diff --git a/Application/learn_static.py b/Application/learn_static.py
index 6e25bc9eb6f1eb7f0f46a458374f91c2be418d62..dbadc254ce4d2864616dbb0f12963dece7e39e5e 100644
--- a/Application/learn_static.py
+++ b/Application/learn_static.py
@@ -4,9 +4,6 @@
#except:
# pass
-import matplotlib
-matplotlib.use("Agg")
-
from keras.layers import Dense, Dropout, Activation, Cropping2D, Flatten, Convolution1D, Convolution2D, MaxPooling1D,MaxPooling2D
from keras.models import Sequential
import keras
@@ -16,7 +13,6 @@ from keras.preprocessing.image import ImageDataGenerator
from keras.utils import np_utils
import TRex
import utils
-import matplotlib.pyplot as plt
import shutil
@@ -163,31 +159,6 @@ def reinitialize_network():
metrics=['accuracy'])
model.summary(print_fn=TRex.log)
-
-def plot_medians(title, train_X, train_Y, classes):
- lengths = {}
- f, axes = utils.axes_for_array(len(classes))
- for ax, c in zip(axes, classes):
- index = train_Y == c
- ax.set_title(str(c)+": "+str(len(train_X[index])))
- lengths[c] = len(train_X[index])
-
- subdata = np.copy(train_X[index])
-
- if len(subdata) == 0:
- #TRex.log("skipping class "+str(c)+" because its empty")
- continue
-
- img = np.max(subdata.astype(float), axis=0)[:,:,0]
- img = np.stack((img, )*3, axis=2).astype(np.uint8)
- ax.imshow(img)
-
- del subdata
-
- TRex.log("#samples per class: "+str(lengths))
- #TRex.imshow(title, utils.figure_as_image())
- #utils.show_figure(title, output_path="")
-
class UserCancelException(Exception):
"""Raised when user clicks cancel"""
pass
@@ -249,105 +220,11 @@ class ValidationCallback(keras.callbacks.Callback):
distance = np.abs(Y - zeros).sum(axis=1)
- if do_plot:
- f, [ax, bx] = plt.subplots(1, 2, figsize=(15, 5))
- ax.scatter(np.arange(len(distance)), distance, label="d", s = 1, c = np.ones(len(distance)), alpha = 0.5)
- pts = np.argmax(Y, axis=1)
- color = Y.sum(axis=1)
-
- for j in range(len(color)):
- color[j] = color[j] * len(Y[j][Y[j] > 0.01])# / len(obj.classes)
-
- pts = np.concatenate((pts, [0, ]), axis=0)
- color = np.concatenate((color, [10, ]), axis=0)
-
- im = bx.scatter(np.arange(len(pts)), pts, label="Y", s = 1, c = color, alpha = 0.5)
- bx.set_yticks(np.arange(len(obj.classes_subset)))
- bx.set_ylim(-1, len(obj.classes_subset))
- #TRex.log(color.max(), color.min())
- plt.colorbar(im)
- plt.suptitle(c)
- ax.legend()
- bx.legend()
- plt.show()
- plt.close(f)
-
result[i, 0] = np.median(np.argmax(Y, axis=1))
result[i, 1] = distance.std()
result[i, 2] = np.median(Y.sum(axis=1)) #distance.mean()
result[i, 3] = (np.argmax(Y, axis=1) == i).sum() / len(Y)
return result, predictions
-
- def plot_confusion_matrix(self, cm, classes,
- normalize,
- title,
- cmap, ax = None):
- """
- This function prints and plots the confusion matrix.
- Normalization can be applied by setting `normalize=True`.
- """
- do_show = False
-
- if normalize:
- cm = cm.astype('float') / (1+cm.sum(axis=1)[:, np.newaxis])
-
- if ax == None:
- f, ax = plt.subplots(1,1)
- do_show = True
-
- im = ax.imshow(cm, interpolation='nearest', cmap=cmap)
- ax.set_title(title)
- cbar = plt.colorbar(im, ax = ax)
- cbar_ticks = np.linspace(0., 1., num=10, endpoint=True)
- cbar.set_ticks(cbar_ticks)
-
- tick_marks = np.arange(len(classes))
- ax.set_xticks(tick_marks)
- ax.set_xticklabels([str(c) for c in classes])
- ax.set_yticks(tick_marks)
- ax.set_yticklabels([str(c) for c in classes])
-
- fmt = '.2f' if normalize else 'd'
- thresh = cm.max() / 2.
-
- ax.set_ylabel('True label')
- ax.set_xlabel('Predicted label')
-
- #if do_show:
- # utils.show_figure("confusion matrix", output_path="")
- #plt.tight_layout()
-
- def plot_confusion(self, X_test, Y_test, predictions = None, title = "", ax = None, cmap = None):
- classes = self.classes
- #TRex.log("shapes:", X_test.shape, Y_test.shape)
-
- y = None
- if type(predictions) == type(None):
- y = self.model.predict(X_test)
- else:
- y = np.concatenate(predictions, axis=0)
- assert len(y) == len(X_test)
-
- result = np.zeros_like(np.ndarray(shape=(len(classes),len(classes))), dtype=float)
-
- wrong = 0
- correct = 0
-
- for i in range(len(X_test)):
- a = np.argmax(y[i,0:len(classes)])
- b = np.argmax(Y_test[i,0:len(classes)])
-
- if a == b:
- correct = correct+1
- else:
- wrong = wrong + 1
-
- result[a, b] = result[a, b] + 1
- #result_x.append(a)
- #result_y.append(b)
-
- title = title + " (accuracy "+str(1-wrong/float(correct+wrong))+")"
- self.plot_confusion_matrix(result, [str(classes[i]) for i in classes], normalize=True, title = title, ax = ax, cmap=cmap)
def update_status(self, print_out = False, logs = {}):
description = "epoch "+str(min(self.epoch+1, self.epochs))+"/"+str(self.epochs)
@@ -375,29 +252,11 @@ class ValidationCallback(keras.callbacks.Callback):
update_work_percent(min(epoch + 1, self.epochs) / self.epochs)
- #fig = plt.figure(figsize=(25,10))
- #gs = fig.add_gridspec(2,2, width_ratios=[1,3])
- #ax = fig.add_subplot(gs[0,0])
- #dx = fig.add_subplot(gs[0,1])
- #bx = fig.add_subplot(gs[1,:])
-
self.epoch = min(epoch + 1, self.epochs)
if np.shape(self.X_test)[-1] > 0:
result, predictions = self.plot_comparison_raw(do_plot = False, length = -1)
- #dx.plot(np.arange(len(result)), result[:, 3], label="% correct")
- #dx.scatter(np.arange(len(result)), result[:, 3])
- #for i in np.arange(len(result)):
- # dx.annotate(str(classes[i])+": "+str(int(result[:, 3][i]*100))+"%", (i + 0.01, result[:, 3][i] + 0.01))
- #dx.set_ylim(0, 1)
- #dx.set_xticks(np.arange(len(result)))
- #dx.set_xticklabels([str(classes[i]) for i in classes])
- #plt.title("epoch "+str(epoch)+" range "+str(np.array(self.settings["global_segment"])))
- #utils.show_figure("classes recognition")
-
- #self.plot_confusion(np.concatenate(self.X_test, axis=0), np.concatenate(self.Y_test, axis=0), predictions = predictions, ax = ax)
-
for i in range(0, len(result[:, 3])):
if not i in self.per_class_accuracy:
self.per_class_accuracy[i] = []
@@ -406,12 +265,6 @@ class ValidationCallback(keras.callbacks.Callback):
self.worst_values.append(np.min(result[:, 3]))
set_per_class_accuracy(result[:, 3].astype(np.float))
-
- #bx.fill_between(np.arange(len(self.worst_values)), self.worst_values, self.mean_values, where=np.array(self.worst_values) <= np.array(self.mean_values), facecolor="red", interpolate = True, alpha= 0.25)
-
- #bx.plot(np.arange(len(self.worst_values)), self.worst_values, label="min", color="blue")
- #bx.plot(np.arange(len(self.mean_values)), self.mean_values, label="mean")
-
worst_acc_per_class = np.min(result[:, 3])
else:
result = None
@@ -424,7 +277,6 @@ class ValidationCallback(keras.callbacks.Callback):
self.uniquenesses.append(unique)
set_uniqueness_history(self.uniquenesses)
- #bx.plot(np.arange(len(self.uniquenesses)), self.uniquenesses, label="uniqueness", color="red")
# check whether our worst value improved, but only use it if it wasnt the first epoch
if unique > self.best_result["unique"] and (self.compare_acc <= 0 or unique >= self.compare_acc**2):
@@ -540,36 +392,7 @@ class ValidationCallback(keras.callbacks.Callback):
set_stop_reason("overfitting")
TRex.log("[STOP]Â overfitting. stopping with loss diffs: "+str(change))
self.model.stop_training = True
-
- #if worst_acc_per_class >= min_acceptable_value and (worst_acc_per_class >= 0.99 or (len(self.worst_values) >= min_number_iterations and self.worst_values[-1] >= min_acceptable_value)):
- # TRex.log("[STOP] stopping training with minimum validation accuracy of "+str(worst_acc_per_class)+" per class")
- # self.model.stop_training = True
-#else:
-# TRex.log("\tcontinuing with worst accuracy of "+str(worst_value))
-
- #if len(self.better_values) > 0:
- # bx.scatter(np.array(self.better_values).T[0], np.array(self.better_values).T[1], label="improvements", color="red")
- # for i,j in zip(np.array(self.better_values).T[0], np.array(self.better_values).T[1]):
- # bx.annotate(str(int(j * 1000) / 10.0)+"%",xy=(i,j), xytext=(1,5), textcoords='offset points')
- #if len(self.much_better_values) > 0:
- # bx.scatter(np.array(self.much_better_values).T[0], np.array(self.much_better_values).T[1], label="eligble", color="green")
-
- #if self.compare_acc > 0:
- # bx.axhline(y=self.compare_acc, label="previous best", color="gray", linestyle="--")
- # bx.axhline(y=self.compare_acc**2, label="previous accept.", color="gray", linestyle="--")
- #bx.set_ylim(-0.05, 1.05)
-
- #if len(self.uniquenesses) < 20:
- # bx.set_xticks(np.arange(len(self.uniquenesses)))
- #else:
- # bx.set_xticks(np.arange(len(self.uniquenesses), 5))
- #bx.set_xlabel("epochs")
- #bx.set_ylabel("accuracy")
- #plt.legend()
- #plt.close(plt.gcf())
- #im = utils.figure_as_image()
- #TRex.imshow("training overview", im)
- #utils.show_figure(self.filename + self.prefix + " " + str(epoch) + " training overview", output_path=self.settings["output_path"], im = im)
+
self.update_status(True, logs=logs)
self.batches = 0
return unique
@@ -675,7 +498,6 @@ def start_learning():
for i, c in zip(np.arange(len(classes)), classes):
mi = max(mi, len(Y_train[np.argmax(Y_train, axis=1) == c]))
- plot_medians(filename + output_prefix + " dataset", X_train, np.argmax(Y_train, axis=1), np.array(classes, dtype=int))
per_epoch = max(settings["min_iterations"], int(len(X_train) // batch_size ) * 0.5 ) #* 2.0) # i am using augmentation
settings["per_epoch"] = per_epoch
TRex.log(str(settings))
diff --git a/Application/src/tracker/CMakeLists.txt b/Application/src/tracker/CMakeLists.txt
index 536c97c79f1369443f4430d3a351f90c8dcd7255..568f659d8f931da15b4f1e1f063642cf67bd9fd6 100644
--- a/Application/src/tracker/CMakeLists.txt
+++ b/Application/src/tracker/CMakeLists.txt
@@ -122,7 +122,6 @@ set(RESOURCE_COPY_FILES
)
set(RESOURCE_COPY_SINGLE_FILES
${CMAKE_SOURCE_DIR}/default.settings
- ${CMAKE_SOURCE_DIR}/utils.py
${CMAKE_SOURCE_DIR}/learn_static.py
)
@@ -228,4 +227,4 @@ install(TARGETS ${targets}
install(DIRECTORY ${CMAKE_SOURCE_DIR}/src/fonts DESTINATION ${CMAKE_INSTALL_PREFIX}/usr/share/trex)
install(DIRECTORY ${CMAKE_SOURCE_DIR}/src/html DESTINATION ${CMAKE_INSTALL_PREFIX}/usr/share/trex)
install(DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/gfx DESTINATION ${CMAKE_INSTALL_PREFIX}/usr/share/trex)
-install(FILES ${CMAKE_SOURCE_DIR}/default.settings ${CMAKE_SOURCE_DIR}/utils.py ${CMAKE_SOURCE_DIR}/learn_static.py DESTINATION ${CMAKE_INSTALL_PREFIX}/usr/share/trex)
+install(FILES ${CMAKE_SOURCE_DIR}/default.settings ${CMAKE_SOURCE_DIR}/learn_static.py DESTINATION ${CMAKE_INSTALL_PREFIX}/usr/share/trex)
diff --git a/Application/src/tracker/python/GPURecognition.cpp b/Application/src/tracker/python/GPURecognition.cpp
index 9042f2d4ea211076d532a92445e179c6b7bcfa34..db27da251f9a146cc8e9bc17d5fcdba61c4c3db1 100644
--- a/Application/src/tracker/python/GPURecognition.cpp
+++ b/Application/src/tracker/python/GPURecognition.cpp
@@ -239,7 +239,7 @@ PYBIND11_EMBEDDED_MODULE(TRex, m) {
namespace track {
namespace py = pybind11;
std::shared_ptr<py::scoped_interpreter> guard = nullptr;
- pybind11::module utils, numpy, TRex, plt, matplotlib, _main;
+ pybind11::module numpy, TRex, _main;
pybind11::dict* _locals = nullptr;
std::mutex module_mutex;
@@ -424,24 +424,7 @@ namespace track {
"\t\tfound = False\nprint('setting version',sys.version,found,physical)\n" \
"set_version(sys.version, found, physical)\n");
- matplotlib = py::module::import("matplotlib");
-
- //#ifdef __linux__
- matplotlib.attr("use")("Agg");
- //#endif
-
- plt = py::module::import("matplotlib.pyplot");
- //main.attr("array") = std::vector<float>{1,2,3,4};
- //plt.attr("plot")(main.attr("array"));
- //plt.attr("show")();
-
numpy = _main.import("numpy");
- try {
- track::utils = _main.import("utils");
- } catch(py::error_already_set &e) {
- Debug("Python runtime error: '%s'", e.what());
- e.restore();
- }
TRex = _main.import("TRex");
_locals = new py::dict("model"_a="None");
@@ -489,7 +472,7 @@ namespace track {
}
try {
- track::utils = track::numpy = track::TRex = track::plt = track::matplotlib = _main = pybind11::module();
+ track::numpy = track::TRex = _main = pybind11::module();
{
std::lock_guard<std::mutex> guard(module_mutex);
diff --git a/Application/src/tracker/python/check_python.cpp b/Application/src/tracker/python/check_python.cpp
index 6b60406af55d7ba4248092ddad609e1d36a461ab..063562d7484db64f5d2cc3060bf99cf548531720 100644
--- a/Application/src/tracker/python/check_python.cpp
+++ b/Application/src/tracker/python/check_python.cpp
@@ -21,10 +21,6 @@ int main(int argc, char**argv) {
printf("loading keras...\n");
#endif
importlib.attr("find_loader")("keras").attr("name");
-#ifndef NDEBUG
- printf("loading matplotlib...\n");
-#endif
- importlib.attr("find_loader")("matplotlib").attr("name");
#ifndef NDEBUG
printf("success.\n");
#endif
diff --git a/Application/src/tracker/tracking/Accumulation.cpp b/Application/src/tracker/tracking/Accumulation.cpp
index 95f444059eb64eeea10780d71d789b8f0a70d320..eafcde5eacd78f1184ff822bd98a15b21d4374fd 100644
--- a/Application/src/tracker/tracking/Accumulation.cpp
+++ b/Application/src/tracker/tracking/Accumulation.cpp
@@ -101,7 +101,7 @@ void Accumulation::setup() {
try {
Recognition::check_learning_module();
} catch(const std::future_error& error) {
- SOFT_EXCEPTION("Checking the learning module failed ('%s'). Most likely one of the required libraries is missing from the current python environment (check for keras, tensorflow, matplotlib and pillow).", error.what());
+ SOFT_EXCEPTION("Checking the learning module failed ('%s'). Most likely one of the required libraries is missing from the current python environment (check for keras and tensorflow).", error.what());
}
}
diff --git a/Application/utils.py b/Application/utils.py
deleted file mode 100644
index f2cc3dd982a9e32e6f16c11c3f718e91221b1903..0000000000000000000000000000000000000000
--- a/Application/utils.py
+++ /dev/null
@@ -1,71 +0,0 @@
-def axes_for_array(length):
- import numpy as np
- import matplotlib.pyplot as plt
-
- rest = 0
- cols = min(16, length)
- if length % cols > 0:
- rest = 1
-
- figsize = (cols * 5, 5 * (length // cols + rest))
- fig, axes = plt.subplots(length // cols + rest, cols, figsize=figsize)
- print(figsize)
-
- axes = np.array(axes).flatten()
- for ax in axes:
- ax.axis('off')
- axes = axes[:length]
-
- return fig, axes
-
-def figure_as_image():
- import io
- from PIL import Image
- import matplotlib.pyplot as plt
- import numpy as np
- import os
-
- plt.gcf().set_tight_layout(True)
- #plt.gcf().patch.set_facecolor('black')
- plt.gcf().patch.set_alpha(0.75)
- buf = io.BytesIO()
- plt.savefig(buf, format='png', transparent=True, facecolor=plt.gcf().get_facecolor(), edgecolor='none')
- buf.seek(0)
- im = Image.open(buf)
- im = np.array(im).astype(np.uint8)
- buf.close()
- plt.close(plt.gcf())
-
- return im
-
-def show_figure(title="plot", output_path="", im = None):
- import TRex
- from PIL import Image
- import numpy as np
- import os
-
- if type(im) == type(None):
- im = Image.fromarray(figure_as_image())
- else:
- im = Image.fromarray(im)
- try:
- path = "/var/www/example.com/html/"+title.replace(" ", "_").replace("/", "-")+".png"
- im.save(path, "PNG")
- TRex.log("saved as"+str(path))
- except Exception as e:
- TRex.warn(str(e))
-
- if len(output_path) > 0:
- try:
- if not output_path.endswith(os.sep):
- output_path = output_path + os.sep
- path = output_path+title.replace(" ", "_").replace("/", "-")+".png"
- im.save(path, "PNG")
- TRex.log("saved as"+str(path))
- except Exception as e:
- TRex.warn(str(e))
-
- im = np.array(im).astype(np.uint8)
-
- #TRex.imshow(title, im)
-
diff --git a/conda/meta.yaml b/conda/meta.yaml
index f8e1a2ae63ff3e54f744d960ea353220e98ec0ef..1afefd101d12ce2668ed621f3670e016e59d2ca3 100644
--- a/conda/meta.yaml
+++ b/conda/meta.yaml
@@ -67,8 +67,6 @@ requirements:
- tensorflow-gpu ==1.13.* [not osx]
- tensorflow ==1.13.* [osx]
- keras
- - matplotlib
- - pillow
- ffmpeg [win]
- python [win]
- ffmpeg ==4.0 [not win]