more_changes

Author: SubhadityaMukherjee

README.md

@@ -4,13 +4,17 @@
 - Faster research and experimentation
 - Based on [fastai](https://github.com/fastai)
 - Not tooo high level and meant only for the "bit" advanced researcher
+- Many many convenience functions that one would need
 - (This library is not stable right now. Everything works of course, but they will be changed and rearranged eventually)
 
+# Features
+- Check features.md for a full list
+
 ## Documentation
-- Check demo.ipynb for initial example
-- There will be blogs on the topics that were required and [more here](https://www.subhadityamukherjee.me/)
+- Check the syntax folder for documentation
+- Check demos folder
 - [Function documentation](https://subhadityamukherjee.github.io/sprintdl/)
-- More will be added soon along with proper documentation
+- There will be blogs on the topics that were required and [more here](https://www.subhadityamukherjee.me/)
 
 ## Why
 - I loved fastai, but in many things didnt agree with the way they went. I also needed some extra features that might have not ever been approved

demos/.ipynb_checkpoints/ArtClass-checkpoint.ipynb

@@ -0,0 +1,150 @@
+# ---
+# jupyter:
+#   jupytext:
+#     text_representation:
+#       extension: .py
+#       format_name: light
+#       format_version: '1.5'
+#       jupytext_version: 1.10.1
+#   kernelspec:
+#     display_name: Python 3
+#     language: python
+#     name: python3
+# ---
+
+# +
+# %load_ext autoreload
+# %autoreload 2
+
+# %matplotlib inline
+
+import os
+os.environ['TORCH_HOME'] = "/media/hdd/Datasets/"
+import sys
+sys.path.append("../")
+# -
+
+from sprintdl.main import *
+from sprintdl.nets import *
+
+device = torch.device('cuda',0)
+from torch.nn import init
+import torch
+import math
+
+# # Define required
+
+# +
+fpath = Path("/media/hdd/Datasets/ArtClass/")
+
+tfms = [make_rgb, ResizeFixed(128), to_byte_tensor, to_float_tensor]
+bs = 256
+# -
+
+# # Actual process
+
+il = ImageList.from_files(fpath, tfms=tfms)
+
+il
+
+tm= Path("/media/hdd/Datasets/ArtClass/Unpopular/mimang.art/69030963_140928767119437_3621699865915593113_n.jpg")
+
+str(tm).split("/")[-3]
+
+sd = SplitData.split_by_func(il, partial(random_splitter, p_valid = .2))
+ll = label_by_func(sd, lambda x: str(x).split("/")[-3], proc_y=CategoryProcessor())
+
+n_classes = len(set(ll.train.y.items))
+
+data = ll.to_databunch(bs, c_in=3, c_out=2)
+
+show_batch(data, 4)
+
+# +
+lr = .001
+pct_start = 0.5
+phases = create_phases(pct_start)
+sched_lr  = combine_scheds(phases, cos_1cycle_anneal(lr/10., lr, lr/1e5))
+sched_mom = combine_scheds(phases, cos_1cycle_anneal(0.95, 0.85, 0.95))
+
+cbfs = [
+    partial(AvgStatsCallback,accuracy),
+    partial(ParamScheduler, 'lr', sched_lr),
+    partial(ParamScheduler, 'mom', sched_mom),
+        partial(BatchTransformXCallback, norm_imagenette),
+    ProgressCallback,
+    Recorder,
+#     MixUp,
+       partial(CudaCallback, device)]
+
+loss_func=LabelSmoothingCrossEntropy()
+# arch = partial(xresnet34, n_classes)
+arch = get_vision_model("resnet34", n_classes=n_classes, pretrained=True)
+
+# opt_func = partial(sgd_mom_opt, wd=0.01)
+opt_func = adam_opt(mom=0.9, mom_sqr=0.99, eps=1e-6, wd=1e-2)
+# opt_func = lamb
+# -
+
+# # Training
+
+clear_memory()
+
+# learn = get_learner(nfs, data, lr, conv_layer, cb_funcs=cbfs)
+learn = Learner(arch,  data, loss_func, lr=lr, cb_funcs=cbfs, opt_func=opt_func)
+
+# +
+# model_summary(learn, data)
+# -
+
+learn.fit(1)
+
+save_model(learn, "m1", fpath)
+
+# +
+temp = Path('/media/hdd/Datasets/ArtClass/Popular/artgerm/10004370_1657536534486515_1883801324_n.jpg')
+
+get_class_pred(temp, learn ,ll, 128)
+# -
+
+temp = Path('/home/eragon/Downloads/Telegram Desktop/IMG_1800.PNG')
+
+get_class_pred(temp, learn ,ll,128)
+
+temp = Path('/home/eragon/Downloads/Telegram Desktop/IMG_20210106_180731.jpg')
+
+get_class_pred(temp, learn ,ll,128)
+
+# # Digging in
+
+# +
+# classification_report(learn, n_classes, device)
+# -
+
+learn.recorder.plot_lr()
+
+learn.recorder.plot_loss()
+
+# # Model vis
+
+run_with_act_vis(1, learn)
+
+# # Multiple runs with model saving
+
+dict_runner = {
+    "xres18":[1, partial(xresnet18, c_out=n_classes)(), data, loss_func, .001, cbfs,opt_func],
+    "xres34":[1, partial(xresnet34, c_out=n_classes)(), data, loss_func, .001, cbfs,opt_func],
+    "xres50":[1, partial(xresnet50, c_out=n_classes)(), data, loss_func, .001, cbfs,opt_func],
+}
+
+learn = Learner(arch(), data, loss_func, lr=lr, cb_funcs=cbfs, opt_func=opt_func)
+
+multiple_runner(dict_runner, fpath)
+
+
+
+
+
+
+
+

demos/.ipynb_checkpoints/ArtClassification-checkpoint.ipynb

@@ -0,0 +1,156 @@
+# ---
+# jupyter:
+#   jupytext:
+#     text_representation:
+#       extension: .py
+#       format_name: light
+#       format_version: '1.5'
+#       jupytext_version: 1.10.1
+#   kernelspec:
+#     display_name: Python 3
+#     language: python
+#     name: python3
+# ---
+
+# +
+# %load_ext autoreload
+# %autoreload 2
+
+# %matplotlib inline
+
+import sys
+sys.path.append("../")
+# -
+
+from sprintdl.main import *
+from sprintdl.nets import *
+
+device = torch.device('cuda',0)
+from torch.nn import init
+import torch
+import math
+
+# # Define required
+
+# +
+# x_train,y_train,x_valid,y_valid = get_mnist("/media/hdd/Datasets/imagenette2-160.tgz")
+
+# +
+fpath = Path("/media/hdd/Datasets/imagenette2-160/")
+
+tfms = [make_rgb, ResizeFixed(128), to_byte_tensor, to_float_tensor]
+bs = 256
+
+
+# -
+
+# # Define model and data
+
+# +
+def prev_pow_2(x): return 2**math.floor(math.log2(x))
+
+def get_cnn_layers(data, nfs, layer, **kwargs):
+    def f(ni, nf, stride=2): return layer(ni, nf, 3, stride=stride, **kwargs)
+    l1 = data.c_in
+    l2 = prev_pow_2(l1*3*3)
+    layers =  [f(l1  , l2  , stride=1),
+               f(l2  , l2*2, stride=2),
+               f(l2*2, l2*4, stride=2)]
+    nfs = [l2*4] + nfs
+    layers += [f(nfs[i], nfs[i+1]) for i in range(len(nfs)-1)]
+    layers += [nn.AdaptiveAvgPool2d(1), Lambda(flatten), 
+               nn.Linear(nfs[-1], data.c_out)]
+    return layers
+
+def get_cnn_model(data, nfs, layer, **kwargs):
+    return nn.Sequential(*get_cnn_layers(data, nfs, layer, **kwargs))
+
+def get_learner(nfs, data, lr, layer, loss_func=F.cross_entropy,
+                cb_funcs=None, opt_func=sgd_opt, **kwargs):
+    model = get_cnn_model(data, nfs, layer, **kwargs)
+    init_cnn(model)
+    return Learner(model, data, loss_func, lr=lr, cb_funcs=cb_funcs, opt_func=opt_func)
+
+
+# +
+nfs = [64,64,128,256]
+# sched = combine_scheds([0.3, 0.7], [sched_cos(0.3, 0.6), sched_cos(0.6, 0.2)]) 
+# sched = combine_scheds([0.3, 0.7], [sched_cos(.1,.3), sched_cos(.3, 0.05)])
+# mnist_view = view_tfm(1,28,28)
+
+lr = 1e-2
+pct_start = 0.5
+phases = create_phases(pct_start)
+sched_lr  = combine_scheds(phases, cos_1cycle_anneal(lr/10., lr, lr/1e5))
+sched_mom = combine_scheds(phases, cos_1cycle_anneal(0.95, 0.85, 0.95))
+
+cbfs = [
+    partial(AvgStatsCallback,accuracy),
+    partial(ParamScheduler, 'lr', sched_lr),
+    partial(ParamScheduler, 'mom', sched_mom),
+        partial(BatchTransformXCallback, norm_imagenette),
+    ProgressCallback,
+    Recorder,
+#     MixUp,
+       partial(CudaCallback, device)]
+
+loss_func=LabelSmoothingCrossEntropy()
+arch = partial(xresnet18, c_out=10)
+epochs = 5
+lr = .4
+# opt_func = partial(sgd_mom_opt, wd=0.01)
+opt_func = adam_opt(mom=0.9, mom_sqr=0.99, eps=1e-6, wd=1e-2)
+# opt_func = lamb
+# -
+
+# # Actual process
+
+il = ImageList.from_files(fpath, tfms=tfms)
+
+il
+
+sd = SplitData.split_by_func(il, partial(grandparent_splitter, valid_name='val'))
+ll = label_by_func(sd, parent_labeler, proc_y=CategoryProcessor())
+
+ll
+
+data = ll.to_databunch(bs, c_in=3, c_out=10)
+
+# # Training
+
+# learn = get_learner(nfs, data, lr, conv_layer, cb_funcs=cbfs)
+learn = Learner(arch(), data, loss_func, lr=lr, cb_funcs=cbfs, opt_func=opt_func)
+
+# +
+# model_summary(learn, data)
+# -
+
+learn.fit(epochs)
+
+# +
+# run_with_act_vis(epochs,learn)
+# -
+
+# # Digging in
+
+learn.avg_stats.valid_stats.avg_stats
+
+learn.recorder.plot_lr()
+
+learn.recorder.plot_loss()
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+