Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Tensorflow slim Resnet101_v1 support #3

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

Tensorflow slim Resnet101_v1 support #3

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f86238d
resnet 101 layer mapping.
insikk May 21, 2018
1d46ba9
tf slim resnet v1 101 visualization draft.
insikk May 25, 2018
1afd658
resnet correct prediction.
insikk May 25, 2018
0fdaffc
add notebook result.
insikk May 25, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Binary file added demo/cat.jpg
View file
Open in desktop
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
1,206 changes: 1,206 additions & 0 deletions demo/demo_res101-tf_support.ipynb
View file
Open in desktop

Large diffs are not rendered by default.

33 changes: 15 additions & 18 deletions demo/demo_res101.ipynb
View file
Open in desktop

Large diffs are not rendered by default.

167 changes: 143 additions & 24 deletions demo/demo_vgg16.ipynb
View file
Open in desktop

Large diffs are not rendered by default.

200 changes: 155 additions & 45 deletions demo/layers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,18 +1,31 @@
import numpy as np
from pprint import pprint


class Fixations:
"""
This code is caffe compatible.
caffe uses (w, h, channel, batch) notation for weights. (Really? cofirmation required)
"""
def __init__(self, net):
self.net = net

def fc(self, points, layer, prevLayer):
for cr in range(5):
"""
In paper, it is decribed in Algorithm 1.

layer: higher layer (closer to final prediction)
prevLayer: lower layer (closer to input data)
"""
for cr in range(5): # loop over 5 final prediction. (multi-crop inference)
if (points[cr] != 0):
layer_out = []
# Blob values for prevous layer
# Blob values (Activation) for prevous layer
data = np.squeeze(self.net.blobs[prevLayer].data[cr, :])
# print("activation of (l-1) layer {} shape: {}".format(prevLayer, data.shape))
# Weights for the current layer
param = self.net.params[layer][0].data
# print("weights of (l) layer {} shape: {}".format(layer, param.shape))
# If previous layer is not fully connected
if (data.ndim == 3):
shape = data.shape
Expand All @@ -23,11 +36,13 @@ def fc(self, points, layer, prevLayer):
position = np.argmax(conv[i, :])
layer_out.append(np.unravel_index(position, shape))
else:
conv = data*param
conv = data*param # Hadamard product. (1000, 1, 1, 1) * (2048,)
# print("conv shape:", conv.shape)
for i in points[cr]:
# Getting Top-num activations for each input
num = np.sum(conv[i, :] > 0)
layer_out.extend(np.argsort(conv[i, :])[::-1][:num])
top_k_list = np.argsort(np.ravel(conv[i, :]))[::-1][:num].tolist()
layer_out.extend(top_k_list)
points[cr] = list(set(layer_out))
return points

Expand Down Expand Up @@ -160,54 +175,149 @@ def inception(self, points, layer, prevLayer, pool=False, out=False, prevLayer2=
return points

# Residual block for ResNet-101
def res(self, points, layer, prevLayer):
def res(self, points, layer, prevLayer, network_name="caffe"):

# print("pervLayer:", prevLayer)
# print("layer:", layer)

for cr in range(5):
if (points[cr] != 0):
layer_out = []
flag = False #flag to check if previous block downsampled the inputs
if 'a' in layer:
flag = True
layer_out = []

flag_caffe_first_layer_of_block = False
flag_tf_slim_resnet_v1_first_layer_of_block = False
flag_tf_slim_resnet_v1_last_layer_of_block = False

flag_handle_feature_map_depth = False # number of channels
flag_handle_feature_map_size = False # spatial w,h size of feature map.

# TODO: replace all flag to specific flag.
if 'a' in layer:
if network_name == "caffe":
# In caffe resnet implementation, first unit of each block downsample the input.
flag_caffe_first_layer_of_block = True
flag_handle_feature_map_depth = True
flag_handle_feature_map_size = True
# Getting activations from previous residual block
branch_skip = self.net.blobs['res'+layer+'_branch1'].data[cr]
else:
branch_skip = self.net.blobs['res'+prevLayer].data[cr]
# Getting activations from all convolution layers inside the residual block
branch_res_blobs = [self.net.blobs['res'+layer+'_branch2a'].data[cr], self.net.blobs['res'+layer+'_branch2b'].data[cr], self.net.blobs['res'+layer+'_branch2c'].data[cr]]
# Pad input to 3x3 block so the size remains same after conv
branch_res_blobs[0] = np.lib.pad(branch_res_blobs[0], 1 ,'constant',constant_values=0)[1:-1,:,:]
for i in points[cr]:
# To check if the higher activation at a point came from skip or delta branch
if branch_skip[i[0], i[1], i[2]] >= branch_res_blobs[2][i[0], i[1], i[2]]:
if flag:
param = self.net.params['res'+layer+'_branch1'][0].data[i[0], :, 0, 0]
if prevLayer == 'pool1':
data = self.net.blobs[prevLayer].data[cr, :, i[1], i[2]]
feature = np.argmax(data*param)
layer_out.append((feature, i[1], i[2]))
else:
data = self.net.blobs['res'+prevLayer].data[cr, :, i[1]*2, i[2]*2]
feature = np.argmax(data*param)
layer_out.append((feature, i[1]*2, i[2]*2))
else:
layer_out.append(i)
elif "tf_slim_resnet_v1" in network_name:
flag_tf_slim_resnet_v1_first_layer_of_block = True
flag_handle_feature_map_depth = True

branch_skip_name = 'res'+layer+'_branch1'
else:
if network_name == "tf_slim_resnet_v1_101" and (layer == "4b22" or layer == "3b3" or layer == "2c"):
# In slim resnet v1 implementation, final unit of each block already downsample the output.
flag_tf_slim_resnet_v1_last_layer_of_block = True
flag_handle_feature_map_size = True
if network_name == "tf_slim_resnet_v1_50" and (layer == "4b5" or layer == "3b3" or layer == "2c"):
# In slim resnet v1 implementation, final unit of each block already downsample the output.
flag_tf_slim_resnet_v1_last_layer_of_block = True
flag_handle_feature_map_size = True

# print("cr:", cr,", branch_skip = ", 'res'+prevLayer)
branch_skip_name = 'res'+prevLayer


# print("branch_skip_name:", branch_skip_name)
branch_skip = self.net.blobs[branch_skip_name].data[cr]
# Getting activations from all convolution layers inside the residual block
branch_res_blobs = [self.net.blobs['res'+layer+'_branch2a'].data[cr], self.net.blobs['res'+layer+'_branch2b'].data[cr], self.net.blobs['res'+layer+'_branch2c'].data[cr]]
# print('res'+layer+'_branch2c' + " shape", self.net.blobs['res'+layer+'_branch2c'].data[cr].shape)

# Pad input to 3x3 block so the size remains same after conv with stride 1
branch_res_blobs[0] = np.lib.pad(branch_res_blobs[0], 1 ,'constant',constant_values=0)[1:-1,:,:]
for idx, i in enumerate(points[cr]):
# print("point index: {}, point: {}".format(idx, i))
# To check if the higher activation at a point came from skip or delta branch
# print("branch skip shape", branch_skip.shape)
# print("branch branch_res_blobs[2] shape", branch_res_blobs[2].shape)
if branch_skip[i[0], i[1], i[2]] >= branch_res_blobs[2][i[0], i[1], i[2]]:
# print("backtrack through skip connection")
if flag_caffe_first_layer_of_block:
param = self.net.params['res'+layer+'_branch1'][0].data[i[0], :, 0, 0]
if prevLayer == 'pool1':
data = self.net.blobs[prevLayer].data[cr, :, i[1], i[2]]
feature = np.argmax(data*param)
layer_out.append((feature, i[1], i[2]))
else:
data = self.net.blobs['res'+prevLayer].data[cr, :, i[1]*2, i[2]*2]
feature = np.argmax(data*param)
layer_out.append((feature, i[1]*2, i[2]*2))

elif flag_tf_slim_resnet_v1_first_layer_of_block:
# NOTE: ON development
param = self.net.params['res'+layer+'_branch1'][0].data[i[0], :, 0, 0]
if prevLayer == "pool1":
data = self.net.blobs[prevLayer].data[cr, :, i[1], i[2]]
feature = np.argmax(data*param)
layer_out.append((feature, i[1], i[2]))
else:
# Just change feature map depth.
data = self.net.blobs['res'+prevLayer].data[cr, :, i[1], i[2]]
feature = np.argmax(data*param) # change depth
layer_out.append((feature, i[1], i[2])) # stride 2 conv.
elif flag_tf_slim_resnet_v1_last_layer_of_block:
# NOTE: ON development
# print("backtrack stride 2, receptive field 1x1 maxpooling. i:", i)
layer_out.append((i[0], i[1]*2, i[2]*2))
else:
layer_out.append(i)
else:
# print("backtrack through delta connection")
# print("param: ", 'res'+layer+'_branch2c')
# Weights
param = self.net.params['res'+layer+'_branch2c'][0].data[i[0], :, 0, 0]
# print("param shape: ", param.shape)

# Activation
# print("branch_res_blobs[1]:", 'res'+layer+'_branch2b')
# print("branch_res_blobs[1] shape:", branch_res_blobs[1].shape)
feature = np.argmax(branch_res_blobs[1][:, i[1], i[2]]*param)
param = self.net.params['res'+layer+'_branch2b'][0].data[feature, :, :, :]
data = branch_res_blobs[0][:, i[1]:i[1]+3, i[2]:i[2]+3]
feature = np.argmax(np.sum(np.sum(data*param, axis=2),axis=1))
x, y = np.unravel_index(np.argmax(data[feature,:,:]),data[feature,:,:].shape)
if flag:
param = self.net.params['res'+layer+'_branch2a'][0].data[feature, :, 0, 0]
if prevLayer =='pool1':
feature = np.argmax(self.net.blobs[prevLayer].data[cr, :, i[1]+x-1, i[2]+y-1]*param)
layer_out.append((feature, i[1]+x-1, i[2]+y-1))
else:
feature = np.argmax(self.net.blobs['res'+prevLayer].data[cr, :, (i[1]+x-1)*2, (i[2]+y-1)*2]*param)
layer_out.append((feature, (i[1]+x-1)*2, (i[2]+y-1)*2))
# print("param shape: ", param.shape)
if flag_tf_slim_resnet_v1_last_layer_of_block:
# NOTE: ON development
# print("tf slim last layer of block")
# Consider stride 2 option to reduce spatial resolution.
data = branch_res_blobs[0][:, (i[1]*2):((i[1]*2)+3), (i[2]*2):((i[2]*2)+3)]
else:
param = self.net.params['res'+layer+'_branch2a'][0].data[feature, :, 0, 0]
feature = np.argmax(branch_skip[:,i[1]+x-1,i[2]+y-1]*param)
data = branch_res_blobs[0][:, i[1]:i[1]+3, i[2]:i[2]+3]
# print("data (receptive field of the current spatial point of {}) shape:".format('res'+layer+'_branch2a'), data.shape)
feature = np.argmax(np.sum(np.sum(data*param, axis=2),axis=1))

# location of where the point come from in prev layer (closer to input)
x, y = np.unravel_index(np.argmax(data[feature,:,:]),data[feature,:,:].shape)
# print("x, y:", x, y)

if flag_caffe_first_layer_of_block:
param = self.net.params['res'+layer+'_branch2a'][0].data[feature, :, 0, 0]
if prevLayer =='pool1':
feature = np.argmax(self.net.blobs[prevLayer].data[cr, :, i[1]+x-1, i[2]+y-1]*param)
layer_out.append((feature, i[1]+x-1, i[2]+y-1))
points[cr] = list(set(layer_out))
else:
feature = np.argmax(self.net.blobs['res'+prevLayer].data[cr, :, (i[1]+x-1)*2, (i[2]+y-1)*2]*param)
layer_out.append((feature, (i[1]+x-1)*2, (i[2]+y-1)*2))

elif flag_tf_slim_resnet_v1_first_layer_of_block:
# NOTE: ON development
param = self.net.params['res'+layer+'_branch2a'][0].data[feature, :, 0, 0]
if prevLayer == 'pool1':
feature = np.argmax(self.net.blobs[prevLayer].data[cr, :, i[1]+x-1, i[2]+y-1]*param)
layer_out.append((feature, i[1]+x-1, i[2]+y-1))
else:
feature = np.argmax(self.net.blobs['res'+prevLayer].data[cr, :, i[1]+x-1, i[2]+y-1]*param)
layer_out.append((feature, i[1]+x-1, i[2]+y-1))
elif flag_tf_slim_resnet_v1_last_layer_of_block:
# NOTE: ON development
# print("tf slim last layer of block")
# print("receptive field data:", data[feature,:,:])
param = self.net.params['res'+layer+'_branch2a'][0].data[feature, :, 0, 0]
feature = np.argmax(self.net.blobs['res'+prevLayer].data[cr, :, (i[1])*2+x-1, (i[2])*2+y-1]*param)
layer_out.append((feature, (i[1])*2+x-1, (i[2])*2+y-1))
else:
param = self.net.params['res'+layer+'_branch2a'][0].data[feature, :, 0, 0]
feature = np.argmax(branch_skip[:,i[1]+x-1,i[2]+y-1]*param)
layer_out.append((feature, i[1]+x-1, i[2]+y-1))

points[cr] = list(set(layer_out))
return points
8 changes: 7 additions & 1 deletion demo/tf_helpers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,7 +86,6 @@ def __init__(self, graph, weights,
:param output_tensor_name: The full tensor name (Include the dev
placement and scope).
"""

self.sess = tf.Session()
new_saver = tf.train.import_meta_graph(graph)
new_saver.restore(self.sess, weights)
Expand Down Expand Up @@ -120,13 +119,15 @@ def add_to_layer(name):

for v in trainable_params:
if 'weight' in v.name:
print("add weight layer:", v.name)
name = v.name.split('/')[0]
params[name] = v
add_to_layer(name)

# Pooling layers usually don't have a nice way of gathering.
for n in tf.get_default_graph().as_graph_def().node:
if 'pool' in n.name:
print("add pool layer:", n.name)
v = get_tensor("{}:0".format(n.name))
name = n.name.split('/')[0]
params[name] = v
Expand All @@ -141,6 +142,8 @@ def add_to_layer(name):

def predict(self, img, oversample=True):
imgs = _oversample(img, (224, 224))
print("imgs:", imgs)
print("imgs shape:", imgs.shape)

predictions = self.sess.run([self.softmax],
feed_dict={self.placeholder: imgs})[0]
Expand All @@ -154,6 +157,9 @@ def predict(self, img, oversample=True):
self.blobs['prob'] = Data(predictions[0].T)

# Gather vars

print("imgs:", imgs)
print("imgs shape:", imgs.shape)
layers, params = self.sess.run([self.layers, self._params],
feed_dict={self.placeholder: imgs})

Expand Down
Loading