NVIDIA DIGITS2 Deep Convolutional Feature Visualization

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.nameif 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].dataif layer.type not in ['InnerProduct']:vis = self.get_layer_visualization(data)else:vis = Nonemean, 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 layersif layer.type == 'Softmax':normalize = Falsevis = 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 NotImplementedErrorreturn (predictions, visualizations)def get_layer_visualization(self, data,normalize = True,max_width = 600,):"""Returns a vis_square for the given layer dataArguments:data -- a np.ndarrayKeyword arguments:normalize -- whether to normalize the data when visualizingmax_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)passelif 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_widthpadsize = 0width = data.shape[2]if width > max_width:data = data[0,:max_width,:max_width]else:if width > 1:padsize = 1width += 1n = max_width/widthn *= ndata = 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 resultKeyword arguments:padsize -- how many pixels go inbetween the tilesnormalize -- 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 mathimages = images.astype('float32')if normalize:images -= images.min()if images.max() > 0:images /= images.max()images *= 255if images.ndim == 3:# they're grayscale - convert to a colormapredmap, 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 togetherimages = 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 uint8images = images.astype('uint8')# Compute the output image matrix dimensionsn = int(np.ceil(np.sqrt(images.shape[0])))ny = nnx = nlength = images.shape[0]if n*(n-1) >= length:nx = n-1# Add padding between the imagespadding = ((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 othertiles = 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 tilesdef 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 noopredmap      = [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.' % nameredmap      = [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