fill(
+ max_num_labels_tns, current_input
+ )
+
+ init = tf.cast(tf.fill([1, label_shape[1]], 0), tf.bool)
+ dense_mask = tf.compat.v1.scan(
+ range_less_than, label_lengths, initializer=init, parallel_iterations=1
+ )
+ dense_mask = dense_mask[:, 0, :]
+
+ label_array = tf.reshape(
+ tf.tile(tf.range(0, label_shape[1]), num_batches_tns), label_shape
+ )
+ label_ind = tf.compat.v1.boolean_mask(label_array, dense_mask)
+
+ batch_array = tf.transpose(
+ tf.reshape(
+ tf.tile(tf.range(0, label_shape[0]), max_num_labels_tns),
+ tf.reverse(label_shape, [0]),
+ )
+ )
+ batch_ind = tf.compat.v1.boolean_mask(batch_array, dense_mask)
+ indices = tf.transpose(
+ tf.reshape(tf.concat([batch_ind, label_ind], axis=0), [2, -1])
+ )
+
+ vals_sparse = tf.compat.v1.gather_nd(labels, indices)
+
+ return tf.SparseTensor(
+ tf.cast(indices, tf.int64), vals_sparse, tf.cast(label_shape, tf.int64)
+ )
+
+
class CTCLayer(layers.Layer):
def __init__(self, name=None):
super().__init__(name=name)
- self.loss_fn = keras.backend.ctc_batch_cost
+ self.loss_fn = ctc_batch_cost
def call(self, y_true, y_pred):
# Compute the training-time loss value and add it
@@ -306,53 +359,80 @@ model = build_model()
model.summary()
```
-
-```
-Model: "ocr_model_v1"
-__________________________________________________________________________________________________
-Layer (type) Output Shape Param # Connected to
-==================================================================================================
-image (InputLayer) [(None, 200, 50, 1)] 0
-__________________________________________________________________________________________________
-Conv1 (Conv2D) (None, 200, 50, 32) 320 image[0][0]
-__________________________________________________________________________________________________
-pool1 (MaxPooling2D) (None, 100, 25, 32) 0 Conv1[0][0]
-__________________________________________________________________________________________________
-Conv2 (Conv2D) (None, 100, 25, 64) 18496 pool1[0][0]
-__________________________________________________________________________________________________
-pool2 (MaxPooling2D) (None, 50, 12, 64) 0 Conv2[0][0]
-__________________________________________________________________________________________________
-reshape (Reshape) (None, 50, 768) 0 pool2[0][0]
-__________________________________________________________________________________________________
-dense1 (Dense) (None, 50, 64) 49216 reshape[0][0]
-__________________________________________________________________________________________________
-dropout (Dropout) (None, 50, 64) 0 dense1[0][0]
-__________________________________________________________________________________________________
-bidirectional (Bidirectional) (None, 50, 256) 197632 dropout[0][0]
-__________________________________________________________________________________________________
-bidirectional_1 (Bidirectional) (None, 50, 128) 164352 bidirectional[0][0]
-__________________________________________________________________________________________________
-label (InputLayer) [(None, None)] 0
-__________________________________________________________________________________________________
-dense2 (Dense) (None, 50, 20) 2580 bidirectional_1[0][0]
-__________________________________________________________________________________________________
-ctc_loss (CTCLayer) (None, 50, 20) 0 label[0][0]
- dense2[0][0]
-==================================================================================================
-Total params: 432,596
-Trainable params: 432,596
-Non-trainable params: 0
-__________________________________________________________________________________________________
-```
-
+Model: "ocr_model_v1"
+
+
+
+
+
+┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┓
+┃ Layer (type) ┃ Output Shape ┃ Param # ┃ Connected to ┃
+┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━┩
+│ image (InputLayer) │ (None, 200, 50, │ 0 │ - │
+│ │ 1) │ │ │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ Conv1 (Conv2D) │ (None, 200, 50, │ 320 │ image[0][0] │
+│ │ 32) │ │ │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ pool1 │ (None, 100, 25, │ 0 │ Conv1[0][0] │
+│ (MaxPooling2D) │ 32) │ │ │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ Conv2 (Conv2D) │ (None, 100, 25, │ 18,496 │ pool1[0][0] │
+│ │ 64) │ │ │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ pool2 │ (None, 50, 12, │ 0 │ Conv2[0][0] │
+│ (MaxPooling2D) │ 64) │ │ │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ reshape (Reshape) │ (None, 50, 768) │ 0 │ pool2[0][0] │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ dense1 (Dense) │ (None, 50, 64) │ 49,216 │ reshape[0][0] │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ dropout (Dropout) │ (None, 50, 64) │ 0 │ dense1[0][0] │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ bidirectional │ (None, 50, 256) │ 197,632 │ dropout[0][0] │
+│ (Bidirectional) │ │ │ │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ bidirectional_1 │ (None, 50, 128) │ 164,352 │ bidirectional[0][0] │
+│ (Bidirectional) │ │ │ │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ label (InputLayer) │ (None, None) │ 0 │ - │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ dense2 (Dense) │ (None, 50, 21) │ 2,709 │ bidirectional_1[0][… │
+├─────────────────────┼───────────────────┼─────────┼──────────────────────┤
+│ ctc_loss (CTCLayer) │ (None, 50, 21) │ 0 │ label[0][0], │
+│ │ │ │ dense2[0][0] │
+└─────────────────────┴───────────────────┴─────────┴──────────────────────┘
+
+
+
+
+
+ Total params: 432,725 (1.65 MB)
+
+
+
+
+
+ Trainable params: 432,725 (1.65 MB)
+
+
+
+
+
+ Non-trainable params: 0 (0.00 B)
+
+
+
+
---
## Training
```python
-epochs = 100
+# TODO restore epoch count.
+epochs = 2
early_stopping_patience = 10
# Add early stopping
early_stopping = keras.callbacks.EarlyStopping(
@@ -371,227 +451,58 @@ history = model.fit(
```
-Epoch 1/100
-59/59 [==============================] - 3s 53ms/step - loss: 21.5722 - val_loss: 16.3351
-Epoch 2/100
-59/59 [==============================] - 2s 27ms/step - loss: 16.3335 - val_loss: 16.3062
-Epoch 3/100
-59/59 [==============================] - 2s 27ms/step - loss: 16.3360 - val_loss: 16.3116
-Epoch 4/100
-59/59 [==============================] - 2s 27ms/step - loss: 16.3318 - val_loss: 16.3167
-Epoch 5/100
-59/59 [==============================] - 2s 27ms/step - loss: 16.3256 - val_loss: 16.3152
-Epoch 6/100
-59/59 [==============================] - 2s 29ms/step - loss: 16.3229 - val_loss: 16.3123
-Epoch 7/100
-59/59 [==============================] - 2s 30ms/step - loss: 16.3119 - val_loss: 16.3116
-Epoch 8/100
-59/59 [==============================] - 2s 27ms/step - loss: 16.2977 - val_loss: 16.3107
-Epoch 9/100
-59/59 [==============================] - 2s 28ms/step - loss: 16.2801 - val_loss: 16.2552
-Epoch 10/100
-59/59 [==============================] - 2s 28ms/step - loss: 16.2199 - val_loss: 16.1008
-Epoch 11/100
-59/59 [==============================] - 2s 28ms/step - loss: 16.1136 - val_loss: 15.9867
-Epoch 12/100
-59/59 [==============================] - 2s 30ms/step - loss: 16.0138 - val_loss: 15.8825
-Epoch 13/100
-59/59 [==============================] - 2s 29ms/step - loss: 15.9670 - val_loss: 15.8413
-Epoch 14/100
-59/59 [==============================] - 2s 29ms/step - loss: 15.9315 - val_loss: 15.8263
-Epoch 15/100
-59/59 [==============================] - 2s 31ms/step - loss: 15.9162 - val_loss: 15.7971
-Epoch 16/100
-59/59 [==============================] - 2s 31ms/step - loss: 15.8916 - val_loss: 15.7844
-Epoch 17/100
-59/59 [==============================] - 2s 31ms/step - loss: 15.8653 - val_loss: 15.7624
-Epoch 18/100
-59/59 [==============================] - 2s 31ms/step - loss: 15.8543 - val_loss: 15.7620
-Epoch 19/100
-59/59 [==============================] - 2s 28ms/step - loss: 15.8373 - val_loss: 15.7559
-Epoch 20/100
-59/59 [==============================] - 2s 27ms/step - loss: 15.8319 - val_loss: 15.7495
-Epoch 21/100
-59/59 [==============================] - 2s 27ms/step - loss: 15.8104 - val_loss: 15.7430
-Epoch 22/100
-59/59 [==============================] - 2s 29ms/step - loss: 15.8037 - val_loss: 15.7260
-Epoch 23/100
-59/59 [==============================] - 2s 29ms/step - loss: 15.8021 - val_loss: 15.7204
-Epoch 24/100
-59/59 [==============================] - 2s 28ms/step - loss: 15.7901 - val_loss: 15.7174
-Epoch 25/100
-59/59 [==============================] - 2s 29ms/step - loss: 15.7851 - val_loss: 15.7074
-Epoch 26/100
-59/59 [==============================] - 2s 27ms/step - loss: 15.7701 - val_loss: 15.7097
-Epoch 27/100
-59/59 [==============================] - 2s 28ms/step - loss: 15.7694 - val_loss: 15.7040
-Epoch 28/100
-59/59 [==============================] - 2s 28ms/step - loss: 15.7544 - val_loss: 15.7012
-Epoch 29/100
-59/59 [==============================] - 2s 31ms/step - loss: 15.7498 - val_loss: 15.7015
-Epoch 30/100
-59/59 [==============================] - 2s 31ms/step - loss: 15.7521 - val_loss: 15.6880
-Epoch 31/100
-59/59 [==============================] - 2s 29ms/step - loss: 15.7165 - val_loss: 15.6734
-Epoch 32/100
-59/59 [==============================] - 2s 27ms/step - loss: 15.6650 - val_loss: 15.5789
-Epoch 33/100
-59/59 [==============================] - 2s 27ms/step - loss: 15.5300 - val_loss: 15.4026
-Epoch 34/100
-59/59 [==============================] - 2s 27ms/step - loss: 15.3519 - val_loss: 15.2115
-Epoch 35/100
-59/59 [==============================] - 2s 27ms/step - loss: 15.1165 - val_loss: 14.7826
-Epoch 36/100
-59/59 [==============================] - 2s 27ms/step - loss: 14.7086 - val_loss: 14.4432
-Epoch 37/100
-59/59 [==============================] - 2s 29ms/step - loss: 14.3317 - val_loss: 13.9445
-Epoch 38/100
-59/59 [==============================] - 2s 29ms/step - loss: 13.9658 - val_loss: 13.6972
-Epoch 39/100
-59/59 [==============================] - 2s 29ms/step - loss: 13.6728 - val_loss: 13.3388
-Epoch 40/100
-59/59 [==============================] - 2s 28ms/step - loss: 13.3454 - val_loss: 13.0102
-Epoch 41/100
-59/59 [==============================] - 2s 27ms/step - loss: 13.0448 - val_loss: 12.8307
-Epoch 42/100
-59/59 [==============================] - 2s 28ms/step - loss: 12.7552 - val_loss: 12.6071
-Epoch 43/100
-59/59 [==============================] - 2s 29ms/step - loss: 12.4573 - val_loss: 12.2800
-Epoch 44/100
-59/59 [==============================] - 2s 31ms/step - loss: 12.1055 - val_loss: 11.9209
-Epoch 45/100
-59/59 [==============================] - 2s 28ms/step - loss: 11.8148 - val_loss: 11.9132
-Epoch 46/100
-59/59 [==============================] - 2s 28ms/step - loss: 11.4530 - val_loss: 11.4357
-Epoch 47/100
-59/59 [==============================] - 2s 29ms/step - loss: 11.0592 - val_loss: 11.1121
-Epoch 48/100
-59/59 [==============================] - 2s 27ms/step - loss: 10.7746 - val_loss: 10.8532
-Epoch 49/100
-59/59 [==============================] - 2s 28ms/step - loss: 10.2616 - val_loss: 10.3643
-Epoch 50/100
-59/59 [==============================] - 2s 28ms/step - loss: 9.8708 - val_loss: 10.0987
-Epoch 51/100
-59/59 [==============================] - 2s 30ms/step - loss: 9.4077 - val_loss: 9.6371
-Epoch 52/100
-59/59 [==============================] - 2s 29ms/step - loss: 9.0663 - val_loss: 9.2463
-Epoch 53/100
-59/59 [==============================] - 2s 28ms/step - loss: 8.4546 - val_loss: 8.7581
-Epoch 54/100
-59/59 [==============================] - 2s 28ms/step - loss: 7.9226 - val_loss: 8.1805
-Epoch 55/100
-59/59 [==============================] - 2s 27ms/step - loss: 7.4927 - val_loss: 7.8858
-Epoch 56/100
-59/59 [==============================] - 2s 28ms/step - loss: 7.0499 - val_loss: 7.3202
-Epoch 57/100
-59/59 [==============================] - 2s 27ms/step - loss: 6.6383 - val_loss: 7.0875
-Epoch 58/100
-59/59 [==============================] - 2s 28ms/step - loss: 6.1446 - val_loss: 6.9619
-Epoch 59/100
-59/59 [==============================] - 2s 28ms/step - loss: 5.8533 - val_loss: 6.3855
-Epoch 60/100
-59/59 [==============================] - 2s 28ms/step - loss: 5.5107 - val_loss: 5.9797
-Epoch 61/100
-59/59 [==============================] - 2s 31ms/step - loss: 5.1181 - val_loss: 5.7549
-Epoch 62/100
-59/59 [==============================] - 2s 31ms/step - loss: 4.6952 - val_loss: 5.5488
-Epoch 63/100
-59/59 [==============================] - 2s 29ms/step - loss: 4.4189 - val_loss: 5.3030
-Epoch 64/100
-59/59 [==============================] - 2s 28ms/step - loss: 4.1358 - val_loss: 5.1772
-Epoch 65/100
-59/59 [==============================] - 2s 28ms/step - loss: 3.8560 - val_loss: 5.1071
-Epoch 66/100
-59/59 [==============================] - 2s 28ms/step - loss: 3.5342 - val_loss: 4.6958
-Epoch 67/100
-59/59 [==============================] - 2s 28ms/step - loss: 3.3336 - val_loss: 4.5865
-Epoch 68/100
-59/59 [==============================] - 2s 27ms/step - loss: 3.0925 - val_loss: 4.3647
-Epoch 69/100
-59/59 [==============================] - 2s 28ms/step - loss: 2.8751 - val_loss: 4.3005
-Epoch 70/100
-59/59 [==============================] - 2s 27ms/step - loss: 2.7444 - val_loss: 4.0820
-Epoch 71/100
-59/59 [==============================] - 2s 27ms/step - loss: 2.5921 - val_loss: 4.1694
-Epoch 72/100
-59/59 [==============================] - 2s 28ms/step - loss: 2.3246 - val_loss: 3.9142
-Epoch 73/100
-59/59 [==============================] - 2s 28ms/step - loss: 2.0769 - val_loss: 3.9135
-Epoch 74/100
-59/59 [==============================] - 2s 29ms/step - loss: 2.0872 - val_loss: 3.9808
-Epoch 75/100
-59/59 [==============================] - 2s 29ms/step - loss: 1.9498 - val_loss: 3.9935
-Epoch 76/100
-59/59 [==============================] - 2s 28ms/step - loss: 1.8178 - val_loss: 3.7735
-Epoch 77/100
-59/59 [==============================] - 2s 29ms/step - loss: 1.7661 - val_loss: 3.6309
-Epoch 78/100
-59/59 [==============================] - 2s 31ms/step - loss: 1.6236 - val_loss: 3.7410
-Epoch 79/100
-59/59 [==============================] - 2s 29ms/step - loss: 1.4652 - val_loss: 3.6756
-Epoch 80/100
-59/59 [==============================] - 2s 27ms/step - loss: 1.3552 - val_loss: 3.4979
-Epoch 81/100
-59/59 [==============================] - 2s 29ms/step - loss: 1.2655 - val_loss: 3.5306
-Epoch 82/100
-59/59 [==============================] - 2s 29ms/step - loss: 1.2632 - val_loss: 3.2885
-Epoch 83/100
-59/59 [==============================] - 2s 28ms/step - loss: 1.2316 - val_loss: 3.2482
-Epoch 84/100
-59/59 [==============================] - 2s 30ms/step - loss: 1.1260 - val_loss: 3.4285
-Epoch 85/100
-59/59 [==============================] - 2s 28ms/step - loss: 1.0745 - val_loss: 3.2985
-Epoch 86/100
-59/59 [==============================] - 2s 29ms/step - loss: 1.0133 - val_loss: 3.2209
-Epoch 87/100
-59/59 [==============================] - 2s 31ms/step - loss: 0.9417 - val_loss: 3.2203
-Epoch 88/100
-59/59 [==============================] - 2s 28ms/step - loss: 0.9104 - val_loss: 3.1121
-Epoch 89/100
-59/59 [==============================] - 2s 30ms/step - loss: 0.8516 - val_loss: 3.2070
-Epoch 90/100
-59/59 [==============================] - 2s 28ms/step - loss: 0.8275 - val_loss: 3.0335
-Epoch 91/100
-59/59 [==============================] - 2s 28ms/step - loss: 0.8056 - val_loss: 3.2085
-Epoch 92/100
-59/59 [==============================] - 2s 28ms/step - loss: 0.7373 - val_loss: 3.0326
-Epoch 93/100
-59/59 [==============================] - 2s 28ms/step - loss: 0.7753 - val_loss: 2.9935
-Epoch 94/100
-59/59 [==============================] - 2s 28ms/step - loss: 0.7688 - val_loss: 2.9940
-Epoch 95/100
-59/59 [==============================] - 2s 27ms/step - loss: 0.6765 - val_loss: 3.0432
-Epoch 96/100
-59/59 [==============================] - 2s 29ms/step - loss: 0.6674 - val_loss: 3.1233
-Epoch 97/100
-59/59 [==============================] - 2s 29ms/step - loss: 0.6018 - val_loss: 2.8405
-Epoch 98/100
-59/59 [==============================] - 2s 28ms/step - loss: 0.6322 - val_loss: 2.8323
-Epoch 99/100
-59/59 [==============================] - 2s 29ms/step - loss: 0.5889 - val_loss: 2.8786
-Epoch 100/100
-59/59 [==============================] - 2s 28ms/step - loss: 0.5616 - val_loss: 2.9697
+Epoch 1/2
+ 59/59 ━━━━━━━━━━━━━━━━━━━━ 8s 78ms/step - loss: 31.6480 - val_loss: 16.4761
+Epoch 2/2
+ 59/59 ━━━━━━━━━━━━━━━━━━━━ 4s 64ms/step - loss: 16.4152 - val_loss: 16.4351
```
---
## Inference
-You can use the trained model hosted on [Hugging Face Hub](https://huggingface.co/keras-io/ocr-for-captcha) and try the demo on [Hugging Face Spaces](https://huggingface.co/spaces/keras-io/ocr-for-captcha).
+You can use the trained model hosted on [Hugging Face Hub](https://huggingface.co/keras-io/ocr-for-captcha)
+and try the demo on [Hugging Face Spaces](https://huggingface.co/spaces/keras-io/ocr-for-captcha).
+
```python
+def ctc_decode(y_pred, input_length, greedy=True, beam_width=100, top_paths=1):
+ input_shape = tf.shape(y_pred)
+ num_samples, num_steps = input_shape[0], input_shape[1]
+ y_pred = tf.math.log(tf.transpose(y_pred, perm=[1, 0, 2]) + keras.backend.epsilon())
+ input_length = tf.cast(input_length, tf.int32)
+
+ if greedy:
+ (decoded, log_prob) = tf.nn.ctc_greedy_decoder(
+ inputs=y_pred, sequence_length=input_length
+ )
+ else:
+ (decoded, log_prob) = tf.compat.v1.nn.ctc_beam_search_decoder(
+ inputs=y_pred,
+ sequence_length=input_length,
+ beam_width=beam_width,
+ top_paths=top_paths,
+ )
+ decoded_dense = []
+ for st in decoded:
+ st = tf.SparseTensor(st.indices, st.values, (num_samples, num_steps))
+ decoded_dense.append(tf.sparse.to_dense(sp_input=st, default_value=-1))
+ return (decoded_dense, log_prob)
+
+
# Get the prediction model by extracting layers till the output layer
prediction_model = keras.models.Model(
- model.get_layer(name="image").input, model.get_layer(name="dense2").output
+ model.input[0], model.get_layer(name="dense2").output
)
prediction_model.summary()
+
# A utility function to decode the output of the network
def decode_batch_predictions(pred):
input_len = np.ones(pred.shape[0]) * pred.shape[1]
# Use greedy search. For complex tasks, you can use beam search
- results = keras.backend.ctc_decode(pred, input_length=input_len, greedy=True)[0][0][
+ results = ctc_decode(pred, input_length=input_len, greedy=True)[0][0][
:, :max_length
]
# Iterate over the results and get back the text
@@ -626,39 +537,67 @@ for batch in validation_dataset.take(1):
plt.show()
```
+
+Model: "functional_1"
+
+
+
+
+
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
+┃ Layer (type) ┃ Output Shape ┃ Param # ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
+│ image (InputLayer) │ (None, 200, 50, 1) │ 0 │
+├─────────────────────────────────┼───────────────────────────┼────────────┤
+│ Conv1 (Conv2D) │ (None, 200, 50, 32) │ 320 │
+├─────────────────────────────────┼───────────────────────────┼────────────┤
+│ pool1 (MaxPooling2D) │ (None, 100, 25, 32) │ 0 │
+├─────────────────────────────────┼───────────────────────────┼────────────┤
+│ Conv2 (Conv2D) │ (None, 100, 25, 64) │ 18,496 │
+├─────────────────────────────────┼───────────────────────────┼────────────┤
+│ pool2 (MaxPooling2D) │ (None, 50, 12, 64) │ 0 │
+├─────────────────────────────────┼───────────────────────────┼────────────┤
+│ reshape (Reshape) │ (None, 50, 768) │ 0 │
+├─────────────────────────────────┼───────────────────────────┼────────────┤
+│ dense1 (Dense) │ (None, 50, 64) │ 49,216 │
+├─────────────────────────────────┼───────────────────────────┼────────────┤
+│ dropout (Dropout) │ (None, 50, 64) │ 0 │
+├─────────────────────────────────┼───────────────────────────┼────────────┤
+│ bidirectional (Bidirectional) │ (None, 50, 256) │ 197,632 │
+├─────────────────────────────────┼───────────────────────────┼────────────┤
+│ bidirectional_1 (Bidirectional) │ (None, 50, 128) │ 164,352 │
+├─────────────────────────────────┼───────────────────────────┼────────────┤
+│ dense2 (Dense) │ (None, 50, 21) │ 2,709 │
+└─────────────────────────────────┴───────────────────────────┴────────────┘
+
+
+
+
+
+ Total params: 432,725 (1.65 MB)
+
+
+
+
+
+ Trainable params: 432,725 (1.65 MB)
+
+
+
+
+
+ Non-trainable params: 0 (0.00 B)
+
+
+
+
```
-Model: "functional_1"
-_________________________________________________________________
-Layer (type) Output Shape Param #
-=================================================================
-image (InputLayer) [(None, 200, 50, 1)] 0
-_________________________________________________________________
-Conv1 (Conv2D) (None, 200, 50, 32) 320
-_________________________________________________________________
-pool1 (MaxPooling2D) (None, 100, 25, 32) 0
-_________________________________________________________________
-Conv2 (Conv2D) (None, 100, 25, 64) 18496
-_________________________________________________________________
-pool2 (MaxPooling2D) (None, 50, 12, 64) 0
-_________________________________________________________________
-reshape (Reshape) (None, 50, 768) 0
-_________________________________________________________________
-dense1 (Dense) (None, 50, 64) 49216
-_________________________________________________________________
-dropout (Dropout) (None, 50, 64) 0
-_________________________________________________________________
-bidirectional (Bidirectional (None, 50, 256) 197632
-_________________________________________________________________
-bidirectional_1 (Bidirection (None, 50, 128) 164352
-_________________________________________________________________
-dense2 (Dense) (None, 50, 20) 2580
-=================================================================
-Total params: 432,596
-Trainable params: 432,596
-Non-trainable params: 0
-_________________________________________________________________
+ 1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 258ms/step
```
-
\ No newline at end of file
+
+
+
+