添加自定义数据集分类部署

GAP8/ai_examples/classification_demo/Makefile

@@ -0,0 +1,93 @@
+# Copyright (C) 2017 GreenWaves Technologies
+# All rights reserved.
+
+# This software may be modified and distributed under the terms
+# of the BSD license.  See the LICENSE file for details.
+
+ifndef GAP_SDK_HOME
+  $(error Source sourceme in gap_sdk first)
+endif
+
+include common.mk
+
+LOAD_QUANTIZATION= #-q #to load a tflite quantized model
+IMAGE=$(CURDIR)/samples/12_11.pgm
+
+io=host
+
+
+RAM_FLASH_TYPE ?= HYPER
+#PMSIS_OS=freertos
+
+ifeq '$(RAM_FLASH_TYPE)' 'HYPER'
+APP_CFLAGS += -DUSE_HYPER
+MODEL_L3_EXEC=hram
+MODEL_L3_CONST=hflash
+else
+APP_CFLAGS += -DUSE_SPI
+CONFIG_SPIRAM = 1
+MODEL_L3_EXEC=qspiram
+MODEL_L3_CONST=qpsiflash
+endif
+
+
+QUANT_BITS?=8
+BUILD_DIR=BUILD
+ifeq ($(QUANT_BITS), 8)
+	MODEL_SQ8=1
+  ifeq '$(RAM_FLASH_TYPE)' 'HYPER'
+	   NNTOOL_SCRIPT=model/nntool_script
+  else
+    NNTOOL_SCRIPT=model/nntool_script_spi
+  endif
+  	MODEL_SUFFIX = _SQ8BIT
+else
+  ifeq ($(QUANT_BITS),16)
+    NNTOOL_SCRIPT=model/nntool_script16
+    MODEL_SUFFIX = _16BIT
+  else
+    $(error Don\'t know how to build with this bit width)
+  endif
+endif
+
+$(info Building GAP8 mode with $(QUANT_BITS) bit quantization)
+
+TRAINED_TFLITE_MODEL=model/$(MODEL_PREFIX).tflite
+
+include model_decl.mk
+
+# Here we set the memory allocation for the generated kernels
+# REMEMBER THAT THE L1 MEMORY ALLOCATION MUST INCLUDE SPACE
+# FOR ALLOCATED STACKS!
+CLUSTER_STACK_SIZE=4028
+CLUSTER_SLAVE_STACK_SIZE=1024
+TOTAL_STACK_SIZE=$(shell expr $(CLUSTER_STACK_SIZE) \+ $(CLUSTER_SLAVE_STACK_SIZE) \* 7)
+MODEL_L1_MEMORY=$(shell expr 60000 \- $(TOTAL_STACK_SIZE))
+MODEL_L2_MEMORY=250000
+MODEL_L3_MEMORY=8388608
+
+pulpChip = GAP
+PULP_APP = model
+
+APP = model
+APP_SRCS += $(MODEL_PREFIX).c $(MODEL_GEN_C) $(MODEL_COMMON_SRCS) $(CNN_LIB)
+
+APP_CFLAGS += -g -O1 -mno-memcpy -fno-tree-loop-distribute-patterns
+APP_CFLAGS += -I. -I$(MODEL_COMMON_INC) -I$(TILER_EMU_INC) -I$(TILER_INC) $(CNN_LIB_INCLUDE) -I$(MODEL_BUILD)
+APP_CFLAGS += -DPERF -DAT_MODEL_PREFIX=$(MODEL_PREFIX) $(MODEL_SIZE_CFLAGS)
+APP_CFLAGS += -DSTACK_SIZE=$(CLUSTER_STACK_SIZE) -DSLAVE_STACK_SIZE=$(CLUSTER_SLAVE_STACK_SIZE)
+APP_CFLAGS += -DAT_IMAGE=$(IMAGE)
+
+READFS_FILES=$(abspath $(MODEL_TENSORS))
+
+# all depends on the model
+all:: model
+
+clean:: clean_model
+
+include train_model.mk
+include model_rules.mk
+$(info APP_SRCS... $(APP_SRCS))
+$(info APP_CFLAGS... $(APP_CFLAGS))
+include $(RULES_DIR)/pmsis_rules.mk
+

GAP8/ai_examples/classification_demo/README.md

@@ -0,0 +1,67 @@
+# Sample MNIST Build using GAPFlow
+
+This project includes a sample based on a simple model graph defined in Keras.
+
+It goes from training right through to working code on GAP8 or the same code running on the PC for
+debugging purposes.
+
+* It first trains the network using keras
+* It then exports the network to H5 format
+* It then converts the H5 file to a TFLITE file using TensorFlow's TFLITE converter
+* It then generates an nntool state file by running an nntool script with commands to adjust the tensor and activation order, fuse certain operations together and automatically quantify the graph
+* It then uses this state file to generate an AutoTiler graph model
+* It then compiles and runs the model to produce GAP8 code
+* Finally it builds and runs the produced code on GAP8
+
+While this example uses Makefiles the same steps could be achieved with any build system.
+
+The process can be run to quantize the model in 16 or 8 bits weights and activations.
+
+## Build and run without Docker
+*This example uses the SDK Version `3.7`.*
+
+To build and run on GAP8:
+```
+make all run
+```
+To build and run on GVSOC
+```
+make all run platform=gvsoc
+```
+The input image is specified in the Makefile and loaded with the functions defined in ${GAP_SDK_HOME}/libs/gap_lib/img_io/ImgIO.c
+
+To clean the generated model and code but not the trained network type
+```
+make clean
+```
+To clean the trained keras save file type
+```
+make clean_train
+```
+To build and run the network compiled on the pc
+```
+make -f emul.mk all
+```
+This will produce an executable, model_emul, that can be used to evaluate files
+```
+e.g. ./model_emul images/5558_6.pgm 
+```
+This mode allows the application to be run with PC tools like valgrind which is very interesting for debugging.
+The cluster only has one core in this mode.
+
+The build defaults to 8 bit quantization. 16 bit quantization can be selected by preceeding the build lines above with QUANT_BITS=16.
+```
+e.g. QUANT_BITS=16 make -f emul.mk all
+```
+
+## Build and run within the docker
+*This example uses the SDK Version `3.7`.*
+
+Navigate to top-level folder of the example
+```
+cd <path/to/ai-deck/repository>/GAP8/ai_examples/gapflow_model_extern/
+```
+Build and execute in Docker similar as the other examples
+```
+docker run --rm -it -v $PWD:/module/data/ --device /dev/ttyUSB0 --privileged -P gapsdk:3.7 /bin/bash -c 'export GAPY_OPENOCD_CABLE=interface/ftdi/olimex-arm-usb-tiny-h.cfg; source /gap_sdk/configs/ai_deck.sh; cd /module/data/;  make clean clean_train clean_model clean_images all run'
+```

GAP8/ai_examples/classification_demo/bridge_stubs.h

@@ -0,0 +1,16 @@
+#ifndef __BRIDGE_STUBS_H__
+#define __BRIDGE_STUBS_H__
+
+#if defined(__PULP_OS__)
+
+#define BRIDGE_Init() {}
+#define BRIDGE_Connect(__wait_bridge, __event)          rt_bridge_connect(__wait_bridge, __event)
+#define BRIDGE_Open(__name, __flags, __mode, __event)   rt_bridge_open(__name, __flags, __mode, __event)
+#define BRIDGE_Read(__file, __buffer, __size, __event)  rt_bridge_read(__file, __buffer, __size, __event)
+#define BRIDGE_Write(__file, __buffer, __size, __event) rt_bridge_write(__file, __buffer, __size, __event)
+#define BRIDGE_Close(__file, __event)                   rt_bridge_close(__file, __event)
+#define BRIDGE_Disconnect(__event)                      rt_bridge_disconnect(__event)
+
+#endif  /* __PULP_OS__ */
+
+#endif  /* __BRIDGE_STUBS_H__ */

GAP8/ai_examples/classification_demo/common.mk

@@ -0,0 +1,10 @@
+# Copyright (C) 2020 GreenWaves Technologies
+# All rights reserved.
+
+# This software may be modified and distributed under the terms
+# of the BSD license.  See the LICENSE file for details.
+
+MODEL_PREFIX=model
+AT_INPUT_WIDTH=28
+AT_INPUT_HEIGHT=28
+AT_INPUT_COLORS=3

GAP8/ai_examples/classification_demo/common_rules.mk

@@ -0,0 +1,9 @@
+# Copyright (C) 2017 GreenWaves Technologies
+# All rights reserved.
+
+# This software may be modified and distributed under the terms
+# of the BSD license.  See the LICENSE file for details.
+
+$(IMAGES):
+	(mkdir -p $(IMAGES); $(MODEL_PYTHON) model/save_samples.py -d $@ -n 5)
+

GAP8/ai_examples/classification_demo/dataloader.py

@@ -0,0 +1,91 @@
+import os
+import numpy as np
+from PIL import Image
+import random
+
+
+def savelist(data_dir):
+	"""
+	从输入根目录获取图像路径、标签列表
+	"""
+	image_paths = []
+	image_labels = []
+
+	print(data_dir)
+
+	image_classes = os.listdir(data_dir)
+	print(image_classes)
+
+	for image_class in image_classes:
+	    sub_dir = os.path.join(data_dir + "\\" + image_class)
+	    # 进入子类目录
+	    image_names = os.listdir(sub_dir)
+	    for image_name in image_names:
+	        sub_path = os.path.join(sub_dir + "\\" + image_name)
+	        # 操作每类图片
+	        image_paths.append(sub_path)
+	        image_labels.append(image_classes.index(sub_dir[38:]))
+
+
+	# 随机打乱
+	random.seed(123)
+	random.shuffle(image_paths)
+	random.seed(123)
+	random.shuffle(image_labels)
+
+	image_count = len(image_paths)
+	split = 0.2
+
+	val_image_paths = image_paths[:int(split*image_count)]
+	train_image_paths = image_paths[int(split*image_count):]
+
+	val_image_labels = image_labels[:int(split*image_count)]
+	train_image_labels = image_labels[int(split*image_count):]
+
+	return train_image_paths, train_image_labels, val_image_paths, val_image_labels
+
+
+
+
+def save_as_numpy(image_paths, image_labels,str):
+	d = len(image_paths)
+	print(d)
+	data = np.empty((d, 28, 28, 3),dtype=np.uint8)
+	print(data)
+	while d>0:
+	    img = Image.open(image_paths[d-1])
+	    img = img.resize((28,28),Image.ANTIALIAS)
+	    img_ndarray = np.array(img)
+	    print(img_ndarray.shape)
+	    data[d - 1] = img_ndarray
+	    d = d - 1
+
+	print(str + "转换完成")
+	print(data)
+
+	d = len(image_labels)
+	label = np.zeros((d),dtype=np.uint8)
+	print(label)
+	while d>0:
+	    label[d - 1] = image_labels[d-1]
+	    d = d - 1
+	print(str + "转换完成")
+	print(label)
+
+	np.save(str + '_images.npy',data)
+	np.save(str + '_labels.npy',label)
+	np.savez(str + '.npz',image=data,label=label)
+
+
+
+
+if __name__=='__main__':
+
+	data_dir = ""
+
+	train_images, train_labels, val_images, val_labels = savelist(data_dir)
+
+
+	save_as_numpy(train_images, train_labels, str='train')
+	save_as_numpy(val_images, val_labels, str='val')
+

GAP8/ai_examples/classification_demo/emul.mk

@@ -0,0 +1,74 @@
+# Copyright (C) 2020 GreenWaves Technologies
+# All rights reserved.
+
+# This software may be modified and distributed under the terms
+# of the BSD license.  See the LICENSE file for details.
+
+include common.mk
+
+QUANT_BITS?=8
+BUILD_DIR=BUILD
+ifeq ($(QUANT_BITS), 8)
+	MODEL_SQ8=1
+	NNTOOL_SCRIPT=model/nntool_script_emul
+  	MODEL_SUFFIX = _SQ8BIT_EMUL
+else
+  ifeq ($(QUANT_BITS),16)
+    NNTOOL_SCRIPT=model/nntool_script_emul16
+    MODEL_SUFFIX = _16BIT_EMUL
+  else
+    $(error Don\'t know how to build with this bit width)
+  endif
+endif
+TRAINED_TFLITE_MODEL=model/$(MODEL_PREFIX).tflite
+
+include model_decl.mk
+
+MODEL_GEN_EXTRA_FLAGS= -f $(MODEL_BUILD)
+MODEL_GENFLAGS_EXTRA+=
+CC = gcc
+CFLAGS += -g -m32 -O0 -D__EMUL__ -DAT_MODEL_PREFIX=$(MODEL_PREFIX) $(MODEL_SIZE_CFLAGS) -DPERF
+INCLUDES = -I. -I$(TILER_EMU_INC) -I$(TILER_INC) $(CNN_LIB_INCLUDE) -I$(MODEL_BUILD) -I$(MODEL_COMMON_INC)
+LFLAGS =
+LIBS =
+SRCS = $(MODEL_PREFIX)_emul.c $(MODEL_GEN_C) $(MODEL_COMMON_SRCS) $(CNN_LIB)
+$(info CNN_LIB++ $(CNN_LIB))
+$(info SRCS++ $(SRCS))
+BUILD_DIR = BUILD_EMUL
+
+OBJS = $(patsubst %.c, $(BUILD_DIR)/%.o, $(SRCS))
+
+MAIN = $(MODEL_PREFIX)_emul
+# Here we set the memory allocation for the generated kernels
+# REMEMBER THAT THE L1 MEMORY ALLOCATION MUST INCLUDE SPACE
+# FOR ALLOCATED STACKS!
+CLUSTER_STACK_SIZE=2048
+CLUSTER_SLAVE_STACK_SIZE=1024
+TOTAL_STACK_SIZE=$(shell expr $(CLUSTER_STACK_SIZE) \+ $(CLUSTER_SLAVE_STACK_SIZE) \* 7)
+MODEL_L1_MEMORY=$(shell expr 60000 \- $(TOTAL_STACK_SIZE))
+MODEL_L2_MEMORY=200000
+MODEL_L3_MEMORY=8388608
+
+# hram - HyperBus RAM
+# qspiram - Quad SPI RAM
+MODEL_L3_EXEC=hram
+# hflash - HyperBus Flash
+# qpsiflash - Quad SPI Flash
+MODEL_L3_CONST=hflash
+
+all: model $(MAIN)
+
+$(OBJS) : $(BUILD_DIR)/%.o : %.c
+	@mkdir -p $(dir $@)
+	$(CC) $(CFLAGS) -c $< $(INCLUDES) -MD -MF $(basename $@).d -o $@
+
+$(MAIN): $(OBJS)
+	$(CC) $(CFLAGS) -MMD -MP $(CFLAGS) $(INCLUDES) -o $(MAIN) $(OBJS) $(LFLAGS) $(LIBS)
+
+clean: clean_model
+	$(RM) -r $(BUILD_DIR)
+	$(RM) $(MAIN)
+
+.PHONY: depend clean
+include train_model.mk
+include model_rules.mk