diff --git a/._README.md b/._README.md
new file mode 100644
index 0000000000000000000000000000000000000000..97192c5d8defd5f096e8bb9f8aaf963384dcb4c2
Binary files /dev/null and b/._README.md differ
diff --git a/._core_utils.py b/._core_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..4c7381b9ee4e5e4d1d3b9404b607a170638952e9
Binary files /dev/null and b/._core_utils.py differ
diff --git a/._deepfind.py b/._deepfind.py
new file mode 100644
index 0000000000000000000000000000000000000000..88716b326631ebe94e2c41209c99c07ba5816901
Binary files /dev/null and b/._deepfind.py differ
diff --git a/._losses.py b/._losses.py
new file mode 100644
index 0000000000000000000000000000000000000000..de5c5f6974bd6dc623366c4daba724ee97acfdf2
Binary files /dev/null and b/._losses.py differ
diff --git a/._models.py b/._models.py
new file mode 100644
index 0000000000000000000000000000000000000000..c8328815dd3b899007feb4f9cc33667374da8cd9
Binary files /dev/null and b/._models.py differ
diff --git a/._utils.py b/._utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..790fa5ca839d74647a268d13cae16cc4126f9448
Binary files /dev/null and b/._utils.py differ
diff --git a/README.md b/README.md
index a232442cc4ab067e330a788ac84a29518805fad4..d553cc8fca6ab2e5ccdc2ce841c77d62f0edea5f 100644
--- a/README.md
+++ b/README.md
@@ -16,16 +16,26 @@ keras (2.1.6)
numpy (1.14.3)
h5py (2.7.1)
lxml (4.3.2)
-sklearn (0.0)
+scikit-learn (0.19.1)
+scikit-image (0.14.2)
+matplotlib (2.2.3)
```
## Installation guide
First install the packages:
```
-pip install numpy tensorflow-gpu keras sklearn h5py lxml
+pip install numpy tensorflow-gpu keras sklearn h5py lxml scikit-learn scikit-image matplotlib
```
For more details about installing Keras, please see [Keras installation instructions](https://keras.io/#installation).
Once the dependencies are installed, the user should be able to run Deep Finder.
-## Instructions for use
\ No newline at end of file
+## Instructions for use
+Instructions for using Deep Finder are contained in folder examples/. The scripts contain comments on how the toolbox should be used. To run a script, first launch ipython:
+```
+ipython
+```
+The launch the script:
+```
+%run script_file.py
+```
\ No newline at end of file
diff --git a/core_utils.py b/core_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..95bd64454ab993c583654c5f79e33ee8997eca6c
--- /dev/null
+++ b/core_utils.py
@@ -0,0 +1,127 @@
+import numpy as np
+import utils
+import h5py
+
+import matplotlib
+matplotlib.use('agg') # necessary else: AttributeError: 'NoneType' object has no attribute 'is_interactive'
+import matplotlib.pyplot as plt
+
+def load_dataset(path_data, path_target):
+ data_list = []
+ target_list = []
+ for idx in range(0,len(path_data)):
+ data_list.append( utils.load_h5array(path_data[idx] ))
+ target_list.append(utils.load_h5array(path_target[idx]))
+ return data_list, target_list
+
+
+def get_bootstrap_idx(objlist,Nbs):
+ # Get a vector containing the object class labels (from objlist):
+ Nobj = len(objlist)
+ label_list = np.zeros((Nobj,))
+ for oo in range(0,Nobj):
+ label_list[oo] = float( objlist[oo].get('class_label') )
+
+ lblTAB = np.unique(label_list) # vector containing unique class labels
+
+ # Bootstrap data so that we have equal frequencies (1/Nbs) for all classes:
+ # ->from label_list, sample Nbs objects from each class
+ bs_idx = []
+ for l in lblTAB:
+ bs_idx.append( np.random.choice(np.squeeze(np.asarray(np.nonzero(label_list==l))), Nbs) )
+ bs_idx = np.concatenate(bs_idx)
+ return bs_idx
+
+def get_patch_position(tomodim, p_in, obj, Lrnd):
+ # sample at coordinates specified in obj=objlist[idx]
+ x = int( obj.get('x') )
+ y = int( obj.get('y') )
+ z = int( obj.get('z') )
+
+ # Add random shift to coordinates:
+ x = x + np.random.choice(range(-Lrnd,Lrnd+1))
+ y = y + np.random.choice(range(-Lrnd,Lrnd+1))
+ z = z + np.random.choice(range(-Lrnd,Lrnd+1))
+
+ # Shift position if too close to border:
+ if (x<p_in) : x = p_in
+ if (y<p_in) : y = p_in
+ if (z<p_in) : z = p_in
+ if (x>tomodim[0]-p_in): x = tomodim[0]-p_in
+ if (y>tomodim[1]-p_in): y = tomodim[1]-p_in
+ if (z>tomodim[2]-p_in): z = tomodim[2]-p_in
+
+ #else: # sample random position in tomogram
+ # x = np.int32( np.random.choice(range(p_in,tomodim[0]-p_in)) )
+ # y = np.int32( np.random.choice(range(p_in,tomodim[0]-p_in)) )
+ # z = np.int32( np.random.choice(range(p_in,tomodim[0]-p_in)) )
+
+ return x,y,z
+
+def save_history(history):
+ h5file = h5py.File('params_train_history.h5', 'w')
+
+ # train and val loss & accuracy:
+ dset = h5file.create_dataset('acc', (len(history['acc']),))
+ dset[:] = history['acc']
+ dset = h5file.create_dataset('loss', (len(history['loss']),))
+ dset[:] = history['loss']
+ dset = h5file.create_dataset('val_acc', (len(history['val_acc']),))
+ dset[:] = history['val_acc']
+ dset = h5file.create_dataset('val_loss', (len(history['val_loss']),))
+ dset[:] = history['val_loss']
+
+ # val precision, recall, F1:
+ dset = h5file.create_dataset('val_f1', np.shape(history['val_f1']))
+ dset[:] = history['val_f1']
+ dset = h5file.create_dataset('val_precision', np.shape(history['val_precision']))
+ dset[:] = history['val_precision']
+ dset = h5file.create_dataset('val_recall', np.shape(history['val_recall']))
+ dset[:] = history['val_recall']
+
+ h5file.close()
+ return
+
+def plot_history(history):
+ Ncl = len(history['val_f1'][0])
+ legend_names = []
+ for lbl in range(0,Ncl):
+ legend_names.append('class '+str(lbl))
+
+ fig = plt.figure(figsize=(15,12))
+ plt.subplot(321)
+ plt.plot(history['loss'] , label='train')
+ plt.plot(history['val_loss'], label='valid')
+ plt.ylabel('loss')
+ plt.xlabel('epochs')
+ plt.legend()
+ plt.grid()
+
+ plt.subplot(323)
+ plt.plot(history['acc'] , label='train')
+ plt.plot(history['val_acc'], label='valid')
+ plt.ylabel('accuracy')
+ plt.xlabel('epochs')
+ plt.legend()
+ plt.grid()
+
+ plt.subplot(322)
+ plt.plot(history['val_f1'])
+ plt.ylabel('F1-score')
+ plt.xlabel('epochs')
+ plt.legend(legend_names)
+ plt.grid()
+
+ plt.subplot(324)
+ plt.plot(history['val_precision'])
+ plt.ylabel('precision')
+ plt.xlabel('epochs')
+ plt.grid()
+
+ plt.subplot(326)
+ plt.plot(history['val_recall'])
+ plt.ylabel('recall')
+ plt.xlabel('epochs')
+ plt.grid()
+
+ fig.savefig('history_train.png')
\ No newline at end of file
diff --git a/data/tomo10_data.h5 b/data/tomo10_data.h5
new file mode 100644
index 0000000000000000000000000000000000000000..c2a2cd29fcb711ecf5e29a38f8ac8ddf0c7fd8f3
Binary files /dev/null and b/data/tomo10_data.h5 differ
diff --git a/data/tomo10_objlist.xml b/data/tomo10_objlist.xml
new file mode 100644
index 0000000000000000000000000000000000000000..2989796109f52d9081a04d32d02ae68b1e482ab2
--- /dev/null
+++ b/data/tomo10_objlist.xml
@@ -0,0 +1,29 @@
+<objlist>
+ <object tomo_idx="0" class_label="1" x="16" y="83" z="81"/>
+ <object tomo_idx="0" class_label="1" x="73" y="31" z="75"/>
+ <object tomo_idx="0" class_label="1" x="41" y="62" z="78"/>
+ <object tomo_idx="0" class_label="2" x="83" y="54" z="31"/>
+ <object tomo_idx="0" class_label="2" x="39" y="62" z="34"/>
+ <object tomo_idx="0" class_label="2" x="71" y="32" z="20"/>
+ <object tomo_idx="0" class_label="3" x="66" y="70" z="51"/>
+ <object tomo_idx="0" class_label="3" x="78" y="24" z="55"/>
+ <object tomo_idx="0" class_label="3" x="78" y="58" z="50"/>
+ <object tomo_idx="0" class_label="4" x="23" y="51" z="18"/>
+ <object tomo_idx="0" class_label="4" x="32" y="41" z="64"/>
+ <object tomo_idx="0" class_label="4" x="62" y="18" z="68"/>
+ <object tomo_idx="0" class_label="5" x="41" y="83" z="67"/>
+ <object tomo_idx="0" class_label="5" x="70" y="50" z="65"/>
+ <object tomo_idx="0" class_label="5" x="21" y="80" z="62"/>
+ <object tomo_idx="0" class_label="6" x="36" y="57" z="50"/>
+ <object tomo_idx="0" class_label="6" x="40" y="24" z="74"/>
+ <object tomo_idx="0" class_label="6" x="84" y="50" z="84"/>
+ <object tomo_idx="0" class_label="7" x="19" y="63" z="42"/>
+ <object tomo_idx="0" class_label="7" x="66" y="21" z="43"/>
+ <object tomo_idx="0" class_label="7" x="47" y="77" z="24"/>
+ <object tomo_idx="0" class_label="8" x="56" y="33" z="34"/>
+ <object tomo_idx="0" class_label="8" x="62" y="54" z="53"/>
+ <object tomo_idx="0" class_label="8" x="76" y="42" z="36"/>
+ <object tomo_idx="0" class_label="9" x="40" y="15" z="34"/>
+ <object tomo_idx="0" class_label="9" x="31" y="35" z="36"/>
+ <object tomo_idx="0" class_label="9" x="17" y="23" z="75"/>
+</objlist>
diff --git a/data/tomo10_target.h5 b/data/tomo10_target.h5
new file mode 100644
index 0000000000000000000000000000000000000000..3d88a3e96e9b0ad4648c62763cc3da319ae44887
Binary files /dev/null and b/data/tomo10_target.h5 differ
diff --git a/data/tomo1_data.h5 b/data/tomo1_data.h5
new file mode 100644
index 0000000000000000000000000000000000000000..3893566a08964ae34675504a09c2406ec46459f0
Binary files /dev/null and b/data/tomo1_data.h5 differ
diff --git a/data/tomo1_objlist.xml b/data/tomo1_objlist.xml
new file mode 100644
index 0000000000000000000000000000000000000000..025bbfa3edc60b9f075569995e175467e2466e12
--- /dev/null
+++ b/data/tomo1_objlist.xml
@@ -0,0 +1,29 @@
+<objlist>
+ <object tomo_idx="0" class_label="1" x="34" y="16" z="64"/>
+ <object tomo_idx="0" class_label="1" x="63" y="72" z="21"/>
+ <object tomo_idx="0" class_label="1" x="45" y="17" z="81"/>
+ <object tomo_idx="0" class_label="2" x="27" y="70" z="68"/>
+ <object tomo_idx="0" class_label="2" x="64" y="59" z="45"/>
+ <object tomo_idx="0" class_label="2" x="55" y="38" z="81"/>
+ <object tomo_idx="0" class_label="3" x="81" y="77" z="63"/>
+ <object tomo_idx="0" class_label="3" x="32" y="25" z="24"/>
+ <object tomo_idx="0" class_label="3" x="31" y="71" z="32"/>
+ <object tomo_idx="0" class_label="4" x="32" y="28" z="39"/>
+ <object tomo_idx="0" class_label="4" x="72" y="39" z="47"/>
+ <object tomo_idx="0" class_label="4" x="34" y="78" z="53"/>
+ <object tomo_idx="0" class_label="5" x="15" y="47" z="54"/>
+ <object tomo_idx="0" class_label="5" x="56" y="26" z="51"/>
+ <object tomo_idx="0" class_label="5" x="20" y="84" z="52"/>
+ <object tomo_idx="0" class_label="6" x="15" y="82" z="22"/>
+ <object tomo_idx="0" class_label="6" x="53" y="55" z="75"/>
+ <object tomo_idx="0" class_label="6" x="18" y="43" z="31"/>
+ <object tomo_idx="0" class_label="7" x="81" y="80" z="23"/>
+ <object tomo_idx="0" class_label="7" x="17" y="15" z="72"/>
+ <object tomo_idx="0" class_label="7" x="60" y="65" z="60"/>
+ <object tomo_idx="0" class_label="8" x="46" y="45" z="48"/>
+ <object tomo_idx="0" class_label="8" x="53" y="76" z="80"/>
+ <object tomo_idx="0" class_label="8" x="36" y="45" z="30"/>
+ <object tomo_idx="0" class_label="9" x="22" y="45" z="83"/>
+ <object tomo_idx="0" class_label="9" x="58" y="48" z="30"/>
+ <object tomo_idx="0" class_label="9" x="76" y="31" z="27"/>
+</objlist>
diff --git a/data/tomo1_target.h5 b/data/tomo1_target.h5
new file mode 100644
index 0000000000000000000000000000000000000000..fa97345ed41ad90d741b010ca8ed7db5e0ba8ab5
Binary files /dev/null and b/data/tomo1_target.h5 differ
diff --git a/data/tomo2_data.h5 b/data/tomo2_data.h5
new file mode 100644
index 0000000000000000000000000000000000000000..6c073bcef2b745cdd327afbd0bb9a037305b17b1
Binary files /dev/null and b/data/tomo2_data.h5 differ
diff --git a/data/tomo2_objlist.xml b/data/tomo2_objlist.xml
new file mode 100644
index 0000000000000000000000000000000000000000..f6c9cd585d9dbb70ab1463881f1f94b63c498cd0
--- /dev/null
+++ b/data/tomo2_objlist.xml
@@ -0,0 +1,29 @@
+<objlist>
+ <object tomo_idx="0" class_label="1" x="45" y="46" z="70"/>
+ <object tomo_idx="0" class_label="1" x="26" y="24" z="20"/>
+ <object tomo_idx="0" class_label="1" x="18" y="70" z="72"/>
+ <object tomo_idx="0" class_label="2" x="60" y="43" z="51"/>
+ <object tomo_idx="0" class_label="2" x="15" y="44" z="35"/>
+ <object tomo_idx="0" class_label="2" x="38" y="49" z="28"/>
+ <object tomo_idx="0" class_label="3" x="66" y="18" z="79"/>
+ <object tomo_idx="0" class_label="3" x="35" y="83" z="43"/>
+ <object tomo_idx="0" class_label="3" x="23" y="27" z="61"/>
+ <object tomo_idx="0" class_label="4" x="83" y="47" z="40"/>
+ <object tomo_idx="0" class_label="4" x="32" y="55" z="41"/>
+ <object tomo_idx="0" class_label="4" x="38" y="65" z="83"/>
+ <object tomo_idx="0" class_label="5" x="76" y="78" z="22"/>
+ <object tomo_idx="0" class_label="5" x="26" y="36" z="44"/>
+ <object tomo_idx="0" class_label="5" x="56" y="21" z="44"/>
+ <object tomo_idx="0" class_label="6" x="44" y="69" z="57"/>
+ <object tomo_idx="0" class_label="6" x="46" y="51" z="45"/>
+ <object tomo_idx="0" class_label="6" x="59" y="80" z="51"/>
+ <object tomo_idx="0" class_label="7" x="61" y="30" z="32"/>
+ <object tomo_idx="0" class_label="7" x="62" y="69" z="38"/>
+ <object tomo_idx="0" class_label="7" x="44" y="47" z="14"/>
+ <object tomo_idx="0" class_label="8" x="27" y="17" z="47"/>
+ <object tomo_idx="0" class_label="8" x="66" y="28" z="57"/>
+ <object tomo_idx="0" class_label="8" x="31" y="64" z="29"/>
+ <object tomo_idx="0" class_label="9" x="22" y="30" z="84"/>
+ <object tomo_idx="0" class_label="9" x="76" y="50" z="84"/>
+ <object tomo_idx="0" class_label="9" x="74" y="18" z="25"/>
+</objlist>
diff --git a/data/tomo2_target.h5 b/data/tomo2_target.h5
new file mode 100644
index 0000000000000000000000000000000000000000..5ae01bdff70bf8d4bac31f715e160a2564228c8e
Binary files /dev/null and b/data/tomo2_target.h5 differ
diff --git a/data/tomo3_data.h5 b/data/tomo3_data.h5
new file mode 100644
index 0000000000000000000000000000000000000000..628b86ba802b3b58e7436f9b7910225e324be651
Binary files /dev/null and b/data/tomo3_data.h5 differ
diff --git a/data/tomo3_objlist.xml b/data/tomo3_objlist.xml
new file mode 100644
index 0000000000000000000000000000000000000000..60e531fafdfec178a631fb74dc4c42955424f787
--- /dev/null
+++ b/data/tomo3_objlist.xml
@@ -0,0 +1,29 @@
+<objlist>
+ <object tomo_idx="0" class_label="1" x="20" y="61" z="32"/>
+ <object tomo_idx="0" class_label="1" x="15" y="55" z="79"/>
+ <object tomo_idx="0" class_label="1" x="73" y="48" z="75"/>
+ <object tomo_idx="0" class_label="2" x="16" y="58" z="53"/>
+ <object tomo_idx="0" class_label="2" x="49" y="18" z="40"/>
+ <object tomo_idx="0" class_label="2" x="25" y="29" z="23"/>
+ <object tomo_idx="0" class_label="3" x="34" y="59" z="17"/>
+ <object tomo_idx="0" class_label="3" x="33" y="75" z="74"/>
+ <object tomo_idx="0" class_label="3" x="43" y="59" z="54"/>
+ <object tomo_idx="0" class_label="4" x="22" y="20" z="81"/>
+ <object tomo_idx="0" class_label="4" x="54" y="58" z="26"/>
+ <object tomo_idx="0" class_label="4" x="80" y="74" z="19"/>
+ <object tomo_idx="0" class_label="5" x="24" y="42" z="27"/>
+ <object tomo_idx="0" class_label="5" x="60" y="47" z="38"/>
+ <object tomo_idx="0" class_label="5" x="18" y="45" z="39"/>
+ <object tomo_idx="0" class_label="6" x="52" y="83" z="23"/>
+ <object tomo_idx="0" class_label="6" x="39" y="27" z="21"/>
+ <object tomo_idx="0" class_label="6" x="27" y="38" z="51"/>
+ <object tomo_idx="0" class_label="7" x="66" y="22" z="78"/>
+ <object tomo_idx="0" class_label="7" x="56" y="27" z="54"/>
+ <object tomo_idx="0" class_label="7" x="77" y="29" z="24"/>
+ <object tomo_idx="0" class_label="8" x="28" y="77" z="15"/>
+ <object tomo_idx="0" class_label="8" x="35" y="37" z="84"/>
+ <object tomo_idx="0" class_label="8" x="72" y="22" z="52"/>
+ <object tomo_idx="0" class_label="9" x="77" y="48" z="53"/>
+ <object tomo_idx="0" class_label="9" x="59" y="22" z="22"/>
+ <object tomo_idx="0" class_label="9" x="23" y="23" z="63"/>
+</objlist>
diff --git a/data/tomo3_target.h5 b/data/tomo3_target.h5
new file mode 100644
index 0000000000000000000000000000000000000000..5bd9e139e3b5bf457b6343b1ecd3441e45982470
Binary files /dev/null and b/data/tomo3_target.h5 differ
diff --git a/data/tomo4_data.h5 b/data/tomo4_data.h5
new file mode 100644
index 0000000000000000000000000000000000000000..c7da220cedbf80a45db154d77de0dc200f5cd20f
Binary files /dev/null and b/data/tomo4_data.h5 differ
diff --git a/data/tomo4_objlist.xml b/data/tomo4_objlist.xml
new file mode 100644
index 0000000000000000000000000000000000000000..1cd3b43b26369fd408291baeda6779c57fe0bfb1
--- /dev/null
+++ b/data/tomo4_objlist.xml
@@ -0,0 +1,29 @@
+<objlist>
+ <object tomo_idx="0" class_label="1" x="53" y="64" z="78"/>
+ <object tomo_idx="0" class_label="1" x="83" y="58" z="26"/>
+ <object tomo_idx="0" class_label="1" x="83" y="42" z="62"/>
+ <object tomo_idx="0" class_label="2" x="41" y="25" z="58"/>
+ <object tomo_idx="0" class_label="2" x="39" y="75" z="67"/>
+ <object tomo_idx="0" class_label="2" x="35" y="38" z="56"/>
+ <object tomo_idx="0" class_label="3" x="53" y="39" z="44"/>
+ <object tomo_idx="0" class_label="3" x="23" y="44" z="44"/>
+ <object tomo_idx="0" class_label="3" x="22" y="66" z="66"/>
+ <object tomo_idx="0" class_label="4" x="21" y="29" z="47"/>
+ <object tomo_idx="0" class_label="4" x="61" y="26" z="27"/>
+ <object tomo_idx="0" class_label="4" x="25" y="63" z="50"/>
+ <object tomo_idx="0" class_label="5" x="79" y="16" z="32"/>
+ <object tomo_idx="0" class_label="5" x="79" y="26" z="80"/>
+ <object tomo_idx="0" class_label="5" x="19" y="30" z="67"/>
+ <object tomo_idx="0" class_label="6" x="71" y="42" z="44"/>
+ <object tomo_idx="0" class_label="6" x="74" y="69" z="71"/>
+ <object tomo_idx="0" class_label="6" x="30" y="84" z="39"/>
+ <object tomo_idx="0" class_label="7" x="27" y="44" z="16"/>
+ <object tomo_idx="0" class_label="7" x="81" y="27" z="54"/>
+ <object tomo_idx="0" class_label="7" x="40" y="30" z="79"/>
+ <object tomo_idx="0" class_label="8" x="60" y="39" z="65"/>
+ <object tomo_idx="0" class_label="8" x="55" y="57" z="41"/>
+ <object tomo_idx="0" class_label="8" x="19" y="83" z="70"/>
+ <object tomo_idx="0" class_label="9" x="77" y="76" z="52"/>
+ <object tomo_idx="0" class_label="9" x="62" y="80" z="83"/>
+ <object tomo_idx="0" class_label="9" x="21" y="77" z="28"/>
+</objlist>
diff --git a/data/tomo4_target.h5 b/data/tomo4_target.h5
new file mode 100644
index 0000000000000000000000000000000000000000..0d667ed0c38f60438cbf34928dddd552fe99d6da
Binary files /dev/null and b/data/tomo4_target.h5 differ
diff --git a/data/tomo5_data.h5 b/data/tomo5_data.h5
new file mode 100644
index 0000000000000000000000000000000000000000..bc40b003298a9b9e3b33986a5a94026119182fd9
Binary files /dev/null and b/data/tomo5_data.h5 differ
diff --git a/data/tomo5_objlist.xml b/data/tomo5_objlist.xml
new file mode 100644
index 0000000000000000000000000000000000000000..f28c232cd999e70ae975e5cb8b6e0c7a43f3529e
--- /dev/null
+++ b/data/tomo5_objlist.xml
@@ -0,0 +1,29 @@
+<objlist>
+ <object tomo_idx="0" class_label="1" x="60" y="65" z="23"/>
+ <object tomo_idx="0" class_label="1" x="58" y="37" z="53"/>
+ <object tomo_idx="0" class_label="1" x="81" y="66" z="73"/>
+ <object tomo_idx="0" class_label="2" x="18" y="60" z="56"/>
+ <object tomo_idx="0" class_label="2" x="65" y="34" z="74"/>
+ <object tomo_idx="0" class_label="2" x="55" y="24" z="73"/>
+ <object tomo_idx="0" class_label="3" x="49" y="50" z="71"/>
+ <object tomo_idx="0" class_label="3" x="27" y="82" z="17"/>
+ <object tomo_idx="0" class_label="3" x="60" y="82" z="20"/>
+ <object tomo_idx="0" class_label="4" x="33" y="23" z="42"/>
+ <object tomo_idx="0" class_label="4" x="18" y="42" z="79"/>
+ <object tomo_idx="0" class_label="4" x="52" y="70" z="66"/>
+ <object tomo_idx="0" class_label="5" x="36" y="18" z="77"/>
+ <object tomo_idx="0" class_label="5" x="19" y="55" z="76"/>
+ <object tomo_idx="0" class_label="5" x="24" y="36" z="60"/>
+ <object tomo_idx="0" class_label="6" x="16" y="24" z="61"/>
+ <object tomo_idx="0" class_label="6" x="55" y="65" z="53"/>
+ <object tomo_idx="0" class_label="6" x="19" y="60" z="29"/>
+ <object tomo_idx="0" class_label="7" x="71" y="80" z="61"/>
+ <object tomo_idx="0" class_label="7" x="82" y="43" z="67"/>
+ <object tomo_idx="0" class_label="7" x="46" y="77" z="25"/>
+ <object tomo_idx="0" class_label="8" x="34" y="43" z="23"/>
+ <object tomo_idx="0" class_label="8" x="63" y="49" z="42"/>
+ <object tomo_idx="0" class_label="8" x="22" y="45" z="45"/>
+ <object tomo_idx="0" class_label="9" x="26" y="20" z="18"/>
+ <object tomo_idx="0" class_label="9" x="58" y="25" z="21"/>
+ <object tomo_idx="0" class_label="9" x="73" y="45" z="16"/>
+</objlist>
diff --git a/data/tomo5_target.h5 b/data/tomo5_target.h5
new file mode 100644
index 0000000000000000000000000000000000000000..51aa5285d51668c5187709738479b52945d4e469
Binary files /dev/null and b/data/tomo5_target.h5 differ
diff --git a/data/tomo6_data.h5 b/data/tomo6_data.h5
new file mode 100644
index 0000000000000000000000000000000000000000..a70fbc2f417535c5c00f60261a93652e142ed653
Binary files /dev/null and b/data/tomo6_data.h5 differ
diff --git a/data/tomo6_objlist.xml b/data/tomo6_objlist.xml
new file mode 100644
index 0000000000000000000000000000000000000000..6389680ee3600fec74a61bc331d7b8bf128d6c6e
--- /dev/null
+++ b/data/tomo6_objlist.xml
@@ -0,0 +1,29 @@
+<objlist>
+ <object tomo_idx="0" class_label="1" x="26" y="50" z="61"/>
+ <object tomo_idx="0" class_label="1" x="55" y="31" z="40"/>
+ <object tomo_idx="0" class_label="1" x="20" y="32" z="24"/>
+ <object tomo_idx="0" class_label="2" x="64" y="49" z="23"/>
+ <object tomo_idx="0" class_label="2" x="42" y="19" z="69"/>
+ <object tomo_idx="0" class_label="2" x="27" y="14" z="30"/>
+ <object tomo_idx="0" class_label="3" x="80" y="77" z="56"/>
+ <object tomo_idx="0" class_label="3" x="81" y="16" z="26"/>
+ <object tomo_idx="0" class_label="3" x="15" y="68" z="82"/>
+ <object tomo_idx="0" class_label="4" x="53" y="59" z="64"/>
+ <object tomo_idx="0" class_label="4" x="42" y="74" z="77"/>
+ <object tomo_idx="0" class_label="4" x="74" y="38" z="56"/>
+ <object tomo_idx="0" class_label="5" x="52" y="17" z="78"/>
+ <object tomo_idx="0" class_label="5" x="72" y="80" z="81"/>
+ <object tomo_idx="0" class_label="5" x="63" y="36" z="14"/>
+ <object tomo_idx="0" class_label="6" x="39" y="50" z="23"/>
+ <object tomo_idx="0" class_label="6" x="40" y="65" z="60"/>
+ <object tomo_idx="0" class_label="6" x="82" y="56" z="34"/>
+ <object tomo_idx="0" class_label="7" x="80" y="38" z="28"/>
+ <object tomo_idx="0" class_label="7" x="58" y="81" z="31"/>
+ <object tomo_idx="0" class_label="7" x="55" y="30" z="65"/>
+ <object tomo_idx="0" class_label="8" x="18" y="50" z="37"/>
+ <object tomo_idx="0" class_label="8" x="78" y="74" z="35"/>
+ <object tomo_idx="0" class_label="8" x="25" y="26" z="54"/>
+ <object tomo_idx="0" class_label="9" x="46" y="71" z="18"/>
+ <object tomo_idx="0" class_label="9" x="80" y="17" z="67"/>
+ <object tomo_idx="0" class_label="9" x="24" y="70" z="15"/>
+</objlist>
diff --git a/data/tomo6_target.h5 b/data/tomo6_target.h5
new file mode 100644
index 0000000000000000000000000000000000000000..ffafda0f417e18bd1fd8c4aee2d78f846ffc025b
Binary files /dev/null and b/data/tomo6_target.h5 differ
diff --git a/data/tomo7_data.h5 b/data/tomo7_data.h5
new file mode 100644
index 0000000000000000000000000000000000000000..eca49e3088eb1ea47953719fb83eb135220243f5
Binary files /dev/null and b/data/tomo7_data.h5 differ
diff --git a/data/tomo7_objlist.xml b/data/tomo7_objlist.xml
new file mode 100644
index 0000000000000000000000000000000000000000..f3cb6c8fcbb57ffd156f9884744d41d08c9a0648
--- /dev/null
+++ b/data/tomo7_objlist.xml
@@ -0,0 +1,29 @@
+<objlist>
+ <object tomo_idx="0" class_label="1" x="44" y="77" z="71"/>
+ <object tomo_idx="0" class_label="1" x="36" y="52" z="52"/>
+ <object tomo_idx="0" class_label="1" x="47" y="21" z="27"/>
+ <object tomo_idx="0" class_label="2" x="14" y="59" z="78"/>
+ <object tomo_idx="0" class_label="2" x="23" y="46" z="29"/>
+ <object tomo_idx="0" class_label="2" x="18" y="45" z="45"/>
+ <object tomo_idx="0" class_label="3" x="82" y="74" z="73"/>
+ <object tomo_idx="0" class_label="3" x="26" y="78" z="59"/>
+ <object tomo_idx="0" class_label="3" x="76" y="45" z="55"/>
+ <object tomo_idx="0" class_label="4" x="84" y="70" z="37"/>
+ <object tomo_idx="0" class_label="4" x="16" y="31" z="23"/>
+ <object tomo_idx="0" class_label="4" x="23" y="81" z="45"/>
+ <object tomo_idx="0" class_label="5" x="63" y="17" z="74"/>
+ <object tomo_idx="0" class_label="5" x="77" y="57" z="78"/>
+ <object tomo_idx="0" class_label="5" x="58" y="72" z="25"/>
+ <object tomo_idx="0" class_label="6" x="81" y="19" z="71"/>
+ <object tomo_idx="0" class_label="6" x="43" y="76" z="17"/>
+ <object tomo_idx="0" class_label="6" x="24" y="25" z="66"/>
+ <object tomo_idx="0" class_label="7" x="31" y="68" z="37"/>
+ <object tomo_idx="0" class_label="7" x="62" y="33" z="53"/>
+ <object tomo_idx="0" class_label="7" x="52" y="41" z="46"/>
+ <object tomo_idx="0" class_label="8" x="67" y="18" z="51"/>
+ <object tomo_idx="0" class_label="8" x="52" y="14" z="43"/>
+ <object tomo_idx="0" class_label="8" x="23" y="66" z="70"/>
+ <object tomo_idx="0" class_label="9" x="59" y="82" z="82"/>
+ <object tomo_idx="0" class_label="9" x="69" y="29" z="19"/>
+ <object tomo_idx="0" class_label="9" x="37" y="17" z="83"/>
+</objlist>
diff --git a/data/tomo7_target.h5 b/data/tomo7_target.h5
new file mode 100644
index 0000000000000000000000000000000000000000..69fcd41e59ac652fe9ac77c0498ae150b7fdadc9
Binary files /dev/null and b/data/tomo7_target.h5 differ
diff --git a/data/tomo8_data.h5 b/data/tomo8_data.h5
new file mode 100644
index 0000000000000000000000000000000000000000..ccea7817e5114b68cea9f313510a4c24ea7a3aaf
Binary files /dev/null and b/data/tomo8_data.h5 differ
diff --git a/data/tomo8_objlist.xml b/data/tomo8_objlist.xml
new file mode 100644
index 0000000000000000000000000000000000000000..07b5b388010a15a2f98a5b96c266f5c76ba80a45
--- /dev/null
+++ b/data/tomo8_objlist.xml
@@ -0,0 +1,29 @@
+<objlist>
+ <object tomo_idx="0" class_label="1" x="16" y="18" z="18"/>
+ <object tomo_idx="0" class_label="1" x="19" y="19" z="57"/>
+ <object tomo_idx="0" class_label="1" x="74" y="22" z="31"/>
+ <object tomo_idx="0" class_label="2" x="50" y="60" z="66"/>
+ <object tomo_idx="0" class_label="2" x="82" y="82" z="16"/>
+ <object tomo_idx="0" class_label="2" x="35" y="60" z="47"/>
+ <object tomo_idx="0" class_label="3" x="65" y="77" z="67"/>
+ <object tomo_idx="0" class_label="3" x="32" y="33" z="81"/>
+ <object tomo_idx="0" class_label="3" x="56" y="35" z="19"/>
+ <object tomo_idx="0" class_label="4" x="31" y="70" z="60"/>
+ <object tomo_idx="0" class_label="4" x="34" y="49" z="79"/>
+ <object tomo_idx="0" class_label="4" x="82" y="50" z="78"/>
+ <object tomo_idx="0" class_label="5" x="63" y="36" z="44"/>
+ <object tomo_idx="0" class_label="5" x="36" y="35" z="42"/>
+ <object tomo_idx="0" class_label="5" x="78" y="17" z="75"/>
+ <object tomo_idx="0" class_label="6" x="43" y="50" z="24"/>
+ <object tomo_idx="0" class_label="6" x="49" y="83" z="51"/>
+ <object tomo_idx="0" class_label="6" x="35" y="17" z="43"/>
+ <object tomo_idx="0" class_label="7" x="50" y="20" z="39"/>
+ <object tomo_idx="0" class_label="7" x="23" y="28" z="71"/>
+ <object tomo_idx="0" class_label="7" x="41" y="74" z="67"/>
+ <object tomo_idx="0" class_label="8" x="73" y="37" z="66"/>
+ <object tomo_idx="0" class_label="8" x="41" y="45" z="66"/>
+ <object tomo_idx="0" class_label="8" x="16" y="47" z="20"/>
+ <object tomo_idx="0" class_label="9" x="19" y="66" z="45"/>
+ <object tomo_idx="0" class_label="9" x="63" y="66" z="27"/>
+ <object tomo_idx="0" class_label="9" x="79" y="52" z="38"/>
+</objlist>
diff --git a/data/tomo8_target.h5 b/data/tomo8_target.h5
new file mode 100644
index 0000000000000000000000000000000000000000..a8f33b5b8f00a0c9ca1f5e3f314f4a5c90638ad7
Binary files /dev/null and b/data/tomo8_target.h5 differ
diff --git a/data/tomo9_data.h5 b/data/tomo9_data.h5
new file mode 100644
index 0000000000000000000000000000000000000000..4dae03be4bd57ba6c3060928e3798a8ece6a664e
Binary files /dev/null and b/data/tomo9_data.h5 differ
diff --git a/data/tomo9_objlist.xml b/data/tomo9_objlist.xml
new file mode 100644
index 0000000000000000000000000000000000000000..0f13054191a23c74938f4aac69685a295e7482f7
--- /dev/null
+++ b/data/tomo9_objlist.xml
@@ -0,0 +1,29 @@
+<objlist>
+ <object tomo_idx="0" class_label="1" x="24" y="80" z="75"/>
+ <object tomo_idx="0" class_label="1" x="77" y="59" z="83"/>
+ <object tomo_idx="0" class_label="1" x="30" y="58" z="15"/>
+ <object tomo_idx="0" class_label="2" x="44" y="31" z="45"/>
+ <object tomo_idx="0" class_label="2" x="32" y="76" z="17"/>
+ <object tomo_idx="0" class_label="2" x="56" y="39" z="21"/>
+ <object tomo_idx="0" class_label="3" x="52" y="18" z="73"/>
+ <object tomo_idx="0" class_label="3" x="49" y="60" z="34"/>
+ <object tomo_idx="0" class_label="3" x="77" y="77" z="16"/>
+ <object tomo_idx="0" class_label="4" x="64" y="27" z="23"/>
+ <object tomo_idx="0" class_label="4" x="65" y="69" z="29"/>
+ <object tomo_idx="0" class_label="4" x="80" y="53" z="33"/>
+ <object tomo_idx="0" class_label="5" x="69" y="81" z="62"/>
+ <object tomo_idx="0" class_label="5" x="48" y="68" z="83"/>
+ <object tomo_idx="0" class_label="5" x="47" y="16" z="41"/>
+ <object tomo_idx="0" class_label="6" x="72" y="47" z="43"/>
+ <object tomo_idx="0" class_label="6" x="66" y="79" z="51"/>
+ <object tomo_idx="0" class_label="6" x="48" y="41" z="59"/>
+ <object tomo_idx="0" class_label="7" x="47" y="25" z="15"/>
+ <object tomo_idx="0" class_label="7" x="77" y="60" z="18"/>
+ <object tomo_idx="0" class_label="7" x="83" y="34" z="45"/>
+ <object tomo_idx="0" class_label="8" x="73" y="73" z="72"/>
+ <object tomo_idx="0" class_label="8" x="75" y="33" z="56"/>
+ <object tomo_idx="0" class_label="8" x="15" y="38" z="19"/>
+ <object tomo_idx="0" class_label="9" x="17" y="39" z="76"/>
+ <object tomo_idx="0" class_label="9" x="79" y="40" z="15"/>
+ <object tomo_idx="0" class_label="9" x="44" y="46" z="84"/>
+</objlist>
diff --git a/data/tomo9_target.h5 b/data/tomo9_target.h5
new file mode 100644
index 0000000000000000000000000000000000000000..895a0bae5febd2df4e2a8486adb2671283d1735c
Binary files /dev/null and b/data/tomo9_target.h5 differ
diff --git a/deepfind.py b/deepfind.py
index a9a6ef774a2c8f1247f37f8c7b8a44717446c4bd..9475ee96d901bea88d1ecb0cbfaae81800dc4244 100644
--- a/deepfind.py
+++ b/deepfind.py
@@ -14,6 +14,8 @@ from lxml import etree
import models
import losses
+import core_utils
+import utils
class deepfind:
def __init__(self, Ncl):
@@ -28,22 +30,35 @@ class deepfind:
self.Nvalid = 100 # number of samples for validation
self.optimizer = Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
+ self.flag_direct_read = 1
+ self.flag_batch_bootstrap = 1
self.Lrnd = 13 # random shifts applied when sampling data- and target-patches (in voxels)
-
- self.net = models.my_model(self.dim_in, self.Ncl)
- self.net.compile(optimizer=self.optimizer, loss=losses.tversky_loss, metrics=['accuracy'])
-
+
# Segmentation, parameters for dividing data in patches:
self.P = 192 # patch length (in pixels) /!\ has to a multiple of 4 (because of 2 pooling layers), so that dim_in=dim_out
self.poverlap = 70 # patch overlap (in pixels)
self.pcrop = 25 # how many pixels to crop from border
- # Parameters:
- # path_data : path to data (i.e. tomograms)
- # path_target : path to targets (i.e. annotated volumes)
- # objlist_train : object list containing coordinates of macromolecules
- # objlist_valid : another object list for monitoring the training (in particular detect overfitting)
+ # This function launches the training procedure. For each epoch, an image is plotted, displaying the progression with different metrics: loss, accuracy, f1-score, recall, precision. Every 10 epochs, the current network weights are saved.
+ # The network is trained on small 3D patches (i.e. sub-volumes), sampled from the larger tomograms (due to memory limitation). The patch sampling is not realized randomly, but is guided by the macromolecule coordinates contained in so-called object lists (objlist).
+ # Concerning the loading of the dataset, two options are possible:
+ # flag_direct_read=0:
+ # flag_direct_read=1:
+ # Arguments:
+ # path_data : a list containing the paths to data files (i.e. tomograms)
+ # path_target : a list containing the paths to target files (i.e. annotated volumes)
+ # objlist_train : an xml structure containing information about annotated objects: origin tomogram (should correspond to the index of 'path_data' argument), coordinates, class. During training, these coordinates are used for guiding the patch sampling procedure.
+ # objlist_valid : same as 'objlist_train', but objects contained in this xml structure are not used for training, but for validation. It allows to monitor the training and check for over/under-fitting. Ideally, the validation objects should originate from different tomograms than training objects.
def train(self, path_data, path_target, objlist_train, objlist_valid):
+ # Build network:
+ net = models.my_model(self.dim_in, self.Ncl)
+ net.compile(optimizer=self.optimizer, loss='categorical_crossentropy', metrics=['accuracy'])
+
+ # Load whole dataset:
+ if self.flag_direct_read == False:
+ print('Loading dataset ...')
+ data_list, target_list = core_utils.load_dataset(path_data, path_target)
+
print('Launch training ...')
# Declare lists for storing training statistics:
@@ -61,18 +76,24 @@ class deepfind:
# TRAINING:
start = time.time()
for it in range(self.steps_per_epoch):
- batch_data, batch_target = self.generate_batch_direct_read(path_data, path_target, objlist_train, self.batch_size)
- loss_train = self.net.train_on_batch(batch_data, batch_target)
+ if self.flag_direct_read:
+ batch_data, batch_target = self.generate_batch_direct_read(path_data, path_target, self.batch_size, objlist_train)
+ else:
+ batch_data, batch_target = self.generate_batch_from_array(data_list, target_list, self.batch_size, objlist_train)
+ loss_train = net.train_on_batch(batch_data, batch_target)
print('epoch %d/%d - it %d/%d - loss: %0.3f - acc: %0.3f' % (e+1, self.epochs, it+1, self.steps_per_epoch, loss_train[0], loss_train[1]))
hist_loss_train.append(loss_train[0])
hist_acc_train.append( loss_train[1])
# VALIDATION (compute statistics to monitor training):
- batch_data_valid, batch_target_valid = self.generate_batch_direct_read(path_data, path_target, objlist_valid, self.Nvalid)
- loss_val = self.net.evaluate(batch_data_valid, batch_target_valid, verbose=0)
+ if self.flag_direct_read:
+ batch_data_valid, batch_target_valid = self.generate_batch_direct_read(path_data, path_target, self.batch_size, objlist_valid)
+ else:
+ batch_data_valid, batch_target_valid = self.generate_batch_from_array(data_list, target_list, self.batch_size, objlist_valid)
+ loss_val = net.evaluate(batch_data_valid, batch_target_valid, verbose=0)
- batch_pred = self.net.predict(batch_data_valid)
+ batch_pred = net.predict(batch_data_valid)
scores = precision_recall_fscore_support(batch_target_valid.argmax(axis=-1).flatten(), batch_pred.argmax(axis=-1).flatten(), average=None)
hist_loss_valid.append(loss_val[0])
@@ -87,16 +108,16 @@ class deepfind:
print('EPOCH %d/%d - valid loss: %0.3f - valid acc: %0.3f - %0.2fsec' % (e+1, self.epochs, loss_val[0], loss_val[1], end-start))
print('=============================================================')
- # Save training history:
+ # Save and plot training history:
history = {'loss':hist_loss_train, 'acc':hist_acc_train, 'val_loss':hist_loss_valid, 'val_acc':hist_acc_valid, 'val_f1':hist_f1, 'val_recall':hist_recall, 'val_precision':hist_precision}
- self.save_history(history)
+ core_utils.save_history(history)
+ core_utils.plot_history(history)
if (e+1)%10 == 0: # save weights every 10 epochs
- self.net.save('params_model_epoch'+str(e+1)+'.h5')
+ net.save('params_model_epoch'+str(e+1)+'.h5')
print "Model took %0.2f seconds to train"%np.sum(process_time)
- self.net.save('params_model_FINAL.h5')
-
+ net.save('params_model_FINAL.h5')
# This function generates training batches:
# - Data and target patches are sampled, in order to avoid loading whole tomograms.
@@ -108,47 +129,47 @@ class deepfind:
# This is usefull when a class is under-represented.
# /!\ TO DO: add a test to ignore objlist coordinate too close to border (using tomodim=h5data['dataset'].shape)
- def generate_batch_direct_read(self, path_data, path_target, objlist, batch_size):
- Nobj = objlist.shape[0]
+ def generate_batch_direct_read(self, path_data, path_target, batch_size, objlist=None):
p_in = np.int( np.floor(self.dim_in /2) )
batch_data = np.zeros((batch_size, self.dim_in, self.dim_in, self.dim_in, 1))
batch_target = np.zeros((batch_size, self.dim_in, self.dim_in, self.dim_in, self.Ncl))
-
- lblTAB = np.unique(objlist[:,0])
- Nbs = 100
-
- # Bootstrap data so that we have equal frequencies (1/Nbs) for all classes:
- bs_idx = []
- for l in lblTAB:
- bs_idx.append( np.random.choice(np.squeeze(np.asarray(np.nonzero(objlist[:,0]==l))), Nbs) )
- bs_idx = np.concatenate(bs_idx)
- for i in range(batch_size):
- # Choose random sample in training set:
- index = np.random.choice(bs_idx)
- tomoID = objlist[index,1]
+ # The batch is generated by randomly sampling data patches.
+ if self.flag_batch_bootstrap: # choose from bootstrapped objlist
+ pool = core_utils.get_bootstrap_idx(objlist,Nbs=batch_size)
+ else: # choose from whole objlist
+ pool = range(0,len(objlist))
- # Add random shift to coordinates:
- x = objlist[index,4] + np.random.choice(range(-self.Lrnd,self.Lrnd+1))
- y = objlist[index,3] + np.random.choice(range(-self.Lrnd,self.Lrnd+1))
- z = objlist[index,2] + np.random.choice(range(-self.Lrnd,self.Lrnd+1))
+
+ for i in range(batch_size):
+ # Choose random object in training set:
+ index = np.random.choice(pool)
+
+ tomoID = int( objlist[index].get('tomo_idx') )
+
+ h5file = h5py.File(path_data[tomoID], 'r')
+ tomodim = h5file['dataset'].shape # get tomo dimensions without loading the array
+ h5file.close()
+
+ x,y,z = core_utils.get_patch_position(tomodim, p_in, objlist[index], self.Lrnd)
# Load data and target patches:
- h5data = h5py.File(path_data[tomoID], 'r')
- patch_data = h5data['dataset'][x-p_in:x+p_in, y-p_in:y+p_in, z-p_in:z+p_in]
- patch_data = (patch_data - np.mean(patch_data)) / np.std(patch_data) # normalize
- h5data.close()
+ h5file = h5py.File(path_data[tomoID], 'r')
+ patch_data = h5file['dataset'][x-p_in:x+p_in, y-p_in:y+p_in, z-p_in:z+p_in]
+ h5file.close()
- h5target = h5py.File(path_target[tomoID], 'r')
- patch_batch = h5target['dataset'][x-p_in:x+p_in, y-p_in:y+p_in, z-p_in:z+p_in]
- #patch_batch[patch_batch==-1] = 0 # /!\ -1 labels from 'ignore mask' could generate trouble
- patch_batch_onehot = to_categorical(patch_batch, self.Ncl)
- h5target.close()
+ h5file = h5py.File(path_target[tomoID], 'r')
+ patch_target = h5file['dataset'][x-p_in:x+p_in, y-p_in:y+p_in, z-p_in:z+p_in]
+ h5file.close()
+
+ # Process the patches in order to be used by network:
+ patch_data = (patch_data - np.mean(patch_data)) / np.std(patch_data) # normalize
+ patch_target_onehot = to_categorical(patch_target, self.Ncl)
# Store into batch array:
batch_data[i,:,:,:,0] = patch_data
- batch_target[i] = patch_batch_onehot
+ batch_target[i] = patch_target_onehot
# Data augmentation (180degree rotation around tilt axis):
if np.random.uniform()<0.5:
@@ -157,38 +178,59 @@ class deepfind:
return batch_data, batch_target
- def save_history(self, history):
- h5trainhist = h5py.File('params_train_history.h5', 'w')
-
- # train and val loss & accuracy:
- dset = h5trainhist.create_dataset('acc', (len(history['acc']),))
- dset[:] = history['acc']
- dset = h5trainhist.create_dataset('loss', (len(history['loss']),))
- dset[:] = history['loss']
- dset = h5trainhist.create_dataset('val_acc', (len(history['val_acc']),))
- dset[:] = history['val_acc']
- dset = h5trainhist.create_dataset('val_loss', (len(history['val_loss']),))
- dset[:] = history['val_loss']
-
- # val precision, recall, F1:
- dset = h5trainhist.create_dataset('val_f1', np.shape(history['val_f1']))
- dset[:] = history['val_f1']
- dset = h5trainhist.create_dataset('val_precision', np.shape(history['val_precision']))
- dset[:] = history['val_precision']
- dset = h5trainhist.create_dataset('val_recall', np.shape(history['val_recall']))
- dset[:] = history['val_recall']
-
- h5trainhist.close()
- return
+ def generate_batch_from_array(self, data, target, batch_size, objlist=None):
+ p_in = np.int( np.floor(self.dim_in /2) )
+
+ batch_data = np.zeros((batch_size, self.dim_in, self.dim_in, self.dim_in, 1))
+ batch_target = np.zeros((batch_size, self.dim_in, self.dim_in, self.dim_in, self.Ncl))
+
+ # The batch is generated by randomly sampling data patches.
+ if self.flag_batch_bootstrap: # choose from bootstrapped objlist
+ pool = core_utils.get_bootstrap_idx(objlist,Nbs=batch_size)
+ else: # choose from whole objlist
+ pool = range(0,len(objlist))
+
+ for i in range(batch_size):
+ # choose random sample in training set:
+ index = np.random.choice(pool)
+
+ tomoID = int( objlist[index].get('tomo_idx') )
+
+ tomodim = data[tomoID].shape
+
+ sample_data = data[tomoID]
+ sample_target = target[tomoID]
+
+ dim = sample_data.shape
+
+ # Get patch position:
+ x,y,z = core_utils.get_patch_position(tomodim, p_in, objlist[index], self.Lrnd)
+
+ # Get patch:
+ patch_data = sample_data[x-p_in:x+p_in, y-p_in:y+p_in, z-p_in:z+p_in]
+ patch_batch = sample_target[x-p_in:x+p_in, y-p_in:y+p_in, z-p_in:z+p_in]
+
+ # Process the patches in order to be used by network:
+ patch_data = (patch_data - np.mean(patch_data)) / np.std(patch_data) # normalize
+ patch_batch_onehot = to_categorical(patch_batch, self.Ncl)
+
+ # Store into batch array:
+ batch_data[i,:,:,:,0] = patch_data
+ batch_target[i] = patch_batch_onehot
+
+ # Data augmentation (180degree rotation around tilt axis):
+ if np.random.uniform()<0.5:
+ batch_data[i] = np.rot90(batch_data[i] , k=2, axes=(0,2))
+ batch_target[i] = np.rot90(batch_target[i], k=2, axes=(0,2))
+
+ return batch_data, batch_target
def segment(self, dataArray, weights_path):
- self.net = models.my_model(self.P, self.Ncl)
- self.net.load_weights(weights_path)
- # Load data:
- #h5data = h5py.File(path_data, 'r')
- #dataArray = h5data['dataset'][:]
- #h5data.close()
+ # Build network:
+ net = models.my_model(self.P, self.Ncl)
+ net.load_weights(weights_path)
+
dataArray = (dataArray[:] - np.mean(dataArray[:])) / np.std(dataArray[:]) # normalize
dim = dataArray.shape
@@ -225,7 +267,7 @@ class deepfind:
print('Segmenting patch ' + str(patchCount) + ' / ' + str(Npatch) + ' ...' )
patch = dataArray[x-l:x+l, y-l:y+l, z-l:z+l]
patch = np.reshape(patch, (1,self.P,self.P,self.P,1)) # reshape for keras [batch,x,y,z,channel]
- pred = self.net.predict(patch, batch_size=1)
+ pred = net.predict(patch, batch_size=1)
predArray[x-lcrop:x+lcrop, y-lcrop:y+lcrop, z-lcrop:z+lcrop, :] = predArray[x-lcrop:x+lcrop, y-lcrop:y+lcrop, z-lcrop:z+lcrop, :] + pred[0, l-lcrop:l+lcrop,l-lcrop:l+lcrop,l-lcrop:l+lcrop, :]
normArray[x-lcrop:x+lcrop, y-lcrop:y+lcrop, z-lcrop:z+lcrop] = normArray[x-lcrop:x+lcrop, y-lcrop:y+lcrop, z-lcrop:z+lcrop] + np.ones((self.P-2*self.pcrop,self.P-2*self.pcrop,self.P-2*self.pcrop))
@@ -240,17 +282,21 @@ class deepfind:
print "Model took %0.2f seconds to predict"%(end - start)
return predArray # predArray is the array containing the scoremaps
-
- # Save scoremaps:
- #path, filename = os.path.split(path_data)
- #scoremap_file = filename[:-3]+'_scoremaps.h5'
- #h5scoremap = h5py.File(scoremap_file, 'w')
- #for cl in range(0,self.Ncl):
- # dset = h5scoremap.create_dataset('class'+str(cl), (dim[0], dim[1], dim[2]), dtype='float16' )
- # dset[:] = np.float16(predArray[:,:,:,cl])
- #h5scoremap.close()
- # For binning: skimage.measure.block_reduce(mat, (2,2), np.mean)
+ def segment_single_block(self, dataArray, weights_path):
+ # Build network:
+ net = models.my_model(self.P, self.Ncl)
+ net.load_weights(weights_path)
+
+ dim = dataArray.shape
+ dataArray = (dataArray[:] - np.mean(dataArray[:])) / np.std(dataArray[:]) # normalize
+ dataArray = np.reshape(dataArray, (1,dim[0],dim[1],dim[2],1)) # reshape for keras [batch,x,y,z,channel]
+
+ pred = net.predict(dataArray, batch_size=1)
+ predArray = pred[0,:,:,:,:]
+
+ return predArray
+
def cluster(self, labelmap, sizeThr, clustRadius):
Nclass = len(np.unique(labelmap)) - 1 # object classes only (background class not considered)
@@ -287,13 +333,7 @@ class deepfind:
labelcount[l] = np.size(np.nonzero( np.array(clustMember)==l+1 ))
winninglabel = np.argmax(labelcount)+1
- # Store cluster infos in array:
- #objlist[c,0] = clustSize
- #objlist[c,1] = centroid[0]
- #objlist[c,2] = centroid[1]
- #objlist[c,3] = centroid[2]
- #objlist[c,4] = winninglabel
-
+ # Store cluster infos in xml structure:
obj = etree.SubElement(objlist, 'object')
obj.set('cluster_size', str(clustSize))
obj.set('class_label' , str(winninglabel))
diff --git a/examples b/examples
new file mode 160000
index 0000000000000000000000000000000000000000..c7220091707ec653bece687be8d61e114269f0b3
--- /dev/null
+++ b/examples
@@ -0,0 +1 @@
+Subproject commit c7220091707ec653bece687be8d61e114269f0b3
diff --git a/utils.py b/utils.py
index aece2163df2bfa0dc345fe22bfc7bcdbfd5b9244..9f564705751fb5c223e30daa22417af1cbe43e61 100644
--- a/utils.py
+++ b/utils.py
@@ -2,9 +2,45 @@ import numpy as np
import h5py
from skimage.measure import block_reduce
from lxml import etree
+from copy import deepcopy
+from sklearn.metrics import pairwise_distances
+
+import matplotlib
+matplotlib.use('agg') # necessary else: AttributeError: 'NoneType' object has no attribute 'is_interactive'
+import matplotlib.pyplot as plt
#def bin_data:
+# Realizes quick visualization of a volume, by plotting its orthoslices, in the same fashion as the matlab function 'tom_volxyz' (TOM toolbox)
+# If volume type is int8, the function assumes that the volume is a labelmap, and hence plots in color scale.
+# Else, it assumes that the volume is tomographic data, and plots in gray scale.
+def plot_volume_orthoslices(vol, filename):
+ # Get central slices along each dimension:
+ dim = vol.shape
+ idx0 = np.round(dim[0]/2)
+ idx1 = np.round(dim[1]/2)
+ idx2 = np.round(dim[2]/2)
+
+ slice0 = vol[idx0,:,:]
+ slice1 = vol[:,idx1,:]
+ slice2 = vol[:,:,idx2]
+
+ # Build image containing orthoslices:
+ img_array = np.zeros((slice0.shape[0]+slice1.shape[0], slice0.shape[1]+slice1.shape[0]))
+ img_array[0:slice0.shape[0], 0:slice0.shape[1]] = slice0
+ img_array[slice0.shape[0]-1:-1, 0:slice0.shape[1]] = slice1
+ img_array[0:slice0.shape[0], slice0.shape[1]-1:-1] = np.flipud(np.rot90(slice2))
+
+ # Drop the plot:
+ fig = plt.figure(figsize=(10,10))
+ if vol.dtype==np.int8:
+ plt.imshow(img_array, cmap='CMRmap', vmin=np.min(vol), vmax=np.max(vol))
+ else:
+ mu = np.mean(vol) # Get mean and std of data for plot range:
+ sig = np.std(vol)
+ plt.imshow(img_array, cmap='gray', vmin=mu-5*sig, vmax=mu+5*sig)
+ fig.savefig(filename)
+
def write_objlist(objlist, filename):
tree = etree.ElementTree(objlist)
tree.write(filename, pretty_print=True)
@@ -13,6 +49,63 @@ def read_objlist(filename):
tree = etree.parse(filename)
objlist = tree.getroot()
return objlist
+
+def print_objlist(objlist):
+ print(etree.tostring(objlist))
+
+# /!\ for now this function does not know how to handle empty objlists
+def get_Ntp_from_objlist(objl_gt, objl_df, tol_pos_err):
+ # tolerated position error (in voxel)
+ Ngt = len(objl_gt)
+ Ndf = len(objl_df)
+ coords_gt = np.zeros((Ngt,3))
+ coords_df = np.zeros((Ndf,3))
+
+ for idx in range(0,Ngt):
+ coords_gt[idx,0] = objl_gt[idx].get('x')
+ coords_gt[idx,1] = objl_gt[idx].get('y')
+ coords_gt[idx,2] = objl_gt[idx].get('z')
+ for idx in range(0,Ndf):
+ coords_df[idx,0] = objl_df[idx].get('x')
+ coords_df[idx,1] = objl_df[idx].get('y')
+ coords_df[idx,2] = objl_df[idx].get('z')
+
+ # Get pairwise distance matrix:
+ D = pairwise_distances(coords_gt, coords_df, metric='euclidean')
+
+ # Get pairs that are closer than tol_pos_err:
+ D = D<=tol_pos_err
+
+ # A detected object is considered a true positive (TP) if it is closer than tol_pos_err to a ground truth object.
+ match_vector = np.sum(D,axis=1)
+ Ntp = np.sum(match_vector==1)
+ return Ntp
+
+def objlist_get_class(objlistIN, label):
+ N = len(objlistIN)
+ label_list = np.zeros((N,))
+ for idx in range(0,N):
+ label_list[idx] = objlistIN[idx].get('class_label')
+ idx_class = np.nonzero(label_list==label)
+ idx_class = idx_class[0]
+
+ objlistOUT = etree.Element('objlist')
+ for idx in range(0,len(idx_class)):
+ objlistOUT.append( deepcopy(objlistIN[idx_class[idx]]) ) # deepcopy is necessary, else the object is removed from objlIN when appended to objlOUT
+ return objlistOUT
+
+def objlist_above_thr(objlistIN, thr):
+ N = len(objlistIN)
+ clust_size_list = np.zeros((N,))
+ for idx in range(0,N):
+ clust_size_list[idx] = objlistIN[idx].get('cluster_size')
+ idx_thr = np.nonzero(clust_size_list>=thr)
+ idx_thr = idx_thr[0]
+
+ objlistOUT = etree.Element('objlist')
+ for idx in range(0,len(idx_thr)):
+ objlistOUT.append( deepcopy(objlistIN[idx_thr[idx]]) ) # deepcopy is necessary, else the object is removed from objlIN when appended to objlOUT
+ return objlistOUT
def bin_scoremaps(scoremaps):
dim = scoremaps.shape
@@ -28,34 +121,40 @@ def scoremaps2labelmap(scoremaps):
return labelmap
def load_h5array(filename): # rename to read_
- h5data = h5py.File(filename, 'r')
- dataArray = h5data['dataset'][:]
- h5data.close()
+ h5file = h5py.File(filename, 'r')
+ dataArray = h5file['dataset'][:]
+ h5file.close()
return dataArray
+def write_h5array(array, filename):
+ h5file = h5py.File(filename, 'w')
+ dset = h5file.create_dataset('dataset', array.shape, dtype='float16' )
+ dset[:] = np.float16(array)
+ h5file.close()
+
def load_scoremaps(filename): # rename to read_
- h5data = h5py.File(filename, 'r')
- datasetnames = h5data.keys()
+ h5file = h5py.File(filename, 'r')
+ datasetnames = h5file.keys()
Ncl = len(datasetnames)
- dim = h5data['class0'].shape
+ dim = h5file['class0'].shape
scoremaps = np.zeros((dim[0],dim[1],dim[2],Ncl))
for cl in range(0,Ncl):
- scoremaps[:,:,:,cl] = h5data['class'+str(cl)][:]
- h5data.close()
+ scoremaps[:,:,:,cl] = h5file['class'+str(cl)][:]
+ h5file.close()
return scoremaps
def write_scoremaps(scoremaps, filename):
- h5scoremap = h5py.File(filename, 'w')
+ h5file = h5py.File(filename, 'w')
dim = scoremaps.shape
Ncl = dim[3]
for cl in range(0,Ncl):
- dset = h5scoremap.create_dataset('class'+str(cl), (dim[0], dim[1], dim[2]), dtype='float16' )
+ dset = h5file.create_dataset('class'+str(cl), (dim[0], dim[1], dim[2]), dtype='float16' )
dset[:] = np.float16(scoremaps[:,:,:,cl])
- h5scoremap.close()
+ h5file.close()
def write_labelmap(labelmap, filename):
dim = labelmap.shape
- h5lblmap = h5py.File(filename, 'w')
- dset = h5lblmap.create_dataset('dataset', (dim[0],dim[1],dim[2]), dtype='int8' )
+ h5file = h5py.File(filename, 'w')
+ dset = h5file.create_dataset('dataset', (dim[0],dim[1],dim[2]), dtype='int8' )
dset[:] = np.int8(labelmap)
- h5lblmap.close()
\ No newline at end of file
+ h5file.close()
\ No newline at end of file