[GitHub] [singa] nudles commented on a change in pull request #703: Refactor sonnx, test cases and examples

[GitBox <gi...@apache.org>]

nudles commented on a change in pull request #703:
URL: https://github.com/apache/singa/pull/703#discussion_r426453900



##########
File path: examples/cnn/onnx/train.py
##########
@@ -0,0 +1,307 @@
+#
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+#
+
+import sys, os
+import json
+from singa import singa_wrap as singa
+from singa import opt
+from singa import device
+from singa import tensor
+from singa import sonnx
+from singa import autograd
+import numpy as np
+import time
+import argparse
+from PIL import Image
+import onnx
+from tqdm import tqdm
+from utils import download_model, update_batch_size, check_exist_or_download, softmax_loss
+import logging
+
+logging.basicConfig(level=logging.INFO, format='%(asctime)s %(message)s')
+sys.path.append(os.path.dirname(__file__) + '/..')
+
+# Data Augmentation
+def augmentation(x, batch_size):
+    xpad = np.pad(x, [[0, 0], [0, 0], [4, 4], [4, 4]], 'symmetric')
+    for data_num in range(0, batch_size):
+        offset = np.random.randint(8, size=2)
+        x[data_num, :, :, :] = xpad[data_num, :,
+                                    offset[0]:offset[0] + x.shape[2],
+                                    offset[1]:offset[1] + x.shape[2]]
+        if_flip = np.random.randint(2)
+        if (if_flip):
+            x[data_num, :, :, :] = x[data_num, :, :, ::-1]
+    return x
+
+
+# Calculate Accuracy
+def accuracy(pred, target):
+    # y is network output to be compared with ground truth (int)
+    y = np.argmax(pred, axis=1)
+    a = y == target
+    correct = np.array(a, "int").sum()
+    # print(correct)
+    return correct
+
+
+# Data partition according to the rank
+def partition(global_rank, world_size, train_x, train_y, val_x, val_y):
+    # Partition training data
+    data_per_rank = train_x.shape[0] // world_size
+    idx_start = global_rank * data_per_rank
+    idx_end = (global_rank + 1) * data_per_rank
+    train_x = train_x[idx_start:idx_end]
+    train_y = train_y[idx_start:idx_end]
+    # Partition evaluation data
+    data_per_rank = val_x.shape[0] // world_size
+    idx_start = global_rank * data_per_rank
+    idx_end = (global_rank + 1) * data_per_rank
+    val_x = val_x[idx_start:idx_end]
+    val_y = val_y[idx_start:idx_end]
+    return train_x, train_y, val_x, val_y
+
+
+# Function to all reduce NUMPY Accuracy and Loss from Multiple Devices
+def reduce_variable(variable, dist_opt, reducer):
+    reducer.copy_from_numpy(variable)
+    dist_opt.all_reduce(reducer.data)
+    dist_opt.wait()
+    output = tensor.to_numpy(reducer)
+    return output
+
+
+def resize_dataset(x, image_size):
+    num_data = x.shape[0]
+    dim = x.shape[1]
+    X = np.zeros(shape=(num_data, dim, image_size, image_size),
+                 dtype=np.float32)
+    for n in range(0, num_data):
+        for d in range(0, dim):
+            X[n, d, :, :] = np.array(Image.fromarray(x[n, d, :, :]).resize(
+                (image_size, image_size), Image.BILINEAR),
+                                     dtype=np.float32)
+    return X
+
+
+def run(global_rank,
+        world_size,
+        local_rank,
+        max_epoch,
+        batch_size,
+        model,
+        data,
+        sgd,
+        graph,
+        dist_option='fp32',
+        spars=None):
+    dev = device.create_cuda_gpu_on(local_rank)
+    dev.SetRandSeed(0)
+    np.random.seed(0)
+
+    if data == 'cifar10':
+        from data import cifar10
+        train_x, train_y, val_x, val_y = cifar10.load()
+    elif data == 'cifar100':
+        from data import cifar100
+        train_x, train_y, val_x, val_y = cifar100.load()
+    elif data == 'mnist':
+        from data import mnist
+        train_x, train_y, val_x, val_y = mnist.load()
+
+    num_channels = train_x.shape[1]
+    image_size = train_x.shape[2]
+    data_size = np.prod(train_x.shape[1:train_x.ndim]).item()
+    num_classes = (np.max(train_y) + 1).item()
+
+    # print content
+    with open(os.path.join(os.path.dirname(__file__), 'models.json')) as json_file:
+        model_config = json.load(json_file)
+        model_config = model_config[model]
+
+    download_model(model_config['url'])
+    onnx_model = onnx.load(os.path.join('/tmp', model_config['path']))
+    onnx_model = update_batch_size(onnx_model, batch_size)
+    sg_ir = sonnx.prepare(onnx_model, device=dev)
+    model = sonnx.create_model(sg_ir, autograd.softmax_cross_entropy, sgd)

Review comment:
       are you going to change to the new SONNXModel API?




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org