相信大家对Digits 这个框架的简单、易上手印象深刻,本人也认为其测试的特征可视化结果较为醒目:
但是该框架是如何实现的呢?
通过https://groups.google.com/forum/#!topic/digits-users/eRGijUu9q30,发现位于digits-2.0/digits/model/tasks中 caffe_train.py
https://groups.google.com/forum/#!topic/digits-users/eRGijUu9q30
def classify_one(self,image,snapshot_eppoch=None,layers=None)
"""
Classify an image
Returns (predictions,visualizations)
..........
.....
<pre name="code" class="python">visualizations = []
if layers and layers != 'none':
if layers == 'all':
added_activations = []
for layer in self.network.layer:
print 'Computing visualizations for "%s"...' % layer.name
if not layer.type.endswith(('Data', 'Loss', 'Accuracy')):
for bottom in layer.bottom:
if bottom in net.blobs and bottom not in added_activations:
data = net.blobs[bottom].data[0]
vis = self.get_layer_visualization(data)
mean, std, hist = self.get_layer_statistics(data)
visualizations.append(
{
'name': str(bottom),
'type': 'Activations',
'mean': mean,
'stddev': std,
'histogram': hist,
'image_html': utils.image.embed_image_html(vis),
}
)
added_activations.append(bottom)
if layer.name in net.params:
data = net.params[layer.name][0].data
if layer.type not in ['InnerProduct']:
vis = self.get_layer_visualization(data)
else:
vis = None
mean, std, hist = self.get_layer_statistics(data)
visualizations.append(
{
'name': str(layer.name),
'type': 'Weights (%s layer)' % layer.type,
'mean': mean,
'stddev': std,
'histogram': hist,
'image_html': utils.image.embed_image_html(vis),
}
)
for top in layer.top:
if top in net.blobs and top not in added_activations:
data = net.blobs[top].data[0]
normalize = True
# don't normalize softmax layers
if layer.type == 'Softmax':
normalize = False
vis = self.get_layer_visualization(data, normalize=normalize)
mean, std, hist = self.get_layer_statistics(data)
visualizations.append(
{
'name': str(top),
'type': 'Activation',
'mean': mean,
'stddev': std,
'histogram': hist,
'image_html': utils.image.embed_image_html(vis),
}
)
added_activations.append(top)
else:
raise NotImplementedError
return (predictions, visualizations)
def get_layer_visualization(self, data,
normalize = True,
max_width = 600,
):
"""
Returns a vis_square for the given layer data
Arguments:
data -- a np.ndarray
Keyword arguments:
normalize -- whether to normalize the data when visualizing
max_width -- maximum width for the vis_square
"""
#print 'data.shape is %s' % (data.shape,)
if data.ndim == 1:
# interpret as 1x1 grayscale images
# (N, 1, 1)
data = data[:, np.newaxis, np.newaxis]
elif data.ndim == 2:
# interpret as 1x1 grayscale images
# (N, 1, 1)
data = data.reshape((data.shape[0]*data.shape[1], 1, 1))
elif data.ndim == 3:
if data.shape[0] == 3:
# interpret as a color image
# (1, H, W,3)
data = data[[2,1,0],...] # BGR to RGB (see issue #59)
data = data.transpose(1,2,0)
data = data[np.newaxis,...]
else:
# interpret as grayscale images
# (N, H, W)
pass
elif data.ndim == 4:
if data.shape[0] == 3:
# interpret as HxW color images
# (N, H, W, 3)
data = data.transpose(1,2,3,0)
data = data[:,:,:,[2,1,0]] # BGR to RGB (see issue #59)
elif data.shape[1] == 3:
# interpret as HxW color images
# (N, H, W, 3)
data = data.transpose(0,2,3,1)
data = data[:,:,:,[2,1,0]] # BGR to RGB (see issue #59)
else:
# interpret as HxW grayscale images
# (N, H, W)
data = data.reshape((data.shape[0]*data.shape[1], data.shape[2], data.shape[3]))
else:
raise RuntimeError('unrecognized data shape: %s' % (data.shape,))
# chop off data so that it will fit within max_width
padsize = 0
width = data.shape[2]
if width > max_width:
data = data[0,:max_width,:max_width]
else:
if width > 1:
padsize = 1
width += 1
n = max_width/width
n *= n
data = data[:n]
#print 'data.shape now %s' % (data.shape,)
return utils.image.vis_square(data,
padsize = padsize,
normalize = normalize,
)
其中,vis_square位于 digits-2.0/digits/digits/utils image.py
def vis_square(images,
padsize=1,
normalize=False,
colormap='jet',
):
"""
Visualize each image in a grid of size approx sqrt(n) by sqrt(n)
Returns a np.array image
(Based on Caffe's filter_visualization notebook)
Arguments:
images -- an array of shape (N, H, W) or (N, H, W, C)
if C is not set, a heatmap is computed for the result
Keyword arguments:
padsize -- how many pixels go inbetween the tiles
normalize -- if true, scales (min, max) across all images out to (0, 1)
colormap -- a string representing one of the suppoted colormaps
"""
assert 3 <= images.ndim <= 4, 'images.ndim must be 3 or 4'
# convert to float since we're going to do some math
images = images.astype('float32')
if normalize:
images -= images.min()
if images.max() > 0:
images /= images.max()
images *= 255
if images.ndim == 3:
# they're grayscale - convert to a colormap
redmap, greenmap, bluemap = get_color_map(colormap)
red = np.interp(images*(len(redmap)-1)/255.0, xrange(len(redmap)), redmap)
green = np.interp(images*(len(greenmap)-1)/255.0, xrange(len(greenmap)), greenmap)
blue = np.interp(images*(len(bluemap)-1)/255.0, xrange(len(bluemap)), bluemap)
# Slap the channels back together
images = np.concatenate( (red[...,np.newaxis], green[...,np.newaxis], blue[...,np.newaxis]), axis=3 )
images = np.minimum(images,255)
images = np.maximum(images,0)
# convert back to uint8
images = images.astype('uint8')
# Compute the output image matrix dimensions
n = int(np.ceil(np.sqrt(images.shape[0])))
ny = n
nx = n
length = images.shape[0]
if n*(n-1) >= length:
nx = n-1
# Add padding between the images
padding = ((0, nx*ny - length), (0, padsize), (0, padsize)) + ((0, 0),) * (images.ndim - 3)
padded = np.pad(images, padding, mode='constant', constant_values=255)
# Tile the images beside each other
tiles = padded.reshape( (ny, nx) + padded.shape[1:]).transpose( (0,2,1,3) + tuple(range(4, padded.ndim + 1)))
tiles = tiles.reshape((ny * tiles.shape[1], nx * tiles.shape[3]) + tiles.shape[4:])
return tiles
def get_color_map(name):
"""
Return a colormap as (redmap, greenmap, bluemap)
Arguments:
name -- the name of the colormap. If unrecognized, will default to 'jet'.
"""
redmap = [0]
greenmap = [0]
bluemap = [0]
if name == 'white':
# essentially a noop
redmap = [0,1]
greenmap = [0,1]
bluemap = [0,1]
elif name == 'simple':
redmap = [0,1,1,1]
greenmap = [0,0,1,1]
bluemap = [0,0,0,1]
elif name == 'hot':
redmap = [0, 0.03968253968253968, 0.07936507936507936, 0.119047619047619, 0.1587301587301587, 0.1984126984126984, 0.2380952380952381, 0.2777777777777778, 0.3174603174603174, 0.3571428571428571, 0.3968253968253968, 0.4365079365079365, 0.4761904761904762, 0.5158730158730158, 0.5555555555555556, 0.5952380952380952, 0.6349206349206349, 0.6746031746031745, 0.7142857142857142, 0.753968253968254, 0.7936507936507936, 0.8333333333333333, 0.873015873015873, 0.9126984126984127, 0.9523809523809523, 0.992063492063492, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
greenmap = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.03174603174603163, 0.0714285714285714, 0.1111111111111112, 0.1507936507936507, 0.1904761904761905, 0.23015873015873, 0.2698412698412698, 0.3095238095238093, 0.3492063492063491, 0.3888888888888888, 0.4285714285714284, 0.4682539682539679, 0.5079365079365079, 0.5476190476190477, 0.5873015873015872, 0.6269841269841268, 0.6666666666666665, 0.7063492063492065, 0.746031746031746, 0.7857142857142856, 0.8253968253968254, 0.8650793650793651, 0.9047619047619047, 0.9444444444444442, 0.984126984126984, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
bluemap = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.04761904761904745, 0.1269841269841265, 0.2063492063492056, 0.2857142857142856, 0.3650793650793656, 0.4444444444444446, 0.5238095238095237, 0.6031746031746028, 0.6825396825396828, 0.7619047619047619, 0.8412698412698409, 0.92063492063492, 1]
elif name == 'rainbow':
redmap = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0.9365079365079367, 0.8571428571428572, 0.7777777777777777, 0.6984126984126986, 0.6190476190476191, 0.53968253968254, 0.4603174603174605, 0.3809523809523814, 0.3015873015873018, 0.2222222222222223, 0.1428571428571432, 0.06349206349206415, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.03174603174603208, 0.08465608465608465, 0.1375661375661377, 0.1904761904761907, 0.2433862433862437, 0.2962962962962963, 0.3492063492063493, 0.4021164021164023, 0.4550264550264553, 0.5079365079365079, 0.5608465608465609, 0.6137566137566139, 0.666666666666667]
greenmap = [0, 0.03968253968253968, 0.07936507936507936, 0.119047619047619, 0.1587301587301587, 0.1984126984126984, 0.2380952380952381, 0.2777777777777778, 0.3174603174603174, 0.3571428571428571, 0.3968253968253968, 0.4365079365079365, 0.4761904761904762, 0.5158730158730158, 0.5555555555555556, 0.5952380952380952, 0.6349206349206349, 0.6746031746031745, 0.7142857142857142, 0.753968253968254, 0.7936507936507936, 0.8333333333333333, 0.873015873015873, 0.9126984126984127, 0.9523809523809523, 0.992063492063492, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0.9841269841269842, 0.9047619047619047, 0.8253968253968256, 0.7460317460317465, 0.666666666666667, 0.587301587301587, 0.5079365079365079, 0.4285714285714288, 0.3492063492063493, 0.2698412698412698, 0.1904761904761907, 0.1111111111111116, 0.03174603174603208, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
bluemap = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.01587301587301582, 0.09523809523809534, 0.1746031746031744, 0.2539682539682535, 0.333333333333333, 0.412698412698413, 0.4920634920634921, 0.5714285714285712, 0.6507936507936507, 0.7301587301587302, 0.8095238095238093, 0.8888888888888884, 0.9682539682539679, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
elif name == 'winter':
greenmap = [0, 1]
bluemap = [1, 0.5]
else:
if name != 'jet':
print 'Warning: colormap "%s" not supported. Using jet instead.' % name
redmap = [0,0,0,0,0.5,1,1,1,0.5]
greenmap = [0,0,0.5,1,1,1,0.5,0,0]
bluemap = [0.5,1,1,1,0.5,0,0,0,0]
return 255.0 * np.array(redmap), 255.0 * np.array(greenmap), 255.0 * np.array(bluemap)
具体实现效果可以在python上参考以上代码