import sys
if '..' not in sys.path:
    sys.path.append('..')

%load_ext autoreload %autoreload 2

import tensorflow as tf
from ultrayolo import YoloV3, datasets
from ultrayolo.helpers import draw
from pathlib import Path
import numpy as np
import logging

from matplotlib import patches
import matplotlib.pyplot as plt

Predict using a custom pretrained model¶

The classes contained in the dataset

classes_dict = datasets.load_classes('./mini_classes.txt', True)
target_shape = (512, 512, 3)
max_objects = 100
num_classes = len(classes_dict)
print(f'number of classes {num_classes}')
classes_dict

number of classes 3

{0: 'bear', 1: 'toaster', 2: 'hair drier'}

anchors = datasets.load_anchors('./mini_anchors.txt')
anchors

array([[ 79.2,  77. ],
       [100.1,  90.2],
       [104.5,  89.1],
       [112.2,  99. ],
       [178.2, 193.6],
       [207.9, 199.1],
       [227.7, 205.7],
       [235.4, 204.6],
       [237.6, 202.4]])

model = YoloV3(target_shape, max_objects, anchors=anchors,
               num_classes=num_classes, score_threshold=0.5, iou_threshold=0.5,
               training=False, backbone='DenseNet121')

tf.keras.utils.plot_model(model.model, show_shapes=True)

Load the weights¶

load a custom model from here

w_path = Path('weights_val.118-7.478.h5')
model.load_weights(w_path)

54385 MainThread loading checkpoint from /Users/fumarolaf/git/ultrayolo/notebooks/weights_val.118-7.478.h5

Predict¶

we predict the objects using an image from the web. You can try with your.

Bear¶

img = datasets.open_image('https://upload.wikimedia.org/wikipedia/commons/thumb/5/5d/Kamchatka_Brown_Bear_near_Dvuhyurtochnoe_on_2015-07-23.jpg/1200px-Kamchatka_Brown_Bear_near_Dvuhyurtochnoe_on_2015-07-23.jpg')
img_pad = datasets.pad_to_fixed_size(img, target_shape)
img_resized = datasets.resize(img, target_shape)
#preprocess the image
x = np.divide(img_pad, 255.)
x = np.expand_dims(x, 0)
plt.imshow(x[0])
x.shape

(1, 512, 512, 3)

../_images/2_predict_custom_dataset_11_1.png

Perform the prediction¶

boxes, scores, classes, sel = model.predict(x)
print(f'found {sel[0]} objects')

found 1 objects

Show the image with the discovered objects

ax = draw.show_img(img_resized, figsize=(8,8))
for i,b in enumerate(boxes[0,:sel[0]]):
    draw.rect(ax, b, color='#9cff1d')
    name_score = f'{classes_dict[classes[0, i]]} {str(round(scores[0,i],2))}'
    draw.text(ax, b[:2], name_score, sz=12)
    print(classes_dict[classes[0, i]], scores[0,i])

plt.show()

bear 0.9999901

../_images/2_predict_custom_dataset_15_1.png

Perform a prediction for the class Hair Drier¶

img = datasets.open_image('https://cdn.mos.cms.futurecdn.net/UA2XKzB6zdzm486hC4TXPf.jpg')
# img = datasets.open_image('https://reviewed-com-res.cloudinary.com/image/fetch/s--IAWIW5ff--/b_white,c_limit,cs_srgb,f_auto,fl_progressive.strip_profile,g_center,q_auto,w_1200/https://reviewed-production.s3.amazonaws.com/1521023219800/toaster-newhhero.jpg')
img_pad = datasets.pad_to_fixed_size(img, target_shape)
img_resized = datasets.resize(img, target_shape)
#preprocess the image
x = np.divide(img_pad, 255.)
x = np.expand_dims(x, 0)
plt.imshow(x[0])
x.shape

(1, 512, 512, 3)

../_images/2_predict_custom_dataset_17_1.png

boxes, scores, classes, sel = model.predict(x)
print(f'found {sel[0]} objects')

found 1 objects

ax = draw.show_img(img_resized, figsize=(8,8))
for i,b in enumerate(boxes[0,:sel[0]]):
    draw.rect(ax, b, color='#9cff1d')
    name_score = f'{classes_dict[classes[0, i]]} {str(round(scores[0,i],2))}'
    draw.text(ax, b[:2], name_score, sz=12)
    print(classes_dict[classes[0, i]], scores[0,i])

plt.show()

hair drier 0.9821011

../_images/2_predict_custom_dataset_19_1.png

Perform a prediction for the class Toaster¶

img = datasets.open_image('https://reviewed-com-res.cloudinary.com/image/fetch/s--IAWIW5ff--/b_white,c_limit,cs_srgb,f_auto,fl_progressive.strip_profile,g_center,q_auto,w_1200/https://reviewed-production.s3.amazonaws.com/1521023219800/toaster-newhhero.jpg')
img_pad = datasets.pad_to_fixed_size(img, target_shape)
img_resized = datasets.resize(img, target_shape)
#preprocess the image
x = np.divide(img_pad, 255.)
x = np.expand_dims(x, 0)
plt.imshow(x[0])
x.shape

(1, 512, 512, 3)

../_images/2_predict_custom_dataset_21_1.png

boxes, scores, classes, sel = model.predict(x)
print(f'found {sel[0]} objects')

found 2 objects

ax = draw.show_img(img_resized, figsize=(8,8))
for i,b in enumerate(boxes[0,:sel[0]]):
    draw.rect(ax, b, color='#9cff1d')
    name_score = f'{classes_dict[classes[0, i]]} {str(round(scores[0,i],2))}'
    draw.text(ax, b[:2], name_score, sz=12)
    print(classes_dict[classes[0, i]], scores[0,i])

plt.show()

toaster 0.89197004
toaster 0.5734626

../_images/2_predict_custom_dataset_23_1.png