pull from remote
Browse files- gcvit/utils/gradcam.py +71 -0
gcvit/utils/gradcam.py
CHANGED
@@ -1,3 +1,4 @@
|
|
|
|
1 |
import tensorflow as tf
|
2 |
import matplotlib.cm as cm
|
3 |
import numpy as np
|
@@ -66,4 +67,74 @@ def get_gradcam_prediction(img, grad_model, process=True, decode=True, pred_inde
|
|
66 |
overlay = array_to_img(overlay)
|
67 |
# decode prediction
|
68 |
preds_decode = tf.keras.applications.imagenet_utils.decode_predictions(preds.numpy())[0] if decode else preds
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
return preds_decode, overlay
|
|
|
1 |
+
<<<<<<< HEAD
|
2 |
import tensorflow as tf
|
3 |
import matplotlib.cm as cm
|
4 |
import numpy as np
|
|
|
67 |
overlay = array_to_img(overlay)
|
68 |
# decode prediction
|
69 |
preds_decode = tf.keras.applications.imagenet_utils.decode_predictions(preds.numpy())[0] if decode else preds
|
70 |
+
=======
|
71 |
+
import tensorflow as tf
|
72 |
+
import matplotlib.cm as cm
|
73 |
+
import numpy as np
|
74 |
+
try:
|
75 |
+
from tensorflow.keras.utils import array_to_img, img_to_array
|
76 |
+
except:
|
77 |
+
from tensorflow.keras.preprocessing.image import array_to_img, img_to_array
|
78 |
+
|
79 |
+
def process_image(img, size=(224, 224)):
|
80 |
+
img_array = tf.keras.applications.imagenet_utils.preprocess_input(img, mode='torch')
|
81 |
+
img_array = tf.image.resize(img_array, size,)[None,]
|
82 |
+
return img_array
|
83 |
+
|
84 |
+
def get_gradcam_model(model):
|
85 |
+
inp = tf.keras.Input(shape=(224, 224, 3))
|
86 |
+
feats = model.forward_features(inp)
|
87 |
+
preds = model.forward_head(feats)
|
88 |
+
return tf.keras.models.Model(inp, [preds, feats])
|
89 |
+
|
90 |
+
def get_gradcam_prediction(img, grad_model, process=True, decode=True, pred_index=None, cmap='jet', alpha=0.6):
|
91 |
+
"""Grad-CAM for a single image
|
92 |
+
|
93 |
+
Args:
|
94 |
+
img (np.ndarray): process or raw image without batch_shape e.g. (224, 224, 3)
|
95 |
+
grad_model (tf.keras.Model): model with feature map and prediction
|
96 |
+
process (bool, optional): imagenet pre-processing. Defaults to True.
|
97 |
+
pred_index (int, optional): for particular calss. Defaults to None.
|
98 |
+
cmap (str, optional): colormap. Defaults to 'jet'.
|
99 |
+
alpha (float, optional): opacity. Defaults to 0.4.
|
100 |
+
|
101 |
+
Returns:
|
102 |
+
preds_decode: top5 predictions
|
103 |
+
heatmap: gradcam heatmap
|
104 |
+
"""
|
105 |
+
# process image for inference
|
106 |
+
if process:
|
107 |
+
img_array = process_image(img)
|
108 |
+
else:
|
109 |
+
img_array = tf.convert_to_tensor(img)[None,]
|
110 |
+
if img.min()!=img.max():
|
111 |
+
img = (img - img.min())/(img.max() - img.min())
|
112 |
+
img = np.uint8(img*255.0)
|
113 |
+
# get prediction
|
114 |
+
with tf.GradientTape(persistent=True) as tape:
|
115 |
+
preds, feats = grad_model(img_array)
|
116 |
+
if pred_index is None:
|
117 |
+
pred_index = tf.argmax(preds[0])
|
118 |
+
class_channel = preds[:, pred_index]
|
119 |
+
# compute heatmap
|
120 |
+
grads = tape.gradient(class_channel, feats)
|
121 |
+
pooled_grads = tf.reduce_mean(grads, axis=(0, 1, 2))
|
122 |
+
feats = feats[0]
|
123 |
+
heatmap = feats @ pooled_grads[..., tf.newaxis]
|
124 |
+
heatmap = tf.squeeze(heatmap)
|
125 |
+
heatmap = tf.maximum(heatmap, 0) / tf.math.reduce_max(heatmap)
|
126 |
+
heatmap = heatmap.numpy()
|
127 |
+
heatmap = np.uint8(255 * heatmap)
|
128 |
+
# colorize heatmap
|
129 |
+
cmap = cm.get_cmap(cmap)
|
130 |
+
colors = cmap(np.arange(256))[:, :3]
|
131 |
+
heatmap = colors[heatmap]
|
132 |
+
heatmap = array_to_img(heatmap)
|
133 |
+
heatmap = heatmap.resize((img.shape[1], img.shape[0]))
|
134 |
+
heatmap = img_to_array(heatmap)
|
135 |
+
overlay = img + heatmap * alpha
|
136 |
+
overlay = array_to_img(overlay)
|
137 |
+
# decode prediction
|
138 |
+
preds_decode = tf.keras.applications.imagenet_utils.decode_predictions(preds.numpy())[0] if decode else preds
|
139 |
+
>>>>>>> 094461a8d383ad2565311ea9a0094b5856887867
|
140 |
return preds_decode, overlay
|