def draw_each_class_ap(epochs, all_losses, classes, the_class):
the_color = ["r", "g", "b", "black"]
Y = [[] for i in range(8)]
for i in range(epochs):
tmp = 0
for j in range(3):
Y[j].append(all_losses[i][j][the_class])
tmp += all_losses[i][j][the_class]
Y[3].append(tmp)
x = [i for i in range(epochs)]
x1 = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]
the_name = []
for i in epoch_ap:
L, name = deal_table(i)
the_name = name
for j in range(4):
Y[4 + j].append(L[(j + 1) % 4][the_class])
# for i in range(4):
# tmp = Y[i][0]
# for j in range(len(Y[i])):
# if tmp == 0.0:
# Y[i][j] = 0
# else:
# Y[i][j] = 1 - (tmp - Y[i][j]) / tmp
# for i in range(4):
# print(Y[i])
fig = plt.figure()
scale_name = ["Small", "Middle", "Large", "ALL"]
ax1 = fig.add_subplot(111)
for i in range(1):
# for i in range(4):
ax1.plot(x,
Y[i],
label=scale_name[i] + "_loss",
color=the_color[i],
linestyle="--")
ax1.set_ylabel('the rest of loss')
ax1.set_title(classes[the_class])
ax2 = ax1.twinx() # this is the important function
for i in range(4, 8):
ax2.plot(x1,
Y[i],
label=scale_name[i - 4] + "_ap",
color=the_color[i - 4],
linestyle="-")
ax2.set_ylabel('the AP')
ax2.set_xlabel('Epoch')
plt.legend()
# print(classes[the_class])
# plt.savefig("pic/classes/" + classes[the_class] + ".png")
# plt.savefig("pic/classes_change/" + classes[the_class] + ".png")
plt.savefig("pic/small_losses/" + classes[the_class] + ".png")
def draw_20_ap(epochs, classes):
x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]
Y = [[] for i in range(len(classes) + 1)]
the_name = []
for i in epoch_ap:
L, name = deal_table(i)
the_name = name
for j in range(len(the_name)):
Y[j].append(L[0][j])
for i in range(len(the_name)):
plt.plot(x, Y[i], label=the_name[i], color=all_color[i])
plt.xlabel("Epoch")
plt.ylabel("AP")
plt.title("20 classes AP")
plt.legend()
plt.show()
def draw_scale_ap(epochs, classes, scale):
x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]
Y = [[] for i in range(len(classes) + 1)]
the_name = []
for i in epoch_ap:
L, name = deal_table(i)
the_name = name
for j in range(len(the_name)):
Y[j].append(L[scale][j])
# print(len(x))
# print(len(Y[0]))
for i in range(len(the_name)):
plt.plot(x, Y[i], label=the_name[i], color=all_color[i])
plt.xlabel("Epoch")
plt.ylabel("AP")
scale_name = ["Small", "Middle", "Large"]
plt.title(scale_name[scale - 1] + " AP")
plt.legend()
plt.show()
def each_loss_num_ap(epochs, classes):
# f = open("each_loss_num_ap.csv","w")
f = open("each_loss_num_nosort_ap.csv", "w")
f.write("类别,尺度,loss")
for i in range(epochs):
f.write(",Epoch" + str(i))
f.write("\n")
result = np.zeros((3, 20))
result = [list(i) for i in list(result)]
for i in range(3):
for j in range(20):
result[i][j] = [0.0 for i in range(epochs)]
import copy
# result_num = np.zeros((3,20))
result_num = copy.deepcopy(result)
each_iter = []
for epoch in range(epochs):
for j in range(2501):
bbox, clss_loss_num, cls_loss, reg_loss_num, reg_loss, all_loss = read_file(
path + str(epoch) + os.sep + filename_part + str(j) + ".txt")
scale = []
each_classes = []
for k in bbox:
size = np.abs(k[2] - k[0]) * np.abs(k[3] - k[1])
if size <= 1024:
scale.append(0)
elif size <= 9216:
scale.append(1)
else:
scale.append(2)
each_classes.append(classes[int(k[4])])
result_num[scale[-1]][int(k[4])][epoch] += 1
for k in range(len(cls_loss)):
result[scale[cls_loss[k][1]]][int(
bbox[cls_loss[k][1]][4])][epoch] += cls_loss[k][2][int(
bbox[cls_loss[k][1]][4])] + sum(reg_loss[k][1])
scale_name = ["S", "M", "L"]
for i in range(len(classes)):
for j in range(3):
the_num = 0
if sum(result_num[j][i]) != 0:
the_num = sum(result[j][i]) / sum(result_num[j][i])
write_content = [
classes[i] + "," + scale_name[j],
sum(result_num[j][i]), result_num[j][i], result[j][i], i, j
]
each_iter.append(write_content)
Y = []
the_name = []
for i in epoch_ap:
L, name = deal_table(i)
the_name = name
Y.append(L)
# each_iter.sort(key = lambda x :x[1])
# for i in each_iter[::-1]:
for i in each_iter:
f.write(i[0])
f.write(",AP:,")
for j in range(epochs - 1):
f.write("," + str(Y[j][i[-1] + 1][i[-2]]))
f.write("\n")
f.write(",,loss:")
for j in range(epochs):
f.write("," + str(i[-3][j]))
f.write("\n")
f.write(",,loss变化:")
for j in range(epochs):
if i[-3][0] == 0:
f.write("," + str(0))
else:
f.write("," + str(1 - (i[-3][0] - i[-3][j]) / i[-3][0]))
f.write("\n")
f.write(",,mean:")
for j in range(epochs):
if i[-4][j] != 0:
f.write("," + str(i[-3][j] / i[-4][j]))
else:
f.write("," + str(0))
f.write("\n")
f.close()
def all_bbox_num_sort(epochs, classes):
f = open("16bbox_num_nosort_loss.csv", "w")
# f = open("16bbox_num_sort_loss.csv","w")
# f = open("16bbox_sort_loss.csv","w")
f.write("类别,尺度,Bboxs数量,样本数量,loss,平均,剩余loss")
for i in range(1, epochs):
f.write(",Epoch" + str(i))
f.write("\n")
result = np.zeros((3, 20))
result_num = np.zeros((3, 20))
each_iter = []
sample_num = np.zeros((3, 20))
start_loss = np.zeros((3, 20))
tmp_loss = np.zeros((3, 20))
end_loss = np.zeros((3, 20))
for epoch in range(epochs):
for j in range(2501):
bbox, clss_loss_num, cls_loss, reg_loss_num, reg_loss, all_loss = read_file(
path + str(epoch) + os.sep + filename_part + str(j) + ".txt")
scale = []
each_classes = []
part_sample_num = np.zeros((3, 20))
for k in bbox:
size = np.abs(k[2] - k[0]) * np.abs(k[3] - k[1])
if size <= 1024:
scale.append(0)
elif size <= 9216:
scale.append(1)
else:
scale.append(2)
each_classes.append(classes[int(k[4])])
result_num[scale[-1]][int(k[4])] += 1
part_sample_num[scale[-1]][int(k[4])] = 1
sample_num += part_sample_num
for k in range(len(cls_loss)):
result[scale[cls_loss[k][1]]][int(
bbox[cls_loss[k][1]][4])] += cls_loss[k][2][int(
bbox[cls_loss[k][1]][4])] + sum(reg_loss[k][1])
if epoch == 0:
start_loss[scale[cls_loss[k][1]]][int(
bbox[cls_loss[k][1]][4])] += cls_loss[k][2][int(
bbox[cls_loss[k][1]][4])] + sum(reg_loss[k][1])
if epoch == epochs - 1:
end_loss[scale[cls_loss[k][1]]][int(
bbox[cls_loss[k][1]][4])] += cls_loss[k][2][int(
bbox[cls_loss[k][1]][4])] + sum(reg_loss[k][1])
# print(end_loss)
# print(start_loss)
the_pos = (start_loss == 0)
end_loss = 1 - (start_loss - end_loss) / start_loss
end_loss[the_pos] = 0.0
scale_name = ["S", "M", "L"]
for i in range(len(classes)):
for j in range(3):
the_num = 0
if sample_num[j][i] != 0:
the_num = result[j][i] / sample_num[j][i]
write_content = [
classes[i] + "," + scale_name[j] + "," +
str(result_num[j][i]) + "," + str(sample_num[j][i]) + "," +
str(result[j][i]) + "," + str(the_num) + "," +
str(end_loss[j][i]), result_num[j][i], result[j][i], i, j
]
each_iter.append(write_content)
Y = []
the_name = []
for i in epoch_ap:
L, name = deal_table(i)
the_name = name
Y.append(L)
# each_iter.sort(key = lambda x :x[1])
# for i in each_iter[::-1]:
for i in each_iter:
f.write(i[0])
for j in range(epochs - 1):
f.write("," + str(Y[j][i[-1] + 1][i[-2]]))
f.write("\n")
f.close()