Caffe 1.1.6 prerelease

scripts/first_conv_force_u8.py

@@ -0,0 +1,79 @@
+import numpy as np
+import os
+import cv2
+import sys
+pycaffe = os.path.split(os.path.realpath(__file__))[0] + '/../python'
+sys.path.insert(0, pycaffe)
+import caffe
+from caffe.proto import caffe_pb2
+import google.protobuf.text_format as txtf
+
+
+def first_conv_u8_input(quantized_prototxt, weights, clx = False, force_u8_prototxt = '', new_weights_path = ''):
+
+    if new_weights_path == '':
+        new_weights_path = weights.split(".caffemodel")[0] + "_new.caffemodel"
+    if force_u8_prototxt == '':
+        force_u8_prototxt = quantized_prototxt.split('.prototxt')[0] + "_force_u8.prototxt"
+
+    if os.path.isfile(quantized_prototxt):
+        print ('caffeNet found.')
+    else:
+        print ('caffenet not found')
+
+    net = caffe.Net(quantized_prototxt, weights, caffe.TEST)
+
+    proto_net = caffe_pb2.NetParameter()
+    with open(quantized_prototxt) as f:
+        s = f.read()
+        txtf.Merge(s, proto_net)
+
+    for index, layer in enumerate(proto_net.layer):
+        if layer.type == "Data":
+            data_layer = layer
+        if layer.type == "Convolution":
+            first_conv_layer = layer
+            break
+
+    first_conv_weight = net.params[first_conv_layer.name][0].data
+    first_conv_bias = net.params[first_conv_layer.name][1].data
+
+    scale_in = first_conv_layer.quantization_param.scale_in[0]
+    bias_scale = 128/scale_in
+    new_bias = np.zeros(first_conv_bias.shape)
+    for i in range(len(first_conv_bias)):
+        new_bias[i] = first_conv_bias[i] - bias_scale * np.sum(first_conv_weight[i])
+
+    net.params[first_conv_layer.name][1].data[...] = new_bias
+
+    net.save(new_weights_path)
+
+    first_conv_layer.quantization_param.is_negative_input = False
+    old_pad = first_conv_layer.convolution_param.pad[0]
+    first_conv_layer.convolution_param.pad[0] = 0L
+    first_conv_layer.quantization_param.force_u8_input = True
+    if not clx:
+        for i in range(len(first_conv_layer.quantization_param.scale_params)):
+            first_conv_layer.quantization_param.scale_params[i] = float(first_conv_layer.quantization_param.scale_params[i]) * 0.5
+
+    data_layer.transform_param.pad = long(old_pad)
+    mean_value = np.array(data_layer.transform_param.mean_value).astype(float)
+    scale = data_layer.transform_param.scale
+    new_scale = scale_in * scale
+    new_mean_value = mean_value - 128/new_scale
+    data_layer.transform_param.mean_value[0] = new_mean_value[0]
+    data_layer.transform_param.mean_value[1] = new_mean_value[1]
+    data_layer.transform_param.mean_value[2] = new_mean_value[2]
+
+    data_layer.transform_param.scale = new_scale
+
+    with open(force_u8_prototxt, 'w') as f:
+        f.write(str(proto_net))
+
+    print("Please use new prototxt in {0}, caffemodel in {1}".format(force_u8_prototxt, new_weights_path))
+
+if __name__=='__main__':
+    if len(sys.argv) < 2:
+        print "Usgae: python first_conv_force_u8.py $prototxt $weights"
+        sys.exit(0)
+    first_conv_u8_input(sys.argv[1], sys.argv[2])