Add transfer learning functionality between TFKG models with .SetLayerWeights() on layer interface

Fix issue where the first row of single_file_dataset was being ignored Add transfer_learning example Save order of model weights in json file in save dir to make transfer learning nicer

Author: CodingBeard

Makefile

@@ -87,6 +87,20 @@ examples-sign-raw:
 go generate ./...
 cd examples/sign && go run main.go
 
+examples-transfer:
+go generate ./...
+docker-compose up -d tf-jupyter-golang
+docker-compose exec tf-jupyter-golang sh -c "cd /go/src/tfkg/examples/transfer_learning && go run main.go"
+
+examples-transfer-gpu:
+go generate ./...
+docker-compose up -d tf-jupyter-golang-gpu
+docker-compose exec tf-jupyter-golang-gpu sh -c "cd /go/src/tfkg/examples/transfer_learning && go run main.go"
+
+examples-transfer-raw:
+go generate ./...
+cd examples/transfer_learning && go run main.go
+
 test-python:
 docker-compose up -d tf-jupyter-golang
 docker-compose exec tf-jupyter-golang sh -c "cd /go/src/tfkg && python test.py"

data/img_folder_dataset.go

@@ -128,24 +128,26 @@ func NewImgFolderDataset(
 },
 }
 
-e = os.MkdirAll(config.CacheDir, os.ModePerm)
-if e != nil && e != os.ErrExist {
-errorHandler.Error(e)
-return nil, e
-}
-
-if _, e := os.Stat(filepath.Join(config.CacheDir, "category-tokenizer.json")); e == nil {
-d.categoryTokenizer = preprocessor.NewTokenizer(
-errorHandler,
-1,
--1,
-preprocessor.TokenizerConfig{IsCategoryTokenizer: true, DisableFiltering: true},
-)
-e = d.categoryTokenizer.Load(filepath.Join(config.CacheDir, "category-tokenizer.json"))
-if e != nil {
+if config.CacheDir != "" {
+e = os.MkdirAll(config.CacheDir, os.ModePerm)
+if e != nil && e != os.ErrExist {
 errorHandler.Error(e)
 return nil, e
 }
+
+if _, e := os.Stat(filepath.Join(config.CacheDir, "category-tokenizer.json")); e == nil {
+d.categoryTokenizer = preprocessor.NewTokenizer(
+errorHandler,
+1,
+-1,
+preprocessor.TokenizerConfig{IsCategoryTokenizer: true, DisableFiltering: true},
+)
+e = d.categoryTokenizer.Load(filepath.Join(config.CacheDir, "category-tokenizer.json"))
+if e != nil {
+errorHandler.Error(e)
+return nil, e
+}
+}
 }
 
 e = d.readFileNames()
@@ -169,11 +171,16 @@ type imgStatsCache struct {
 
 func (d *ImgFolderDataset) readFileNames() error {
 cacheFileName := "file-stats.json"
-cacheFileBytes, e := ioutil.ReadFile(filepath.Join(d.cacheDir, cacheFileName))
-if e != nil && !errors.Is(e, os.ErrNotExist) {
-d.errorHandler.Error(e)
-return e
-} else if e == nil {
+var cacheFileBytes []byte
+var e error
+if d.cacheDir != "" {
+cacheFileBytes, e = ioutil.ReadFile(filepath.Join(d.cacheDir, cacheFileName))
+if e != nil && !errors.Is(e, os.ErrNotExist) {
+d.errorHandler.Error(e)
+return e
+}
+}
+if len(cacheFileBytes) > 0 {
 var cache imgStatsCache
 e = json.Unmarshal(cacheFileBytes, &cache)
 if e != nil {

data/single_file_dataset.go

@@ -225,7 +225,7 @@ func (d *SingleFileDataset) readLineOffsets() error {
 
 lastPrint := time.Now().Unix()
 progress, lastProgress := 0, 0
-skippedHeaders := false
+skippedHeaders, zeroAdded := false, false
 swg := sizedwaitgroup.New(128)
 var errs []error
 for true {
@@ -238,6 +238,9 @@ func (d *SingleFileDataset) readLineOffsets() error {
 if !skippedHeaders && d.skipHeaders {
 skippedHeaders = true
 continue
+} else if !d.skipHeaders && !zeroAdded {
+zeroAdded = true
+d.lineOffsets = append(d.lineOffsets, 0)
 }
 
 if len(errs) > 0 {