Script for converting minigo weights.

Pull request #1538.

training/minigo/convert_minigo.py

@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+import tensorflow as tf
+import numpy as np
+import sys
+
+if len(sys.argv) < 2:
+    print('Model filename without extension needed as an argument.')
+    exit()
+
+sess = tf.Session()
+saver = tf.train.import_meta_graph(sys.argv[1]+'.meta')
+saver.restore(sess, sys.argv[1])
+
+if 0:
+    # Exports graph to tensorboard
+    with tf.Session() as sess:
+        writer = tf.summary.FileWriter('logs', sess.graph)
+        writer.close()
+
+trainable_names = []
+for v in tf.trainable_variables():
+    trainable_names.append(v.name)
+
+weights = []
+for v in tf.global_variables():
+    if v.name in trainable_names:
+        weights.append(v)
+    elif 'batch_normalization' in v.name:
+        # Moving mean and variance are not trainable, but are needed for the model
+        if 'moving_mean' in v.name or 'moving_variance' in v.name:
+            weights.append(v)
+
+if 0:
+    for w in weights:
+        print(w.name)
+
+def merge_gammas(weights):
+    out_weights = []
+    skip = 0
+    for e, w in enumerate(weights):
+        if skip > 0:
+            skip -= 1
+            continue
+        if 'kernel' in w.name and 'conv2d' in w.name and 'gamma' in weights[e+2].name:
+            kernel = w
+            bias = weights[e+1]
+            gamma = weights[e+2]
+            beta = weights[e+3]
+            mean = weights[e+4]
+            var = weights[e+5]
+
+            new_kernel = kernel * tf.reshape(gamma, (1, 1, 1, -1))
+            new_bias = gamma * bias + beta * tf.sqrt(var + tf.constant(1e-5)) 
+            new_mean = mean * gamma
+
+            out_weights.append(new_kernel)
+            out_weights.append(new_bias)
+            out_weights.append(new_mean)
+            out_weights.append(var)
+
+            skip = 5
+        elif 'dense' in w.name and 'kernel' in w.name:
+            # Minigo uses channels last order while LZ uses channels first,
+            # Do some surgery for the dense layers to make the output match.
+            planes = w.shape[0].value//361
+            if planes > 0:
+                w1 = tf.reshape(w, [19, 19, planes, -1])
+                w2 = tf.transpose(w1, [2, 0, 1, 3])
+                new_kernel = tf.reshape(w2, [361*planes, -1])
+                out_weights.append(new_kernel)
+            else:
+                out_weights.append(w)
+        else:
+            out_weights.append(w)
+
+    return out_weights
+
+def save_leelaz_weights(filename, weights):
+    with open(filename, "w") as file:
+        # Version tag
+        # Minigo outputs winrate from blacks point of view (same as ELF)
+        file.write("2")
+        for e, w in enumerate(weights):
+            # Newline unless last line (single bias)
+            file.write("\n")
+            work_weights = None
+            if w.shape.ndims == 4:
+                # Convolution weights need a transpose
+                #
+                # TF (kYXInputOutput)
+                # [filter_height, filter_width, in_channels, out_channels]
+                #
+                # Leela/cuDNN/Caffe (kOutputInputYX)
+                # [output, input, filter_size, filter_size]
+                work_weights = tf.transpose(w, [3, 2, 0, 1])
+            elif w.shape.ndims == 2:
+                # Fully connected layers are [in, out] in TF
+                #
+                # [out, in] in Leela
+                #
+                work_weights = tf.transpose(w, [1, 0])
+            else:
+                # Biases, batchnorm etc
+                work_weights = w
+            nparray = work_weights.eval(session=sess)
+            if e == 0:
+                # Fix input planes
+
+                # Add zero weights for white to play input plane
+                nparray = np.pad(nparray, ((0, 0), (0, 1), (0, 0), (0, 0)), 'constant', constant_values=0)
+
+                # Permutate weights
+                p = [0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15, 16, 17]
+
+                nparray = nparray[:, p, :, :]
+            wt_str = [str(wt) for wt in np.ravel(nparray)]
+            file.write(" ".join(wt_str))
+
+save_leelaz_weights(sys.argv[1]+'_converted.txt', merge_gammas(weights))