setup development on GPU

.dvc/config

@@ -0,0 +1,4 @@
+[core]
+    remote = gdrive
+['remote "gdrive"']
+    url = gdrive://1TkUf745P5c4aHe1Xt2WFEDLBsUEyiPuA

Dockerfile

@@ -1,4 +1,4 @@
-FROM tensorflow/tensorflow:latest-py3 as base
+FROM tensorflow/tensorflow:2.2.0-gpu as base
 
 RUN apt update && \
     apt install -y make git
@@ -13,7 +13,6 @@ RUN python -m spacy download en_core_web_sm && \
 FROM base as second
 
 RUN pip install \
-    tensorflow==2.2 \
     tensorflow_datasets \
     dvc pydrive2
 

Makefile

@@ -43,6 +43,6 @@ docker-build:
 
 # Will start in docker develoment environment
 docker-console:
-	docker run -it --rm -v ${PWD}:/work -w /work -p 8888:8888 $(DOCKER_IMAGE_TAG) bash
+	docker run --gpus all -it --rm -v ${PWD}:/work -w /work -p 8888:8888 $(DOCKER_IMAGE_TAG) bash
 
 

README.md

@@ -19,6 +19,12 @@ Can be run in docker enviroment, or if you on linux:
 
 All Python libraries and models described in Dockerfile
 
+### Hardware
+
+* NVidia videocard with DLSS version >= 10 - actually GPU optional, and learning can be run on CPU,
+    but models and enviroment configurated to run on GPU,
+    in base case tenserflow can fallback to CPU, so not need change anything for start development
+
 ## Development
 
 If you on Windows, build and run in-Docker development enviroment

data.dvc

@@ -1,3 +1,3 @@
 outs:
-- md5: a9204af33b6043bc2e0f4752452db9cc.dir
+- md5: 5e1e1637ef6e7039a40a0ee22dd6f7e8.dir
   path: data

dvc.lock

@@ -4,27 +4,27 @@ train:
   - path: ./src/model.py
     md5: f2d2fe914061e31ab9a9b126273b0032
   - path: ./src/normalize.py
-    md5: 1e468f96edf82078bafa4fd6afe105b7
+    md5: 941426ea99ad2a81dee04ab8dba4b994
   - path: ./src/train.py
     md5: d93c994f261cf04723f5a0b373e5166b
   - path: data
-    md5: a9204af33b6043bc2e0f4752452db9cc.dir
+    md5: 5e1e1637ef6e7039a40a0ee22dd6f7e8.dir
   outs:
   - path: ./metrics/training.csv
-    md5: af1be0b754ba0cb8a6f1a97ddb9bfa70
+    md5: f93e951f0d028511427dca798c6460ce
   - path: ./saved_models/main.h5
-    md5: 04263a7f31aababf4949702b3f5ceb1a
+    md5: b1aee890497513c89cc7a736cb43d6e3
 test:
   cmd: python -m src.test
   deps:
   - path: data
-    md5: a9204af33b6043bc2e0f4752452db9cc.dir
+    md5: 5e1e1637ef6e7039a40a0ee22dd6f7e8.dir
   - path: saved_models/main.h5
-    md5: 04263a7f31aababf4949702b3f5ceb1a
+    md5: b1aee890497513c89cc7a736cb43d6e3
   - path: src/normalize.py
-    md5: 1e468f96edf82078bafa4fd6afe105b7
+    md5: 941426ea99ad2a81dee04ab8dba4b994
   - path: src/test.py
-    md5: 1b04bc50552b3b16345ffe4e2dffe6db
+    md5: 2fd1d3c454d60d2cc0d5e1979667505d
   outs:
   - path: metrics/test.json
-    md5: c02593eb341008ae5f3d27e1190f0d98
+    md5: d4233d255efe51f0e8235b78f7072299

metrics/test.json

@@ -1,4 +1,4 @@
 {
-    "loss": 0.5036352872848511,
-    "accuracy": 0.7742400169372559
+    "loss": 0.5321113467216492,
+    "accuracy": 0.7470800280570984
 }

metrics/training.csv

@@ -1,11 +1,11 @@
 epoch,accuracy,loss,val_accuracy,val_loss
-0,0.6493499875068665,0.6274378299713135,0.7731999754905701,0.5212063789367676
-1,0.7305999994277954,0.548753559589386,0.7833999991416931,0.49326327443122864
-2,0.7640500068664551,0.5078993439674377,0.6394000053405762,0.6567526459693909
-3,0.6689500212669373,0.6107459664344788,0.7024000287055969,0.5796387195587158
-4,0.7591500282287598,0.5226651430130005,0.8190000057220459,0.44037047028541565
-5,0.7748500108718872,0.4964852035045624,0.8181999921798706,0.43456318974494934
-6,0.8187500238418579,0.43161866068840027,0.7376000285148621,0.492388516664505
-7,0.8090000152587891,0.43665245175361633,0.8040000200271606,0.41991737484931946
-8,0.7987499833106995,0.44822627305984497,0.5127999782562256,0.8335568308830261
-9,0.6500499844551086,0.6251434683799744,0.7833999991416931,0.4926334619522095
+0,0.6239500045776367,0.6509515047073364,0.7278000116348267,0.5713796615600586
+1,0.7107999920845032,0.5822036862373352,0.7093999981880188,0.5925928950309753
+2,0.7537000179290771,0.521277666091919,0.8041999936103821,0.44496265053749084
+3,0.786050021648407,0.48704293370246887,0.8172000050544739,0.4280240833759308
+4,0.7549999952316284,0.5206849575042725,0.5929999947547913,0.6776124238967896
+5,0.6494500041007996,0.6242504119873047,0.7720000147819519,0.4978633224964142
+6,0.7963500022888184,0.4657124876976013,0.758400022983551,0.5457758903503418
+7,0.8179500102996826,0.4302661120891571,0.8230000138282776,0.4051404595375061
+8,0.8256499767303467,0.4140107333660126,0.8015999794006348,0.44847530126571655
+9,0.8198999762535095,0.41230010986328125,0.7512000203132629,0.5304372310638428

src/normalize.py

@@ -8,7 +8,7 @@
 
 VECTOR_SIZE = 300
 
-nlp = spacy.load("en_core_web_md")
+nlp = spacy.load("en_core_web_lg")
 
 def extract_sentences(text):
     doc = nlp(text)

src/test.py

@@ -5,6 +5,12 @@
 
 print("Tensorflow:", tf.__version__)
 
+# fix issue with "cannot find dnn implementation"
+# https://github.com/tensorflow/tensorflow/issues/36508
+physical_devices = tf.config.list_physical_devices('GPU')
+tf.config.experimental.set_memory_growth(physical_devices[0], enable=True) 
+print("Enabled experimental memory growth for", physical_devices[0])
+
 metrics_file='metrics/test.json'
 
 # Load normalised datasets