layer 3 - for patterns of layer 2 - thus can sense geometric shapes, lines of text or barcode.Thus can detect window, table corners, circles such as clocks layer 2 - activates for two edges, curves.For instance, these are how the layers in resnet look like Resnet34 has 34 layers, resnet50 has 50 layers. However, to improve this furture, we need to retrain the whole model, meaning, all its layers. So far, we took a resnet34 model, added a few layers to the end and trained. It is tradecraft to make the end learning rate about 10 times smaller than rate at which errors start to increase. What the slice suggests is, train the initial layers at start value specified and last layer at the end value specified and interpolate for the rest of the layers. With this new information retrain the model. We need to be more nuanced here.įind the optimal learning rate: Now load the original model using learn.load('stage-1'), then run learn.lr_find() and find the highest learning rate that has the lowest loss. This happens because you have a reckless learning rate which makes the model lose it original learning. Sometimes, the error goes up when doing this. Next, you repeat the learn.fit_one_cycle(numepochs). To improve this better, you need to call learn.unfreeze() to unfreeze the model. Generally, when you call fit_one_cycle it only trains the last or last few layers.Fastai also has interp.most_confused(min_val=2) which will return the top losses. You can also plot ot_confusion_matrix() to view the CF matrix. Find the biggest losses using ot_top_losses(9, figsize=(15,11)).The learn object so far knows the data and the model used to train. Validation - create an interpreter object using om_learner(learn).look at the results and if good, save by calling learn.save('filename').Pass the epoch number (also called cycles) You can use fit or fit_one_cycle methods, but recommended is to use latter. create a ConvLearner object by passing the data bunch, specifying the model architecture and metrics to use to evaluate training stats.run show_batch to see the classes and labels.This process automatically creates a validation set. We pick a model that already knows something about images and tune it to our case study. data.classes -> gives the names of the classes. normalizing images means turning them to (mean 0, 1 SD).The image dimensions used here is 224.Thus if you type help(fastai.untar_data), you notice type hints. PostGIS - SQLAlchemy, GeoAlchemy, GeoPandasĪs of 2019, fast.ai supports 4 types of DL applicationsįast.ai uses type hinting introduced in Python 3.5 quite heavily.Reading multi-dimensional data using open geo tools.GeoPandas - GP, IO, interactive plotting, geocoding.Seaborn grids & custom - pair, facet grids.Geographical plotting with Basemap - matplotlib toolkit.Matplotlib log scales, ticks, scientific.

Python memory, ref counts, garbage collection.

Fast.ai tabular data pass np.array