From 8fca28579c59f45c83cdb77a27d9bad2a4225cf6 Mon Sep 17 00:00:00 2001 From: Yanzhan Yang Date: Fri, 21 Jun 2019 21:53:08 +0800 Subject: [PATCH] enhance auto debug tool (#1698) --- src/common/types.h | 1 + src/framework/executor.cpp | 4 +- test/net/test_net.cpp | 46 +++++++++++------- tools/python/caffetools/run.py | 30 ++++++++++++ tools/python/fluidtools/run.py | 87 +++++++++++++++++++++++----------- 5 files changed, 123 insertions(+), 45 deletions(-) create mode 100644 tools/python/caffetools/run.py diff --git a/src/common/types.h b/src/common/types.h index 9f0e792299..8e793e959b 100644 --- a/src/common/types.h +++ b/src/common/types.h @@ -133,6 +133,7 @@ enum PowerMode { struct PaddleMobileConfigInternal { bool load_when_predict = false; + bool enable_memory_optimization = true; }; extern const char *G_OP_TYPE_CONV; diff --git a/src/framework/executor.cpp b/src/framework/executor.cpp index 57af9c62de..47ee22f323 100644 --- a/src/framework/executor.cpp +++ b/src/framework/executor.cpp @@ -65,7 +65,9 @@ Executor::Executor(const Program &program, "program_desc_ should not be nullptr"); #if !defined(PADDLE_MOBILE_FPGA) && !defined(PADDLE_MOBILE_FPGA_KD) && \ !defined(PADDLE_MOBILE_CL) - pass::MemoryOptPass()(program_desc_.get(), program_.scope.get()); + if (config_.enable_memory_optimization) { + pass::MemoryOptPass()(program_desc_.get(), program_.scope.get()); + } #endif // resize feed and fetch list // should init feed and fetch variables before infer shape diff --git a/test/net/test_net.cpp b/test/net/test_net.cpp index f20db14a0d..24a5002111 100644 --- a/test/net/test_net.cpp +++ b/test/net/test_net.cpp @@ -17,39 +17,51 @@ limitations under the License. */ #include "../test_helper.h" #include "../test_include.h" -void test(int argc, char *argv[], bool fuse); +void test(int argc, char *argv[]); int main(int argc, char *argv[]) { - test(argc, argv, false); - test(argc, argv, true); + test(argc, argv); return 0; } -void test(int argc, char *argv[], bool fuse) { - paddle_mobile::PaddleMobile paddle_mobile; + +void test(int argc, char *argv[]) { + int arg_index = 1; + bool fuse = std::stoi(argv[arg_index]) == 1; + arg_index++; + bool enable_memory_optimization = std::stoi(argv[arg_index]) == 1; + arg_index++; + paddle_mobile::PaddleMobileConfigInternal config; + config.enable_memory_optimization = enable_memory_optimization; + paddle_mobile::PaddleMobile paddle_mobile(config); paddle_mobile.SetThreadNum(1); - std::string tag = fuse ? "-fuse" : ""; - int dim_count = std::stoi(argv[1]); + int dim_count = std::stoi(argv[arg_index]); + arg_index++; int size = 1; std::vector dims; for (int i = 0; i < dim_count; i++) { - int64_t dim = std::stoi(argv[2 + i]); + int64_t dim = std::stoi(argv[arg_index + i]); size *= dim; dims.push_back(dim); } + arg_index += dim_count; - int var_count = std::stoi(argv[1 + dim_count]); + int var_count = std::stoi(argv[arg_index]); + arg_index++; + int sample_step = std::stoi(argv[arg_index]); + arg_index++; std::vector var_names; for (int i = 0; i < var_count; i++) { - std::string var_name = argv[1 + dim_count + 1 + 1 + i]; + std::string var_name = argv[arg_index + i]; var_names.push_back(var_name); } + arg_index += var_count; auto time1 = time(); if (paddle_mobile.Load("./checked_model/model", "./checked_model/params", fuse, false, 1, true)) { auto time2 = time(); - std::cout << "auto-test" << tag + std::cout << "auto-test" << " load-time-cost :" << time_diff(time1, time1) << "ms" << std::endl; @@ -73,8 +85,9 @@ void test(int argc, char *argv[], bool fuse) { auto out = paddle_mobile.Predict(input_data, dims); } auto time4 = time(); - std::cout << "auto-test" << tag << " predict-time-cost " - << time_diff(time3, time4) / 50 << "ms" << std::endl; + std::cout << "auto-test" + << " predict-time-cost " << time_diff(time3, time4) / 50 << "ms" + << std::endl; // 测试正确性 auto out = paddle_mobile.Predict(input_data, dims); @@ -88,13 +101,12 @@ void test(int argc, char *argv[], bool fuse) { continue; } auto data = out->data(); - int step = len / 20; std::string sample = ""; - for (int i = 0; i < len; i += step) { + for (int i = 0; i < len; i += sample_step) { sample += " " + std::to_string(data[i]); } - std::cout << "auto-test" << tag << " var " << var_name << sample - << std::endl; + std::cout << "auto-test" + << " var " << var_name << sample << std::endl; } std::cout << std::endl; } diff --git a/tools/python/caffetools/run.py b/tools/python/caffetools/run.py new file mode 100644 index 0000000000..914ec83f0f --- /dev/null +++ b/tools/python/caffetools/run.py @@ -0,0 +1,30 @@ +import caffe +import numpy as np + +prototxt_path = "" +caffemodel_path = "" +input_path = "input.txt" +input_name = "" +output_name = "" + +shape = (1, 3, 64, 64) + +data = np.loadtxt(input_path).astype("float32").reshape(shape) + +net = caffe.Net(prototxt_path, caffemodel_path, caffe.TEST) + +# view inputs blob names +print(net.inputs) + +# view outputs blob names +print(net.outputs) + +# set input data +net.blobs[input_name].reshape(*shape) +net.blobs[input_name].data[...] = data + +# predict +net.forward() + +# view output data +print(net.blobs[output_name].data) diff --git a/tools/python/fluidtools/run.py b/tools/python/fluidtools/run.py index 19f185bbff..8eb67b3c38 100644 --- a/tools/python/fluidtools/run.py +++ b/tools/python/fluidtools/run.py @@ -1,3 +1,4 @@ +# -*- coding: utf-8 -* import os import sys import math @@ -9,6 +10,9 @@ model_path = "model" checked_model_path = "checked_model" feed_path = "feeds" output_path = "outputs" +diff_threshold = 0.01 + +np.set_printoptions(linewidth=150) mobile_exec_root = "/data/local/tmp/bin" mobile_src_root = os.path.abspath("../../../") @@ -16,11 +20,11 @@ if mobile_src_root.endswith("/"): mobile_src_root = mobile_src_root[:-1] dot = "•" -black = lambda x: "\033[30m" + str(x) -red = lambda x: "\033[31m" + str(x) -green = lambda x: "\033[32m" + str(x) +black = lambda x: "\033[30m" + str(x) + "\033[0m" +red = lambda x: "\033[31m" + str(x) + "\033[0m" +green = lambda x: "\033[32m" + str(x) + "\033[0m" +yellow = lambda x: "\033[33m" + str(x) + "\033[0m" reset = lambda x: "\033[0m" + str(x) -yellow = lambda x: "\033[33m" + str(x) def pp_tab(x, level=0): header = "" @@ -61,6 +65,7 @@ def resave_model(): # 强制所有var为可持久化 p_names = [] for name in vars: + name = str(name) v = fluid.framework._get_var(name, prog) if not v.persistable: v.persistable = True @@ -69,6 +74,7 @@ def resave_model(): has_found_wrong_shape = False # 修正每个var的形状 for name in vars: + name = str(name) v = vars[name] if v.persistable: v1 = fluid.global_scope().find_var(name) @@ -117,6 +123,8 @@ last_feed_var_name = None last_feed_file_name = None # 加载feed的key-value对 def load_feed_kv(): + if not os.path.exists(feed_path): + return None global last_feed_var_name global last_feed_file_name feed_kv = {} @@ -128,7 +136,16 @@ def load_feed_kv(): file_name = feed_name.replace("/", "_") last_feed_var_name = feed_name last_feed_file_name = file_name - data = np.loadtxt(feed_path + "/" + file_name).reshape(feed_shape).astype("float32") + feed_file_path = feed_path + "/" + file_name + if not os.path.exists(feed_file_path): + return None + data = np.loadtxt(feed_file_path) + expected_len = 1 + for dim in feed_shape: + expected_len *= dim + if len(data) != expected_len: + return None + data = data.reshape(feed_shape).astype("float32") feed_kv[feed_name] = data return feed_kv @@ -166,10 +183,11 @@ def get_var_data(var_name, feed_kv=None): return output output_var_cache = {} +sample_step = 1 def tensor_sample(tensor): - step = math.floor(len(tensor) / 20) + # step = math.floor(len(tensor) / 20) sample = [] - for i in range(0, len(tensor), step): + for i in range(0, len(tensor), sample_step): sample.append(tensor[i]) return sample op_cache = {} @@ -209,19 +227,21 @@ for op in ops: op_types.add(op.type) pp_tab("op types : {}".format(op_types), 1) -def check_mobile_results(lines, fuse): - pp_yellow(dot + dot + " checking {} paddle mobile results".format("fusion" if fuse else "non fusion")) +def check_mobile_results(args, fuse, mem_opt): + args = "{} {} {}".format("1" if fuse else "0", "1" if mem_opt else "0", args) + res = sh("adb shell \"cd {} && export LD_LIBRARY_PATH=. && ./test-net {}\"".format(mobile_exec_root, args)) + lines = res.split("\n") + for line in lines: + if line.startswith("auto-test-debug"): + print(line) + pp_yellow(dot + dot + " checking paddle mobile results for {} -- {} ".format(green("【fusion】" if fuse else "【non fusion】"), green("【memory-optimization】" if mem_opt else "【non-memory-optimization】"))) mobile_var_cache = {} for line in lines: parts = line.split(" ") - if len(parts) <= 0: + if len(parts) < 2: + continue + if "auto-test" != parts[0]: continue - if fuse: - if "auto-test-fuse" != parts[0]: - continue - else: - if "auto-test" != parts[0]: - continue if parts[1] == "load-time-cost": pp_green("load time cost : {}".format(parts[2]), 1) elif parts[1] == "predict-time-cost": @@ -235,6 +255,14 @@ def check_mobile_results(lines, fuse): error_values2 = None for index in op_cache: op_output_var_name, op = op_cache[index] + if mem_opt: + found_in_fetch = False + for fetch in fetches: + if op_output_var_name == fetch.name: + found_in_fetch = True + break + if not found_in_fetch: + continue if not op_output_var_name in output_var_cache: continue if not op_output_var_name in mobile_var_cache: @@ -247,7 +275,7 @@ def check_mobile_results(lines, fuse): for i in range(len(values1)): v1 = values1[i] v2 = values2[i] - if abs(v1 - v2) > 0.01: + if abs(v1 - v2) > diff_threshold: error_index = index break if error_index != None: @@ -257,19 +285,23 @@ def check_mobile_results(lines, fuse): if error_index == None: pp_green("outputs are all correct", 1) else: + error_values1 = np.array(error_values1) + error_values2 = np.array(error_values2) pp_red("{} op's output is not correct, op's type is {}".format(error_index, op_cache[error_index][1].type), 1) - pp_red("fluid results are : {}".format(error_values1), 1) - pp_red("paddle mobile results are : {}".format(error_values2), 1) + pp_red("fluid results are : ", 1) + pp_red(str(error_values1).replace("\n", "\n" + "\t" * 1), 1) + pp_red("paddle mobile results are : ", 1) + pp_red(str(error_values2).replace("\n", "\n" + "\t" * 1), 1) # print(output_var_cache) # print(mobile_var_cache) def main(): - # 如果feed_path不存在，则需要生成并保存feed的键值对 - if not os.path.exists(feed_path): - feed_kv = gen_feed_kv() - save_feed_kv(feed_kv) # 加载kv feed_kv = load_feed_kv() + if feed_kv == None: + feed_kv = gen_feed_kv() + save_feed_kv(feed_kv) + feed_kv = load_feed_kv() pp_yellow(dot + dot + " checking fetch info") for fetch in fetches: pp_tab("fetch var name : {}".format(fetch.name), 1) @@ -297,12 +329,13 @@ def main(): for dim in last_feed_var_shape: args += " " + str(dim) args += " " + str(len(output_var_cache)) + args += " " + str(sample_step) for var_name in output_var_cache.keys(): args += " " + var_name - res = sh("adb shell \"cd {} && export LD_LIBRARY_PATH=. && ./test-net {}\"".format(mobile_exec_root, args)) - lines = res.split("\n") - check_mobile_results(lines, False) - check_mobile_results(lines, True) + check_mobile_results(args, False, False) + check_mobile_results(args, False, True) + check_mobile_results(args, True, False) + check_mobile_results(args, True, True) if __name__ == "__main__": main() -- GitLab