Merge pull request Kaggle#860 from Kaggle/pytorch1-6

rosbo · web-flow · commit 5b01c8343ee4 · 2020-09-14T18:07:04.000-07:00
Upgrade to Pytorch 1.6
diff --git a/Dockerfile b/Dockerfile
@@ -47,7 +47,7 @@ RUN conda config --add channels conda-forge && \
     conda install matplotlib basemap cartopy python-igraph imagemagick "pysal==2.1.0" && \
     # b/142337634#comment22 pin required to avoid torchaudio downgrade.
     # b/162357958##comment7 Upgrade once new versions of torch* libs are released for pytorch 1.6.
-    conda install "pytorch=1.5" "torchvision=0.6" "torchaudio=0.5" "torchtext=0.6" cpuonly && \
+    conda install "pytorch=1.6" "torchvision=0.7" "torchaudio=0.6" "torchtext=0.7" cpuonly && \
     /tmp/clean-layer.sh
 
 # The anaconda base image includes outdated versions of these packages. Update them to include the latest version.
@@ -343,8 +343,7 @@ RUN pip install bcolz && \
     pip install nbformat && \
     pip install notebook==5.5.0 && \
     pip install olefile && \
-    # b/162850432 unpin once we upgrade to pytorch 1.6
-    pip install kornia==0.3.2 && \
+    pip install kornia && \
     pip install pandas_summary && \
     pip install pandocfilters && \
     pip install pexpect && \
@@ -374,8 +373,7 @@ RUN pip install bcolz && \
     pip install widgetsnbextension && \
     pip install pyarrow && \
     pip install feather-format && \
-    # b/162850432 unpin once we upgrade to pytorch 1.6
-    pip install fastai~=1.0.61 && \
+    pip install fastai && \
     pip install allennlp && \
     python -m spacy download en && python -m spacy download en_core_web_lg && \
     apt-get install -y ffmpeg && \
@@ -426,7 +424,8 @@ RUN pip install flashtext && \
     pip install plotly_express && \
     pip install albumentations && \
     pip install catalyst && \
-    pip install osmnx && \
+    # b/162850432 prevent matplotlib upgrade.
+    pip install osmnx==0.15.1 && \
     apt-get -y install libspatialindex-dev && \
     pip install pytorch-ignite && \
     pip install qgrid && \
diff --git a/gpu.Dockerfile b/gpu.Dockerfile
@@ -61,7 +61,7 @@ RUN apt-get install -y ocl-icd-libopencl1 clinfo libboost-all-dev && \
 # However, because this image is based on the CPU image, this isn't possible but better
 # to put them at the top of this file to minize conflicts.
 RUN conda remove --force -y pytorch torchvision torchaudio cpuonly && \
-    conda install "pytorch=1.5" "torchvision=0.6" "torchaudio=0.5" "torchtext=0.6" cudatoolkit=$CUDA_VERSION && \
+    conda install "pytorch=1.6" "torchvision=0.7" "torchaudio=0.6" "torchtext=0.7" cudatoolkit=$CUDA_VERSION && \
     /tmp/clean-layer.sh
 
 # Install LightGBM with GPU
diff --git a/tests/test_fastai.py b/tests/test_fastai.py
@@ -1,21 +1,12 @@
 import unittest
 
 import fastai
-import pandas as pd
-import torch
 
-from fastai.tabular import *
-from fastai.core import partition
-from fastai.torch_core import tensor
+from fastai.tabular.all import *
 
 class TestFastAI(unittest.TestCase):
-    def test_partition(self):
-        result = partition([1,2,3,4,5], 2)
-
-        self.assertEqual(3, len(result))
-
     def test_has_version(self):
-        self.assertGreater(len(fastai.__version__), 1)
+        self.assertGreater(len(fastai.__version__), 2)
     
     # based on https://github.com/fastai/fastai/blob/master/tests/test_torch_core.py#L17
     def test_torch_tensor(self):
@@ -25,18 +16,12 @@ def test_torch_tensor(self):
         self.assertTrue(torch.all(a == b))
 
     def test_tabular(self):
-        df = pd.read_csv("/input/tests/data/train.csv")
-        procs = [FillMissing, Categorify, Normalize]
-
-        valid_idx = range(len(df)-5, len(df))
-        dep_var = "label"
-        cont_names = []
-        for i in range(784):
-            cont_names.append("pixel" + str(i))
-
-        data = (TabularList.from_df(df, path="", cont_names=cont_names, cat_names=[], procs=procs)
-            .split_by_idx(valid_idx)
-            .label_from_df(cols=dep_var)
-            .databunch())
-        learn = tabular_learner(data, layers=[200, 100])
-        learn.fit(epochs=1)
+        dls = TabularDataLoaders.from_csv(
+            "/input/tests/data/train.csv",
+            cont_names=["pixel"+str(i) for i in range(784)],
+            y_names='label',
+            procs=[FillMissing, Categorify, Normalize])       
+        learn = tabular_learner(dls, layers=[200, 100])
+        learn.fit_one_cycle(n_epoch=1)
+        
+        self.assertGreater(learn.smooth_loss, 0)
